JP7119349B2 - Determination device, determination program, determination method - Google Patents

Description

The present invention relates to an apparatus, program, and method for determining the state of target data based on a sparse structure model.

2. Description of the Related Art Conventionally, as one method of analyzing dynamically changing multivariate data (time series data), a method of determining the state of multivariate data based on a vector consisting of a combination of variables is known. For example, a technique is known in which a normal model is created by extracting a feature vector whose elements are the feature values from multivariate data for learning, and the state of the multivariate data is determined using this model. . Also, it is known to set a plurality of operation modes according to the operating state of the target device and create a normal model for each operation mode (see, for example, Patent Document 1). There is also known a technique of performing clustering based on bias in data distribution when setting a plurality of operation modes (see Patent Document 2, for example).

On the other hand, correlations between variables included in multivariate data are often vulnerable to disturbances. Therefore, in the determination method based on the feature vector as described above, a noise component (noise) is mixed in the feature value itself, and the signs of abnormality may not be accurately grasped. Therefore, a technique has been proposed in which only essential connections between variables are extracted from multivariate data as a sparse structure, and the state of the multivariate data is determined based on the structural change. For example, calculate the correlation coefficient for each of the multivariate data, and remove unnecessary components (components showing low correlation) from the variance-covariance matrix (also simply called the covariance matrix) whose elements are the correlation coefficients. A technique for determining the degree of abnormality of a state using a zeroed one (sparse variance-covariance matrix) is known (see, for example, Non-Patent Document 1 and Patent Document 1).

Japanese Patent Application Laid-Open No. 2011-070635 JP 2016-224915 A JP 2010-078467 A Tsuyoshi Ide et al., “Proximity-Based Anormaly Detection using Sparse Structure Learning”, The Society for Industrial and Applied Mathematics (SIAM), Proceedings of the 2009 SIAM International Conference on Data Mining, 2009, ISBN: 978 -0-89871-682-5

State determination based on changes in sparse structure can provide good accuracy for highly unimodal multivariate data (multivariate data with unimodal distributions). However, when dealing with highly multimodal multivariate data (multivariate data having a multimodal distribution), it may not be possible to accurately grasp the collapse of the correlation between variables. For example, when analyzing multivariate data acquired in rock excavation work using a hydraulic drifter, the pattern of correlation between variables changes depending on the type and quality of the rock. This fact means that fluctuations in correlations between variables can be attributed not only to hydraulic drifter anomalies but also to changes in lithology. Therefore, a correlation that should be normal may be erroneously determined to be "abnormal", and it is difficult to improve the accuracy of determination of the state of the target data. In addition, for member assembly work in an assembly factory, for example, when multivariate data acquired in connector fitting work is analyzed, variables can be compared depending on the number and size of connector pins and the direction in which the connector is inserted. changes in the pattern of correlations. Therefore, a correlation that should be normal may be erroneously determined to be "abnormal", and it is difficult to improve the accuracy of determining the state of target data.

An object of one aspect is to improve the determination accuracy of the state of target data.

In one embodiment, a sparse structure model is generated from time-series learning data obtained by a plurality of sensors in the learning stage, and time-series target data obtained by the plurality of sensors and the sparse structure model are generated in the operation stage. A determination device is disclosed for determining the state of the target data based on. This determination device includes a clustering unit that clusters the learning data into a plurality of sets corresponding to a plurality of normal modes in the learning stage. A model creation unit is provided for creating, for each set, the sparse structure model representing the relationship between parameters in the learning data included in each of the plurality of sets using a Markov random field. Further, in the operation stage, a target set to be compared with the target data is estimated from among the plurality of sets, and each time the target data is acquired, the plurality of clusters clustered by the clustering unit are estimated. A cluster determination unit is provided for estimating, according to a selection condition, which set among the sets is appropriate to use as the target set to perform state determination. Further, the state of the target data is determined based on the structural change amount of the target data with respect to the sparse structural model for the target set, and the larger the structural change amount, the higher the degree of abnormality of the target data. The selection condition is to select, from among the plurality of sets, a set having the smallest distance from the target data to the barycenter position of the set as the target set .

In one aspect, it is possible to improve the determination accuracy of the state of the target data.

It is a block diagram which shows the hardware constitutions of a determination apparatus. 4 is a graph illustrating time-series data obtained by multiple sensors; 4 is a block diagram showing the functional configuration of a determination program; FIG. (A) to (D) are schematic diagrams for explaining a clustering method (DP-means method). (A) and (B) are schematic diagrams showing clustering results. It is a visualization example of the correlation between sensor values, (A) corresponds to a non-sparse graphical model, and (B) to (D) correspond to sparsified graphical models (sparse structure models). 4 is a flow chart illustrating control contents in a learning stage; 4 is a flowchart illustrating control details in an operation stage;

Hereinafter, a determination device, a determination program, and a determination method as embodiments will be described with reference to the drawings. These determination device, determination program, and determination method provide a function of detecting whether the state of the determination target is normal or abnormal based on time-series multivariate data obtained by a plurality of sensors. Also, the state determination process includes two stages, a learning stage and an operation stage. The learning stage is the stage of creating a sparse structure model for decision making using multivariate data that is considered normal in the majority. On the other hand, the operation stage is the stage of judging the state of multivariate data, which is unknown whether it is normal or abnormal, using the sparse structure model created in the learning stage.

The sparse structure model means a model having a structure obtained by extracting only essential connections between variables from multivariate data. In other words, it is not a “dense structure” in which all variables are considered to be connected by some kind of correlation, but a “sparse structure” obtained by omitting weak correlations and leaving only strong correlations. structure”. In other words, a sparse structure model is a structure in which most components are 0 in a matrix-expressed correlation (for example, a precision matrix for a variance-covariance matrix of multivariate data). It is a model with

The determination device, determination program, and determination method described in this embodiment generate a sparse structure model from time-series learning data obtained by multiple sensors in the learning stage, and when obtained by multiple sensors in the operation stage The state of the target data is determined based on the series target data and the sparse structure model. Devices to be determined in this embodiment include industrial machines, construction machines, construction vehicles, automatic machines (machine tools), etc., to which a plurality of sensors are attached. Each of the plurality of sensors referred to here can acquire time-series data. A large amount of data detected by multiple sensors within a predetermined period is used to create a sparse structural model and determine its state.

[1. Hardware configuration]
FIG. 1 is a block diagram illustrating the hardware configuration of a determination device 11 as an embodiment. This determination device 11 detects anomalies in various devices. For example, hydraulic drifters in rock drilling operations have the ability to drill through various rock types. A plurality of sensors and an output device 15 are connected to the determination device 11 . As a specific example of the sensors, taking the case of a hydraulic drifter for rock excavation as an example, a feed pressure sensor 12 for detecting the feed pressure (thrust force for pressing the drilling rod against the rock), and the impact pressure (impact hydraulic pressure of the drilling rod). ), and a rotational pressure sensor 14 for detecting rotational pressure (rotation hydraulic pressure of the drilling rod).

Examples of other sensors (not shown) include a piercing speed sensor that detects the piercing speed (amount of piercing per unit time), a rotation speed sensor that detects the rotation speed of the piercing rod, and a flushing pressure (compression jetted from the tip of the piercing rod). air pressure), a time sensor that measures working hours and continuous operation time, and an oil temperature sensor that detects the operating oil temperature. In addition to rock excavation work, it can also be applied to assembly work using industrial robots. For example, a force sensor, a pressure sensor, a proximity sensor, a tactile sensor, and a visual sensor such as a camera for sensing a connector fitting operation can be used. Time-series data detected by these sensors are transmitted to the determination device 11 . Also, the output device 15 is a device for outputting the determination result, and is, for example, a display device, a speaker device, a printer device, or the like. The connection form of the sensors 12 to 14 and the output device 15 is arbitrary, and may be a wired connection or a wireless connection.

A processor 21 (central processing unit), a memory 22 (main memory, main memory), an auxiliary storage device 23, an interface device 24, and a recording medium drive 25 are built in the determination device 11. communicatively connected to each other. The processor 21 is a general-purpose CPU (Central Processing Unit) or DSP (Digital Signal Processor) containing a control unit (control circuit), an arithmetic unit (arithmetic circuit), a cache memory (register group), and the like. The memory 22 is a storage device that stores programs and data during operation, and includes ROM (Read Only Memory), RAM (Random Access Memory), and the like.

The auxiliary storage device 23 is a storage device that stores data and firmware that are held longer than the memory 22, and includes non-volatile memories such as flash memory and EEPROM (Electrically Erasable Programmable Read-Only Memory). be The interface device 24 controls input and output (I/O) between the determination device 11 and the outside thereof. The recording medium drive 25 is a reading device (or reading/writing device) having a function of reading information recorded and stored in a recording medium 16 such as an optical disk or a semiconductor memory (portable flash drive conforming to the Universal Serial Bus standard). is.

FIG. 2 is a graph illustrating changes over time in time-series data detected by various sensors including sensors 12-14. Each sensor value changes greatly according to the operating state of the hydraulic drifter, and the correlation between the sensor values can also change. Therefore, good judgment accuracy cannot be expected if the state is judged based on a feature vector having individual sensor values as elements. On the other hand, in the present embodiment, the correlation between each sensor value is extracted as a sparse structure, and whether the operating state of the hydraulic drifter is normal or abnormal is determined based on the structural change. In addition, since the correlation between sensor values is not always the same, we extract multiple sparse structures in advance, select the appropriate one from among them, and then calculate the operating state. judge. These functions are realized by causing a computer (for example, processor 21) to execute processing described in determination program 17, which will be described below.

[2. Software configuration]
FIG. 3 is a block diagram showing the functional configuration of the determination program 17 executed by the determination device 11. As shown in FIG. This determination program 17 is recorded and saved in the memory 22 and the auxiliary storage device 23 . Alternatively, the determination program 17 is recorded and stored on the recording medium 16, and the determination program 17 written in the recording medium 16 is read into the determination device 11 via the recording medium drive 25 and executed.

The functions realized by the determination program 17 are broadly divided into a function of creating a state determination model in advance and a function of performing state determination using the model. The learning unit 1 in FIG. 3 provides the former function, and the operation unit 7 provides the latter function. The learning unit 1 has a function of generating a sparse structure model from time-series data obtained by a plurality of sensors 12-14 in a learning stage. On the other hand, the operation unit 7 has a function of judging the state of the target data based on the time-series data and the sparse structure model obtained by the plurality of sensors 12 to 14 in the operation stage.

Hereinafter, among the time-series data obtained by the plurality of sensors 12-14, the data used for creating the sparse structure model in the learning stage will be referred to as "learning data". Also, the data used for setting the threshold for judgment after the sparse structure model is created in the learning stage is called "test data". On the other hand, the data used for state determination in the operation stage is called "target data" and handled separately from learning data and test data.

The learning unit 1 is provided with a statistical model estimation unit 2 , an abnormality degree setting unit 5 and a threshold value setting unit 6 . The statistical model estimator 2 uses learning data acquired by the sensors 12 to 14 to create a statistical model (sparse structure model) corresponding to a normal state. The statistical model estimation unit 2 is provided with a clustering unit 3 and a model creation unit 4 . Also, the operation unit 7 is provided with a cluster determination unit 8 , an abnormality degree calculation unit 9 , and a state determination unit 10 .

The clustering unit 3 has a function of clustering learning data into a plurality of sets (clusters) corresponding to a plurality of normal modes in the learning stage. Here, a vector containing as elements a plurality of sensor values detected at the same time is regarded as one piece of point data, and clusters are classified based on the state of distribution of the plurality of point data. A specific clustering method can be arbitrarily selected, and for example, a k-means method, a fuzzy c-means method, a spectral clustering method, or the like can be used.

The learning unit 1 of this embodiment clusters learning data using the DP-means method. The DP-means method is a clustering method that is an extension of the k-means method based on the Dirichlet Process mixture model. In the k-means method, the number of clusters is specified prior to clustering, whereas in the DP-means method, a reference value λ (granularity parameter) for increasing the number of clusters is specified.

For example, as shown in FIG. 4(A), the distance d between one learning data and each cluster (the distance to the center of gravity of the cluster) is calculated, and the distance to the cluster with the smallest distance d (nearest neighbor cluster) is calculated. The distance d is compared with a reference value λ. Here, when the distance d is equal to or greater than the reference value λ, a new cluster including the learning data is generated as shown in FIG. 4(B). Conversely, if the distance d is less than the reference value λ, the learning data is determined to be included in the nearest neighbor cluster, as shown in FIG. 4(C). In this case, the centroid position of the nearest neighbor cluster is recalculated as shown in FIG. 4(D).

By repeating such calculations, even if the number of clusters is unknown, clustering can be performed while automatically estimating an appropriate number of clusters as shown in FIG. 5(A). The number of clusters estimated here increases as the reference value λ decreases, and decreases as the reference value λ increases. In the graphs shown in FIGS. 4 and 5, the number of dimensions of the learning data is three, but the actual learning data has dimensions corresponding to the number of types of sensor values input to the determination device 11 .

In the DP-means method, hard clustering is realized by asymptotically approaching the parameter corresponding to the variance in the nonparametric basis method to 0, and the following objective function with a penalty term attached to the number of clusters in the k-means method algorithm is obtained. (See Equation 1). In addition, for more detailed application methods of the DP-means method, it is possible to refer to known literature (for example, JP 2016-224915, Brian Kulis and Michael I. Jordan, Revisiting k-means: New Algorithms via Bayesian Nonparametrics, Proceedings of the 29th International Conference on Machine Learning (ICML-12), 2012, p.513-520, etc.).

Each cluster clustered by the clustering unit 3 means that the learning data included in the cluster corresponds to different rock quality (clay layer, sand layer, gravel layer, hard layer, soft layer, etc.). Therefore, it is expected that the number of clusters will increase as the types of rocks excavated by the hydraulic drifter increase. On the other hand, the clustering of this embodiment is "unsupervised learning", and it is not necessary to explicitly specify the lithology, and differences in lithology are automatically identified during clustering. In other words, it is not necessary to confirm in advance what kind of lithology the learning data contained in each cluster corresponds to, and each cluster is simply set to the first normal mode as shown in FIG. 5(B). or second normal mode.

The model creation unit 4 creates a sparse structure model for each of the plurality of sets (each cluster) clustered by the clustering unit 3 based on the learning data included in each cluster. Here, a sparse structure model is generated for each cluster in which the relationship between parameters in the learning data is represented by a Markov Random Field (MRF). A Markov random field is a network-like graphical model in which a plurality of nodes representing random variables are connected by nodes representing dependencies between the random variables. A specific sparse structure model can be arbitrarily selected, and for example, an Ising model or a Potts model can be used. The model creation unit 4 of this embodiment uses a Gaussian Graphical Model (GGM).

For each cluster generated by the clustering unit 3, the model creation unit 4 calculates the variance-covariance matrix _Σp of the learning data included in each cluster. The value of each element included in the variance-covariance matrix Σ _p corresponds to the degree of correlation between two pieces of learning data. An example of the variance-covariance matrix Σ _p when the number of types of sensor values (the number of dimensions of learning data) is six is given below (see Equation 2).

When visualizing the correlation between sensor values represented by this variance-covariance matrix Σ _p , each sensor value corresponds to a vertex, and the value of each element of the variance-covariance matrix Σ _p corresponds to an edge. can be considered. On the other hand, in practice, it is rare that the correlation between multiple sensor values is not detected at all due to the influence of noise. Therefore, the overall structure tends to have a shape in which each sensor value has some correlation with each of the other sensor values, as shown in FIG. becomes complicated.

がスパース化され、クラスタD_p毎のスパース構造モデルが作成される。なお、統計モデル

の表現での

は省略されることがあり、

として表現することもあるが、同一の表現である。 On the other hand, the model creation unit 4 uses a Gaussian graphical model to perform processing for extracting only essential connections between sensor values as a sparse structure. That is, a normal distribution model is applied to the variance-covariance matrix Σ _p , a likelihood function with a regularization term added is calculated (see Equation 3), and maximum likelihood estimation is performed for the variance-covariance matrix Σ _p . By performing such parameter optimization, a statistical model representing the correlation structure of training data in each cluster D _p (p is the cluster number)

is sparsified, and a sparse structure model for each cluster D _p is created. Note that the statistical model

in the expression of

may be omitted,

Although it is sometimes expressed as, it is the same expression.

Graphs of sparsified structures (sparse structure models) are illustrated in FIGS. 6(B)-(D). These structures are structures from which weak correlations due to the influence of noise have been removed, and can be said to be structures from which only essential connections (correlations) have been extracted. For a more detailed method of creating a sparse structure model using a Gaussian graphical model, reference can be made to known literature (for example, see Non-Patent Literature 1 mentioned above).

The degree-of-abnormality setting unit 5 sets an abnormality score value (degree of abnormality) for evaluating the degree of “deviation” from the sparse structure model created by the statistical model estimation unit 2 . Here, the abnormal score value s ^c _k,m of the m-th element (sensor value) in the data vector x _k of the test data given in the cluster c is expressed by the conditional probability p(x It is defined by the natural logarithm (reverse sign) of _k,m |x _k ,D _p ).

Based on the anomaly score value s ^c _k,m defined by the anomaly degree setting unit 5, the threshold value setting unit 6 determines whether or not it can be said that the sparsity structure is significantly different (that is, the state is abnormal). It sets a threshold th ^c for Here, the threshold th ^c is set based on learning data (test data) different from that used by the clustering unit 3 . Let a be the number of types of sensor values, j _c be the test data, and M _c be the number of test data. In addition, the anomaly score value s ^c _k,m of the test data j _c for the cluster c is calculated for each sensor value, and the sum of the anomaly score values s ^c _k,m for all sensor values is the “sum S ^c _jc ”. Further, the threshold value setting unit 6 sets the maximum value of the sum S ^c _jc of the abnormality score values s ^c _k,m for the cluster c as the threshold value th ^c , as shown in Equation 5 below.

The cluster determination unit 8 estimates a cluster (target set) to be compared with target data in the operation stage. Here, each time one target data is acquired, it is estimated which cluster should be used for the state determination from among the plurality of clusters clustered by the clustering unit 3, and one cluster is Selected as the "target cluster". The target cluster selection condition preferably includes a condition regarding the distance d between the target data and each cluster (the distance to the centroid position of the cluster), and for example, the cluster with the smallest distance d is selected. However, the condition of distance d is not essential. For example, regardless of the size of the distance d, the target cluster in the current calculation cycle may be selected after considering the selection results (history) in the previous calculation cycles.

The degree-of-abnormality calculation unit 9 calculates an abnormality score value (degree of abnormality) of the target data for the target cluster selected by the cluster determination unit 8 . Here, the anomaly score value of the target data is calculated based on the same method as the anomaly degree setting unit 5, and the total sum _Skp is calculated. The degree-of-abnormality calculation unit 9 of the present embodiment, as shown in the following equation 6, _the natural _logarithm ( _{reverse sign} ) is calculated as the total sum _Skp . This total sum S _kp corresponds to the structural change amount of the target data with respect to the sparse structure model for the target cluster.

The state determination unit 10 determines the state of the target data based on the total sum S _kp calculated by the abnormality degree calculation unit 9 . Here, it is determined whether or not the state of the target data is normal by comparing the threshold th ^c and the total sum S _kp regarding the target cluster selected by the cluster determination unit 8 . The state determination unit 10 of the present embodiment determines "normal" when the sum S _kp is equal to or less than the threshold th ^c , and determines "abnormal" when the sum S _kp is greater than the threshold th ^c . do.

As described above, the sum S _kp corresponds to the structural change amount of the target data, so here the magnitude of the structural change amount of the target data is determined. For example, assuming that the sparse structure model generated in the learning stage has the structure shown in FIG. ), the total S _kp increases due to the structural change. Therefore, when the summation _Skp exceeds the threshold ^thc , it can be considered that the sparse structure has changed significantly, and it can be determined that some abnormality has occurred in the operating state of the hydraulic drifter.

[3. flowchart]
[3-1. learning stage]
FIG. 7 is a flowchart for explaining the determination method (learning stage) of this embodiment. The content of this flow mainly corresponds to the content of control executed by the learning section 1 . In the learning stage, sensor values detected by the sensors 12 to 14, etc. within a predetermined period are acquired as learning data x _i (step A1). Each learning data x _i is a vector whose elements are sensor values. The subscript i is an ordinal number corresponding to the detection time of the sensor value, i=1 to L (the number of learning data xi is L).

Subsequently, clustering is performed by the DP-means method to classify the learning data x _i into a plurality of clusters (step A2). By using the DP-means method, the number of clusters of the learning data x _i is automatically estimated, and a plurality of clusters as shown in FIGS. 5(A) and 5(B) are generated. After that, a variance-covariance matrix Σ _p is created for each cluster (step A3). As a result, the correlation between sensor values can be grasped after being classified for each cluster.

Also, based on the variance-covariance matrix Σ _p created in the previous step, a sparse structural model p(x|D _p ) is created for each cluster (step A4). Here, a sparse structure with a different shape can be set for each cluster. For example, when the sparse structure of the first normal mode shown in FIG. 5B has the shape of FIG. 6(C) and FIG. 6(D). That is, the correlation between sensor values is grasped in units of clusters.

When the sparse structure model of each cluster is created by the above steps, test data x _j different from learning data x _i is obtained (step A5). Each test data x _j is a vector whose elements are sensor values. Also, the subscript j is an ordinal number corresponding to the sensor value detection time, and j=1 to M (the number of learning data x _j is M). Subsequently, the test data x _j are classified into a plurality of clusters like the learning data x _i (step A6).

Also, an abnormality score value s _jp is calculated for each cluster and for each sensor (step A7), and for each cluster, a sum S ^c _jp of the abnormality score values s _jp in each cluster c is calculated (step A8). . The sum S ^c _jp calculated here includes the abnormality score value s _jp (abnormality score value s _jp of all sensors) for each element of the test data x _j belonging to the cluster c. After that, the maximum value of the sum S ^c _jp of the abnormality score values s _jp is calculated and set as the threshold th ^c in the cluster c (step A9). The threshold th ^c calculated here is used for determining the state of the target data in the operation stage.

[3-2. Operation stage]
FIG. 8 is a flowchart for explaining the determination method (operation stage) of this embodiment. The content of this flow mainly corresponds to the content of control executed by the operation unit 7 . When acquisition of target data x _k is started in the operation stage, a variable k corresponding to a timer for measuring determination work time is set to k=0 (step B1). Further, on condition that the variable k is less than the predetermined value T (step B2), the sensor values detected by the sensors 12 to 14 are acquired as the target data x _k (step B3).

After that, a target cluster to be compared with the target data x _k is selected (step B4), an abnormality score value s _kp is calculated for each sensor in the selected cluster c (step B5), and the sum S ^c _kp is calculated (step B6). This total sum S ^c _kp corresponds to the amount of structure change of the target data with respect to the sparse structure model for the target cluster. After k+1 is substituted for the variable k (step B7), it is determined whether or not the sum S ^c _kp exceeds the threshold th ^c (step B8).

Here, if S ^c _kp > th ^c , it is determined that the structural change amount of the target data is relatively large, and the state of the target data (or the operating state of the hydraulic drifter) is determined to be "abnormal". is determined (step B9). This determination result is transmitted from the determination device 11 to the output device 15, and information to that effect is output from, for example, a display device, a speaker device, a printer device, or the like. On the other hand, if ^{Sckp ≤} _thc , it is determined that the structural change amount of the target data is relatively small, and the process proceeds to step B2. Such determination is continued until the variable k becomes equal to or greater than the predetermined value T (for example, until a predetermined time elapses).

[4. action, effect]
(1) In the present embodiment, learning data is separated into clusters corresponding to a plurality of normal modes in the learning stage, and a sparse structure model of the learning data included in each cluster is created. On the other hand, in the operation stage, which cluster is to be compared with the target data is estimated, and the state determination is performed based on the structural change amount of the target data with respect to the sparse structure model of the cluster. With such a control configuration, for example, as shown in FIG. 5B, it is possible to accurately determine a work state including a plurality of normal modes. That is, it is possible to accurately grasp the state of the target data whose distribution has multiple peaks, and to improve the state determination accuracy.

(2) In this embodiment, learning data is clustered using the DP-means method. Thereby, even if the number of normal modes is unknown (that is, the number of clusters is unknown), clustering can be performed while automatically estimating an appropriate number of clusters. Therefore, workability in the learning stage can be improved, and the time required for learning can be shortened.

(3) In this embodiment, a Gaussian graphical model is used to create a sparse structural model. As a result, for example, as shown in FIG. 2, it is possible to accurately model the distribution of learning data in which the time-series data varies greatly over time, and the state determination accuracy can be improved. There is also the advantage that analytical handling is possible regardless of the number of dimensions of learning data (number of elements, number of types of sensor values).

(4) In this embodiment, a unique threshold th ^c is set for each cluster. That is, when determining the state based on the structural change amount of the target data, the magnitude of the structural change amount for determining "abnormal" can be made different for each cluster. For example, even if there is a characteristic that the correlation between the feed pressure and the impact pressure is weaker when excavating a clay layer than when excavating a sand layer, it is possible to set the threshold value ^thc in consideration of such characteristic. . Therefore, it is possible to further improve the accuracy of determining the state of work that includes a plurality of normal modes.

(5) In the present embodiment, the sum S _kp of the natural logarithms of the conditional probability p(x _k,m |x _k ,D _p ) is used as the value corresponding to the structural change amount. As a result, changes in the correlation of the entire sensor values can be accurately grasped, and the state determination accuracy can be further improved.

[5. Modification]
This embodiment is merely an example, and there is no intention to exclude various modifications and application of techniques not explicitly described in the above embodiment. That is, it is possible to carry out various modifications (for example, by combining the embodiments and modifications) without departing from the scope of the above embodiments.
For example, a construction vehicle equipped with a hydraulic drifter may incorporate the determination device 11, and a control configuration may be adopted in which determination results are reported to an operator (worker) of the construction vehicle in real time. In this case, it is possible to quickly discover an abnormality in the operating state of the hydraulic drifter, and to urge the operator to perform prompt inspection and maintenance, thereby improving workability.

In addition, when excavation work is carried out by an unmanned work vehicle equipped with a hydraulic drifter, the judgment device 11 is installed in the supervision building where the supervision engineer is stationed, and the detection information of the sensors 12 to 14 is acquired by wireless communication. control configuration. In this case, the operating state of the hydraulic drifter can be accurately grasped without relying on visual information regarding the excavation work position (for example, camera information that captures the soil and bedrock at the excavation position), improving workability. be able to.

Alternatively, when hydraulic drifters of the same type are used at a plurality of construction sites, the determination device 11 is arranged on a known network (Internet, wireless communication network for mobile phones, other digital wireless communication networks, etc.). can be considered. The information detected by the sensors 12 to 14 is transmitted to the determination device 11 via the network. With such a configuration, it becomes easy to utilize a sparse structure model created with a certain hydraulic drifter to determine the state of hydraulic drifters at other sites, and construction workability can be further improved.

Although the determination device 11 for detecting an abnormality in the hydraulic drifter for rock excavation work has been exemplified in the above-described embodiment, the device to be subjected to state determination is not limited to this. For example, it can be used for the state detection and abnormality detection of devices that perform general drilling work on the ground, such as tunnel excavation work, oil field excavation work, and building piling work. In addition to drilling work, it can also be used to detect the state and abnormality of equipment that presses existing piles and sheet piles into the ground.

It can also be applied to industrial robots and machine tools that assemble various products. For example, it is conceivable to use multiple sensors to detect the torque, force, and amount of actuation that act on a work machine during screw fastening work, and handle this as time-series multivariate data. In this case, the correlation between data fluctuates depending on the screwing position, the type of screw, the material, location, diameter, etc. of the part to be screwed, and a plurality of normal modes can be assumed. Therefore, similar effects can be obtained by performing control similar to that of the above-described embodiment. It is also conceivable to detect the torque, force, and amount of work that act during connector fitting work with multiple sensors and treat them as time-series multivariate data.

In the above-described embodiment, the control configuration for determining "abnormality" when the total sum ^Skp of the target data exceeds the threshold value _thc was described in detail. A control configuration may be adopted in which it is determined that the larger the sum S _kp is, the higher the degree of abnormality is. For example, a state in which the total sum S _kp is equal to the threshold th ^c is defined as a “state of 100% suspicion of abnormality”, and the percentage of the sum S _kp with respect to the threshold th ^c is defined as the “degree of suspicion of abnormality”. By performing such a determination, it becomes possible to grasp the state of the target data in more detail, and it is possible to improve the determination accuracy of the state.

[6. Note]
The following notes are disclosed with respect to embodiments including the above variations.
(Determination device: Supplementary note 1 to Supplementary note 5)
(Appendix 1)
In the learning stage, a clustering unit that clusters time-series learning data obtained by multiple sensors into multiple sets corresponding to multiple normal modes;
a model creation unit that creates a sparse structure model for each of the plurality of sets, in which the relationship between parameters in the learning data included in each of the plurality of sets is represented by a Markov random field;
a cluster determination unit for estimating, in an operation stage, a target set to be compared with time-series target data obtained by the plurality of sensors among the plurality of sets;
and a state determination unit that determines a state of the target data based on a structural change amount of the target data with respect to the sparse structure model for the target set.

(Appendix 2)
The determination device according to appendix 1, wherein the clustering unit clusters the learning data using a DP-means method.

(Appendix 3)
3. The determination device according to appendix 1 or 2, wherein the model creation unit creates the sparse structure model using a Gaussian graphical model.

(Appendix 4)
4. The determination device according to any one of appendices 1 to 3, further comprising a threshold value setting unit that individually sets a threshold value of the structural change amount for each of the plurality of sets.

(Appendix 5)
Any one of Appendices 1 to 4, characterized by comprising an anomaly degree calculation unit that calculates the sum of the natural logarithms of the conditional probabilities of the target data for the target set as the value corresponding to the structural change amount. The determination device according to .

(Judgment program: appendix 6 to appendix 10)
(Appendix 6)
A sparse structure model is generated from time-series learning data obtained by a plurality of sensors in a learning stage, and the target data based on the time-series target data obtained by the plurality of sensors and the sparse structure model in an operation stage. A determination program for determining the state of
clustering the learning data into a plurality of sets corresponding to a plurality of normal modes in the learning stage;
creating for each of the plurality of sets the sparse structure model in which the relationship between parameters in the learning data included in each of the plurality of sets is represented by a Markov random field;
estimating a target set to be compared with the target data among the plurality of sets in the operation stage;
A determination program that causes a computer to execute a process of determining a state of the target data based on a structural change amount of the target data with respect to the sparse structure model for the target set.

(Appendix 7)
7. The determination program according to appendix 6, causing a computer to execute a process of clustering the learning data using the DP-means method.

(Appendix 8)
8. The determination program according to appendix 6 or 7, causing a computer to execute a process of creating the sparse structure model using a Gaussian graphical model.

(Appendix 9)
9. The determination program according to any one of attachments 6 to 8, causing a computer to execute a process of individually setting the threshold value of the structural change amount for each of the plurality of sets.

(Appendix 10)
The determination program according to any one of Appendices 6 to 9, causing a computer to execute a process of calculating the sum of natural logarithms of the conditional probabilities of the target data for the target set as the value corresponding to the structural change amount. .

(Determination method: Supplementary note 11 to Supplementary note 15)
(Appendix 11)
A sparse structure model is generated from time-series learning data obtained by a plurality of sensors in a learning stage, and the target data based on the time-series target data obtained by the plurality of sensors and the sparse structure model in an operation stage. A determination method for determining the state of
clustering the learning data into a plurality of sets corresponding to a plurality of normal modes in the learning stage;
creating for each of the plurality of sets the sparse structure model in which the relationship between parameters in the learning data included in each of the plurality of sets is represented by a Markov random field;
estimating a target set to be compared with the target data among the plurality of sets in the operation stage;
A determination method, wherein the state of the target data is determined based on a structural change amount of the target data with respect to the sparse structure model for the target set.

(Appendix 12)
12. The determination method according to appendix 11, wherein the learning data is clustered using a DP-means method.

(Appendix 13)
13. The determination method according to appendix 11 or 12, wherein the sparse structure model is created using a Gaussian graphical model.

(Appendix 14)
14. The determination method according to any one of Appendices 11 to 13, wherein a threshold for the amount of structural change is individually set for each of the plurality of sets.

(Appendix 15)
15. The determination method according to any one of Appendices 11 to 14, wherein the sum of natural logarithms of the conditional probabilities of the target data for the target set is calculated as the value corresponding to the structural change amount.

(supplement)
In the determination device according to Supplementary Note 1, it is preferable that the state determination unit determines that the degree of abnormality of the target data is higher as the structural change amount is larger.
In the determination device according to appendix 2, the clustering unit classifies the learning data into the nearest neighbor cluster when the distance between the learning data and the nearest neighbor cluster is less than a reference value, and the distance is equal to or greater than the reference value. It is preferable to generate a new cluster containing the learning data when .

In the determination device according to Supplementary Note 3, the model creation unit applies a normal distribution model to the variance-covariance matrix of the learning data, calculates a likelihood function with a regularization term added, and calculates the variance-covariance matrix for the variance-covariance matrix Preferably, the sparse structural model is created by performing maximum likelihood estimation.
In the determination device according to supplementary note 4, it is preferable that the state determination unit determines that the target data is abnormal when the structural change amount exceeds the threshold value of the target set.
In the determination device according to Supplementary Note 5, it is preferable that the degree-of-abnormality calculator calculates the sum based on the following equation.

1 learning unit 2 statistical model estimation unit 3 clustering unit 4 model creation unit 5 abnormality degree setting unit 6 threshold value setting unit 7 operation unit 8 cluster determination unit 9 abnormality degree calculation unit 10 state determination unit 11 determination device 12 feed pressure sensor 13 Impact pressure sensor 14 Rotational pressure sensor 15 Output device 16 Recording medium 17 Judgment program

Claims (7)

A sparse structure model is generated from time-series learning data obtained by a plurality of sensors in a learning stage, and the target data based on the time-series target data obtained by the plurality of sensors and the sparse structure model in an operation stage. A determination device for determining the state of
a clustering unit that clusters the learning data into a plurality of sets corresponding to a plurality of normal modes in the learning stage;
a model creation unit that creates, for each set, the sparse structure model that expresses the relationship between parameters in the learning data included in each of the plurality of sets using a Markov random field;
A target set to be compared with the target data among the plurality of sets is estimated in the operation stage, and each time the target data is acquired, the plurality of sets clustered by the clustering unit a cluster determination unit for estimating, according to a selection condition, which set is appropriate to use as the target set to perform state determination;
Determining the state of the target data based on the structural change amount of the target data with respect to the sparse structural model for the target set, and determining that the greater the structural change amount, the higher the degree of abnormality of the target data a state determination unit ;
The determination device , wherein the selection condition is to select, from among the plurality of sets, a set that minimizes the distance from the target data to the center of gravity of the set as the target set . 2. The determination device according to claim 1, wherein said clustering unit clusters said learning data using a DP-means method. 3. The determination device according to claim 1, wherein said model creation unit creates said sparse structure model using a Gaussian graphical model. 4. The determination device according to any one of claims 1 to 3, further comprising a threshold value setting unit for individually setting a threshold value of said structural change amount for each of said plurality of sets. Any one of claims 1 to 4, further comprising an anomaly degree calculation unit that calculates the sum of natural logarithms of the conditional probabilities of the target data for the target set as the value corresponding to the structural change amount. The determination device according to the item. A sparse structure model is generated from time-series learning data obtained by a plurality of sensors in a learning stage, and the target data based on the time-series target data obtained by the plurality of sensors and the sparse structure model in an operation stage. A determination program for determining the state of
clustering the learning data into a plurality of sets corresponding to a plurality of normal modes in the learning stage;
creating for each of the plurality of sets the sparse structure model in which the relationship between parameters in the learning data included in each of the plurality of sets is represented by a Markov random field;
In order to estimate a target set to be compared with the target data among the plurality of sets in the operation stage, each time the target data is acquired, which set of the clustered sets is selected as the target set Estimate according to the selection condition whether it is appropriate to perform state determination using as,
a process for judging the state of the target data based on the amount of structure change of the target data with respect to the sparse structural model for the target set so that the degree of abnormality of the target data is higher as the amount of structural change is larger; let it run ,
The determination program according to claim 1, wherein the selection condition is to select, from among the plurality of sets, a set that minimizes the distance from the target data to the position of the center of gravity of the set as the target set . A sparse structure model is generated from time-series learning data obtained by a plurality of sensors in a learning stage, and the target data based on the time-series target data obtained by the plurality of sensors and the sparse structure model in an operation stage. A determination method for determining the state of
clustering the learning data into a plurality of sets corresponding to a plurality of normal modes in the learning stage;
creating for each of the plurality of sets the sparse structure model that expresses the relationship between parameters in the learning data contained in each of the plurality of sets using a Markov random field;
Each time the target data is acquired, in order to estimate a target set to be compared with the target data among the plurality of sets in the operation stage, which set of the clustered sets is selected as the target set Estimate according to the selection condition whether it is appropriate to perform state determination using as,
determining the state of the target data based on the amount of structure change of the target data with respect to the sparse structure model for the target set, determining that the greater the amount of structural change, the higher the degree of abnormality of the target data ;
The determination method, wherein the selection condition is to select, from among the plurality of sets, a set that minimizes the distance from the target data to the center of gravity of the set as the target set .

Images