JP6606003B2 - Process diagnostic apparatus, process diagnostic method, and computer program - Google Patents

Description

  Embodiments described herein relate generally to a process diagnosis apparatus, a process diagnosis method, and a computer program.

  Process plants such as sewage treatment processes, sludge digestion processes, water purification processes and water supply / distribution processes, and process plants such as petrochemical processes, steel processes, and semiconductor manufacturing processes, can monitor the process status from multiple perspectives. Various sensors are installed for monitoring. A process monitoring device (SCADA: Supervisory Control And Data Acquisition) is measurement data (for example, data indicating flow rate, temperature, water quality, manipulated variable, etc.) acquired by a sensor group installed in a process system. Data ”) is acquired in time series. Plant managers and operators grasp the state of the process by monitoring this time-series process data, and change or control the operation of the process. Each process data acquired in time series usually has upper and lower limits called control limits, and an alarm is issued when the control limits are exceeded. Based on this alarm, the plant manager and the operator confirm the plant state and review the plant operation. Operation management based on such alarms is fundamental to plant operation.

  In the operation management of the plant, which is further advanced, it is required to improve productivity and avoid risks by quickly detecting abnormal signs related to predetermined performance indicators of the plant and taking early measures. The predetermined performance index mentioned here is, for example, the quality of discharged water in a sewage treatment process, the quality of water distribution in a water purification process, quality performance (for example, yield) in a petrochemical plant or a steel plant, and the like. In plant operation, it is required to control the process so that this performance index is maintained at a desired target value and does not deviate from the target value. Therefore, the process status related to the performance index of the plant is monitored, the state change or abnormality that impedes the realization of the target performance index is quickly detected, and a prompt action is taken against the detected state change or abnormality. This is an important point in operation management. In addition, in order to perform operations that lead to energy savings and cost savings while limiting the realization of the target performance index, the process status related to the target performance, energy savings, and cost savings must always be kept in a good state. It is also required to quickly detect changes in the process state that are likely to deviate from.

MSPC (Multi-Variate Statistical Process Control) is based on the multivariate statistical analysis method that has been used mainly in the fields of petrochemical processes and steel plants as a method for diagnosing such changes and abnormalities in the process state. A method called process management is known. PCA (Principal Component Analysis), latent variable projection method, PLS (Projection to Latent Structure / Partial Least Square), etc. are known as the most commonly used methods in MSPC. . The main purpose of MSPC is to detect fault signs in the plant (Fault Detection) and the type of process data that causes the anomaly (hereinafter referred to as “process variables”) using multivariate analysis methods such as PCA and PLS. (Fault Isolation). Regarding the detection of abnormal signs, it is possible to detect signs of minor abnormalities that cannot be detected by one process variable by using correlation information of a plurality of process variables. Further, the abnormality factor can be estimated based on the contribution of each process variable to the abnormality detection data. The abnormality detection data is data used for abnormality detection, and is acquired based on a plurality of process variables. For example, Q statistics, Hotelling's T ² statistics, or the like is used for abnormality detection data. In such a monitoring method using MSPC, a conventional monitoring method based on the control limit of each individual process variable (for example, a monitoring method for a production line or the like is called SPC (Statistical Process Control) in contrast to MSPC). Monitoring and diagnosis that are more useful to plant managers and operators.

However, in order to diagnose a process using MSPC, first, a mathematical model (hereinafter referred to as Q statistic, T ² statistic, etc.) for calculating anomaly detection data (for example, Q statistic or T ² statistic) using a statistical method such as PCA or PLS , "Abnormality diagnosis model") must be constructed based on process data during normal operation. This abnormality diagnosis model defines a standard for determining the acquired process data as abnormal. For this reason, the abnormality diagnosis model is generally constructed using process data acquired in a situation where the process is operated in a normal state (during normal operation). However, in plant operation, there are many cases where a lot of abnormal data called outliers are included in the process data to be monitored. Therefore, in the construction of an abnormality diagnosis model, it is necessary to use a method for constructing an abnormality diagnosis model that is not easily affected by the outlier, using preprocessing for removing outliers or robust statistics.

However, in actual plant operation, it is often difficult to uniquely define the normal operation of the plant even if an abnormality diagnosis model in which the influence of the outlier is removed or reduced can be constructed. This is because in an actual plant operation, it is not uncommon for the operation of the plant to change, and the state of the plant may change depending on the season. For this reason, in actual plant operation, it is often inherently difficult to uniquely define what operating state is normal operation. In such a case, if the abnormality diagnosis model is not updated at an appropriate timing, there is a possibility that the plant state changed according to the operation change or the season cannot be properly diagnosed. Specifically, the abnormality when the diagnosis model becomes unsuitable for actual plant state, becomes as shown always outliers Q statistic and T ² statistics, can no longer be able to continue the appropriate diagnosis There is sex.

  As a method for solving such a problem, a method of updating an abnormality diagnosis model sequentially or adaptively and a method called JIT (Just in Time) modeling have been studied. JIT modeling is a local method for constructing an abnormality diagnosis model by extracting only appropriate data from the latest process data. In many cases, a sequential method or a local method is effective, but if it relies excessively on such a method, there is a possibility that the process state that should be detected as an abnormality is erroneously detected as normal. For example, there is a possibility that a sensor drift or a gradual transition to an abnormal state of the plant cannot be properly detected. This is because the abnormality detection data is generated based on the local process data. In other words, the normal state of the plant is constantly updated based on the local process data. In other words, the abnormal state of the plant should be defined relative to the normal state, and constantly updating the reference normal state may hinder appropriate detection of the abnormal state. .

JP 2004-303007 A Japanese Patent Laid-Open No. 08-241121

  The problem to be solved by the present invention is to provide a process diagnosis apparatus, a process diagnosis method, and a computer program capable of updating an abnormality diagnosis model at a more appropriate timing.

  The process diagnosis apparatus according to the embodiment detects a process abnormality at a predetermined diagnosis cycle based on an abnormality diagnosis model constructed by multivariate analysis of various time-series data indicating the state of a process to be diagnosed. Device. The process diagnosis apparatus includes an abnormality diagnosis model construction unit, an update determination unit, and a model update unit. The abnormality diagnosis model construction unit constructs the abnormality diagnosis model at a predetermined construction cycle based on the time series data. The update determination unit includes an active model that is an abnormality diagnosis model used for abnormality detection of the process, and a pre-application model that is an abnormality diagnosis model constructed after the active model is constructed by the abnormality diagnosis model construction unit Is calculated at an update cycle shorter than the construction cycle, and whether or not the active model needs to be updated is determined based on the calculated difference. The model update unit updates the active model with the pre-application model according to the determination result of the update determination unit.

The figure which shows the outline of a structure of the process diagnostic apparatus 2 of 1st Embodiment. The block diagram which shows the detail of a function structure of the process diagnostic apparatus 2 of 1st Embodiment. The block diagram which shows the detail of a function structure of the process diagnostic apparatus 2a of 2nd Embodiment. The block diagram which shows the detail of a function structure of the process diagnostic apparatus 2b of 3rd Embodiment. The figure which shows the specific example of the change of the process data by an operation change. The block diagram which shows the detail of a function structure of the process diagnostic apparatus 2c of 4th Embodiment.

  Hereinafter, a process diagnosis apparatus, a process diagnosis method, and a computer program according to embodiments will be described with reference to the drawings.

(First embodiment)
FIG. 1 is a diagram illustrating an outline of a configuration of a process diagnostic apparatus 2 according to the first embodiment. FIG. 1 shows a specific example in which the diagnosis target of the process diagnostic apparatus 2 is a sewage advanced treatment process 1. Sewage advanced treatment process 1 is a process aimed at removing nitrogen and phosphorus from sewage. The advanced sewage treatment process 1 includes a first sedimentation tank 101, an anaerobic tank 102, an oxygen-free tank 103, an aerobic tank 104, and a final sedimentation tank 105. Sewage to be treated (hereinafter referred to as “treated water”) is fed and processed in the order of the first sedimentation basin 101, the anaerobic tank 102, the anaerobic tank 103, the aerobic tank 104, and the final sedimentation basin 105.

  The first sedimentation basin 101 is a reservoir for treated water sent to the advanced sewage treatment process 1. In the first sedimentation basin 101, solid matter having a high specific gravity is separated from the water to be treated by precipitation. The anaerobic tank 102 is a water tank for introducing microorganisms that decompose organic substances into the water to be treated and for dephosphorizing the microorganisms. In the anaerobic tank 102, the water to be treated is stirred in a state where air is not supplied. This causes the microorganisms to exhale phosphorus in the body. This process is generally called phosphorus discharge. Dephosphorization of a series of processes to accumulate phosphorus in the body of microorganisms by phosphorus discharge in an anaerobic tank and absorption of phosphorus in an aerobic tank described later, and to remove phosphorus by separation of sludge containing microorganisms that have accumulated phosphorus That's it. The anoxic tank 103 is a water tank for removing nitrogen from the water to be treated. Specifically, in the anaerobic tank 103, the water to be treated returned from the aerobic tank 104 in the subsequent stage is mixed with the water to be treated sent from the anaerobic tank 102, and is stirred without being supplied with air. In the anoxic tank 103, nitric acid in the water to be treated is decomposed into nitrogen by the action of microorganisms and released to the atmosphere. This process is generally called denitrification.

  The aerobic tank 104 is a water tank for decomposing organic substances in the water to be treated, removing phosphorus, and nitrifying ammonia. Specifically, air is supplied to the water to be treated to activate the microorganisms, causing the microorganisms to decompose organic matter and causing the microorganisms to absorb phosphorus in the water to be treated. Since the microorganisms in the dephosphorized state are activated to absorb more phosphorus than is exhaled by dephosphorization, the phosphorus in the water to be treated is removed. In the aerobic tank 104, ammonia is decomposed into nitric acid by supplying air to the water to be treated. This treatment is generally called nitrification.

  The final sedimentation basin 105 is a reservoir for treated water from which phosphorus has been removed and ammonia has been nitrified. In the final sedimentation basin 105, solid matter remaining in the water to be treated is separated by precipitation, and the supernatant clear water is discharged as treated water.

  The first sedimentation basin excess sludge extraction pump 111 is a pump that extracts and removes the sludge precipitated from the first sedimentation basin 101. The first settling pond excess sludge extraction pump 111 has a flow rate sensor for measuring the flow rate of the extracted sludge.

  The blower 112 is a blower that supplies oxygen to the aerobic tank 104. The blower 112 has a flow rate sensor that measures the flow rate of the supplied air.

  The circulation pump 113 is a pump that returns the water to be treated from the aerobic tank 104 to the anoxic tank 103. The circulation pump 113 has a flow rate sensor that measures the flow rate of the returned treated water.

  The return sludge pump 114 is a pump that extracts a part of the sludge that has settled from the final sedimentation basin 105 and returns it to the anaerobic tank 102. The return sludge pump 114 has a flow rate sensor that measures the flow rate of the returned sludge.

  The final sedimentation basin excess sludge extraction pump 115 is a pump that extracts and removes the sludge precipitated from the final sedimentation basin 105. The final sedimentation pond surplus sludge extraction pump 115 has a flow rate sensor for measuring the flow rate of the extracted sludge.

  The rainfall sensor 121 is a sensor that measures the rainfall in the vicinity that flows into the sewage altitude treatment process 1. The sewage inflow sensor 122 is a sensor that measures the flow rate of sewage flowing into the sewage altitude treatment process 1 (hereinafter referred to as “inflow sewage”). The inflow TN sensor 123 is a sensor that measures the total nitrogen amount (TN) contained in the inflow sewage. The inflow TP sensor 124 is a sensor that measures the total phosphorus amount (TP) contained in the inflow sewage. The inflow organic matter sensor 125 is a UV (absorbance) sensor or a COD (chemical oxygen demand) sensor that measures the amount of organic matter contained in the inflow sewage.

  The ORP sensor 126 is a sensor that measures the ORP (oxidation-reduction potential) of the anaerobic tank 102. The anaerobic tank pH sensor 127 is a sensor that measures the pH of the anaerobic tank 102. The oxygen-free tank ORP sensor 128 is a sensor that measures the ORP of the oxygen-free tank 103. The oxygen-free tank pH sensor 129 is a sensor that measures the pH of the oxygen-free tank 103. The phosphoric acid sensor 130 is a sensor that measures the phosphoric acid concentration in the aerobic tank 104. The DO sensor 131 is a sensor that measures the dissolved oxygen concentration (DO) in the aerobic tank 104. The ammonia sensor 132 is a sensor that measures the ammonia concentration in the aerobic tank 104. The MLSS sensor 133 is a sensor that measures the activated sludge concentration (MLSS) in at least one location of the anaerobic tank 102, the anoxic tank 103, or the aerobic tank 104.

  The water temperature sensor 134 is a sensor that measures the water temperature in at least one location of the anoxic tank 103 or the aerobic tank 104. The excess sludge SS sensor 135 is a sensor that measures the solid matter (SS) concentration of the sludge drawn from the final sedimentation basin 105. The discharge SS sensor 136 is a sensor that measures the SS concentration of water discharged from the final sedimentation basin 105. The sludge interface sensor 137 is a sensor that measures the sludge interface level of the final sedimentation basin 105. The sewage discharge flow rate sensor 138 is a sensor that measures the flow rate of discharged water. The discharge TN sensor 139 is a sensor that measures the total amount of nitrogen contained in the discharge water. The discharge TP sensor 140 is a sensor that measures the total amount of phosphorus contained in the discharge water. The discharged organic matter sensor 141 is a UV sensor or a COD sensor that measures the amount of organic matter contained in the discharged water.

  The first sedimentation pond surplus sludge extraction pump 111, the blower 112, the circulation pump 113, the return sludge pump 114, and the final sedimentation basin surplus sludge extraction pump 115 operate in a predetermined cycle. In addition, each of the above sensors including the flow rate sensors of the first sedimentation pond surplus sludge extraction pump 111, the blower 112, the circulation pump 113, the return sludge pump 114, and the final sedimentation pond surplus sludge extraction pump 115 is a sensing target at a predetermined cycle. Measure. Hereinafter, the flow rate sensors of the first settling tank excess sludge extraction pump 111, the blower 112, the circulation pump 113, the return sludge pump 114, and the final settling tank excess sludge extraction pump 115 are collectively referred to as operation amount sensors, and other sensors. Are collectively referred to as process sensors. Each manipulated variable sensor and each process sensor transmits measurement data obtained by sensing at a predetermined cycle to the process diagnosis apparatus 2 as process data.

  The process diagnostic device 2 includes a CPU (Central Processing Unit), a memory, an auxiliary storage device, and the like connected by a bus, and executes a diagnostic device program. The process diagnostic apparatus 2 includes a data collection unit 201, a collected data storage unit 202, a data extraction unit 203, an abnormality diagnosis model construction unit 204, a latest model buffer unit 205, a diagnostic model storage unit 206, an active model buffer by executing a diagnostic device program. It functions as an apparatus including a unit 207, an abnormality diagnosis unit 208, a diagnosis result storage unit 209, an abnormality diagnosis model update unit 210, and a diagnosis result provision unit 211. Note that all or a part of each function of the process diagnostic apparatus 2 may be realized by using hardware such as an application specific integrated circuit (ASIC), a programmable logic device (PLD), or a field programmable gate array (FPGA). . The diagnostic device program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built in the computer system. The diagnostic device program may be transmitted via a telecommunication line.

The data collection unit 201 acquires process data from each operation amount sensor and each process sensor. The data collection unit 201 stores the acquired process data in the collection data storage unit 202.
The collected data storage unit 202 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The collected data storage unit 202 stores the process data collected by the data collection unit 201.

  The data extraction unit 203 uses, from the process data stored in the collected data storage unit 202, data necessary for abnormality diagnosis (hereinafter referred to as “diagnosis data”) and data necessary for construction of an abnormality diagnosis model (hereinafter referred to as “data for diagnosis”). "Construction data"). Specifically, the diagnostic data is process data extracted for each past predetermined period, and the construction data is process data from which process data for the past predetermined period is extracted at a period closer to real time. The data extraction unit 203 outputs the extracted construction data to the abnormality diagnosis model construction unit 204, and outputs the extracted diagnosis data to the abnormality diagnosis unit 208.

  The abnormality diagnosis model construction unit 204 constructs an abnormality diagnosis model based on the construction data output from the data extraction unit 203. The abnormality diagnosis model construction unit 204 outputs the constructed abnormality diagnosis model to the latest model buffer unit 205.

  The latest model buffer unit 205 includes a storage device such as a magnetic hard disk device or a semiconductor storage device, and stores the latest abnormality diagnosis model (hereinafter referred to as “latest model”) constructed by the abnormality diagnosis model construction unit 204. And stored in the diagnostic model storage unit 206. The latest model buffer unit 205 outputs the latest model stored therein to the abnormality diagnosis model update unit 210.

  The diagnostic model storage unit 206 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The diagnostic model storage unit 206 stores the latest model output from the latest model buffer unit 205. The diagnostic model storage unit 206 accumulates the latest model constructed based on the construction data for each predetermined period in time series.

  The active model buffer unit 207 includes a storage device such as a magnetic hard disk device or a semiconductor storage device, and stores an abnormality diagnosis model (hereinafter referred to as an “active model”) used for current abnormality diagnosis. The active model to be stored is output to the abnormality diagnosis unit 208 and the abnormality diagnosis model update unit 210. The initial value of the active model may be set to the abnormality diagnosis model that is first constructed after the operation of the own device is started, or an abnormality diagnosis model that is registered in advance as the initial value of the active model in the own device is set. May be.

  The abnormality diagnosis unit 208 diagnoses the process state to be diagnosed at a predetermined monitoring cycle based on the real-time diagnosis data extracted by the data extraction unit 203 and the active model output from the active model buffer unit 207. . The abnormality diagnosis unit 208 stores the process state diagnosis result in the diagnosis result storage unit 209.

  The diagnosis result storage unit 209 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The diagnosis result storage unit 209 stores the diagnosis result of the abnormality diagnosis unit 208. The diagnosis result storage unit 209 accumulates diagnosis results based on real-time diagnosis data in time series.

  The abnormality diagnosis model update unit 210 determines whether or not to update the active model based on the latest model output from the latest model buffer unit 205 and the active model output from the active model buffer unit 207. When it is determined that the active model should be updated, the abnormality diagnosis model update unit 210 updates the active model with the latest model.

  The diagnostic result providing unit 211 provides the user (plant manager or operator) with the diagnostic results accumulated in the diagnostic result storage unit 209. When the process diagnosis apparatus 2 includes a display unit such as a CRT (Cathode Ray Tube) display, a liquid crystal display, or an organic EL (Electro-Luminescence) display, the diagnosis result providing unit 211 may display the diagnosis result on the display unit. Further, when the process diagnosis apparatus 2 includes a communication unit capable of communicating with the user terminal, the diagnosis result providing unit 211 may transmit the diagnosis result to the user terminal via the communication unit. In this case, the diagnosis result providing unit 211 may transmit the diagnosis result at a predetermined timing, or may transmit the diagnosis result in response to a request from the user terminal.

  FIG. 2 is a block diagram illustrating details of a functional configuration of the process diagnostic apparatus 2 according to the first embodiment. The abnormality diagnosis model construction unit 204 includes a normalization unit 221 and a model construction unit 222. The normalization unit 221 converts the construction data extracted by the data extraction unit 203 for each predetermined period into position parameters (average, median, etc.) and scale parameters (standard deviation or MAD (Median Absolute Deviation) for each process variable. ) Etc.) to normalize.

  The model construction unit 222 constructs an abnormality diagnosis model based on the construction data normalized for each process variable by a multivariate statistical process management (MSPC) technique. Specifically, the abnormality diagnosis model is constructed as a set of mathematical expressions for obtaining a loading matrix, a score matrix, and the like for the construction data.

  The abnormality diagnosis model update unit 210 includes an update determination unit 231 and a model update unit 232. The update determination unit 231 determines whether or not the active model should be updated based on the latest model and the active model. When the update determination unit 231 determines that the active model should be updated, the model update unit 232 updates the active model with the latest model.

  Specifically, the update determination unit 231 determines whether or not the active model should be updated by determining four determination units (a first determination unit 241, a second determination unit 242, a third determination unit 243, and a fourth determination unit 244). Prepare. The four determination units included in the update determination unit 231 determine whether or not the active model needs to be updated based on the degree of divergence between the latest model and the active model calculated based on a predetermined index value related to the abnormality diagnosis model.

  The first determination unit 241 calculates the divergence degree based on the deviation of the position parameter of each process variable. The second determination unit 242 calculates the divergence degree based on the deviation of the scale parameter of each process variable. The third determination unit 243 calculates the degree of divergence based on the loading matrix deviation calculated by the multivariate analysis method. The fourth determination unit 244 calculates the degree of divergence based on the deviation of the contribution amount vector (corresponding to the score matrix) calculated by the multivariate analysis method. The update determination unit 231 determines whether the active model needs to be updated based on the divergence calculated by the four determination units, and outputs the determination to the model update unit 232. The model update unit 232 updates the active model with the latest model in accordance with the necessity of updating the active model output from the update determination unit 231.

  Hereinafter, a method for updating the abnormality diagnosis model in the process diagnosis apparatus 2 of the first embodiment will be described in detail.

[Construction of abnormal diagnosis model]
First, a method for constructing an abnormality diagnosis model will be described. First, the data extraction unit 203, for each predetermined period T _L, to obtain the process data acquired in a predetermined period L from the stored process data. For example, when the process is controlled every day based on the trend of the past week, T _L = 1 day and L = 7 days (one week). When the construction data extraction is specified to be executed at 0:00 every day, process data for the past week is extracted as construction data at 0:00 every day. The data set extracted in this way is identified as Z _k (k = 1, 2,...). Here, k is an identification number of the process data collected in time series with the period _TL . For example, the first day of operation corresponds to k = 1, and the next day corresponds to k = 2.

For example, Z _k is represented by a matrix and has the number of columns corresponding to the number of variables and the number of rows corresponding to the time series. When the diagnosis target is the above-described sewage advanced treatment process 1, the number of columns of the matrix is the sum of the number of process variables (5) measured by the manipulated variable sensor and the number of process variables (21) measured by the process sensor. 26. For example, when the measurement cycle of each sensor is 1 minute, the number of rows of the matrix is 10080 (= 60 [minutes] × 24 [hours] × 7 [days]). In the following description, the number of rows and the number of columns of the matrix Z _k are described as n.
In the following description, when there is no need to identify the process data collected at the period _TL , the matrix Z _k may be described as Z with the subscript k omitted.

Subsequently, the abnormality diagnosis model construction unit 204 constructs an abnormality diagnosis model from the construction data Z by a multivariate analysis method (for example, PCA or PLS). Specifically, abnormality detection data such as Q statistics and T ² statistics is generated from the construction data Z by PCA or PLS. When the construction data Z includes a lot of abnormal data such as outliers, a method such as robust PCA or robust PLS expanded in consideration of robustness may be used.

  In addition, when a strong nonlinear correlation is assumed between the process variables, a method such as kernel PCA or kernel PLS extended in consideration of nonlinearity between data may be used. In addition, when the number m of process variables becomes extremely large, a method such as sparse PCA or sparse PLS that can select process variables necessary for diagnosis may be used. As described above, PCA and PLS have various variations depending on their purposes. As a multivariate analysis method used in the present embodiment, a loading matrix and a score matrix described below can be defined. Any other method may be used. Hereinafter, as an example, a method of generating abnormality detection data using a general PCA will be described.

  First, as pre-processing for PCA application, the normalization unit 221 normalizes the construction data Z for each process variable. For example, normalization of the construction data Z is expressed by the following equation (1).

In Expression (1), i represents the number of columns of the matrix Z, and x _i represents a column vector of the matrix after normalization. z _i represents a column vector of the matrix Z. That is, the matrix Z = [z ₁ , z ₂ ,..., Z _m ] can be expressed. μ _i and σ _i are the position and scale parameters of the column vector z _i , typically the mean and standard deviation. When it is assumed that the construction data includes abnormal data, a median (median) or the like may be used as the position parameter, or a median absolute deviation (MAD) or the like as the scale parameter. It may be used. In this case, a trimmed average or a trimmed standard deviation for obtaining an average or standard deviation by removing a certain ratio of data may be used. The construction data X normalized as described above is defined by the following equation (2).

Then, by applying the PCA matrix X to get the loading matrix P _a and score matrix T _a satisfy the following equation (3).

The score matrix T _a is a real number set (T _a εR) and is a matrix of n rows and m columns. The number of rows n score matrix T _a corresponds to the number of samples of constructing data, the number of columns m corresponds to the number of principal components. The loading matrix P _a is a real number set (P _a εR) and is a matrix of m rows and m columns. Line number m of the loading matrix P _a corresponds to the number of process variables, the number of columns m corresponds to the number of principal components. Loading matrix P _a shows the relationship between process variables and the principal component. Matrix T is a submatrix consisting p (<< m) pieces of m column scores matrix T _a. Similarly, the matrix P is p (<< m) consisting of pieces submatrix of the m columns of the loading matrix P _a. The p, among the m columns the loading matrix P _a has a parameter for adjusting whether to use the process data of how much amount the calculation of Q statistic and T ² statistics. Specifically, in this case, the process data for the p column is used for calculating the Q statistic and the T ² statistic, and the process data for the mp column is truncated. In this way, by calculating by cutting off part of the acquired process data, it is possible to reduce the calculation load required for calculating the Q statistic and the T ² statistic. Such data truncation is a method generally used in the field of multivariate analysis, and the adjustment parameter p is called a truncation order.

Further, the error matrix E is a real number set (EεR) and is a matrix of n rows and m columns. The number of rows n of the error matrix E corresponds to the number of samples, and the number of columns m corresponds to the number of process variables. In the following, in order to distinguish between the score matrix _Ta and its submatrix T, the former is referred to as a total score matrix and the latter is referred to as a score matrix. Similarly, the matrix P _a is referred to as a total loading matrix, the matrix P is referred to as a loading matrix.

Here, the Q statistic and the T ² statistic that are components of the abnormality diagnosis model are expressed by the following equations (4) and (5) using the loading matrix P and Λ derived from the score matrix T. ).

Q (x (t)) in Equation (4) represents a Q statistic regarding x (t) that is the t (1 ≦ t ≦ m) -th element of the matrix X, and I is a unit matrix of an appropriate size. . T ² (x (t)) in equation (5) represents a T ² statistic relating to x (t), and Λ is a diagonal matrix composed of a set of eigenvalues whose diagonal elements are T ^T × T. . These eigenvalues represent the variance in each principal component direction in the principal component analysis. t (t) and S are values obtained by the following equations (6) and (7).

Thus, the Q statistic and the T ² statistic are calculated based on the loading matrix P and the score matrix T. On the other hand, as shown in Equation (5), the T ² statistic can be represented by the construction data x (t), the loading matrix P, and Λ. In general, in principal component analysis, eigenvalues are said to represent the level of explanatory power (also referred to as explanatory power) of each principal component, and based on this eigenvalue, each component contributes to the phenomenon. Contribution rate is required. In order to calculate the contribution rate, a contribution amount vector λ obtained by extracting only the diagonal component of Λ representing the eigenvalue is defined as the following equation (6).

  From the above, it is understood that it is sufficient that at least the loading matrix P and the contribution amount vector λ can be calculated for the abnormality diagnosis model. For this reason, as long as the loading matrix P and the score matrix T can be calculated, any other multivariate analysis method may be used as a method for constructing the abnormality diagnosis model. In the process diagnosis apparatus 2 of the present embodiment, the position parameter and the scale parameter of the process data are used for determining whether or not the active model needs to be updated. Therefore, the abnormality diagnosis model construction unit 204 acquires the loading matrix P, the contribution vector λ, the position parameter μ, and the scale parameter σ as components of the abnormality diagnosis model. Here, the position parameter μ and the scale parameter σ are defined as in the following equations (7) and (8).

In Expression (7), μ _t (1 ≦ t ≦ m) represents a position parameter for the t-th element of the construction data X. In Expression (8), σ _t (1 ≦ t ≦ m) represents a scale parameter for the t-th element of the construction data X.

[Abnormal diagnosis using an abnormal diagnosis model]
Next, an abnormality diagnosis method using the abnormality diagnosis model will be described. The abnormality diagnosis unit 208 performs abnormality diagnosis of the target process based on the diagnostic data output from the data extraction unit 203 in real time and the active model held in the active model buffer unit 207. Specifically, the abnormality diagnosis unit 208 diagnoses a process abnormality by comparing a Q statistic and a T ² statistic acquired from the diagnostic data with a predetermined threshold value set in advance. Since this threshold is a parameter that is largely related to detection of process state changes and detection of abnormal signs, its setting method is important. Hereinafter, this threshold value setting method will be described.

As a typical setting method of this threshold value, there is a method using a statistical confidence limit value of the Q statistic and a statistical confidence limit value of the T ² statistic. In this case, the statistical confidence limit value Q _limit of the Q statistic is expressed as the following equation (9).

In Expression (9), p is the truncation order of the construction data. _cα is the deviation of the standard deviation of the standard normal distribution when the limit of the confidence interval is 1-α. For example, c _α when α = 0.01 is 2.53, and c _α when α = 0.05 is 1.96. Further, λ _i represents each element of the contribution vector described above. That is, Θ _i is the sum of powers of the elements of the contribution vector corresponding to the censored construction data. On the other hand, the statistical confidence limit value T ² _limit of the T ² statistic is expressed as the following equation (10).

  Similarly to equation (9), p in equation (10) is the order of truncation of the construction data. n is the total number of data for construction. F (p, n−p, α) is an F distribution when the degree of freedom is (p, n−p) and the reliability limit is α. In general, α is often set to 0.01 or 0.05.

In the process abnormality diagnosis using MSPC, the Q statistic and the T ² statistic obtained by applying the diagnostic data to Equations (4) and (5) instead of the construction data are represented by Equation (9) and Anomalies in the process are detected by comparison with the statistical confidence limit value previously determined by equation (10). Further, when an abnormality is detected, the process variable used for the abnormality diagnosis is detected by calculating a contribution amount indicating the magnitude of influence (that is, explanatory ability) exerted on the detected abnormality. Process variables that have a large influence on the abnormality are estimated, and the validity of the abnormality detection is determined. The contribution amount can be acquired from the equations (4) and (5), but such acquisition of the contribution amount is a general technique in the field of multivariate analysis. Therefore, detailed description of the contribution amount calculation method is omitted here.

As described above, the abnormality diagnosis model based on the MSPC is based on the statistical values of the position parameter and the scale parameter, the Q statistic and the T ² statistic, and the Q statistic and the T ² statistic that are the normalization parameters of each process variable. It can be seen that it is composed of the confidence limit value and the contribution amount of each process variable to each statistical confidence limit value. For obtaining the Q statistic, the T ² statistic, the statistical confidence limit value of each statistic, and the contribution amount of each process variable, the equations (4), (5), (9), and (10) are used. It can be seen that the loading matrix P, Λ, the number of data n, the number m of process variables, and the truncation order p are necessary. Of these, the number n of data and the number m of process variables are usually fixed values that do not change. Generally, the truncation order p is determined based on Λ. That is, the information necessary for the abnormality diagnosis model is the loading matrices P and Λ.

[Update abnormality diagnosis model]
The latest model buffer unit 205 holds the latest abnormality diagnosis model (latest model) constructed for each period _TL and also stores it in the diagnosis model storage unit 206. The diagnostic model storage unit 206 stores the abnormality diagnosis models constructed in the past by sequentially storing the constructed latest models. The abnormality diagnosis model update unit 210 acquires the latest model at that time from the latest model buffer unit 205, acquires the active model at that time from the active model buffer unit 207, and outputs it to the update determination unit 231. The update determination unit 231 outputs the acquired latest model and active model to the first determination unit 241 to the fourth determination unit 244, respectively.

<1. Judgment based on position parameter>
The first determination unit 241 calculates the degree of divergence between the position parameter of the latest model and the position parameter of the active model, and determines whether the active model needs to be updated by comparing the calculated degree of divergence with a predetermined threshold. To do. In the position parameter, the size and sign of each element may differ depending on the physical dimension of each process variable. Therefore, the first determination unit 241 first normalizes the position parameter. Here, for the following explanation, the position parameter of the active model is defined as μ _a , and the position parameter of the latest model is defined as μ _b .

There are various methods for normalization. For example, a method based on the position parameter of the active model can be mentioned. In this case, the normalized position parameters μ _a ′ and μ _b ′ are expressed by the following equations (11) and (12).

Another normalization method is a method based on the scale parameter of each process variable. In this case, the normalized position parameters μ _a ′ and μ _b ′ are expressed by the following equations (13) to (15).

In the formula _(13), σ _{a_i} ( "_" is i represents that the subscript subscripts sigma _a, and so forth.) Represents the i-th element of scale parameter number sigma _a, sigma _{b_i} represents the i-th element of the scale parameter σ _b . Based on the normalized position parameters μ _a ′ and μ _b ′, the position parameter divergence dμ is defined as, for example, the distance (or inner product) of each position parameter vector. In this case, the divergence degree dμ is expressed as the following equation (16).

The expression “|| vector ||” in equation (16) represents the norm of the vector (for example, Euclidean norm, L2 norm, etc.). Expression (16) is an example in which dμ is defined as the L2 norm of the difference between μ _a ′ and μ _b ′. In this case, if μ _a ′ and μ _b ′ coincide, dμ = 0. Further, when the number of location parameter normalized by Equation (11) and Equation (12), mu _b becomes d [mu] = K-1 when a K times mu _a. Further, when the position parameter is normalized by Expressions (13) and (14), dμ = K when | μ _a −μ _b | is K times σ. Thus, the threshold value of the location parameter of the deviation degree can be set by the size of mu _b deviation relative to the mu _a and sigma (i.e. K).

  However, since the L2 norm is evaluated by the sum of square squares of each element of the vector, it may be difficult to set a threshold value. In such a case, the threshold value of the divergence degree may be defined by an L1 norm that represents the divergence degree by the sum of absolute values of each element, or an L∞ norm that represents the divergence degree by the maximum absolute value of each element. In this case, the deviation degree dμ of the position parameter is expressed by the following equations (17) and (18).

When the divergence degree is defined by the L1 norm as in Expression (17), the divergence degree is evaluated by the sum of absolute values of the differences between the elements of the vector, so the threshold is set in consideration of the number m of process variables. It is necessary to On the other hand, when the divergence degree is defined by the L∞ norm as in the equation (18), the maximum difference value evaluated for each element of the vector is the divergence degree. Therefore, the threshold value can be set more directly. Specifically, the threshold is set to 1 when detecting that μ _b has deviated more than twice μ _a or when detecting that | μ _a −μ _b | is deviated by σ. can do.

  In this way, the first determination unit 241 determines whether or not the active model needs to be updated by comparing the degree of deviation of the position parameter between the active model and the latest model with a preset threshold value. Specifically, the first determination unit 241 determines that the active model needs to be updated when the deviation of the position parameter exceeds a threshold, and the deviation of the position parameter is equal to or less than the threshold. It is determined that updating of the active model is unnecessary. The first determination unit 241 determines whether or not the active model needs to be updated based on the position parameter in accordance with the latest model construction cycle, and outputs the determination result to the model update unit 232.

<2. Judgment based on scale parameter>
The second determination unit 242 calculates the divergence between the scale parameter of the latest model and the scale parameter of the active model, and determines whether the active model needs to be updated by comparing the calculated divergence with a predetermined threshold value. To do. The determination of the divergence degree based on the scale parameter by the second determination unit 242 can be performed in the same manner as the first determination unit 241 by replacing the position parameter with the scale parameter. In this case, the divergence degree of the scale parameter defined by the L2 norm, the L1 norm, or the L∞ norm is expressed by the following equations (19) to (21).

  Similar to the first determination unit 241, the second determination unit 242 determines that the active model needs to be updated when the scale parameter divergence exceeds a threshold, and the scale parameter divergence is equal to or less than the threshold. If it is, it is determined that the update of the active model is unnecessary. The second determination unit 242 determines whether or not the active model needs to be updated based on the scale parameter in accordance with the construction cycle of the latest model, and outputs the determination result to the model update unit 232.

<3. Determination based on loading matrix>
The third determination unit 243 calculates the degree of divergence between the loading matrix of the latest model and the loading matrix of the active model, and determines whether or not the active model needs to be updated by comparing the calculated degree of divergence with a predetermined threshold. One method for evaluating the similarity of loading matrices is to define the degree of divergence based on the canonical angles of the two loading matrices. A loading matrix P _a in the active model, canonical angle between the loading matrix P _b of the latest model, it is known that can be obtained by singular value decomposition matrix A is generated as A = P _a ^T P _b ing. When the canonical angle thus obtained is θ _i (i = 1, 2,..., P), the divergence degree of the loading matrix can be defined by a weighted sum of the canonical angles θ _i . Here, if the weight of each canonical angle θ _i is w _i (i = 1, 2,..., P), the divergence dP of the loading matrix is expressed by the following equation (22).

Here, the weight w _i corresponding to each canonical angle θ _i can be defined by a weight proportional to the contribution vector of the principal component, for example. Specifically, the weight w _i can be defined as w _i = k × λ _i using a predetermined coefficient k. In Equation (22), it is desirable to normalize the loading matrix so that the range of dP is 0-1. In addition, in the definition of the degree of divergence using such a canonical angle, there is a possibility that the calculation load increases due to the complexity of the singular value decomposition of the matrix A. In such a case, the divergence degree can be defined by a norm like a position parameter or a scale parameter. In this case, as the simplest example, one vector in which p column vectors (hereinafter referred to as “loading vectors”) of the loading matrix are arranged is constituted by the active model and the latest model, respectively, and the norm between these vectors is set. Can be defined as the degree of divergence. The divergence degree in this case can be defined as the following equation (23) using, for example, the L2 norm.

In Expression (23), p _a ′ and p _b ′ are vectors obtained by normalizing vectors in which column vectors of P _a and P _b are vertically arranged so that the norm is 1. When the divergence degree is defined by a norm as in Expression (23), the divergence degree can be calculated only by simple algebraic calculation without performing complicated processing such as singular value decomposition. Incidentally, similarly to Equation (22) in the case of defined norm deviance as equation (23), by weighting corresponding to the contribution amount vector λ for each column vector of P _a and P _b Also good.

  Similar to the first determination unit 241 and the second determination unit 242, the third determination unit 243 determines that the active model needs to be updated when the deviation degree of the loading matrix exceeds the threshold, and the deviation of the loading matrix When the degree is less than or equal to the threshold value, it is determined that the update of the active model is unnecessary. The third determination unit 243 determines whether or not the active model needs to be updated based on the loading matrix in accordance with the latest model construction cycle, and outputs the determination result to the model update unit 232.

<4. Judgment based on contribution vector>
The fourth determination unit 244 determines the necessity of updating the active model by calculating a divergence between the contribution vector of the latest model and the contribution vector of the active model and comparing the calculated divergence with a predetermined threshold. To do. In this case, the divergence degree of the contribution amount vector can be defined by a norm similarly to the first determination unit 241 and the third determination unit 243. However, each element of the contribution amount vector corresponds to each loading vector direction in the loading matrix. In some cases, evaluation by the norm defined by the vector difference is not preferable.

  By the way, as described above, the contribution amount vector is also used for determining the truncation order p of the construction data used for construction of the abnormality diagnosis model. For example, the truncation order is determined by the sum of contribution amounts of process variables corresponding to the process data to be used. This is because the total sum of contribution amounts (hereinafter referred to as “cumulative contribution amount”) is considered to represent the satisfaction of the process variables necessary for the analysis (explanation) of the event. In this way, the contribution amount is a meaningful parameter that is compared with the sum rather than the difference. Therefore, here, the divergence degree based on the contribution amount vector is defined as the following Expression (24).

In Equation (24), λ _a ′ represents the contribution vector of the active model, and λ _b ′ represents the contribution vector of the latest model. When such a divergence degree is defined, the value range of the divergence degree is not known, and it may be difficult to set a threshold value. Therefore, in such a case, the divergence degree may be defined by a cumulative contribution rate instead of the contribution amount vector. The degree of divergence in this case is expressed as the following equation (25). The cumulative contribution rate is a ratio of the cumulative contribution amount of process variables used for analysis to the cumulative contribution amount of all process variables.

In equation (25), C _{λ — a} (_a means that a is a subscript of λ) represents the cumulative contribution ratio for the loading matrix of the active model, and C _{λ —} b is the loading matrix of the latest model. Represents the cumulative contribution rate for. Usually, the truncation order p is set so that the cumulative contribution ratio is about 0.8 to 0.9 (80% to 90%). Since the cumulative contribution rate defined in this way is a value between 0 and 1, it is easy to set a threshold value. In this case, the threshold value α of the divergence degree of the cumulative contribution rate can be set as a value in the range of 0-1. For example, α = 0.1 may be set in order to detect a deviation in the cumulative contribution rate of 10%.

  Similar to the first determination unit 241 to the third determination unit 243, the fourth determination unit 244 determines that the active model needs to be updated when the divergence degree based on the contribution amount vector exceeds the threshold, and the contribution amount When the divergence degree based on the vector is equal to or less than the threshold value, it is determined that the active model is not required to be updated. The fourth determination unit 244 determines whether or not the active model needs to be updated based on the contribution amount vector in accordance with the construction cycle of the latest model, and outputs the determination result to the model update unit 232.

  As described above, the method of evaluating the degree of divergence between the active model and the latest model by mainly defining the vector norm has been described. However, the definition of the degree of divergence is not limited by the vector norm. The divergence degree may be defined by any other method as long as it is defined as an index value indicating the degree of divergence between data sets. For example, the degree of divergence may be defined using index values indicating various similarities and abnormalities known in the fields of machine learning and multivariate statistical analysis. In any case, the degree of divergence may be arbitrarily designed so as to satisfy various items required according to the process to be diagnosed, such as accuracy and performance required for processing, and ease of understanding for the user.

  The model update unit 232 makes a final determination (hereinafter referred to as “final determination”) as to whether or not the active model needs to be updated based on the determination results of the first determination unit 241 to the fourth determination unit 244. The model update unit 232 updates the active model with the latest model when it is determined in the final determination that the active model needs to be updated. Note that what kind of final determination result the model update unit 232 outputs based on the determination results of the first determination unit 241 to the fourth determination unit 244 is arbitrary. For example, the model update unit 232 may update the active model when all of the first determination unit 241 to the fourth determination unit 244 determine that the active model needs to be updated, or the first determination unit 241. The active model may be updated when at least one of the fourth determination unit 244 determines that the active model needs to be updated.

  The process diagnosis apparatus 2 according to the first embodiment configured as described above has a difference between the model construction unit 222 that constructs an abnormality diagnosis model at a predetermined construction cycle based on the construction data, the active model, and the latest model. The update determination unit 231 that determines the necessity of updating the active model based on the calculated divergence degree, and the latest model according to the determination result of the update determination unit 231. And a model updating unit 232 for updating. By providing such a configuration, the process diagnosis apparatus 2 can update the abnormality diagnosis model at a more appropriate timing.

(Second Embodiment)
FIG. 3 is a block diagram illustrating details of a functional configuration of the process diagnostic apparatus 2a according to the second embodiment. The process diagnostic apparatus 2a includes a pre-application model buffer unit 212 and a previous model buffer unit 213, a point provided with an update determination unit 231a instead of the update determination unit 231, and a model update unit 232a instead of the model update unit 232. It differs from the process diagnostic apparatus 2 of 1st Embodiment by the point provided.

  The pre-application model buffer unit 212 and the previous model buffer unit 213 include a storage device such as a magnetic hard disk device or a semiconductor storage device. The pre-application model buffer unit 212 stores the latest model that satisfies a predetermined condition as a pre-application model. As the initial value of the pre-application model, the latest model constructed in the first construction cycle after the start of the operation of the device itself may be set, or the initial value of the active model may be set. The previous model buffer unit 213 stores the active model immediately before being updated to the current active model as the previous model. The initial value of the active model is set as the initial value of the previous model.

  The update determination unit 231a determines whether the pre-application model needs to be updated based on the latest model and the pre-application model, and determines whether the active model needs to be updated based on the pre-application model and the active model. Similar to the first embodiment, the update determination unit 231a uses the divergence degree of each index value (specifically, the position parameter, the scale parameter, the loading matrix, and the contribution vector) between the models to be compared. Based on the above, it is determined whether or not the pre-application model and the active model need to be updated.

  The model update unit 232a updates the pre-application model with the latest model according to the determination result of whether or not the pre-application model needs to be updated, and updates the active model with the pre-application model according to the determination result of whether or not the active model needs to be updated. To do. The model updating unit 232a has a function of updating the active model with the previous model based on a predetermined condition. Conditions for the model updating unit 232a to update the active model to the previous model may be arbitrarily set according to the necessity of returning the active model to the previous model. For example, the model update unit 232a may execute the update of the active model to the previous model according to the update operation of the user, or the abnormality detection frequency of the abnormality diagnosis unit 208 after the update of the active model in the pre-application model May be executed when a predetermined threshold is exceeded.

Process diagnostic device 2a of the second embodiment constructed in this manner, with determining the update necessity before apply model in accordance with the deviation degree of the applied pre-model the latest model, and before application model, active An update determination unit 231a that determines whether the active model needs to be updated based on the degree of deviation from the model is provided. By providing such a configuration, the process diagnosis apparatus 2a can update the abnormality diagnosis model in accordance with a longer-term process state change. Furthermore, the process diagnostic apparatus 2a can update the active model with the pre-application model instead of the latest model, so that the abnormality diagnosis standard based on the abnormality diagnosis model can be more gently changed with respect to the process state change. . Therefore, the process diagnostic apparatus 2a can suppress an excessive detection of abnormality of the process to be diagnosed.

  In addition, the process diagnosis apparatus 2a includes a model update unit 232a that updates the active model with the previous model based on a predetermined condition. By providing such a configuration, the process diagnosis apparatus 2a can quickly resolve the problems caused by the update of the active model by returning the active model to the previous model.

(Third embodiment)
FIG. 4 is a block diagram illustrating details of a functional configuration of the process diagnostic apparatus 2b according to the third embodiment. The process diagnostic apparatus 2b is different from the process diagnostic apparatus 2a of the third embodiment in that it further includes an operation change monitoring unit 214. The operation change monitoring unit 214 has a function of monitoring a change in operation for a diagnosis target process. Specifically, the operation change monitoring unit 214 detects a change in operation of the target process by monitoring a change in the position parameter.

  For example, when an operation change is intentionally performed in a plant or the like, a process variable serving as a process operation amount often changes abruptly. In particular, in the above-described advanced sewage treatment process, the measured values of the operation amount sensors of the first settling pond excess sludge extraction pump 111, the blower 112, the circulation pump 113, the return sludge pump 114, and the final settling pond excess sludge extraction pump 115 are greatly changed. Often appears. For example, when the set value of the sludge return rate is changed in the advanced sewage treatment process, the return flow rate measured by the flow rate sensor of the return sludge pump 114 greatly changes in a stepped manner. The process state gradually changes by such setting change. Therefore, the process abnormality diagnosis model is updated according to a change in the state of the process by the update method described in the first embodiment or the second embodiment.

  However, immediately after the operation is changed, only the return flow rate changes, and the state of the process may not change so much. For example, in a plant that realizes a water treatment process such as the above-described advanced sewage treatment process, the treatment status of the water to be treated is often determined by the properties of sludge measured by MLSS or the like. The change in is very gradual. Specifically, sludge properties and the like change daily or weekly. Therefore, in a plant that takes a long time to change the state of the process, there is no large change in the state of the process, and the state in which only the position parameter of the manipulated variable (return flow rate) is greatly changed continues for a while.

Therefore, in such a plant, if the same active model is continuously used, the Q statistic and the T ² statistic may be detected as abnormal continuously. Abnormalities detected due to such operational changes are not detected as time passes by updating the abnormality diagnosis model with the above update method. Since the position parameter μ used in is an average or representative value of the process data in the predetermined data extraction period L, at least the time of the data extraction period L is required until no abnormality is detected. Therefore, the abnormality is continuously detected from the time when the operation is changed until the time of about period L elapses. For these reasons, in the diagnosis of the process state, it is desirable to be able to detect whether or not there is an operation change in addition to detecting an abnormality. Further, in order to determine whether or not the detection of such an abnormality is caused by the operation change, it is desirable that the presence or absence of the operation change is detected in a cycle shorter than the period L.

Therefore, the operation changing monitor 214, a change of position parameter shorter than build period T _L of the latest models, and to monitor the period T _H or more changes in the position parameter length of the period T _M of the abnormality diagnosis. That is, T _H ≦ T _M <T _L. For example, when T _H = 1 minute and T _L = 1 day, the position parameter is monitored in a cycle ranging from 1 minute to 1 day. Incidentally, the monitoring period, it is preferable close to the period T _H of the abnormality diagnosis as possible. Therefore, the following _description assumes that _T M = T _H.

Specifically, when the latest model is constructed, the operation change monitoring unit 214 adds a period _TL that is the construction cycle of the latest model in addition to the position parameter μ for the period L that is the construction data extraction period. The position parameter μ _{c of} is calculated. For example, when the position parameter μ is an average value of L = 7 days, the position parameter μ _c is calculated as T _L = an average value of one day.

On the other hand, the operation change monitoring unit 214 determines, for each abnormality diagnosis period T _H (= monitoring period T _M ), the degree of divergence between the process data acquired during the period T _H and the position parameter μ _c. . Specifically, the operation change monitoring unit 214 determines the divergence degree of the position parameter by the following equation (26).

In the equation (26), x (t) means t-th process data among the process data acquired in the cycle T _H. Here, since x (t) is an instantaneous value and is considered to have a large variation, for example, the value of k is preferably set to about k = 3 to 6.

  When the process data and the position parameter are sufficiently deviated (that is, the equation (26) is true), the operation change monitoring unit 214 determines that there is a possibility that the operation change has been performed. However, since x (t) is an instantaneous value, it is likely to be affected by an abnormal value or the like, and a situation may occur where Expression (26) is satisfied instantaneously. For this reason, if the occurrence of an operation change is estimated using only Equation (26), the occurrence of an operation change may be erroneously detected. Therefore, the operation change monitoring unit 214 determines whether or not an operation change has occurred depending on whether or not a situation satisfying the equation (26) (hereinafter referred to as “divergence state”) continues for a predetermined period or longer.

Here, the threshold of the duration time of the divergence state is defined as T _J = C _max × T _M. This means that it is determined that an operation change has occurred when the divergence state is continuously detected for C _max diagnosis periods T _M (= T _H ). In this case, the operation change monitoring unit 214 initializes the counter C to C = 1 when the divergence state is detected for the first time from the undetected state, and every time the divergence state is detected continuously in the subsequent diagnosis cycle. Counter C is incremented. The operation change monitoring unit 214 determines the determination of the divergence state after the initial detection by the following equation (27) different from the equation (26).

In the formula (27), μ _d is the location parameter of the process data in each monitoring period T _M. What is important here is that in Equation (26), it is determined that the deviation degree of the position parameter exceeds the threshold value k, whereas in Equation (27), the deviation degree of the position parameter is a predetermined threshold value k. It is judging that 'is not exceeded. This predetermined threshold value k ′ needs to be set to be equal to or less than the threshold value k in Expression (26), and is preferably set to about k ′ = 0.5 k. The reason for this will be described with reference to FIG.

FIG. 5 is a diagram illustrating a specific example of a change in process data due to operation change. The horizontal axis in FIG. 5 represents time, and the vertical axis represents process data values (measured values). As described above, the position parameter greatly changes before and after the operation change. Figure 5 is a location parameter number mu by changing operation at time t ₀ is represents a state in which changes to the position mother several mu _d. Times t ₁ , t _2, and t ₃ shown on the time axis represent examples of the update timing (that is, the construction cycle) of the position parameter μ. In this case, the occurrence of the offset state by operational changes made at time t _0, based on the equation (26), is detected at the timing, for example, time t _4. In this case, if the divergence degree of the measured value is determined with the same position parameter μ after the first detection of the divergence state, the next update timing (t _{3 in} FIG. 5) of the position parameter μ is reached. Therefore, it is always determined that an operation change has occurred. Therefore, after the first detection of the divergence state, the operation change monitoring unit 214 determines the divergence degree of the measurement value by comparison with the position parameter μ _d for each monitoring period.

If the threshold value k ′ of the divergence degree after the initial detection is k ′ = k, the divergence degree of the measured values after the initial detection is set to the same fluctuation width (that is, kσ) as that before the operation change. Will be judged. In this case, there is a possibility that an instantaneous measurement value fluctuation is erroneously detected as an operation change. Therefore, in order to more reliably determine that the operation change has been performed (that is, the position parameter has changed), it is preferable to set k ′ to about k ′ = 0.5k. By threshold k after the first time detection 'is set in this manner, the operation changing monitor 214 does not continue to exist at the position closer to the position mother several mu _d than the measured value after the first time the detection position base number mu As long as the operation change is not determined, it is possible to detect the operation change more accurately. Note that μ _d updated by the operation change monitoring unit 214 for each monitoring period is calculated as the following equation (28).

Expression (28) represents that the position parameter of the measurement value after the first detection is sequentially updated with the measurement value after the first detection. Operation Monitoring unit 214 determines that the result of executing repeatedly the above determination for each monitoring period T _M, operational changes when a C = C _max were made. Then, the operation change monitoring unit 214 updates the position parameter μ of the active model with the position parameter μ _d at the time of detection in response to the detection of the operation change.

Note that the monitoring cycle T _M can be arbitrarily set between T _H <T _M <T _L. In this case, by replacing x (t) of formula (26) and (27) to the average value of the monitoring period T _M, it is possible to detect the change operation in the same manner. Here, when the monitoring cycle T _M is set such that T _M > T _L / 2, the counter C may not change from the initial value. Therefore, in such a case, the divergence degree may be determined only by the equation (26) without performing the divergence degree determination by the equation (27).

In addition, the position parameter may be monitored by setting μ _c calculated for each construction cycle _TL to be μ _c = μ. However, in this case, the update cycle of μ _c becomes long, and thus the deviation of the measured value is Therefore, it is difficult to distinguish between a phenomenon that changes slowly, such as a drift phenomenon, and a phenomenon that changes suddenly, such as a change in operation. The operational change is a change in the intended process, whereas the drift phenomenon is generally considered abnormal. Therefore, by close to update period T _M possible unless diagnostic period T _H as described above, the mutual detectable separating operation changes from the drift phenomena is desirable. However, when it is not necessary to detect the operation change separately from the drift phenomenon, the position parameter may be monitored by setting μ _c = μ.

The drift phenomenon can be detected by the abnormality diagnosis unit 208 by updating the abnormality diagnosis model at the construction cycle _TL . Therefore, in the process diagnosis apparatus 2b capable of such an abnormality diagnosis, the operation change does not necessarily have to be detected separately from the drift phenomenon. However, abnormality when it is desired to detect a drift phenomenon at a timing earlier than the update of the diagnostic model, the diagnostic period T shift the following equation with _H each of the acquired are mu _c and mu is obtained for each construct period T _L The drift phenomenon may be detected by determining in (29).

The determination of the formula (29), by executing in accordance with the monitoring period T _M of operational changes, it is possible to detect the drift phenomenon at an earlier timing. When the drift phenomenon is detected by the equation (29), the operation change monitoring unit 214 may update the position parameter μ of the active model with the position parameter μ _c when the drift phenomenon is detected.

  The process diagnostic apparatus 2b according to the third embodiment configured as described above has a position parameter of the latest model updated every construction period and a position parameter of time series data obtained every period shorter than the construction period. And an operation change monitoring unit 214 that detects a sudden change in process data due to a process operation change or the like based on the degree of deviation. By providing such a configuration, the process diagnosis apparatus 2b can detect a change in the process state due to an operation change or the like earlier than the construction cycle.

  It is desirable that the active model is updated when the operation change or the drift phenomenon is detected, also for the pre-application model. By updating the pre-application model in accordance with the update of the active model, it is possible to prevent the position parameter of the active model from being returned to the original position parameter when the abnormality diagnosis model is updated.

(Fourth embodiment)
FIG. 6 is a block diagram illustrating details of a functional configuration of the process diagnostic apparatus 2c according to the fourth embodiment. The process diagnostic apparatus 2c is different from the process diagnostic apparatus 2b of the third embodiment in that it further includes a validity determination unit 215. The validity determination unit 215 determines the validity of the active model, and updates the active model with the previous model according to the determination result.

Specifically, the validity determination unit 215 acquires a diagnosis result for the construction cycle _TL from the abnormality diagnosis result accumulated in the diagnosis result storage unit 209 at the same cycle as the construction cycle _TL of the latest model. Specifically, the diagnosis result is a Q statistic and a T ² statistic generated from the diagnosis data in the past diagnosis. The validity determination unit 215 calculates the Q parameter and the position parameter of the T ² statistic for the acquired construction period _TL . As described above, the position parameter acquired here is also a representative value such as a median value or a pruning average. The validity determination unit 215 calculates the position parameter of the Q statistic as Q _med and the position parameter of the T ² statistic as T ² _med .

On the other hand, the validity determination unit 215 includes the position parameters of the Q statistic and the T ² statistic calculated in the latest construction cycle _TL , and thresholds (Q _lim and A comparison with T ² _lim ) is performed in response to the update of the active model. When at least one of Q _{med and} T ² _med exceeds the threshold value, it is considered that a lot of process data indicates an abnormal value. Therefore, in this case, the validity determination unit 215 determines that the active model being applied to the abnormality diagnosis is not appropriate. When it is determined that the active model being applied is not appropriate, the validity determination unit 215 updates the active model with the previous model. In addition, the validity determination unit 215 prevents the previous active model from being updated to an inappropriate abnormality diagnosis model again by the pre-application model, so that when the active model is reverted, the previous model is also applied to the previous model. Update with.

The process diagnostic apparatus 2c of the fourth embodiment configured as described above determines the validity of the active model based on the divergence between the position parameters of the Q statistic and the T ² statistic before and after the update of the active model. A validity determination unit 215 for determination is provided. By providing such a configuration, the process diagnosis apparatus 2c can diagnose the validity of the updated active model.

Hereinafter, modified examples of the process diagnosis apparatus of the embodiment will be described.
The operation change monitoring unit 214 may be provided in the process diagnostic apparatus 2 according to the first embodiment. Similarly, the validity determination unit 215 may be provided in the process diagnostic apparatus 2 of the first embodiment. Further, the process diagnosis apparatus 2c of the fourth embodiment may be configured as an apparatus that does not include the operation change monitoring unit 214. The process diagnosis apparatus of each of the above-described embodiments does not necessarily have to be configured as a single apparatus, and may be configured as a process diagnosis system by a plurality of apparatuses having each function unit of the process diagnosis apparatus. In addition, the process to be diagnosed by the process diagnostic apparatus of the embodiment may be any process other than the above-described advanced sewage treatment process 1.

  According to at least one embodiment described above, a model construction unit that constructs an abnormality diagnosis model at a predetermined construction cycle based on construction data, an update between the active model and the latest model is shorter than the construction cycle. An update determination unit that calculates the period and calculates whether or not the active model needs to be updated based on the calculated degree of divergence, and a model update unit that updates the active model with the latest model according to the determination result of the update determination unit Thus, the abnormality diagnosis model can be updated at a more appropriate timing.

  Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 1 ... Advanced sewage treatment process, 101 ... First sedimentation tank, 102 ... Anaerobic tank, 103 ... Anoxic tank, 104 ... Aerobic tank, 105 ... Final sedimentation tank, 105 ... Final sedimentation tank, 111 ... First sedimentation tank excess sludge extraction Pump, 112 ... Blower, 113 ... Circulation pump, 114 ... Return sludge pump, 115 ... Final sedimentation basin excess sludge extraction pump, 121 ... Rainfall sensor, 122 ... Sewage inflow sensor, 123 ... Inflow TN sensor, 124 ... Inflow TP sensor 125 ... Inflow organic substance sensor, 126 ... ORP sensor, 127 ... Anaerobic tank pH sensor, 128 ... Oxygen-free tank ORP sensor, 129 ... Oxygen-free tank pH sensor, 130 ... Phosphate sensor, 131 ... DO sensor, 132 ... Ammonia sensor DESCRIPTION OF SYMBOLS 133 ... MLSS sensor, 134 ... Water temperature sensor, 135 ... Excess sludge SS sensor, 136 ... Discharge SS sensor DESCRIPTION OF SYMBOLS 137 ... Sludge interface sensor, 138 ... Sewage discharge sensor, 139 ... Release TN sensor, 140 ... Release TP sensor, 141 ... Release organic sensor, 2, 2a, 2b, 2c ... Process diagnostic device, 201 ... Data collection unit, 202 ... Collected data storage unit, 203 ... Data extraction unit, 204 ... Abnormal diagnosis model construction unit, 221 ... Normalization unit, 222 ... Model construction unit, 205 ... Latest model buffer unit, 206 ... Diagnostic model storage unit, 207 ... Active Model buffer unit 208 ... Abnormality diagnosis unit 209 ... Diagnosis result storage unit 210 ... Abnormality diagnosis model update unit 211 ... Diagnosis result provision unit 212 ... Pre-application model buffer unit 213 ... Previous model buffer unit 214 ... Operation Change monitoring unit, 215... Validity determination unit, 231, 231 a... Update determination unit, 232, 232 a. 241 ... first determination unit, 242 ... second determination unit, 243 ... third determining unit, 244 ... fourth determination unit

Claims (9)

Based on an abnormality diagnosis model constructed by multivariate analysis of various time-series data indicating the state of a process to be diagnosed, a process diagnosis apparatus that detects an abnormality of the process at a predetermined diagnosis cycle,
An abnormality diagnosis model construction unit that constructs the abnormality diagnosis model at a predetermined construction cycle based on the time series data;
Degree of divergence between an active model that is an abnormality diagnosis model used for abnormality detection of the process and a pre-application model that is an abnormality diagnosis model constructed after the active model is constructed by the abnormality diagnosis model construction unit An update determination unit that calculates an update cycle shorter than the construction cycle, and determines whether or not the active model needs to be updated based on the calculated degree of divergence;
According to the determination result of the update determination unit, a model update unit that updates the active model with the pre-application model;
A process diagnostic apparatus comprising: The update determination unit determines whether or not the active model needs to be updated based on a degree of deviation between the active model and the latest model that is the latest pre-application model.
The process diagnostic apparatus according to claim 1. The update determination unit determines whether or not the pre-application model needs to be updated according to the degree of divergence between the pre-application model and the latest model, and based on the degree of divergence between the pre-application model and the active model. Determine whether the active model needs to be updated ,
The model update unit updates the pre-application model with the latest model according to the determination result of the update determination unit, and updates the active model with the pre-application model.
The process diagnostic apparatus according to claim 2. The abnormality diagnosis model includes a position parameter of time-series data for a predetermined period,
Based on the degree of divergence between the position parameter of the latest model updated for each construction period and the position parameter of time series data acquired for each period shorter than the construction period, the process change, etc. It further includes an operation change monitoring unit that detects abrupt changes in time series data.
The process diagnostic apparatus according to any one of claims 1 to 3. The operation change monitoring unit, when the sudden change is detected, updates the position parameter of the active model or the pre-application model with the position parameter of the time series data,
The process diagnosis apparatus according to claim 4. The abnormality diagnosis model includes a Q statistic and a T2 statistic based on time-series data for a predetermined period,
A validity determination unit that determines the validity of the active model based on the degree of divergence between the position parameters of the Q statistic and the T2 statistic before and after the update of the active model;
The process diagnostic apparatus according to any one of claims 1 to 5. It said model updating unit, in accordance with the judgment result of the validity determination unit, the current active model and current applied prior model, to update the previous model is active model before updating,
The process diagnosis apparatus according to claim 6. A process diagnosis method performed by a process diagnosis apparatus that detects an abnormality of the process at a predetermined diagnosis cycle based on an abnormality diagnosis model constructed by multivariate analysis of various time-series data indicating the state of a process to be diagnosed. And
A model construction step of constructing the abnormality diagnosis model at a predetermined construction cycle based on the time series data;
An active model is abnormal diagnosis model used in the anomaly detection of the process, the degree of separation between the application predecessor active model is abnormal diagnosis model built after being built in the previous liver del building steps An update determination step for calculating an update cycle shorter than the construction cycle and determining whether the active model needs to be updated based on the calculated degree of divergence;
According to the determination result in the update determination step, a model update step for updating the active model with the pre-application model;
A process diagnostic method comprising: For a computer that functions as a process diagnosis device that detects an abnormality of the process at a predetermined diagnosis cycle based on an abnormality diagnosis model constructed by multivariate analysis of various time-series data indicating the state of a process to be diagnosed ,
A model construction step of constructing the abnormality diagnosis model at a predetermined construction cycle based on the time series data;
An active model is abnormal diagnosis model used in the anomaly detection of the process, the degree of separation between the application predecessor active model is abnormal diagnosis model built after being built in the previous liver del building steps An update determination step for calculating an update cycle shorter than the construction cycle and determining whether the active model needs to be updated based on the calculated degree of divergence;
According to the determination result in the update determination step, a model update step for updating the active model with the pre-application model;
A computer program for running.

Images