JP7344935B2 - Manhole pump monitoring system - Google Patents

Description

The present invention relates to a manhole pump monitoring system for detecting abnormalities in manhole pumps and effectively utilizing deterioration prediction information.

Approximately 47,000 manhole pumps for discharging sewage are installed across the country. Small and medium-sized local governments that own many manhole pumps are facing issues such as lack of human resources and financial difficulties in maintenance and stock management. In order to fundamentally improve these, it is necessary to realize appropriate stock management, and management of the three elements of "people,""materials," and "money" is essential.
Manhole pumps often exhibit unique behavior depending on where they are installed. For this reason, skilled technicians with special skills are on hand 24 hours a day, 365 days a year to maintain and manage the facilities, including identifying and implementing operational controls according to the actual inflow volume and responding to emergencies. We are responding accordingly. The number of such engineers is decreasing due to the aging of the population due to the recent population decline. In addition, it is necessary to secure and train next-generation engineers, but it is difficult to secure next-generation engineers in regional cities. Therefore, a system for sustainable sewerage management that ensures technical precision is required.
In the maintenance and management of manhole pumps, it is difficult to predict abnormalities and failures even if routine inspections are performed, so the main method is to respond to the occurrence of an event (corrective maintenance type). Therefore, by implementing early detection of abnormalities and failures of manhole pumps and equipment by installing measuring equipment, etc., risks can be mitigated through efficient operation of equipment and materials, and management toward condition monitoring maintenance (preventive maintenance type) can be implemented. need to shift to.
In the current state of stock management planning, it is difficult to identify facilities to be inspected, determine the inspection period, and predict the renewal period, so it is difficult to equalize the renewal of facilities developed during the baby boom period, and systematic renewal is difficult. and difficulty in securing project costs. For local governments that are having difficulty securing project costs, there is an increasing need to take measures such as restricting the use of sewage systems.
In order to avoid such a situation, it is necessary to smooth the budget and make business plans easier.
Daily and periodic inspection information is reported on paper by the private contractor to the administrator (maintenance department), and is shared and stored only in the maintenance department, so it is necessary to formulate systematic inspection and update plans. It is not being usefully used as basic information. As a result, it is assumed that the sewerage project is being operated inefficiently and it will be difficult to secure project costs, so it is necessary to aim for efficient management in the future.

JP2019-020264 JP2016-018526 JP2016-012158 JP2012-038298

In view of the above circumstances, it is an object of the present invention to provide a manhole pump monitoring system that improves efficiency and reduces maintenance costs, and further reduces the burden on field workers.

In order to achieve the above object, the manhole pump monitoring system according to the present invention includes an integrated monitoring cloud system that centrally manages information using a cloud system, and an abnormality detection discriminator using AI that utilizes the information to prevent abnormalities from occurring. The integrated monitoring cloud system includes a detection prediction system equipped with an AI engine that detects and predicts deterioration, and a stock management support system that supports the formulation of a stock management plan based on the information accumulated in the integrated monitoring cloud system. Both the AI engine-equipped detection prediction system and the stock management support system are implemented and constructed on the cloud, and the AI engine-equipped detection prediction system uses real-time measurement data from IoT devices to detect abnormalities in real time. The integrated monitoring cloud system receives and accumulates information on the abnormality detection and deterioration prediction, and the stock management support system performs condition monitoring based on the information accumulated in the integrated monitoring cloud system. The stock management plan and information on abnormality detection and deterioration prediction will be notified to maintenance companies and local governments as necessary , and the AI engine-equipped detection prediction system will A machine that uses learning data that includes abnormality data that is measurement data from the IoT device when an abnormality is detected within a predetermined detection period based on the length of the time period that is a night time period for maintenance and management contractors. The abnormality detection is performed by inputting the real-time measurement data to the abnormality detection classifier, which is a trained model generated by learning, and inferring whether or not an abnormality will occur within the detection period. That is.

According to the present invention, abnormality detection information before an abnormality occurs is sent to the maintenance management company and the local government that is ultimately responsible through the integrated monitoring cloud system that centrally manages the data. The maintenance company can then secure a predetermined amount of time before an abnormality occurs, reduce the burden on site workers, and perform repairs on site as necessary. In addition, the deterioration prediction information is accumulated and managed in the integrated monitoring cloud system, and furthermore, by receiving data related to deterioration prediction in the stock management support system, this information is utilized for pump health evaluation, risk consideration, etc. By sharing information with each local government, it is possible to provide efficient and continuous support, such as improving the accuracy of renovation plans, resolving human resource shortages, and reducing costs.

FIG. 1 is a schematic diagram of a manhole pump monitoring system. It is an explanatory diagram showing a flow of an abnormality detection classifier. FIG. 2 is a network diagram of an autoencoder. It is a graph showing the relationship between restoration error by an abnormality detection discriminator and warning/emergency dispatch. FIG. 3 is an explanatory diagram of training data of an abnormality determiner. It is a graph showing the number of abnormal and normal correct answers in the evaluation data of the trained neural network. It is a flowchart of pump deterioration prediction. FIG. 2 is an explanatory diagram showing the structure of a decision tree. It is a graph showing classification results using a decision tree model. It is a graph showing the remaining life of the pump. It is an explanatory diagram of a stock management support system. FIG. 2 is an explanatory diagram of data exchange using an API. It is a graph of an operation example of a manhole pump monitoring system.

Hereinafter, a manhole pump monitoring system according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic explanatory diagram showing an overview of this system. This monitoring system consists of three element systems: an integrated monitoring cloud system 1 that centrally manages information using a cloud system, an AI engine-equipped detection prediction system 2 that uses the information to construct an anomaly detection and discriminator 21 using AI, and A system consisting of stock management support system 3 will be constructed on a cloud environment. Although each elemental system is independent, it functions as a single "cloud AI system" by having a mechanism for exchanging necessary data with each other. The AI engine-equipped detection and prediction system 2 also includes a mechanism for importing and utilizing data from external systems and devices such as the IoT device 5 that measures the current and insulation resistance (leakage current) conditions attached to existing manhole pumps. , real-time measurement data from the IoT device 5 is used to detect abnormalities in the manhole pump and predict deterioration. Then, the integrated monitoring cloud system 1 receives and accumulates information on abnormality detection and deterioration prediction by the AI engine-equipped detection and prediction system 2. The stock management support system 3 then supports the appropriate formulation of a stock management plan through condition monitoring based on the abnormality detection and deterioration prediction information accumulated in the integrated monitoring cloud system 1. The stock management plan and information on the abnormality detection and deterioration prediction are notified to maintenance companies and local governments as necessary.

The abnormality detection of the manhole pump by the AI engine-equipped detection and prediction system 2 is performed as follows. In order to detect anomalies before they occur, it is necessary to construct an anomaly detection discriminator. The abnormality detection discriminator is for automatically distinguishing between abnormal current value data and normal current value data. FIG. 2 is an explanatory diagram showing the flow of the abnormality detection discriminator.

In order to construct an anomaly detection classifier, it is necessary to define an "abnormal" state. In general, "abnormalities" do not occur often, and it is complex and difficult to define the characteristics of "abnormalities" from time-series data obtained through real-time monitoring. Even if "abnormal" could be defined, most states would be biased toward "normal," and the proportion of "abnormal" to all states (hereinafter referred to as abnormal probability) would be extremely small. Therefore, when creating a classifier that distinguishes between "normal" and "abnormal" from data obtained through real-time monitoring, the probability of misidentification is greater than the probability of abnormality, and it is better to set all outputs as "normal" as a system. There is a high possibility that this will result in an evaluation. Therefore, instead of forcing the definition of "abnormality" from the time-series data obtained through real-time monitoring, we define "normal" from the daily report data or time-series data obtained from real-time monitoring, and identify conditions that deviate from that state. We adopted a method to avoid these problems by defining "abnormality." As a way to define "normal", we used an autoencoder, which is a neural network (a network consisting of a combination of layers and activation functions) that reproduces input data. Autoencoders are classified as unsupervised learning because they use the input data itself as the correct answer data. Then, the autoencoder uses machine learning to determine weighting so that the input feature vector and the output vector have the same value in a neural network with three or more layers that has an intermediate layer with a smaller dimension than the input layer and the output layer. It is a method of extracting the features of input data by dimensional compression, and is used to obtain the essence of data, such as by removing noise.

<Selection of number of operations 24 and operation time 25, creation of data set>
When selecting data to be input into the input layer of the autoencoder, we focused on and collected existing daily report data. However, it is not possible to use all the daily report data uniformly because the attributes of the pumps and other entries differ in each manhole pump facility. For example, if attributes such as new pump formats and outputs arise during nationwide expansion, the same system cannot be applied and versatility cannot be ensured. Therefore, items that may impede versatility such as pump attributes may be changed into input values. I decided not to include it. Therefore, we identified common data items in the various daily reports created during the general operation of manhole pumps, and identified the data items that affect whether the pump is normal or abnormal. We decided to select "Time 25" and use these two indicators. The data to be input into the autoencoder was data of the number of times of operation 24 and the operation time 25 for each hour for 24 consecutive hours. Then, in order to estimate whether the state is normal or abnormal for one hour after the target time, a data set of the number of driving times 24 and the driving time 25 for a total of 24 hours from the target time to 23 hours before is used. Created. Note that in order to provide versatility to the results calculated by the autoencoder using only two indicators, the number of driving times 24 and the driving time 25, daily report data over the past many years was collected and organized. This data set is input to the autoencoder and the input is restored.

<Learning by autoencoder, calculation of restoration error A>
A network diagram of the autoencoder is shown in FIG. This autoencoder consists of the following six parts. In addition, since the number of operations 24 and the operation time 25 have a difference in the range of possible numerical values by about two digits, the input processing for each was separated.

<Number of operations time series standardization unit 27>
Let the number of operations input vector CI be a 48-dimensional vector {CI1_23, CI2_23, CI1_22, CI2_22, CI1_21, CI2_21, ..., CI1_n, CI2_n, ..., CI1_2, CI2_2, CI1_1, CI2_1, CI1_0, CI2_0}. Here, CI1 and CI2 represent pump No. 1 and pump No. 2 at the same pump station, respectively, and _n represents n hours before the restoration target time. For example, if the restoration target time is around 21:00, CI2_23 is the number of times pump No. 2 was operated at around 22:00 the previous day (23 hours before 21:00), and CI1_0 is the number of times pump No. 2 was operated at around 21:00 on the previous day. .1 represents the number of operations 4 (times). Normally, pump No. 1 and pump No. 2 at the same pumping station are operated alternately, so it is thought that the values are usually equal. For this reason, regardless of pump No. 1 and pump No. 2, the overall average number of operations (times) is subtracted from each value to find the deviation, and the deviation is normalized by dividing by the maximum absolute value of the deviation. did. As a result, the range of each element of the normalized operation count input vector CI' becomes {-1≦CI'(1,2)_n≦1}.

<Operating time time series standardization unit 28>
Let the driving time input vector TI be a 48-dimensional vector {TI1_23,TI2_23,TI1_22,TI2_22, TI1_21,TI2_21,...,TI1_n,TI2_n,...,TI1_2,TI2_2,TI1_1,TI2_1,TI1_0,TI2_0}. Here, TI1 and TI2 represent pump No. 1 and pump No. 2, respectively, and _n represents n hours before the restoration target time. For example, if the restoration target time is around 21:00, TI2_23 is the operation time 5 (seconds) of pump No. 2 at around 22:00 the previous day (23 hours before 21:00), and TI1_0 is the operation time of pump No. 2 at around 21:00 on that day. .1 represents the operating time 5 (seconds). Normally, pump No. 1 and pump No. 2 at the same pumping station are operated alternately, so it is thought that the values are usually equal. For this reason, regardless of pump No. 1 and pump No. 2, the overall average operating time (seconds) is subtracted from each value to find the deviation, and the deviation is normalized by dividing by the maximum absolute value of the deviation. did. As a result, the range of the normalized operation time input vector TI' becomes {-1≦TI'(1,2)_n≦1}.

<Encoder section 29>
The normalized operation count input vector CI' and operation time input vector TI' are combined using the concatinate function to create a 2 × 48 matrix P', and a 4-layer convolution layer (kernel size: 3, number of maps :32) and activation function (LeakyReLU, Alpha: 0.1)) and two convolutional layers (convolution (kernel size: 3, number of maps: 16) and activation function (LeakyReLU, Alpha: 0.1)). The data was compressed (encoded) into a 16x2 matrix by passing it through a total of 6 layers (compression rate: 1/3).

<Decoder section 210>
The 16×2 matrix output from the encoder unit 29 is processed through two deconvolution layers (set of deconvolution (kernel size: 5, number of maps: 16) and activation function (LeakyReLU, Alpha: 0.1)), deconvolution layer ( Set of deconvolution (kernel size: 4, number of maps: 32) and activation function (LeakyReLU, Alpha: 0.1)), deconvolution layer (set of deconvolution (kernel size: 3, number of maps: 32) and activation function ( LeakyReLU, Alpha: 0.1)), deconvolution layer (set of deconvolution (kernel size: 4, number of maps: 32) and activation function (LeakyReLU, Alpha: 0.1)) and deconvolution layer (set of deconvolution (kernel size: 4, number of maps: 32) and activation function (LeakyReLU, Alpha: 0.1)) Outputs a 48x2 matrix through a total of 6 layers (size: 6, number of maps: 2)).

<Driving frequency time series restoration unit 211>
The first half of the 48×2 matrix generated by the decoder unit 210 is restored as the 48-dimensional normalized operation count output vector CO' (processing opposite to normalization: multiply by the maximum absolute value of the residual and add the average value) Then, obtain the 48-dimensional operation count output vector CO.

<Driving time time series restoration unit 212>
The latter half of the 48×2 matrix generated by the decoder unit 210 is used as the 48-dimensional normalized operating time output vector TO', and restored (processing opposite to normalization: multiply by the maximum absolute value of the residual and add the average value. ) and obtain the 48-dimensional operation count output vector TO.

A loss function (HuberLoss) was applied to the above autoencoder network, and Adam (α=0.001, β1=0.9, β2=0.999, ε=1e-8) was used as the optimizer. Training was performed with a batch size of 64,2000 epochs using training data (4,346,997) and evaluation data (711,258). For learning, Nvidia's DGX-Staion (GPU: Tesla V100) was used, and the deep learning framework was SONY's Neural Network Console version 1.6.7263.14761 and NNabla (Version 1.1.0, Build 190820052242).

When you input the number of operations 24 and operation time 25 for the past 24 hours into the above autoencoder and let it learn, the autoencoder will detect the difference between the input data and the autoencoder output (representing normal). The restoration error A is calculated, and the time period in which the restoration error A exceeds the threshold value D is defined as "abnormal" operation, which is different from normal operation, and the time period is automatically extracted as abnormal state time period data. did. Note that the restoration error A may be a ratio between input data to the autoencoder and output of the autoencoder. Further, the threshold value D of the restoration error A for distinguishing this abnormality is also a hyperparameter. Regarding the method for determining the threshold value D, a realistic threshold value D was determined by first determining a temporary "abnormality" and comparing it with past data (emergency dispatch records, etc.).

Here, if you input the time-series data (operation count 24 and operation time 25) obtained by real-time monitoring using an ammeter with IoT function into the autoencoder, the time-series data obtained by real-time monitoring will indicate that each time period is "normal". ” or “Abnormal” is output. In other words, the anomaly detection classifier calculates the restoration error A not only for daily report data but also for time-series data obtained through real-time monitoring, and determines the time period in which the restoration error A exceeds the threshold value D. It is defined as an "abnormal" operation that is different from normal operation, and the time period in question is automatically extracted as abnormal time period data.

In this way, using an autoencoder, we built an abnormality detection discriminator that can easily determine whether it is normal or abnormal using only two indicators: the number of operations 24 and the operation time 25, which can be easily obtained from the daily report. did. It is difficult to determine which data is normal or abnormal from a huge amount of past data by manually comparing it over a huge amount of time, but if you build an anomaly detection classifier like the one above, By simply inputting the above two indicators into the autoencoder, it is now possible to automatically determine whether something is normal or abnormal.

Incidentally, typical examples of alarms caused by pump abnormalities include "pump overload alarm" when a value exceeding 1 to 1.1 times the rated current continues for 3 to 5 seconds or more, The three main alarms are ``pump operation alarm'' when the pump is operated continuously for more than 20 minutes (standard), and ``pump failure alarm'' detected by the motor breaker for each pump, which accounts for 95% of all abnormalities. occupies . When this past warning/emergency dispatch data is input into the anomaly detection discriminator, as shown in Figure 4, the portion where the reconstruction error A exceeds 0.3, which is set as the threshold D, results in a warning/emergency dispatch. It was found that the results almost matched when an abnormality occurred, and that it was almost possible to detect an abnormality in a timely manner. Here, in FIG. 4, the horizontal axis represents the passage of time, and the vertical axis represents the restoration error A. A broken line A in the figure represents a threshold value D of the restoration error A. Furthermore, a black circle B indicates an abnormality that resulted in an emergency dispatch, and a white circle C indicates an abnormality that resulted in only a warning that did not result in an emergency dispatch.
In addition, in the built anomaly detection discriminator, the auto encoder detects abnormalities that are not linked to emergency dispatch or alarms due to phenomena that differ from normal pump movement (for example, inflow of unknown water, irregular drainage, etc.) It is thought that the image captured an abnormality that had been overlooked.

The abnormality detection discriminator detects an abnormality from the number of operations and operation time in the daily report data, and the current value at the time of abnormality is determined by linking the detected abnormality data and the current value data acquired in real time from the IoT device 5. Use this data as data to detect pump abnormalities before they occur. The method for detecting abnormalities is as follows.

The main purpose of the abnormality detection method of this embodiment is to reduce the need for field workers to be dispatched at night (assuming 15 hours from 18:00 to 9:00 the next day). To do this, it is necessary to detect signs of pump abnormalities occurring within 15 hours from now.
In this example, by creating training data and inputting it to the neural network for learning, an abnormality accuracy calculator within the predictive detection period is created, and the starting point is the evaluation reference time R2 calculated by the abnormality accuracy calculator. Pump abnormality is detected in advance with a probability J that the pump abnormality will occur within the predicted detection period E. The predicted detection period E is a fixed period starting from the evaluation reference time R2, during which the occurrence of pump abnormality is predicted, and the length of the predicted detection period E is determined when creating the training data. It is. In addition, hereinafter, the reference time for learning R1 and the reference time for evaluation R2 both refer to the current time, arbitrary judgment time, pump abnormality occurrence time, etc., which represent various events, and the reference time for learning R1 is the time used when creating the training data, and the evaluation reference time R2 is the time specified by the user when using the abnormality accuracy calculator. Furthermore, in creating the training data, the pump abnormality occurrence time refers to the occurrence time of the abnormality detected by the abnormality detection and discriminator.

In this embodiment, current value data obtained in real time using an ammeter with an IoT function (hereinafter referred to as an IoT ammeter) is used.
The IoT ammeter in this embodiment transmits 1200 pieces of current value data sampled every 0.05 seconds every minute.
In addition, hereinafter, current value time series data refers to a data string in which the current value data is arranged by time, and in this example, the current value data from the IoT ammeter is calculated from the current value data at a minimum interval of 1 minute (IoT The data are collected with the end point at an arbitrary learning reference time R1 or evaluation reference time R2 (the shortest interval of current value data transfer by an ammeter). One current value time-series data is generated every 1 minute or every 1 minute during a predetermined period P (24 hours in this embodiment) with the learning reference time R1 or any evaluation reference time R2 as the end point. It is configured as a sequence of current value data averaged over the pump operating time for each second (in this embodiment, a total of 24 hours worth of 1438 minutes and 120 seconds).
Here, the average current value data of the pump operating time for each 1 minute or 1 second is the 1 minute (in this example, 59.95 seconds from 0 second of the minute) measured by the IoT ammeter. The time at which the pump was operated (minute or second) out of 1,200 pieces of current value data (up to ) or 20 pieces of current value data for 1 second (from 0 seconds to 0.95 seconds in this example) This is the current value data that is averaged by dividing the total of the current values by the total number of data for only the current value data with a value other than zero, and it is calculated every minute or every second. Means representative current value data.
The sequence of average current value data (current value time series data) of the pump operating time every minute or every second for a predetermined period of P minutes means This is a data string in which the average current value data of the operating time is continued for a predetermined time P minutes. (In this example, a continuous 24-hour data string was used, with an average of 1438 pieces of operating time for each 1-minute increment, and 120 pieces of data for an average of 120 pieces of operating time for each 1-second increment.) In other words, the length of the predetermined period P is the data length of the current value time series data that is the input data to the abnormality accuracy calculator.

The abnormality accuracy calculator is an accuracy that is the probability that the pump will be in an abnormal state within the predictive detection period E starting from the evaluation reference time R2, based on current value time series data before the time when the pump abnormality occurs. This outputs J.
The abnormality accuracy calculator is constructed as follows using a neural network. First, create training data for the neural network to learn. FIG. 5 is an explanatory diagram of the teacher data.

The teacher data consists of input data, which is a set of current value time series data, and the state of the pump at that time (normal or abnormal), and the input data consists of normal time data 55 and abnormal time data 56. . Both the normal time data 55 and the abnormal time data 56 are selected from the current value time series data collected with the learning reference time R1 as the end point of the predetermined period P.
The abnormal time data 56 is current value time series data in which the end point of the predetermined period P (learning reference time R1) is included in the predicted detection period E (15 hours in this example) up to the time when the pump abnormality occurs. It is about. Therefore, for each actual pump abnormality case, it is possible to collect abnormality data 56 with 15 hours x 60 minutes = 900 times as the reference time R1 for learning. However, if all of these 900 pieces of data are used, it is thought that the contribution of the abnormal case to the judged abnormality will be too high, so about 25% (225 pieces in this example) of randomly selected data will be used. adopt.
The normal time data 55 includes all collected current value time series data other than the abnormal time data 56, that is, from the end point of the predetermined period P (learning reference time R1) to the predicted detection period E (15 in this embodiment). This refers to current value time-series data when no pump abnormality occurs until the time period (time) has elapsed. Normally, a large amount of normal data 55 can be collected because the number of abnormalities is small, but in order to obtain the ability to distinguish it from abnormal data 56, it is necessary to collect about twice the number of abnormal data (in this implementation). In this example, 450 normal data items are used for each abnormal data item.
In other words, the teacher data includes the state of whether an abnormality has occurred between the end point of the predetermined period P (learning reference time R1) and the elapsed time of the predicted detection period E, and the information extracted for the predetermined period P. This is a data matrix that includes multiple sets of current value time series data.

The pump facilities targeted for construction of this abnormality accuracy calculator are three facilities in Shimahongo, Tajima, and Eirakucho, and the abnormality data at each facility used for learning and verification is as shown in Tables 1 and 2. It is. Furthermore, as shown in Tables 1 and 2, the time of pump abnormality occurrence obtained from emergency dispatch records, etc., and the current value obtained in real time using an ammeter with IoT function that matches that time. By comparing the data, it can be determined that the real-time current value at the abnormality occurrence time is the current value data at the time of the abnormality occurrence.

The training data is used to train the neural network as follows.
First, we started training with a neural network that is easy to acquire features on a multi-scale due to the multi-layered SE structure with multiple stages of Squeeze-and-Excitation blocks, and used AutoML/NAS to create about 300 to 400 trained neural networks. Create automatically. Learning with AutoML/NAS refers to the use of software to automatically set parameters generally called hyperparameters set by engineers (AutoML), or to automatically change the network structure (NAS: Network This is a general term for the functions that automatically proceed with tuning work by performing automatic structure search).
The input items to the neural network are learning data including three or more abnormalities at the same pump station, the rated current of the pump, and the alert setting value. The alert setting value is a coefficient that indicates how many times the rated current of the pump is to be issued for an alert in the existing system. The abnormality accuracy calculator of the present invention can be constructed if there are three or more abnormal data items for learning per target pump, so it can be used in situations where there is little abnormal data in deep learning, which is generally said to require a large amount of data. However, since the amount of data is easy to secure, it is extremely useful.

The evaluation data shown in Table 2 is input into the 300 to 400 trained neural networks created in this way, and the neural network with the optimal number of normal/abnormal correct answers is selected. Note that when inputting the evaluation data, each period including the abnormality occurrence date and time that is not included in the teacher data was input as the evaluation reference time R2. FIG. 6 is a graph showing the number of abnormal and normal correct answers in the evaluation data of the trained neural network. Generally, it is impossible to find a complete neural network that gives all the correct answers, so a realistic neural network is selected using Pareto optimal solutions. Taking into account the characteristics of the pump facility and the demands of the site, we can create a neural network that emphasizes reducing misses (inside the circle with a dashed-dotted line in Figure 6) or a neural network that emphasizes reducing misses (within the circle with a dashed-double line in Figure 6). Select one of the networks. Note that when visualizing the characteristics of each neural network using a graph when selecting a neural network, displaying it on a logarithmic axis may eliminate the difficulty in viewing due to the difference in the number of normal/abnormal data. As described above, the abnormality accuracy calculator is constructed.
In addition, since the selected neural network is a Pareto optimal solution, when it is actually incorporated into an abnormality reporting application, it is necessary to detect abnormalities before they occur due to temporal variations in accuracy J caused by noise contained in input data. In order to suppress this problem, it is necessary to stabilize the accuracy J by removing noise using a method such as a moving average when operating the abnormality reporting application. The abnormality reporting application has functions such as calling and displaying the accuracy J calculated by the abnormality accuracy calculator, passing it to other sections, and reporting abnormality detection information.

The selected trained neural network (abnormality accuracy calculator) is loaded into the abnormality reporting application as a so-called weight file along with current value time series data for a predetermined period P up to evaluation reference time R2, which is input data. Then, the probability J that an abnormality will occur in the manhole pump between the evaluation reference time R2 and the elapse of the predicted detection period E is calculated. The final abnormality accuracy K is obtained by removing noise from this accuracy J using a method such as a moving average as described above using an abnormality reporting application. In this embodiment, the final abnormality accuracy K is obtained by taking a 15-hour moving average of the accuracy J. When this final abnormality accuracy K exceeds a threshold value T (60% in this embodiment), the abnormality notification application issues abnormality detection information. In this example, the final abnormality accuracy K was calculated every 30 minutes using real-time measurement results by an ammeter with an IoT function as input data.

Next, the accuracy rate of the constructed abnormality accuracy calculator will be described. The definition for calculating the hit rate is as follows. If there is an abnormality detection (forecast notification) with a final abnormality accuracy K of 60% or more in the time period before the existing system alert is issued, it is determined to be accurate. The accuracy rate is defined as the ratio of the number of cases in which the abnormality accuracy calculator correctly detects an abnormality compared to the alerts issued by the existing system.
Table 3 shows the accuracy rate of the abnormality accuracy calculator for the verification data in Table 2 above. Of the eight existing alerts issued, seven were accurate, for an accuracy rate of 87.5%.

Note that the various setting values for the abnormality detection described above are not limited to the above embodiments. For example, the predicted detection period E is not limited to 15 hours, but may be divided into multiple periods, such as 1 hour units each until 15 hours later. It's okay.

The abnormality detection information obtained as described above is sent to the maintenance management company 6 and the local government 7, which is ultimately responsible, through the integrated monitoring cloud system 1 that centrally manages the data. Then, the maintenance management company 6 can secure a predetermined period P before an abnormality occurs and perform repairs on-site as necessary.

Next, prediction of deterioration of manhole pumps will be explained.
Flowcharts S1 to S9 for predicting deterioration of manhole pumps (hereinafter referred to as pumps) are shown in FIG. In this embodiment, the lifespan of a pump is assumed to vary depending on operating conditions, environment, etc. even for the same type of equipment, so it is assumed that "the lifespan of a pump depends on the amount of work done in operation, not simply on the passage of time." Then, the pump life is predicted from the changes in the cumulative workload of each pump. In addition, by utilizing IoT devices, it is possible to monitor the condition of pumps in real time, saving labor on site inspection work, and using accumulated measurement data to predict changes in cumulative workload. A calculation model 73 is constructed. This remaining life calculation model 73 calculates the number of years remaining (remaining life) until the pump reaches its limit cumulative workload (cumulative workload at the time of pump life), and predicts the future renewal timing.

Here, in order to accurately calculate the remaining life from the cumulative amount of work, we use AI to calculate the remaining life of the pump, because as mentioned above, the lifespan of pumps varies depending on the equipment, and even for the same equipment, the lifespan varies depending on the operating conditions, environment, etc. A decision tree model 72 constructed through analysis was used. The decision tree model 72 classifies the operating status of each pump into “operating status that leads to short life (hereinafter referred to as “short life”)” or “operating status that leads to long life (hereinafter referred to as “long life”).” It is something. For each pump classified as long-lived or short-lived by the decision tree model 72, real-time measurement data by the IoT device is input into the remaining life calculation model 73, which is a model that returns the remaining life until the pump breaks down. , calculate the remaining life of the pump. This makes it possible to predict deterioration based on the actual operating conditions of the pump. The decision tree model 72 and remaining life calculation model 73 will be explained in detail below.

<Decision tree model 72>
The decision tree model 72 is an AI-based identification method with training data for clarifying and defining ambiguous characteristics of long-life or short-life judgments made by maintenance contractors (skilled engineers) based on experience. There are various methods for AI analysis to solve classification problems, but machine learning is used to predict pump deterioration, taking into account existing data items and uses (use of remaining life calculation results for stock management, optimal operation, etc.) Among these, we perform an analysis using a decision tree, which corresponds to the identification method with training data. The characteristics of this decision tree analysis are that it can handle a mixture of categorical variables and numerical variables;
Also, the tree structure allows for concise and clear output.

維持管理業者へのヒアリング結果として、ポンプ120基（60施設）に対する長命／短命の峻別データ（長命78基，短命42基）を得た。 In order to construct the decision tree model 72 for each regional management category (municipality, etc.), the training data 74 includes information on long-life/short-life judgments by maintenance companies for 60 pumps in Toyama City, which is one of the regional management categories. The results were learned by AI. The results of qualitative interviews with maintenance and management contractors were used to determine whether a pump was long-lived or short-lived. By creating this decision tree, maintenance and management contractors who are skilled engineers can easily grasp ambiguous features used when performing classification intuitively. The hearing on long-lived and short-lived life (creation of teacher data 74) was conducted based on the conditions shown in Table 4.

As a result of interviews with maintenance and management contractors, we obtained data on 120 pumps (60 facilities) to clearly distinguish between long-life and short-life pumps (78 long-lived pumps, 42 short-lived pumps).

Furthermore, in order to extract items and their values necessary for long-life/short-life classification, AI is made to learn a data set of element data 75 consisting of pump specifications and past daily report data shown in Table 5 as learning data. Features were extracted to classify each pump as long-lived or short-lived. In addition, we have also referred to items such as environmental conditions that are affected by special wastewater from factories, commercial facilities, hospitals, nursing care facilities, new housing, etc. Among the data set items, weighting factors were calculated using the AHP method for special drainage effects. The social impact X[1], which is the weighting coefficient, was quantified by interviewing maintenance management companies, and is 0.00 (not affected) < 0.03 (factory) < 0.05 (commercial facility including hot springs) <0.08 (hospitals/nursing facilities). Note that the number of types of data included in the data set is better as it increases, and is not limited to the data items listed in Table 5.

Based on the results of interviews with maintenance and management companies as described above, an analysis was conducted by regional management category, and the characteristics of "operating conditions that lead to short lifespans" were analyzed using decision trees (AI). As a result of aggregating the items of analysis data and creating a decision tree model 72, we found that "driving time", which has a particularly large feature value, and "special drainage effects (especially external factors from commercial businesses such as restaurants, hospitals, nursing care facilities, etc.)" It was found that the classification is based on two items: 1. Effect of clogging due to foreign matter, etc.). In addition, in facilities with long "operating hours" (high operating rates), the inflow volume is large and they are susceptible to the effects of scum in the tank, so this is considered to be the most important feature in determining deterioration. In addition, it is assumed that facilities that are affected by special drainage are particularly affected by external factors (the influence of clogging with foreign objects, etc.) from businesses such as restaurants, hospitals, nursing care facilities, etc.

As described above, the AI analyzes and extracts feature data from the long-lived/short-lived classification results by the maintenance contractor and each data set item, and the decision tree model 72 is constructed. Decision trees (tree structure) T1 to T9 of the decision tree model 72 obtained in this way are shown in FIG. The characteristics of the short-lived pump in the decision tree model 72 are as follows, and the higher the elements in the decision tree, the greater the characteristics.
Features of short-life pump:
(1) Annual operating time X[0] is 761.5 hours (approximately 2.1 hours/day) or more.
(2) The plant is affected by special drainage and the annual operating time X[0] is 210 hours (approximately 0.6 hours/day) or more.
(3) The facility is affected by wastewater from hospitals, nursing care facilities, and commercial facilities, with an annual operating time X[0] of 209.5h (approximately 0.6h/day) or less and a large social impact X[1]. .
Note that gini in FIG. 8 is the Gini coefficient, which is an index indicating impurity or inequality in decision tree analysis, and the decision tree is constructed using the Gini coefficient as an index. Specifically, when decision tree analysis is performed on many elements, the decision tree is constructed so that the value of the Gini coefficient that is automatically quantified becomes smaller. Furthermore, elements with large Gini coefficient values represent characteristics and common points that maintenance management companies classify into long-lived/short-lived. As shown by points P1 and P2 in the graph showing the decision tree analysis results in FIG. Classified.

Next, the accuracy of the decision tree model 72 was verified by extending the decision tree model 72 to other pumps within the same regional management division and calculating the correct answer rate. A method of verifying the decision tree model 72 created as the training data 74 using the 60 locations of IoT device installed pumps as described above will be described below. First, for 300 pumps where the IoT device 5 was not installed, a decision tree model 72 was used to clearly distinguish between long-lived and short-lived pumps. For the pumps without the IoT device installed, we conducted another interview with the maintenance company to determine whether the pumps had a long life or a short life. Then, the decision tree model 72 and the correct answer rate of the maintenance contractor were calculated and evaluated.

The above verification results are shown below. As shown in decision tree model O in Table 6, decision tree model 72 was able to clearly distinguish between long life and short life with 85% accuracy. In addition, as a result of case-by-case analysis of the three decision tree models M, N, and O as shown below, we found that, like models M and N, abnormalities (number of dispatches and special drainage effects) caused by clogging with scum etc. We found that it is related to the driving conditions that lead to short lifespans.
・Decision tree model M: Cases classified only by driving time.
・Decision tree model N: Cases classified by driving time and dispatch frequency.
・Decision tree model O: Cases classified based on operating time and special drainage effects.

As described above, by constructing the decision tree model 72, it is possible to know the items to be noted in the maintenance and management of each pump, so for short-lived pumps in particular, for example, those that operate for a long time may need to review the float water level settings, etc. It is possible to take advantage of this and take measures to extend its lifespan. Furthermore, in the renewal plan, it becomes possible to formulate an improvement plan so that the operation can be shifted to a long-life pump. In addition, it is considered that feature data for classifying pumps as long-lived or short-lived can be extracted by similarly constructing the decision tree model 72 in other regional management classifications, and therefore has excellent versatility.

<Remaining life calculation model 73>
Next, using the results of the long-life/short-life classification of each pump using the decision tree model 72, the remaining life calculation model 73 calculates the remaining life.

The pump deterioration prediction method in this example is based on the pump workload being defined as (voltage) x (current) x (operating time). The remaining life H, which is the number of years remaining until the pump reaches the end of its life, is calculated by assuming the limit cumulative work G of the pump based on past results. By installing the IoT device 5 in each pump equipment and monitoring the current value in real time, changes in the amount of work (characteristics) of each pump can be grasped. The remaining life is calculated by constructing a remaining life calculation model 3, which is a prediction formula for future changes in workload, for each pump's characteristics and classification.

Pump deterioration is thought to be caused by frequent current overloads in facilities that are affected by special drainage problems that are prone to clogging with sludge, etc., and the accumulation of load on the pump, which shortens the period of failure due to wear and deterioration. I can do it. Therefore, assuming that ``the life of a pump depends on the amount of work done in operation, not simply on the passage of time,'' we will focus on and calculate the ``work elapsed rate y.'' Based on the above assumption, since the voltage is constant at 200V, the load appears directly in changes in the current, so it can be considered that the work done by the pump depends on the "current value" and "operating time."
The amount of work of the pump is defined below.

[Equation 1] Cumulative work of the pump W = Work rate w x Total operating time t = Voltage V (constant) x Current consumption I x Total operating time t

The limit cumulative work G, which is the cumulative work W at the end of the pump's life, was calculated from the average operating time (6 hours/day) and the standard service life of the pump (15 years) based on past results. It is something.

[Math 2] Limit operating time = Standard service life 15 years x 365 days x 6 hours/day

[Equation 3] Limit cumulative work G (pump life) = Voltage V (constant) x Planned current Ip x Limit operating time

Note that the "planned current Ip" is a current value at which the power is most efficient, which is 70 to 80% of the maximum current value.

[Equation 4] Work progress rate y (%) = Cumulative workload W/Limit cumulative workload G

It is.

Then, an approximate expression for the work progress rate y is obtained as follows.
Since the lifespan of a pump varies depending on the environment and regional characteristics, and the operating conditions differ depending on each pump, the results of calculating the remaining lifespan of each group of long-life/short-life pumps classified by the above decision tree model 72 are shown below. .

[Math 5] Remaining life (years) = Lifespan years - Operating years

Life prediction formula (remaining life calculation model 73)

[Math 6] y=f(x)

shall be. Here, y is the work progress rate, x is the number of years of operation, and when y=100%, x is the number of years of service life.

As mentioned above, the number of years the pump has been in operation and the cumulative amount of work W up to a certain point can be calculated based on the accumulated real-time current value data. Plot on a graph with time (years of work) x on the axis and work progress rate y on the vertical axis. Note that although the vertical axis of the graph may be the cumulative amount of work W, in this embodiment, the vertical axis is the work progress rate y to make it easier to understand.
And the inverse function [Math. 7] x=g(y)
Then, if the life prediction formula (remaining life calculation model 73) is approximated from the work progress rate (coordinate points) plotted on the graph,
Remaining life calculation model 73 is for short-lived pumps [Equation 8] y = 0.0831e^ (0.1375x)
For a long-life pump [Math. 9] y=0.0157x
It became.

As shown in FIG. 10, the limit cumulative workload G is indicated by a broken line on the graph. The intersection of the graph of the life prediction formula (remaining life calculation model 73) of each of the short-life/long-life pumps and the straight line indicating the limit cumulative work G is the life point L of the pump, and The period becomes the remaining life H of the pump, and deterioration can be predicted.

The technical feature of this remaining life calculation model 73 is that real-time current value data sent from the IoT device 5 provided in the manhole pump is used for calculating the remaining life.
The "consumption current I" and "total operating time t" in the cumulative workload W of the pump are calculated based on the accumulated real-time current value data including the pump operating time that is constantly sent by the IoT device 5. It can be calculated from the current value data.

By predicting the deterioration of pumps as described above, we can manage multiple scattered facilities, and conduct detailed inspections of individual pumps that are difficult to inspect on a regular basis and grasp the deterioration status by utilizing real-time measurement data from the IoT device 5. By calculating the remaining service life without carrying out maintenance, it is possible to estimate the amount of future renovation of the pump, which can be expected to improve the efficiency of maintenance management and stock management plans. In addition, in the conventional stock management plan, there was no difference in priority order in time-planned management, and there was a large deviation from the actual situation in the deterioration state, but in the decision tree model 72, the operation is more than planned. By extracting manhole pumps by long-life/short-life classification, creating a prediction formula for remaining life, and calculating the remaining life, we can prioritize (screen) inspection areas, improve the efficiency of inspection and investigation plans, and improve the overall efficiency of pumps. It will be possible to support the creation of long-term renovation project scenarios. Furthermore, by constructing the decision tree model 72, it is possible to know the items to be noted in the maintenance and management of each pump, so it can be used to review the float water level settings, especially for short-lived pumps, for example, those that operate for a long time. It is possible to take measures to extend lifespan. Furthermore, in the renewal plan, it becomes possible to formulate an improvement plan so that the operation can be shifted to a long-life pump. In addition, in other regional management categories (municipalities, etc.), it is thought that feature data for classifying pumps as long or short-lived can be extracted by constructing the decision tree model 72 in the same way, so it is also highly versatile. Excellent.

The deterioration prediction information obtained as described above is accumulated and managed in the integrated monitoring cloud system 1, and furthermore, the stock management support system 3 receives data related to the deterioration prediction and uses it to determine the health of the pump. It can be used for evaluations, risk studies, etc., and by sharing information with each local government7, it is possible to provide efficient and continuous support, such as improving the accuracy of renovation plans and reducing costs.

Next, the integrated monitoring cloud system 1 will be explained. The integrated monitoring cloud system 1 receives the linkage of facility operation data (measurements and daily reports) and maintenance data (emergency dispatch history) from the existing monitoring system, and the linkage of the equipment ledger from the stock management support system 3. Centrally manage information necessary for detecting abnormalities and predicting deterioration. In addition, by linking with the detection and prediction system equipped with an AI engine, the IoT device 5 monitors the condition of the manhole pump (detects abnormalities and notifies maintenance departments, etc.) and accumulates pump deterioration prediction information obtained from facility operation information. The results will be linked to the "Stock Management Support System 3", which will play the role of a data management center for the maintenance and management of manhole pumps. The purpose of constructing the integrated monitoring cloud system 1 is as follows.
Purpose 1: By importing daily report data from existing systems and automatically linking it to the detection and prediction system equipped with an AI engine, the training data will be passed to the AI engine.
Purpose 2: In response to the abnormality detection cooperation from the AI engine-equipped detection and prediction system, alert emails are sent to maintenance managers, and the abnormality detection status can be monitored on the screen in real time. This enables early response to abnormalities. Additionally, by inputting the response to a manhole in which an abnormality has been detected, it is possible to manage maintenance for the manhole pump.
Purpose 3: By displaying the abnormality detection status of manhole pumps in a map and list, it is possible to collectively understand the manhole pumps that are currently detecting abnormalities.
Purpose 4: By keeping a record of the maintenance contractor's response to the content of abnormality detection, young engineers will be able to learn how experienced engineers respond to the content of abnormality detection, equipment specifications of manhole pumps, and current value waveforms. make it possible to convey. In addition, by feeding back to the detection and prediction system equipped with an AI engine, it is possible to accumulate training data that is useful for improving prediction accuracy.
Purpose 5: By making it possible to input inspection and repair work carried out by maintenance and management contractors, it is possible to accumulate input inspection and repair data and to share information with maintenance and management departments.
The functions of the integrated monitoring cloud system 1 for the above five purposes are as follows.

<Equipment specifications import function>
The stock management support system 3 receives the facility information of the current/leakage current measurement points and reflects it in the database. As a database synchronization method, the API on the stock management support system 3 side is called to obtain the data to be synchronized. The data to be synchronized shall be the facility specifications shown in Table 7.

IoTデバイス５で測定したリアルタイム電流値を統合監視クラウドシステム１に取り込むことにより、運転日報データとしてマンホールポンプの運転記録を蓄積することができる。蓄積した運転記録の各項目は「運転日報」画面にて閲覧することができる。また、「運転日報」のデータは「帳票出力」ボタンを選択することにより帳票出力することができる。 <Daily driving report data linkage function>
Imports daily driving report data from the existing remote monitoring system and stores it as daily driving report data. The daily driving report data of the existing remote monitoring system targeted in this embodiment is based on the daily driving report data items (Excel format) held by the existing remote monitoring system widely adopted in the field. Daily report data from other existing remote monitoring systems will be linked after being converted to a standard format. Table 8 shows standard data items.

By importing the real-time current values measured by the IoT device 5 into the integrated monitoring cloud system 1, it is possible to accumulate operation records of the manhole pump as daily operation report data. Each item of the accumulated driving record can be viewed on the "Daily Driving Report" screen. Furthermore, the data of the "daily driving report" can be output as a form by selecting the "form output" button.

<Alert notification function for abnormality detection>
The integrated monitoring cloud system 1 receives notification of abnormality detection (pump ID, abnormality detection date and time, abnormality prediction date and time, abnormality type, probability of abnormality occurrence within 15 hours) from the AI engine-equipped detection and prediction system 2.
When receiving a notification that an abnormality has been detected in which the probability of abnormality occurrence exceeds a predetermined threshold, an alert email of abnormality detection is sent to the maintenance management company. The maintenance and management contractor who receives the alert email selects the URL included in the email to display a screen where they can check the status of abnormality detection and a map of the manhole pump facility.

<Compatibility status input function>
When receiving a notification of abnormality detection in which the probability of abnormality occurrence exceeds a threshold for a manhole pump facility where an abnormality has been detected, an abnormality detection alert email is sent to the maintenance and management company. The maintenance and management contractor who receives the alert email selects the URL included in the email to display a screen where they can check the status of abnormality detection.
To display detailed abnormality detection contents on this screen, select "Details" from the menu in "Abnormality detection contents" on the screen, and the time change of the final abnormality accuracy K will be displayed in a graph.
In addition, by selecting "Current value graph" from the menu within "Abnormality detection contents", it is possible to graphically display the temporal change in the current value around the time when the notification of abnormality detection is received.
The maintenance and management contractor then takes a response to the manhole pump abnormality based on the equipment specifications, current value graph, and the content of the abnormality detection, enters the details of the response in the ``Response input'' field, and selects the ``Response status''. ” to “Complete.” If the response has been completed, the record of the emergency dispatch can be output as a form by selecting the "Output Form" button.

<Anomaly detection history function>
By displaying on the screen the history of accumulated abnormality detections and the actions taken by maintenance contractors in the past in response to notified abnormality detections, it is possible to check the details of abnormality detections, equipment specifications of manhole pumps, and current values. It is possible to convey to young engineers the responses of experienced engineers to waveforms.

<Driving overview monitor function>
For manhole pump maintenance and management companies, the results of abnormalities detected by the AI engine-equipped detection and prediction system 2 can be viewed on a screen that lists them on a map. Facilities where an abnormality has been detected are highlighted on the screen, and maintenance contractors conduct on-site inspections based on this information. Furthermore, by having the maintenance and management company register the results of the investigation, such as whether there are any abnormalities detected or whether the pump is clogged, information on actual phenomena can be stored as training data for the AI.

<Inspection input/viewing function>
By inputting the results of periodic inspections and regular inspections conducted by the manhole pump maintenance and management contractor 6, it is possible to accumulate the inspection results and share information with the maintenance and management department. Periodic inspection results for manhole pumps can be entered from the inspection registration screen. The registered inspection results can be displayed as a list on the inspection history screen, and the details of the inspection can be viewed by clicking the details button.

<Repair input/view function>
Enter the details of the manhole pump repairs that were carried out after an alert was raised in the existing system and an abnormality was detected, repairs for abnormalities found during periodic inspections, and repairs conducted in response to investigations based on contacts from surrounding residents, etc. By doing so, records of repairs can be accumulated and information can be shared with maintenance departments, etc. The details of the repair can be entered from the repair registration screen, and the entered repair details can be displayed in a list from the repair history screen. Furthermore, by selecting the "form output" button, the repair details can be output as a form.
The functions of the integrated monitoring cloud system 1 are as described above.

Next, the stock management support system 3 will be explained. Stock management support system 3 clarifies the scale of specific renovation demand based on the latest facility ledger specification data, maintenance implementation status, and predicted deterioration status, and provides appropriate information for medium- to long-term renovation planning. This is a system that does this. The stock management support system 3 receives data related to maintenance information (inspection/survey) and deterioration prediction accumulated in the integrated monitoring cloud system 1, and utilizes this for pump health evaluation, risk consideration, etc. The aim is to support efficient and continuous stock management, such as visualizing various information, improving the accuracy of renovation plans, and reducing LCC. In this embodiment, the stock management support system 3 uses a cloud-type information system (service name: Blitz GROW, Nissui Con) that has already been demonstrated and commercialized. This system is highly reliable as it has a track record of implementation and operation in local governments.

As shown in Figure 11, the functions of the stock management support system 3 are broadly divided into equipment ledger, construction ledger, current status map management, maintenance inspection management, and renovation planning, and many people involved in facility management use this system in the course of their work. Achieve efficient and continuous stock management by utilizing functions. Additionally, the equipment specifications of the stock management support system 3 and the integrated monitoring cloud system 1 are linked using Web API. Table 9 shows the specifications of equipment linked to the integrated monitoring cloud system 1.

Furthermore, the stock management support system 3 acquires deterioration prediction data from the integrated monitoring cloud system 1 using the Web API. A user with system administrator authority can obtain deterioration prediction data from the integrated monitoring cloud system 1 at any timing. The deterioration prediction data includes prediction date and time, long-life/short-life classification data, and remaining life data. The deterioration prediction data reflected in the stock management support system 3 can be confirmed from the equipment ledger details screen of ledger management. Furthermore, the deterioration prediction data accumulated in the stock management support system 3 and associated with equipment ledger information can be output as versatile EXCEL data. It is possible to easily extract target facilities for stock management plans (updates) from the EXCEL output, long-life/short-life classification data, and remaining life data.

In addition, the linkage of repair and renovation work history will be done using Web API, similar to the linkage of equipment specifications. Note that linking construction information requires the equipment ID listed in the equipment specifications, so it should be performed after linking the equipment specifications. In addition, materials such as completed books registered in the stock management support system 3 can also be linked. The construction information linked to the integrated monitoring cloud system 1 is shown in Table 10.

Next, a method for constructing the integrated monitoring cloud system 1 will be explained. First, a batch program for acquiring equipment specifications is executed to acquire equipment specifications. Then, the daily report data output from the existing remote monitoring system is saved in the daily report data import folder. Upload to the S3 area (where data is saved) of the integrated monitoring cloud by executing the command at the command prompt. Next, when batch processing is executed on the server of the integrated monitoring cloud system 1, daily report data is registered in the database of the integrated monitoring cloud system 1. In addition to the daily report data imported from the existing remote monitoring system, real-time current values measured by the IoT device 5 are also imported through monthly batch processing and used as driving records. When the real-time current value capture is completed, daily report data up to the month in which the real-time current value was captured can be displayed on the screen of the integrated monitoring cloud system 1.

Then, by calling the API for anomaly detection from the AI engine, you will receive a notification of anomaly detection. Data for abnormality detection is in JSON format. Upon receiving this notification of abnormality detection, an abnormality detection email is sent. By selecting the URL included in the email, the details of the abnormality detection will be displayed on the screen. In addition, the AI engine acquires current value data in order to display a current value graph when there is a notification that an abnormality has been detected. To obtain current value data, make an API call using the following steps. First, call the API to obtain a list of URLs where current value data can be downloaded. Next, download the current value data stored in the URL. The current value data is saved in the AWS S3 area in parquet format.

Next, the stock management support system 3 and the integrated monitoring cloud system 1 are linked. As shown in FIG. 12, an API is constructed for linking the stock management support system 3 and the integrated monitoring cloud system 1. In linking the stock management support system 3 and the integrated monitoring cloud system 1, the following two cases were assumed, and data items required for each were determined.
(Case 1) Cooperation for both systems to synchronize and share attribute information such as basic equipment ledger specification data of manhole pumps (cooperation for the integrated monitoring cloud system 1 to receive data).
(Case 2) Cooperation for the stock management support system to receive deterioration prediction result data accumulated and managed by the integrated monitoring cloud system.
The API will be a versatile Web-API that can respond in common data formats (JSON format, etc.), allowing for flexible response during widespread deployment. The flow of data acquisition and response in each case is shown in FIGS. 12(a) and 12(b), respectively.
(Case 1)
Call the equipment ledger API from the integrated monitoring cloud system 1.
The called equipment ledger API passes data such as equipment ledger ID and equipment name to the integrated monitoring cloud system 1.
(Case 2)
Call the API of the integrated monitoring cloud system 1 from the stock management support system 3.
The called API of the integrated monitoring cloud system 1 passes data such as equipment ledger ID and remaining life to the stock management support system 3.
As described above, the stock management support system 3 and the integrated monitoring cloud system 1 were linked.

As described above, the integrated monitoring cloud system 1 receives and accumulates information on abnormality detection and deterioration prediction by the AI engine-equipped detection and prediction system 2. The stock management support system 3 then supports the appropriate formulation of a stock management plan through condition monitoring based on the abnormality detection and deterioration prediction information accumulated in the integrated monitoring cloud system 1. The stock management plan, abnormality detection, and deterioration prediction information are then notified to maintenance management companies and local governments, which are ultimately in charge, as necessary, through an integrated monitoring cloud system that centrally manages this information. Ru. As a result, maintenance and management contractors can use information about abnormality detection to secure a predetermined time before an abnormality occurs, reduce the burden on site workers, and perform repairs on-site as necessary. Furthermore, by receiving data related to deterioration prediction through the stock management support system, maintenance companies and local governments can utilize this data to evaluate pump health and consider risks, and share information with each local government. It is possible to provide efficient and continuous support by improving the accuracy of renovation plans, resolving human resource shortages, and reducing costs.

Next, we will show the results of actual operation of the manhole pump monitoring system. This result is an example in which abnormality detection information was notified to the maintenance and management contractor 6 by the integrated monitoring cloud system 1, and the maintenance and management contractor 6 took action. FIG. 13 is a graph showing the relationship between the abnormality accuracy and time of the No. 1 manhole pump at Tayakita Sewage Relay Pump Station, calculated by the AI engine-equipped detection prediction system 2 and notified by the integrated monitoring cloud system 1. As the final abnormality accuracy K (15-hour moving average of accuracy J) reached 60% at 5:30 on July 8, 2021, a notification of abnormality detection was issued from the integrated monitoring cloud system 1 to the maintenance management company 6. reported. After receiving this notification, maintenance contractor 6 responded to the scene at 14:00 on July 8, 2021. During this on-site response, it was discovered that there was a foreign substance (scum) inside the casing, so when the foreign substance was removed, the current value of the pump fell below the rated current. As a result, the final abnormality probability K fell below 60% at 6:30 on July 9, 2021, and maintenance contractor 6's standby system was lifted. In this way, the integrated monitoring cloud system 1 notifies the maintenance contractor 6 of abnormality detection information in advance during the daytime, making it possible to secure time during the daytime for repairs before an abnormality occurs. As a result, maintenance contractor 6 was able to complete the response without having to be dispatched during the night (18:00 to 9:00 the next day). This greatly contributed to reducing the burden on field workers.

1 Integrated monitoring cloud system 2 AI engine-equipped detection prediction system 3 Stock management support system 5 IoT device 6 Maintenance management company 7 Local government

Claims (1)

An integrated monitoring cloud system that centrally manages information using a cloud system,
An AI engine-equipped detection prediction system that utilizes the information to construct an abnormality detection and discriminator using AI to detect abnormalities before they occur and to predict deterioration;
A stock management support system that supports the formulation of a stock management plan based on the information accumulated in the integrated monitoring cloud system,
The integrated monitoring cloud system, the AI engine-equipped detection prediction system, and the stock management support system are all implemented and constructed on the cloud,
The AI engine-equipped detection and prediction system uses real-time measurement data from IoT devices to detect abnormalities and predict deterioration in real time.
The integrated monitoring cloud system receives and accumulates information on the abnormality detection and deterioration prediction,
The stock management support system aims to support the appropriate formulation of a stock management plan by monitoring conditions based on the information accumulated in the integrated monitoring cloud system,
The stock management plan and information on abnormality detection and deterioration prediction are notified to maintenance companies and local governments as necessary ,
The AI engine-equipped detection and prediction system detects an abnormality, which is measured data from the IoT device, when an abnormality is detected within a predetermined detection period based on the length of the nighttime period for the maintenance company. Inputting the real-time measurement data into the anomaly detection classifier, which is a trained model generated by machine learning using learning data, to infer whether or not an anomaly occurs within the detection period. A manhole pump monitoring system, characterized in that the above-mentioned abnormality detection is performed in advance .

Images