Abnormality diagnostic system for object structure

Abstract

To provide an abnormality diagnostic system of an object structure capable of accurately and stably performing an abnormality diagnosis of the object structure.SOLUTION: An abnormality diagnostic system performs eigenvalue analysis with respect to a shape model obtained by modeling a structure of an object structure 200, and constructs a degradation model based on a result of the eigenvalue analysis. The system includes: an estimation section 420 for estimating, based on the degradation model and an acceleration measurement value 11, an acceleration estimation value and a stress estimation value at at least one evaluation object point being a part of the object structure 200 and also estimating an estimation value of a natural vibration number and an estimation value of an attenuation ratio which are a vibration parameter of the object structure 200; and an abnormality determination section 430 for determining a normal / abnormal state of the object structure 200 by machine learning based on the acceleration estimation value, the stress estimation value, the estimation value of the natural vibration number, and the estimation value of the attenuation ratio.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present application relates to an abnormality diagnosis system for a target structure. [Background technology]

[0002] Conventionally, as a diagnostic system for a target structure, for example, there is a vehicle fatigue damage diagnostic system shown in Patent Document 1 below. This vehicle fatigue damage diagnostic system includes a sensor and a vehicle ECU, and the sensor is attached to the target structure. The target structure is a device or member attached to the vehicle. The vehicle ECU calculates the fatigue damage level of the target structure based on the detection value obtained from the sensor (for example, see Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2013-79920 A Summary of the Invention [Problem to be solved by the invention]

[0004] For example, in a target structure that is continuously subjected to vibration loads, such as equipment mounted on a train, there is a possibility that bolts of the target structure may loosen or abnormalities such as cracks may occur. As a method for dealing with this, a method is considered in which an acceleration sensor is attached to the target structure, an acceleration waveform is calculated from the electric signal obtained from the acceleration sensor, and an evaluation is performed by performing statistical classification such as machine learning. However, a problem with this method is that the acceleration is affected by load variations due to the train's running period or riding conditions, and even in normal cases, acceleration exceeding a threshold indicating an abnormality may occur, making it difficult to accurately distinguish between normal and abnormal states.

[0005] The present application discloses technology for solving the above-mentioned problems, and aims to provide an abnormality diagnosis system for a target structure that can accurately and stably diagnose abnormalities in the target structure. [Means for solving the problem]

[0006] The abnormality diagnosis system for a target structure disclosed in the present application comprises: Acquire acceleration measurement values ​​measured by an acceleration sensor installed on the target structure, performing an eigenvalue analysis on a shape model that models the structure of the target structure, constructing a reduced model based on the result of the eigenvalue analysis, and a control unit that estimates an acceleration estimate and a stress estimate at at least one evaluation point that is a part of the target structure, based on the reduced model and the acceleration measurement value, as well as an estimate of a natural frequency and an estimate of a damping ratio, which are vibration parameters of the target structure; and an abnormality judgment unit that judges whether the target structure is in a normal or abnormal state through machine learning based on the acceleration estimated value, the stress estimated value, the natural frequency estimated value, and the damping ratio estimated value. Effect of the Invention

[0007] According to the target structure anomaly diagnosis system disclosed in the present application, anomaly diagnosis of the target structure can be performed accurately and stably. [Brief description of the drawings]

[0008] [Figure 1] 1 is a block diagram showing an abnormality diagnosis system for a target structure according to an embodiment. [Diagram 2] 4 is a flowchart showing a process for creating an analysis program by the abnormality diagnosis device of FIG. 1. [Diagram 3] FIG. 1 is a diagram illustrating a method for discretizing a degenerate model of a continuous system. [Figure 4] 4 is a flowchart showing an acceleration measurement process performed by the acceleration sensor of FIG. 1. [Diagram 5] 4 is a flowchart showing a stress estimated value calculation process performed by the abnormality diagnosis device of FIG. 1. [Figure 6] FIG. 2 is a block diagram showing a first example of a learning device in the abnormality determination unit; [Figure 7] 7 is a flowchart showing a learning process of the learning device of FIG. 6. [Figure 8] FIG. 2 is a block diagram showing a first example of an inference device in an abnormality determination unit. [Figure 9] 9 is a flowchart showing the processing performed by the inference device of FIG. 8. [Figure 10] FIG. 13 is a block diagram showing a second example of a learning device in the abnormality determination unit. [Figure 11] 11 is a flowchart showing a learning process of the learning device of FIG. 10. [Figure 12] FIG. 13 is a block diagram showing a second example of an inference device in the abnormality determination unit. [Figure 13] 13 is a flowchart showing the processing performed by the inference device of FIG. 12. [Figure 14] FIG. 4 is a diagram illustrating an example of a neural network in an abnormality determination unit. [Figure 15] 2 is a configuration diagram showing a first example of a processing circuit that realizes each function of the abnormality diagnosis device according to the embodiment. FIG. [Figure 16] FIG. 4 is a diagram illustrating a second example of a processing circuit that realizes each function of the abnormality diagnostic device according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] The present application relates to an abnormality diagnosis system for a target structure that is continuously subjected to vibration.

[0010] In the embodiment described below, an abnormality diagnosis of a target structure that is subjected to a vibration load, such as an on-board device of a train, will be described as an example. However, the present invention can also be applied to an abnormality diagnosis of a target structure that is subjected to a vibration load other than an on-board device of a train.

[0011] Description of the embodiment. FIG. 1 is a block diagram showing an abnormality diagnosis system for a target structure according to an embodiment.

[0012] 1, a target structure abnormality diagnosis system 1000 is, for example, a system that performs abnormality diagnosis on a target structure 200 provided on a vehicle body 100. Here, the vehicle body 100 is, for example, the main body of a railway vehicle. The target structure 200 is, for example, an inverter device. The target structure 200 is provided, for example, under the vehicle body 100.

[0013] The target structure abnormality diagnosis system 1000 includes an acceleration sensor 300 and an abnormality diagnosis device 400 . The acceleration sensor 300 is attached, for example, to the end of the lower surface of the target structure 200. The acceleration sensor 300 includes a sensor body 310 and an amplifier 320. The sensor body 310 detects vibrations at the location where it is attached and converts the detected vibrations into an electrical signal. The amplifier 320 amplifies the electrical signal converted by the sensor body 310, and converts the amplified signal into acceleration. FIG. 1 shows an example in which the acceleration sensor 300 is installed at one location on the target structure 200, but the acceleration sensor 300 may be installed at multiple locations on the target structure 200. In the present disclosure, even if the number of acceleration sensors 300 installed is small, the accuracy of abnormality diagnosis can be improved.

[0014] The acceleration sensor 300 measures the acceleration value at a location where the acceleration sensor 300 is attached at regular intervals. The acceleration sensor 300 transmits data to the abnormality diagnosis device 400 for each detection cycle in which the acceleration value is measured. The data transmitted here is referred to as the acceleration measurement value 11. The acceleration sensor 300 sequentially transmits the acceleration measurement value 11 via a communication means (not shown). Alternatively, the acceleration sensor 300 may accumulate the acceleration measurement values ​​11 for a regular period in a storage device (not shown) and transmit the accumulated data set to the abnormality diagnosis device 400.

[0015] The abnormality diagnostic device 400 includes a control unit 460, an abnormality determination unit 430, and an output unit 440. The control unit 460 includes an acceleration acquisition unit 410, an estimation unit 420, an analysis program 13, and an analysis program creation unit 450. The acceleration acquisition unit 410 acquires the acceleration measurement value 11 measured by the acceleration sensor 300 . The estimation unit 420 estimates a stress estimated value and an acceleration estimated value at at least one evaluation point that is a part of the target structure 200. The estimation unit 420 also estimates a natural frequency and a damping ratio as vibration parameters of the target structure 200. The estimation unit 420 acquires an acceleration measurement value 11 from the acceleration acquisition unit 410. The estimation unit 420 acquires an analysis program 13 that has been created in advance. The estimation unit 420 then incorporates the acceleration measurement value 11 into the analysis program 13 and executes the analysis program 13. By executing this analysis program 13, the estimation unit 420 estimates a stress estimated value and an acceleration estimated value at the evaluation point, and estimates a natural frequency and a damping ratio as vibration parameters of the target structure 200.

[0016] In the embodiment, evaluation points P1, P2, and P3 shown in Fig. 1 are positions at which stress estimate values ​​and acceleration estimate values ​​are obtained. Here, evaluation points P1 and P2 are located, for example, at a joint between vehicle body 100 and target structure 200. Evaluation point P3 is located, for example, at the center of the upper surface of target structure 200. Note that, although three evaluation points P1, P2, and P3 are shown in Fig. 1, one or more arbitrary points can be selected as the target points to be evaluated on target structure 200.

[0017] The abnormality determination unit 430 determines an abnormality in the target structure 200 based on the stress estimated value and the acceleration estimated value, and the natural frequency estimated value and the damping ratio estimated value, both of which are estimated by the estimation unit 420. The determination result for the target structure 200 is output to the output unit 440. The abnormality determination unit 430 determines whether the target structure 200 is abnormal based on multiple variables, including stress estimates, acceleration estimates, and natural frequency estimates and damping ratio estimates as vibration parameters. That is, abnormalities are diagnosed by machine learning such as a support vector machine based on a combination of multiple variables or principal component analysis of the variables, or by AI (Artificial Intelligence) such as deep learning. By using multiple variables and separating features based on principal components, the system is less susceptible to disturbances, improving the accuracy of abnormality diagnosis. The abnormality determination by abnormality determination section 430 will be described in detail later.

[0018] The output unit 440 is composed of a display device, an alarm device, etc. The output unit 440 displays the determination result of the abnormality determination unit 430 on a display device, and outputs an alarm from the alarm device when an abnormality is determined. The output unit 440 may be provided outside the abnormality diagnosis device 400.

[0019] The analysis program creation unit 450 creates the analysis program 13. The analysis program creation unit 450 acquires the sensor position information, the evaluation target point position information, and other necessary parameters set by the worker. The analysis program creation unit 450 creates the analysis program 13 based on the various set data. Here, the sensor position information is the position information of the acceleration sensor 300 in the target structure 200. The evaluation target point position information is the position information of the evaluation target points P1 to P3 in the target structure 200.

[0020] FIG. 2 is a flowchart showing the process of creating the analysis program 13 by the abnormality diagnostic device 400 of FIG.

[0021] The analysis program creation unit 450 of the abnormality diagnosis device 400 constructs a finite element model that reproduces the target structure 200 in step S101. Here, the operator specifies the type of element to be used (solid element, shell element, beam element, etc.) and the type of model dimension (3D model, 2D model, etc.). The operator also inputs the material properties of the target structure 200. The material properties include the Young's modulus, Poisson's ratio, and density. The material properties used may be either isotropic or anisotropic. The analysis program creation unit 450 acquires the specified element type, the specified dimension type, and material properties, and constructs a finite element model using a CAD (Computer Aided Design) tool.

[0022] Next, in step S102, the analysis program creation unit 450 performs eigenvalue analysis on the constructed finite element model. At this time, the analysis program creation unit 450 acquires boundary conditions set by the operator to reproduce the installation state of the target structure 200. In other words, the boundary conditions refer to the fixed conditions of the target structure 200, and are conditions under which the displacement at the fixed position of the target structure 200 is set to zero, for example.

[0023] The analysis program creation unit 450 extracts natural frequencies, modal vectors, and modal stresses by performing eigenvalue analysis. Natural frequency is a numerical value that indicates the resonant frequency. When a force is applied to an object repeatedly with a certain period, for example every 0.1 seconds, the object resonates and vibrates greatly. The reciprocal of the period at that time is the natural frequency. A mode vector is the distribution of motion of an object when it resonates at its natural frequency. The mode vector of parts that vibrate strongly due to resonance becomes large. Modal stress is the distribution of stress that occurs in an object when it resonates at its natural frequency. Modal stress is high in the parts of the component that support the vibration.

[0024] In step S103, the analysis program creation unit 450 creates a degenerate model of a continuous system based on the extracted natural frequencies, mode vectors, and mode stresses for the finite element model prepared as described above.

[0025] When constructing the reduced model, the analysis program creation unit 450 acquires data for each of the first, second, ..., nth natural vibration modes. The data for the natural vibration modes includes the information on the natural frequencies, mode vectors, and mode stresses extracted in step S102.

[0026] That is, when constructing the reduced model, the analysis program creation unit 450 acquires from the worker the natural vibration modes (primary, secondary, ..., nth order), the damping ratios (modal damping ratios) for each natural vibration mode, the load input points, the number of acceleration sensors, acceleration sensor position information, the number of evaluation target points, and evaluation target point position information. Here, the load input points are points where a load is applied. Specifically, the load input points are the joints between the vehicle body 100 and the target structure 200, and are the evaluation target points P1 and P2 in Fig. 1. Further, the damping ratio here means the mode damping ratio (set for each of the first mode, second mode, . . . , nth mode).

[0027] Based on these data, the analysis program creation unit 450 creates a degenerate model of the continuous system. The following formula (1) expresses the degenerate model of the continuous system in state space representation. Here, there is a mode superposition method as a method for reproducing vibrations using the finite element method. In this method, the vibration state occurring in reality is expressed as a state in which mode vectors at multiple natural frequencies are superimposed, and the sum of the vibration ratios for each mode is expressed as the state variable z = {x}. x is called the modal coordinate displacement. For example, when a structure is excited at 15 Hz, the vibration occurring in the real world is expressed as a deformation state (x = x1 + x2 + x3 + ···) obtained by adding up all the mode vectors of the state variable x1 = 0.3 for the first mode, the state variable x2 = 0.2 for the second mode, the state variable x3 = 0.01 for the third mode, and so on. Note that the above 0.3, 0.2, and 0.01 are coefficients for the magnitude of the mode vectors obtained from the eigenvalue analysis.

[0028]

number

[0029] The state space expression of Equation (1) is composed of a state equation shown in the upper part and an observation equation shown in the lower part. Furthermore, matrices A to D are matrices obtained by the process of step S103.

[0030] The determinants of matrix A are the natural frequencies and modal damping ratios. The determinants of matrix B are the modal displacements of the load input points. The determinants of matrix C are the natural frequencies, modal damping ratios, modal displacements of the response points, and modal stresses of the response points. The determinants of matrix D are the modal displacements of the load input points, and modal displacements of the response points.

[0031] In the formula (1), the state quantity of the displacement and the state quantity of the stress in the real world are represented by {W, σ}. That is, the acceleration acquired by the acceleration sensor 300 is represented on the left side of the observation equation as W (two dots) obtained by second-order differentiation of the displacement W. Also, the estimated value of the desired stress is represented on the left side of the observation equation as σ.

[0032] In formula (1), as the number of evaluation points increases, the number of rows in matrices C and D increases. Also, as the number of acceleration sensors 300 increases, the number of rows in matrices C and D increases.

[0033] Returning to the explanation of the flowchart in Fig. 2, in step S104, the analysis program creation unit 450 discretizes the created continuous system degenerate model. The acceleration measurement values ​​11 are data that are discrete for each detection cycle of the acceleration sensor 300. For this reason, the analysis program creation unit 450 discretizes the continuous system degenerate model with respect to time.

[0034] As shown in formula (1), the degenerated model of the continuous system is composed of two equations, a state equation and an observation equation. Of these two equations, only the state equation is a continuous system equation. Therefore, the analysis program creation unit 450 performs discretization on the state equation.

[0035] The following equation (2) expresses a degenerate model of a discrete system in state space representation.

[0036]

number

[0037] Fig. 3 is a diagram for explaining a method for discretizing a degenerate model of a continuous system. To discretize, it is necessary to find matrices A2 and B2 in the above formula (2). Here, matrices A2 and B2 are found using, for example, the fourth-order Runge-Kutta method.

[0038] Formula F1 in Fig. 3 is a discretized formula of the fourth-order Runge-Kutta method. If we assume that the weight u is constant in this formula F1, the discrete matrix shown in formula F2 in Fig. 3 is calculated. Although the discretization formula based on the Runge-Kutta method is used in the description here, other discretization formulas such as the Euler method and the Hoyne method may also be used.

[0039] In formula F2, by replacing AdT with k1a, BdT with k1b, etc., matrix A2 and matrix B2 become as shown in formula F3, where I in formula F3 represents a unit matrix.

[0040] Returning to the explanation of the flowchart in Fig. 2, in step S105, analysis program creation unit 450 creates analysis program 13. Analysis program 13 is a program that combines a degenerate model after discretization, that is, a state space expression after discretization, with a Kalman filter. The Kalman filter used is a nonlinear Kalman filter such as an extended Kalman filter or an unscented Kalman filter.

[0041] The analysis program 13 thus created is called by the estimation unit 420 during the estimated value calculation process described below.

[0042] FIG. 4 is a flowchart showing an acceleration measurement process performed by the acceleration sensor 300 of FIG.

[0043] In step S201, the acceleration sensor 300 starts measuring acceleration. For example, when a start operation by a worker is received or when a specified time arrives, the acceleration sensor 300 starts measuring acceleration.

[0044] In step S202, the sensor body 310 of the acceleration sensor 300 converts vibrations occurring at the position where it is attached into an electrical signal.

[0045] In step S203, the amplifier 320 amplifies the signal from the sensor main body 310 and converts it into a voltage signal. The amplifier 320 converts the amplified voltage signal into an acceleration by coefficient calculation.

[0046] In step S204, the acceleration sensor 300 transmits the measured acceleration to the abnormality diagnosis device 400 as an acceleration measurement value 11. This acceleration measurement value 11 is acquired by the acceleration acquisition unit 410 of the abnormality diagnosis device 400.

[0047] In step S205, the acceleration sensor 300 determines whether to end the acceleration measurement. For example, when the acceleration sensor 300 receives an end operation from the worker or when a specified time arrives, the acceleration sensor 300 ends the acceleration measurement. If the acceleration sensor 300 does not end the acceleration measurement, the process returns to step S202.

[0048] Fig. 5 is a flowchart showing the calculation process of an estimated value by the abnormality diagnosis device 400 in Fig. 1. Here, load control on the target structure 200 is not taken into consideration. That is, in the flowchart in Fig. 5, the matrix D shown in the formula (2) is not used. Note that when the load on the target structure 200 is known, the matrix D is also used.

[0049] The estimation unit 420 of the abnormality diagnostic device 400 calls the acceleration measurement value 11 and the analysis program 13 to perform an estimated value calculation process.

[0050] In step S301, the estimation unit 420 sets initial values ​​for the state estimate value X and the error covariance matrix P. The state estimate value X here means the above-mentioned state variable z.

[0051] Since there is no vibration in the initial state, the state estimate value X is 0. In addition, for the error covariance matrix P, a unit matrix is ​​set as the initial value I here.

[0052] In addition to the above initial values, constant parameters of the system noise Q and the observation noise R are also set. The system noise Q and the observation noise R are values ​​appropriately determined by the operator according to the noise level.

[0053] In step S302, the estimation unit 420 calculates the prior state estimate X using the state estimate X and the matrix A2. - Here, the prior state estimate X - is the estimate of the state variable z at time t based on the data available up to time t-1.

[0054] In step S303, the estimation unit 420 calculates a prior error covariance matrix P using the error covariance matrix P, the matrix A2, the matrix B2, the system noise Q, the matrix A2′, and the matrix B2′. - Here, matrix A2' is a matrix obtained by transposing matrix A2, and matrix B2' is a matrix obtained by transposing matrix B2.

[0055] In step S304, the estimation unit 420 calculates the prior error covariance matrix P - , matrix C, matrix C', and observation noise R are used to calculate the Kalman gain matrix G. Here, matrix C' is a matrix obtained by transposing matrix C.

[0056] In step S305, the estimation unit 420 calculates the Kalman gain matrix G and the prior state estimate X -, matrix C′, and acceleration measurement value 11 at time t to calculate state estimate value X. Estimation unit 420 updates the state estimate value X that has been used up until now to the calculated one.

[0057] In step S306, the estimation unit 420 estimates the natural frequency and modal damping ratio based on the updated state estimate value X and formula (2), and calculates the acceleration estimate value W (two dots) and the stress estimate value σ at the evaluation point. The estimation unit 420 substitutes the state estimate value X as a state variable z(t+Δt) into the left side of the state equation shown in formula (2). If the load u(t) is not taken into consideration or is constant, the estimation unit 420 can obtain the state variable z(t) from one cycle ago from the state equation of formula (2).

[0058] That is, the estimation unit 420 estimates the natural frequency and the modal damping ratio based on the updated state estimate value X (obtains an estimated value of the natural frequency and an estimated value of the modal damping ratio), and further obtains an estimated acceleration value W (two dots) and an estimated stress value σ at the evaluation points P1 to P3 based on the updated state estimate value X and formula (2).

[0059] Furthermore, the estimation unit 420 additionally writes the obtained acceleration estimate W (two dots), stress estimate σ, eigenvalue estimate, and modal damping ratio estimate to the data stored up to that point. In this manner, the acceleration estimate, stress estimate, eigenfrequency estimate, and modal damping ratio estimate are generated.

[0060] In step S307, the estimation unit 420 calculates the initial value I of the error covariance matrix, the Kalman gain matrix G, the matrix C′, and the prior error covariance matrix P - The estimation unit 420 uses the above to calculate the error covariance matrix P. The estimation unit 420 updates the error covariance matrix P that has been used so far to the calculated one.

[0061] In step S308, the estimation unit 420 determines whether or not to acquire a new acceleration measurement value 11 within a specified time. If the estimation unit 420 acquires an acceleration measurement value 11 within the specified time, the process returns to step S302. Then, the estimation unit 420 performs the processes from step S302 onwards on the newly acquired acceleration measurement value 11. If the estimation unit 420 does not acquire an acceleration measurement value 11 within the specified time, the process ends.

[0062] Next, the abnormality determination of the target structure 200 by the abnormality determination unit 430 will be described.

[0063] [Unsupervised learning] First, unsupervised learning for realizing abnormality judgment of the target structure 200 will be described. Here, as input A1, the historical values ​​of the estimated stress values ​​and the historical values ​​of the estimated acceleration values ​​for each of the evaluation points P1, P2, and P3, as well as the estimated natural frequency and the estimated damping ratio as vibration parameters of the target structure 200, are input. Then, according to learning data created based on the input A1, a C output is output by so-called unsupervised learning, which indicates normal and abnormal states of the target structure 200 for the input A1, which is data for inference (for example, a normal state with no abnormalities, an abnormal state with loose bolts, an abnormal state with a crack, an abnormal state with a break).

[0064] <Learning Phase> 6 is a block configuration diagram showing a first learning device 500A in the abnormality determination unit 430. The first learning device 500A includes a 1-1 data acquisition unit 510A, a 1-1 model generation unit 520A, and a first learned model storage unit 530A.

[0065] The 1-1st data acquiring unit 510A acquires an input A1 as learning data. Here, the input A1 is the historical values ​​of the estimated stress values ​​and the historical values ​​of the estimated acceleration values ​​at the evaluation points P1, P2, and P3, as well as the estimated natural frequency and the estimated damping ratio as vibration parameters of the target structure 200. Here, the history value of the stress estimated value refers to the history value of the stress estimated value during a preset period, and the history value of the acceleration estimated value refers to the history value of the stress estimated value during a preset period.

[0066] The 1-1 model generation unit 520A learns the C output (the normal or abnormal state of the target structure 200) based on the learning data created based on the input A1 output from the 1-1 data acquisition unit 510A. That is, it generates a learned model that infers the C output from the input A1 of the target structure 200. Here, the learning data is data in which the inputs A1 are associated with each other.

[0067] The learning algorithm used by the 1-1 model generation unit 520A may be a known algorithm. Here, a case where the K-means method (clustering), which is unsupervised learning, is applied will be described. Unsupervised learning refers to a method of learning features in learning data by providing the learning device with learning data that does not include results (labels).

[0068] The 1-1st model generating unit 520A learns the C output by so-called unsupervised learning, for example, in accordance with a grouping technique using the K-means method.

[0069] K-means is a non-hierarchical clustering algorithm that uses the mean of the clusters to classify a given number of clusters into k.

[0070] Specifically, the K-means algorithm is processed as follows: First, a cluster is randomly assigned to each data xi. Next, the center Vj of each cluster is calculated based on the assigned data. Next, the distance between each xi and each Vj is calculated, and xi is reassigned to the closest central cluster. Then, if the cluster assignment for all xi remains unchanged through the above process, or if the amount of change falls below a certain threshold set in advance, it is determined that convergence has occurred and the process ends.

[0071] In this embodiment, the C output is learned by so-called unsupervised learning in accordance with learning data created based on a combination of inputs A1 acquired by the 1-1st data acquisition unit 510A.

[0072] The 1-1 model generation unit 520A executes the above-mentioned learning to generate and output a trained model.

[0073] The first trained model storage unit 530A stores the trained model output from the 1-1 model generation unit 520A.

[0074] Next, the learning process of the first learning device 500A will be described with reference to Fig. 7. Fig. 7 is a flowchart relating to the learning process of the first learning device 500A.

[0075] In step S510A, the 1-1st data acquisition unit 510A acquires the input A1. Note that, although the input A1 is acquired simultaneously, it is sufficient that the input A1 is input in a related manner, and the data of the input A1 may be acquired at different timings.

[0076] In step S520A, the 1-1st model generation unit 520A learns the C output by so-called unsupervised learning in accordance with the learning data created based on the input A1 acquired by the 1-1st data acquisition unit 510A, and generates a learned model.

[0077] In step S530A, the first trained model storage unit 530A stores the trained model generated by the 1-1 model generation unit 520A.

[0078] <Utilization phase> 8 is a block diagram showing a first inference device 600A in the abnormality determination section 430. The first inference device 600A includes a first-second data acquisition section 610A and a first-second inference section 620A.

[0079] The 1-2 data acquisition unit 610A acquires input A1 for inference, which is the historical stress estimate values ​​and historical acceleration estimate values ​​at the evaluation points P1, P2, and P3, respectively, as well as the estimated natural frequency and the estimated damping ratio as vibration parameters of the target structure 200.

[0080] The 1-2 inference unit 620A infers the C output (normal or abnormal state of the target structure 200) obtained by using the learned model stored in the first learned model storage unit 530A. That is, by inputting the inference input A1 acquired by the 1-2 data acquisition unit 610A to this learned model, it is possible to infer to which cluster the inference input A1 belongs and output the inference result as the C output.

[0081] In this embodiment, the C output is described as being output using a trained model trained by the 1-1 model generation unit 520A of the abnormality diagnosis device 400. However, it is also possible to obtain a trained model from an external source, such as another abnormality diagnosis device 400, and output the C output based on this trained model.

[0082] In this manner, the first-second inference section 620A outputs the C output obtained based on the input A1 to the output section 440 of the abnormality diagnosis device 400.

[0083] Next, the process for obtaining the C output using the first-second inference unit 620A will be described with reference to FIG.

[0084] In step S610A, the 1-2 data acquisition unit 610A acquires an input A1.

[0085] In step S620A, the first-2 inference unit 620A inputs an input A1 to the learned model stored in the first learned model storage unit 530A, and obtains a C output.

[0086] In step S630A, the first-second inference unit 620A outputs the C output obtained by the trained model to the output unit.

[0087] In step S640A, the output unit 440 displays on a display device based on the output C output (normal or abnormal state of the target structure 200), or outputs an alarm from an alarm device if an abnormality is determined.

[0088] When implementing the unsupervised learning in this embodiment, the method is not limited to the non-hierarchical clustering using the K-means method described above, and any other known method capable of clustering may be used. For example, hierarchical clustering such as the shortest distance method may be used.

[0089] In this embodiment, the first learning device 500A and the first inference device 600A may be connected to the abnormality diagnosis device 400 via a network, for example, and may be devices separate from the abnormality diagnosis device 400. The first learning device 500A and the first inference device 600A may be built into the abnormality diagnosis device 400. Furthermore, the first learning device 500A and the first inference device 600A may exist on a cloud server.

[0090] The 1-1 model generation unit 520A may learn the C output according to learning data created for a plurality of abnormality diagnosis devices 400. The 1-1 model generation unit 520A may acquire learning data from a plurality of abnormality diagnosis devices 400 used in the same area, or may learn the C output using learning data collected from a plurality of abnormality diagnosis devices 400 operating independently in different areas. It is also possible to add or remove an abnormality diagnosis device 400 that collects learning data to or from the target midway. Furthermore, a learning device that has learned the C output for a certain abnormality diagnosis device may be applied to another abnormality diagnosis device, and the C output for the other abnormality diagnosis device may be re-learned and updated.

[0091] [Supervised learning using neural networks] Next, supervised learning using a neural network for realizing abnormality determination of the target structure 200 will be described.

[0092] As input B1 of the neural network, the history values ​​of the stress estimates and the history values ​​of the acceleration estimates of the evaluation points P1, P2, and P3, as well as the natural frequency estimates and the damping ratio estimates as vibration parameters, are input. In addition, as input B2 (correct answer), the normal and abnormal states of the target structure 200 (e.g., a normal state without any abnormalities, an abnormal state with loose bolts, an abnormal state with a crack, and an abnormal state with a break) are input. Then, according to learning data created based on a combination of inputs B1 and B2, the normal and abnormal states of the target structure 200 (e.g., a normal state without any abnormalities, an abnormal state with loose bolts, an abnormal state with a crack, and an abnormal state with a break), which are output 1, are learned by so-called supervised learning. A specific explanation will be given below with reference to the drawings.

[0093] <Learning Phase> 10 is a block configuration diagram showing the second learning device 500B in the abnormality determination unit 430. The second learning device 500B includes a 2-1 data acquisition unit 510B, a 2-1 model generation unit 520B, and a second learned model storage unit 530B.

[0094] The 2-1 data acquisition unit 510B acquires the input B1 and the input B2 (correct answer) as learning data. As described above, the input B1 is the history values ​​of the estimated stress values ​​and the history values ​​of the estimated acceleration values ​​of each of the evaluation points P1, P2, and P3, as well as the estimated natural frequency and the estimated damping ratio as vibration parameters. Here, the history value of the stress estimated value refers to the history value of the stress estimated value during a preset period, and the history value of the acceleration estimated value refers to the history value of the stress estimated value during a preset period. The input B2 is the normal or abnormal state of the target structure 200 (for example, a normal state, an abnormal state in which a bolt is loose, an abnormal state in which a crack is present, an abnormal state in which a break is present, etc.).

[0095] The 2-1 model generation unit 520B learns the C output (the normal or abnormal state of the target structure 200) based on learning data created based on a combination of the input B1 and the input B2 (correct answer) output from the 2-1 data acquisition unit 510B. That is, it generates a learned model that infers the optimal C output from the input B1 and the input B2 (correct answer) of the second learning device 500B. Here, the learning data is data in which the input B1 and the input B2 (correct answer) are associated with each other.

[0096] The 2-1st model generation unit 520B learns the C output by so-called supervised learning, for example, according to a neural network model. Here, supervised learning refers to a method of providing a learning device with a set of input and result (label) data, learning the features of the learning data, and inferring the result from the input.

[0097] A neural network is composed of an input layer made up of multiple neurons, an intermediate layer (hidden layer) made up of multiple neurons, and an output layer made up of multiple neurons. The intermediate layer may be one layer, or two or more layers.

[0098] For example, in a three-layered neural network as shown in Figure 14, when multiple inputs are input to the input layer (X1-X3), the values ​​are multiplied by weight W1 (w11-w16) and input to the intermediate layer (Y1-Y2), and the result is further multiplied by weight W2 (w21-w26) and output from the output layer (Z1-Z3). This output result changes depending on the values ​​of weights W1 and W2.

[0099] In this embodiment, the neural network learns the C output by so-called supervised learning according to learning data created based on a combination of input B1 and input B2 (correct answer) acquired by 2-1 data acquisition unit 510B.

[0100] In other words, the neural network learns by inputting an input B1 to the input layer and adjusting the weights W1 and W2 so that the result output from the output layer approaches the input B2 (correct answer).

[0101] The 2-1 model generation unit 520B executes the above-mentioned learning to generate and output a trained model.

[0102] The second trained model storage unit 530B stores the trained model output from the 2-1 model generation unit 520B.

[0103] Next, the learning process of the second learning device 500B will be described with reference to Fig. 11. Fig. 11 is a flowchart relating to the learning process of the second learning device 500B.

[0104] In step S510B, the 2-1 data acquisition unit 510B acquires the input B1 and the input B2 (correct answer). Note that, although the input B1 and the input B2 (correct answer) are acquired simultaneously, it is sufficient that the input B1 and the input B2 (correct answer) are associated with each other and the data of the input B1 and the input B2 (correct answer) may be acquired at different times.

[0105] In step S520B, the 2-1 model generation unit 520B learns the C output by so-called supervised learning according to learning data created based on a combination of input B1 and input B2 (correct answer) acquired by the 2-1 data acquisition unit 510B, and generates a learned model.

[0106] In step S530B, the second trained model storage unit 530B stores the trained model generated by the 2-1 model generation unit 520B.

[0107] <Utilization phase> 12 is a block diagram showing the configuration of a second inference device 600B in the abnormality determination section 430. The second inference device 600B includes a 2-2 data acquisition section 610B and a 2-2 inference section 620B.

[0108] The 2-2 data acquisition unit 610B acquires an input B1.

[0109] The 2-2 inference unit 620B infers the C output (normal and abnormal states of the target structure) obtained by using the trained model. That is, by inputting the input B1 acquired by the 2-2 data acquisition unit 610B to this trained model, the C output inferred from the input B1 can be output.

[0110] In this embodiment, the C output is described as being output using a learned model learned by 2-1 model generation unit 520B of abnormality judgment unit 430. However, it is also possible to obtain a learned model from an external source, such as another abnormality judgment unit, and output the C output based on this learned model.

[0111] Next, the process for obtaining the C output (normal and abnormal states of the target structure) using the second inference device 600B will be described with reference to FIG.

[0112] In step S610B, the 2-2 data acquisition unit 610B acquires an input B1.

[0113] In step S620B, the 2-2 inference unit 620B inputs the input B1 to the learned model stored in the second learned model storage unit 530B, and obtains the C output.

[0114] In step S630B, the 2-2 inference unit 620B outputs the C output obtained by the learned model to the abnormality determination unit 430.

[0115] In step S640B, the output unit 440 displays on a display device based on the output C output (normal or abnormal state of the target structure 200), or outputs an alarm from an alarm device if an abnormality is determined.

[0116] In the present embodiment, the case where supervised learning is applied to the learning algorithm used by the 2-1 model generation unit 520B has been described, but the present invention is not limited to this. As for the learning algorithm, reinforcement learning and the like can also be applied in addition to supervised learning.

[0117] The 2-1 model generation unit 520B may learn the C output according to learning data created for a plurality of abnormality diagnosis devices. The 2-1 model generation unit 520B may acquire learning data from a plurality of abnormality diagnosis devices used in the same area, or may learn the C output using learning data collected from a plurality of abnormality diagnosis devices operating independently in different areas. It is also possible to add an abnormality diagnosis device that collects learning data to the target or remove it from the target midway. Furthermore, a learning device that has learned the C output for a certain abnormality diagnosis device may be applied to another abnormality diagnosis device, and the C output for the other abnormality diagnosis device may be re-learned and updated. Furthermore, the learning data may be generated by performing an abnormality reproduction experiment on the target structure in advance or by a simulation using a finite element model or the like.

[0118] In addition, the learning algorithm used in the 2-1 model generation unit 520B may be deep learning, which learns to extract the features themselves, or machine learning may be performed according to other known methods, such as genetic programming, functional logic programming, and support vector machines.

[0119] In addition, in the above embodiment, a large number of explanatory variables (estimated stress value, estimated acceleration value, estimated natural frequency value, and estimated damping ratio value) are input as input A1 and input B1 to make them multivariable, but it is also possible to perform principal component analysis to reduce the many explanatory variables (multiple variables) to fewer dimensions, and to speed up abnormality diagnosis by principal component analysis of input A1 and input B1.

[0120] Each function of the abnormality diagnosis device 400 of this embodiment is realized by a processing circuit. Fig. 15 is a configuration diagram showing a first example of a processing circuit that realizes each function of the abnormality diagnosis device 400 of this embodiment. The processing circuit 700 of the first example is dedicated hardware.

[0121] The processing circuit 700 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination of these. Each function of the abnormality diagnosis device 400 may be realized by an individual processing circuit 700. Alternatively, each function of the abnormality diagnosis device 400 may be realized by the processing circuit 700 collectively.

[0122] 16 is a diagram showing a second example of a processing circuit that realizes each function of the abnormality diagnostic device 400 according to the embodiment. The processing circuit 800 of the second example includes a processor 810 and a memory 820.

[0123] In the processing circuit 800, each function of the abnormality diagnosis device 400 is realized by software, firmware, or a combination of software and firmware. The software and firmware are written as programs. The software and firmware are then stored in the memory 820. The processor 810 realizes the function of each unit by reading and executing the programs stored in the memory 820.

[0124] It can be said that the programs stored in the memory 820 cause the computer to execute the procedures or methods of each of the above-mentioned sections. Here, the memory 820 corresponds to non-volatile or volatile semiconductor memories such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable and Programmable Read Only Memory), etc. In addition, the memory 820 also corresponds to magnetic disks, flexible disks, optical disks, compact disks, mini disks, DVDs, etc.

[0125] The functions of each of the above-mentioned units may be partially realized by dedicated hardware and partially realized by software or firmware.

[0126] In this manner, the processing circuitry can realize the functions of each of the above-described units through hardware, software, firmware, or a combination of these.

[0127] Although exemplary embodiments are described herein, the various features, aspects, and functions described in the embodiments are not limited to application to a particular embodiment, but may be applied to the embodiments alone or in various combinations. Therefore, countless modifications not illustrated are expected within the scope of the technology disclosed in this application, including, for example, modifying, adding, or omitting at least one component.

[0128] Various aspects of the present disclosure are summarized below as appendices. [Appendix 1] Acquire acceleration measurement values ​​measured by an acceleration sensor installed on the target structure, performing an eigenvalue analysis on a shape model that models the structure of the target structure, constructing a reduced model based on the result of the eigenvalue analysis, and a control unit that estimates an acceleration estimate and a stress estimate at at least one evaluation point that is a part of the target structure, based on the reduced model and the acceleration measurement value, as well as an estimate of a natural frequency and an estimate of a damping ratio, which are vibration parameters of the target structure; and an abnormality determination unit that determines whether the target structure is in a normal or abnormal state through machine learning based on the acceleration estimated value, the stress estimated value, the natural frequency estimated value, and the damping ratio estimated value. [Appendix 2] The shape model is a finite element model obtained by reproducing the target structure by a finite element method, The control unit performs an eigenvalue analysis on the finite element model to extract a natural frequency and set a damping ratio, which are vibration parameters of the target structure; constructing the degenerated model including at least an acceleration element, a stress element, a natural frequency element, and a damping ratio element; 2. An abnormality diagnosis system for a target structure as described in claim 1, which estimates the acceleration estimated value, the stress estimated value, the natural frequency estimated value, and the damping ratio estimated value based on the degenerated model and the acceleration measurement value. [Appendix 3] 3. An abnormality diagnosis system for a target structure as described in Appendix 1 or Appendix 2, wherein the degenerated model is represented by a state space representation composed of a state equation and an observation equation including at least an acceleration element, a stress element, a natural frequency element, and a damping ratio element. [Appendix 4] the control unit executes an analysis program to calculate the acceleration estimated value, the stress estimated value, the natural frequency estimated value, and the damping ratio estimated value; 4. An abnormality diagnosis system for a target structure as described in Appendix 3, wherein the analysis program is a program created in advance based on a combination of the state space representation composed of the state equation and the observation equation and a Kalman filter. [Appendix 5] The control unit includes an analysis program creation unit that creates the analysis program, The analysis program creation unit acquiring sensor position information, which is position information of the acceleration sensor in the target structure, and evaluation target point position information, which is position information of the evaluation target point in the target structure; The abnormality diagnosis system for a target structure described in Appendix 4, wherein the analysis program is created based on the sensor position information and the evaluation target point position information. [Appendix 6] The abnormality determination unit a first-first data acquisition unit that acquires learning data including a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio; a first-first model generation unit that generates a trained model for inferring normal and abnormal states of the target structure from a history value of the acceleration estimate value, a history value of the stress estimate value, an estimate of the natural frequency, and an estimate of the damping ratio, using the learning data; An abnormality diagnosis system for a target structure as described in any one of Supplementary Note 1 to Supplementary Note 5, comprising a first learning device having the above structure. [Appendix 7] The abnormality determination unit a first-second data acquisition unit that acquires a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio; a first-second inference unit that uses a trained model for inferring normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate of the natural frequency, and the estimate of the damping ratio, and outputs normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate of the natural frequency, and the estimate of the damping ratio acquired by the first-second data acquisition unit; An abnormality diagnosis system for a target structure as described in Appendix 6, comprising a first inference device having the above-mentioned features. [Appendix 8] The abnormality determination unit a 2-1 data acquisition unit that acquires learning data including a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio, as well as normal and abnormal states of the target structure; a 2-1 model generation unit that generates a trained model for inferring normal and abnormal states of the target structure from a history value of the acceleration estimate value, a history value of the stress estimate value, an estimate of the natural frequency, and an estimate of the damping ratio, using the learning data; and An abnormality diagnosis system for a target structure as described in any one of Supplementary Note 1 to Supplementary Note 5, comprising a second learning device having the above structure. [Appendix 9] The abnormality determination unit is a 2-2 data acquisition unit that acquires a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio; a first-second inference unit that uses a trained model for inferring normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate value of the natural frequency, and the estimate value of the damping ratio, and outputs normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate value of the natural frequency, and the estimate value of the damping ratio acquired by the second-second data acquisition unit; An abnormality diagnosis system for a target structure as described in appendix 8, comprising a second inference device having the above-mentioned. [Explanation of symbols]

[0129] 11 Acceleration measurement value, 13 Analysis program, 200 Target structure, 300 acceleration sensor, 400 abnormality diagnosis device, 410 acceleration acquisition unit, 420 Estimation unit, 430 Abnormality determination unit, 440 Output unit, 450 analysis program creation unit, 460 control unit, 500A first learning device, 510A 1-1st data acquisition unit, 520A 1-1st model generation unit, 530A first trained model storage unit, 600A first inference device, 610A first-second data acquisition unit, 620A first-second inference unit, 500B second learning device, 510B second-1 data acquisition unit, 520B 2-1st model generation unit, 530B 2nd trained model storage unit, 600B second inference device, 610B second-2 data acquisition unit, 620B 2nd-2nd inference unit, 1000 Abnormality diagnosis system for target structure.

Claims

1. Acquire acceleration measurement values ​​measured by an acceleration sensor installed on the target structure, performing an eigenvalue analysis on a shape model that models the structure of the target structure, constructing a reduced model based on the result of the eigenvalue analysis, and a control unit that estimates an acceleration estimate and a stress estimate at at least one evaluation point that is a part of the target structure, based on the reduced model and the acceleration measurement value, as well as an estimate of a natural frequency and an estimate of a damping ratio, which are vibration parameters of the target structure; and an abnormality determination unit that determines whether the target structure is in a normal or abnormal state through machine learning based on the acceleration estimated value, the stress estimated value, the natural frequency estimated value, and the damping ratio estimated value.

2. The shape model is a finite element model obtained by reproducing the target structure by a finite element method, The control unit performs an eigenvalue analysis on the finite element model to extract a natural frequency and set a damping ratio, which are vibration parameters of the target structure; constructing the degenerated model including at least an acceleration element, a stress element, a natural frequency element, and a damping ratio element; 2. The abnormality diagnosis system for a target structure according to claim 1, wherein the estimated acceleration value, the estimated stress value, the estimated natural frequency value, and the estimated damping ratio value are estimated based on the reduced model and the acceleration measurement value.

3. 3. The abnormality diagnosis system for a target structure according to claim 1 or 2, wherein the degenerated model is represented by a state space representation composed of a state equation and an observation equation, the state equation and the observation equation including at least an acceleration element, a stress element, a natural frequency element, and a damping ratio element.

4. the control unit executes an analysis program to calculate the acceleration estimated value, the stress estimated value, the natural frequency estimated value, and the damping ratio estimated value; 4. The abnormality diagnosis system for a target structure according to claim 3, wherein the analysis program is a program that is created in advance based on a combination of the state space representation constituted by the state equation and the observation equation and a Kalman filter.

5. The control unit includes an analysis program creation unit that creates the analysis program, The analysis program creation unit acquiring sensor position information, which is position information of the acceleration sensor in the target structure, and evaluation target point position information, which is position information of the evaluation target point in the target structure; 5. The target structure abnormality diagnosis system according to claim 4, wherein the analysis program is created based on the sensor position information and the evaluation target point position information.

6. The abnormality determination unit is a first-first data acquisition unit that acquires learning data including a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio; a first-first model generation unit that generates a trained model for inferring normal and abnormal states of the target structure from a history value of the acceleration estimate value, a history value of the stress estimate value, an estimate of the natural frequency, and an estimate of the damping ratio, using the learning data; and 3. An abnormality diagnosis system for a target structure according to claim 1 or 2, comprising a first learning device having the following:

7. The abnormality determination unit is a first-second data acquisition unit that acquires a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio; a first-second inference unit that uses a trained model for inferring normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate of the natural frequency, and the estimate of the damping ratio, and outputs normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate of the natural frequency, and the estimate of the damping ratio acquired by the first-second data acquisition unit; 7. The target structure abnormality diagnosis system according to claim 6, comprising a first inference device having the following:

8. The abnormality determination unit is a second-1 data acquisition unit that acquires learning data including a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio, as well as normal and abnormal states of the target structure; a second-1 model generation unit that generates a trained model for inferring normal and abnormal states of the target structure from a history value of the acceleration estimate value, a history value of the stress estimate value, an estimate of the natural frequency, and an estimate of the damping ratio, using the learning data; and The abnormality diagnosis system for a target structure according to claim 1 or 2, further comprising a second learning device having the above-mentioned.

9. The abnormality determination unit is a second-2 data acquisition unit that acquires a history value of the acceleration estimation value, a history value of the stress estimation value, an estimate of the natural frequency, and an estimate of the damping ratio; a first-2 inference unit that outputs normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate of the natural frequency, and the estimate of the damping ratio acquired by the second-2 data acquisition unit, using a trained model for inferring normal and abnormal states of the target structure from the history values ​​of the acceleration estimate value, the history values ​​of the stress estimate value, the estimate of the natural frequency, and the estimate of the damping ratio; The abnormality diagnosis system for a target structure according to claim 8, further comprising a second inference device having the following:

Images