A performance status evaluation method, device and electronic device based on spatiotemporal attention mechanism

Abstract

The present invention discloses a performance status evaluation method, device and electronic device based on a spatiotemporal attention mechanism, the method comprising: using a detection data sliding window reconstruction method to convert equipment operation monitoring performance degradation data into three-dimensional time series data; graphically expressing the performance monitoring sensor as a graph neural network adjacency matrix according to the relationship between the sensor physical layout and the equipment structure; constructing a STAGCN model based on spatiotemporal attention graph convolution, and inputting the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; and performing performance status evaluation on the equipment to be evaluated based on the trained STAGCN model.

Description

Technical Field

[0001] The present invention relates to the technical field of performance status evaluation, and in particular to a performance status evaluation method, device and electronic device based on a spatiotemporal attention mechanism. Background Art

[0002] In the field of equipment performance status assessment and performance degradation prediction, with the development of signal acquisition and transmission technology and the widespread use of status monitoring software and hardware systems, compared with the past, the current equipment status monitoring data has the characteristics of a large number of signal channels, high sampling frequency, and long sampling time. In order to integrate the complete degradation information of the equipment, time series monitoring data obtained by different sensors are usually used as the input of the network. However, there are differences in the degradation characterization capabilities of each sensor. In addition, as the degree of fault deepens, the description of the performance degradation evolution process at different time series positions is also quite different. A large amount of status monitoring data covers more and more complex equipment degradation information. However, the increase in the number of acquisition channels also reduces the degradation information density. Therefore, the accurate identification and positioning of degradation information and efficient characterization have become new capabilities that performance status assessment models urgently need to have.

[0003] After the equipment is put into use after leaving the factory, the performance status of the equipment gradually degrades as the running time continues to accumulate. In this process, the characterization of the equipment performance status by the data monitored by different sensors and at different time points is not independent, but dynamically related to each other. This phenomenon brings about the problem of extracting the degradation features of the correlation between the monitoring signal channels and between the time before and after in the process of equipment performance status evaluation. This process has certain similarities with the process of human observation and thinking. When humans receive visual, auditory and other input information, they can actively focus their attention on the input information with higher information density through the experience learned in the past. Therefore, inspired by the highly intelligent thinking process of human beings, the attention mechanism is introduced in the process of equipment performance status evaluation to improve the efficiency of mining performance degradation information in monitoring data and direct the attention of model learning to the data space with higher information density of performance status degradation.

[0004] Existing attention mechanisms, such as the transformer model, pay more attention to the mapping relationship between previous data and subsequent data in time series data, which is called temporal attention. However, in the process of predicting device performance degradation, there is often more than one sampled monitoring data, that is, the input dimension is large, but different monitoring data are interrelated and have different performance representation capabilities. Therefore, paying attention to the input dimension, that is, spatial attention, is of great significance for deconstructing multi-dimensional monitoring signals and improving the ability to extract performance degradation features. Summary of the invention

[0005] The present invention provides a performance status evaluation method, device and electronic device based on the spatiotemporal attention mechanism, so as to solve the technical problem of how to achieve accurate identification and positioning of degradation information and efficient characterization requirements under the actual background of increasing equipment complexity and increasing monitoring signal parameters.

[0006] An embodiment of the present invention provides a performance status evaluation method based on a spatiotemporal attention mechanism, including: using a detection data sliding window reconstruction method to convert equipment operation monitoring performance degradation data into three-dimensional time series data; graphically expressing the performance monitoring sensor as a graph neural network adjacency matrix according to the relationship between the sensor physical layout and the equipment structure; constructing a STAGCN model based on spatiotemporal attention graph convolution, and inputting the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; and performing performance status evaluation on the equipment to be evaluated based on the trained STAGCN model.

[0007] Preferably, the STAGCN model includes a spatiotemporal attention model and an attention graph convolution model; the three-dimensional time series data includes monitoring dimension time series data, time dimension time series data and operation dimension time series data.

[0008] Preferably, the three-dimensional time series data and the graph neural network adjacency matrix are input into the STAGCN model for training, and the trained STAGCN model is obtained, including: inputting the three-dimensional time series data into the time attention machine in the spatiotemporal attention model to calculate the time attention matrix; inputting the time attention matrix and the three-dimensional time series data into the space attention machine in the spatiotemporal attention model to calculate the space attention matrix; multiplying the space attention matrix with the graph neural network adjacency matrix to construct an attention graph convolution model, and inputting the three-dimensional time series data into the attention graph convolution model to obtain the evaluation health index of the device; calculating the loss value of the evaluation health index and the evaluation label, and based on the loss value, obtaining the trained STAGCN model.

[0009] Preferably, the temporal attention matrix comprises:

[0010]

[0011] Where E is the temporal attention matrix, is the first r The input of the spatiotemporal network block, , , , These are the parameters of the temporal attention matrix to be learned. is the activation function.

[0012] Preferably, the spatial attention matrix includes:

[0013]

[0014] in, S is the spatial attention matrix, is the first r The input of the spatiotemporal network block, , , , , These are the parameters of the spatial attention matrix to be learned. is the activation function.

[0015] Preferably, according to the relationship between the physical layout of the sensor and the device structure, the performance monitoring sensor is graphically expressed as a graph neural network adjacency matrix including:

[0016] in, S is the spatial attention matrix, is the first r The input of the spatiotemporal network block, , , , , These are the parameters of the spatial attention matrix to be learned. is the activation function.

[0017]

[0018] in, Indicates The monitoring parameters and The edge weights between monitoring parameters are is the correlation coefficient of sensors at the same point, and its range is , m and n They represent different collection points respectively.

[0019] Preferably, the sensors are multiple types of sensors, and one or more different types of sensors are installed at each position of the equipment structure;

[0020] The number of sensor acquisition channels is:

[0021]

[0022] in, Indicates the number of acquisition channels; Indicates m The number of sensors at each collection point; Indicates the number of sensor locations.

[0023] Preferably, the detection data sliding window reconstruction method is used to convert the equipment operation monitoring performance degradation data into three-dimensional time series data, including: using a sliding time window with a fixed length and a fixed step size to divide the data in each start-up and shutdown working cycle in the entire life cycle from operation to failure of the training into small multidimensional time series, and stacking the multidimensional time series into three-dimensional time series data.

[0024] An embodiment of the present invention also provides a performance status evaluation device based on a spatiotemporal attention mechanism, including: a data preprocessing module, used to convert equipment operation monitoring performance degradation data into three-dimensional time series data using a detection data sliding window reconstruction method; a graphical expression module, used to graphically express the performance monitoring sensor as a graph neural network adjacency matrix according to the relationship between the sensor physical layout and the equipment structure; a construction and training module, used to construct a STAGCN model based on spatiotemporal attention graph convolution, and input the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; an evaluation module, used to perform performance status evaluation on the equipment to be evaluated based on the trained STAGCN model.

[0025] An embodiment of the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the performance status evaluation method based on the spatiotemporal attention mechanism are implemented.

[0026] The beneficial effect of the present invention is that, by designing the time and space attention machines, the performance status evaluation model actively focuses on the input information with higher information density during the training iteration process, and connects the graph convolutional neural network through the time and space attention matrix to mine the performance degradation characteristics in the noticed degradation information, thereby improving the performance degradation trajectory fitting effect and evaluation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of a performance status evaluation method based on a spatiotemporal attention mechanism provided by the present invention;

[0028] Figure 2 It is a flow chart of a performance status evaluation method based on spatiotemporal attention graph convolution provided by an embodiment of the present invention;

[0029] Figure 3 Schematic diagram of a spatiotemporal attention graph convolution model provided by an embodiment of the present invention;

[0030] Figure 4 It is a flow chart of a method for predicting continuous learning performance degradation based on elastic weight consolidation provided by an embodiment of the present invention;

[0031] Figure 5 It is a flow chart of elastic weight consolidation continuous learning for performance degradation prediction tasks provided by an embodiment of the present invention;

[0032] Figure 6 is a schematic diagram of a task accuracy matrix provided by an embodiment of the present invention;

[0033] Figure 7 It is a flow chart of the DP-GRUC performance degradation prediction model structure and training method provided by an embodiment of the present invention;

[0034] Figure 8 is a schematic diagram of elastic weight consolidation training provided by an embodiment of the present invention;

[0035] FIG. 9 is a schematic diagram of the Laplace approximation probability density provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In the subsequent description, the use of suffixes such as "module", "component" or "unit" used to represent elements is only to facilitate the description of the present invention and has no specific meaning in itself. Therefore, "module", "component" or "unit" can be used in a mixed manner.

[0037] The embodiment of the present invention combines the attention mechanism with the graph convolutional neural network, takes multi-channel and long-period performance status monitoring as the background, starts from the sensor channel dimension and the time series position dimension, combines the advantages of the attention mechanism model in association relationship mining and key focus, locates the position with high correlation with device performance degradation from high-latitude and long-sequence input data, and obtains high-density degradation information distribution. At the same time, combined with the advantages of deep graph convolutional neural networks in feature mining, the device performance degradation state is mined and quantitatively characterized from the located high degradation information density channels, providing a basis for device performance degradation prediction.

[0038] Figure 1 is a flow chart of a performance status evaluation method based on a spatiotemporal attention mechanism provided by the present invention. Figure 1As shown, it includes: step S1: using the detection data sliding window reconstruction method to convert the equipment operation monitoring performance degradation data into three-dimensional time series data; step S2: according to the relationship between the physical layout of the sensor and the equipment structure, the performance monitoring sensor is graphically expressed as a graph neural network adjacency matrix; step S3: constructing a STAGCN model based on spatiotemporal attention graph convolution, and inputting the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; step S4: based on the trained STAGCN model, the performance status of the equipment to be evaluated is evaluated.

[0039] In an embodiment of the present invention, the STAGCN model includes a spatiotemporal attention model and an attention graph convolution model; the three-dimensional time series data includes monitoring dimension time series data, time dimension time series data and operation dimension time series data.

[0040] In an embodiment of the present invention, the three-dimensional time series data and the graph neural network adjacency matrix are input into the STAGCN model for training to obtain a trained STAGCN model, including: inputting the three-dimensional time series data into the time attention machine in the spatiotemporal attention model to calculate the time attention matrix; inputting the time attention matrix and the three-dimensional time series data into the space attention machine in the spatiotemporal attention model to calculate the space attention matrix; multiplying the space attention matrix with the graph neural network adjacency matrix to construct an attention graph convolution model, and inputting the three-dimensional time series data into the attention graph convolution model to obtain an evaluation health index of the device; calculating the loss value of the evaluation health index and the evaluation label, and based on the loss value, obtaining a trained STAGCN model.

[0041] In an embodiment of the present invention, the temporal attention matrix includes:

[0042]

[0043] Where E is the temporal attention matrix, is the first r The input of the spatiotemporal network block, , , , , These are the parameters of the temporal attention matrix to be learned. is the activation function.

[0044] In an embodiment of the present invention, the spatial attention matrix includes:

[0045]

[0046] in, S is the spatial attention matrix, is the first r The input of the spatiotemporal network block, , , , , These are the parameters of the spatial attention matrix to be learned. is the activation function.

[0047] In an embodiment of the present invention, according to the relationship between the physical layout of the sensor and the device structure, the performance monitoring sensor is graphically expressed as a graph neural network adjacency matrix including:

[0048]

[0049] in, Indicates The monitoring parameters and The edge weights between monitoring parameters are is the correlation coefficient of sensors at the same point, and its range is , m and n They represent different collection points respectively.

[0050] In an embodiment of the present invention, the sensors are sensors of various types, and one or more sensors of different types are installed at various positions of the device structure;

[0051] The number of sensor acquisition channels is:

[0052]

[0053] in, Indicates the number of acquisition channels; Indicates m The number of sensors at each collection point; Indicates the number of sensor locations.

[0054] In an embodiment of the present invention, a detection data sliding window reconstruction method is used to convert equipment operation monitoring performance degradation data into three-dimensional time series data, including: using a sliding time window with a fixed length and a fixed step size to divide the data in each start-stop working cycle in the entire life cycle from operation to failure of the training into small multidimensional time series, and stacking the multidimensional time series into three-dimensional time series data.

[0055] An embodiment of the present invention also provides a performance status evaluation device based on a spatiotemporal attention mechanism, including: a data preprocessing module, used to convert equipment operation monitoring performance degradation data into three-dimensional time series data using a detection data sliding window reconstruction method; a graphical expression module, used to graphically express the performance monitoring sensor as a graph neural network adjacency matrix according to the relationship between the sensor physical layout and the equipment structure; a construction and training module, used to construct a STAGCN model based on spatiotemporal attention graph convolution, and input the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; an evaluation module, used to perform performance status evaluation on the equipment to be evaluated based on the trained STAGCN model.

[0056] An embodiment of the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the performance status evaluation method based on the spatiotemporal attention mechanism are implemented.

[0057] The present invention uses complex equipment system status monitoring data as the information source and quantitatively characterizes its HI (Health Index) through a performance status evaluation method. In this context, a complex equipment system status monitoring network is constructed. G In total N Monitoring points, the monitoring point records are ; Record at each work start-stop cycle monitoring point of the equipment L measurements, .by Indicated in t Moment, n Among the status monitoring points, l Condition monitoring measurement value. Indicated in t Moment, n All status monitoring measurements of nodes; express t All status monitoring measurements at all times; express All state monitoring measurement values ​​in the time series. Based on the above assumptions, the objective function of equipment state monitoring is ,in, HI=(h 1 ,h 1 ,… h t ,…) by h t Characterization tQuantitative value of the device performance status at a moment.

[0058] In view of the above problem description and model fitting objectives, combined with the engineering characteristics of equipment operation and data collection in the performance status evaluation process, this paper proposes a performance status evaluation method based on Spatial-TemporalAttention based on Graph Convolutional Networks (STAGCN). The method architecture is as follows: Figure 2 As shown in the figure. First, in the data preprocessing process, a three-dimensional sliding window reconstruction method for monitoring data is designed to convert the equipment operation monitoring data into training and test data for the performance status evaluation model in view of the differences in the duration of a single start-stop cycle of the equipment. Secondly, a graphical expression method for performance monitoring sensors is designed based on the relationship between the physical layout of equipment sensors and the equipment structure, and the equipment physical structure design information is input into the STAGCN model through the graph neural network adjacency matrix. On this basis, the STAGCN model is trained based on the input data and the adjacency matrix to output the performance status evaluation results.

[0059] (1) Data preprocessing

[0060] In the process of engineering application, due to the differences in frequency, physical dimensions, sampling quality, etc. between different sensors, equipment operation monitoring data cannot usually be directly used for data-driven performance status evaluation. Therefore, it is necessary to preprocess the data first, and the processed data is used as the input of the performance status evaluation model.

[0061] Since most equipment has different start-up and shutdown time, t Monitoring data collected at all times X t The corresponding length L In order to solve this problem, the present invention designs a sliding window reconstruction method for monitoring data based on the two conventional preprocessing methods of resampling and normalization and the characteristics of equipment performance status evaluation. The method flow is as follows: Figure 2 As shown, using a fixed length of Win , a sliding time window with a step size of 1, splits the data of each startup and shutdown cycle in the training life cycle from operation to failure into small multidimensional time series, and stacks them into three-dimensional time series data to ensure that the data contained in each sliding window is dimensionally regular input data, that is, X t Convert to . In addition, the corresponding health status of each sliding window data is used as the evaluation label. The original working cycle data of different lengths are converted into a regular three-dimensional data structure through the data sliding window reconstruction method, providing standardized data input for the performance status evaluation model.

[0062] (2) Graphical expression of performance monitoring sensors

[0063] The operation monitoring sensors of complex system equipment are installed in different positions of the equipment structure. Some positions may be equipped with one or more sensors of different types. G In the equation, the number of sensor locations is M, then:

[0064] (0.1)

[0065] in, N m is the number of sensors at the mth collection point, N is the number of acquisition channels. The sensors are numbered according to the physical layout. The data collected at the same sampling point have a certain correlation. Therefore, the performance monitoring sensor is graphically expressed as an adjacency matrix as shown in the formula:

[0066] (0.2)

[0067] in, Indicates The monitoring parameters and The edge weights between monitoring parameters are is the correlation coefficient of sensors at the same point, and its range is , m and n They represent different collection points respectively.

[0068] (3) Construction and training of spatiotemporal attention graph convolution model

[0069] like Figure 2 As shown in the figure, the spatiotemporal attention graph convolution model is the core of the equipment performance status evaluation method proposed in the present invention. It identifies the time dimension and space dimension performance degradation information in the equipment status monitoring data step by step through the stacked model structure, fits the equipment performance status degradation characteristics through the graph neural network, and fits the degradation characteristics into a one-dimensional vector based on the feature decoder, finally realizing the conversion of a large amount of performance status information covered in the monitoring data into a quantitative health indicator HI.

[0070] The specific structure of the STAGCN model is as follows Figure 3As shown in the figure, it consists of two parts: the spatiotemporal attention model and the attention graph convolution model. The spatiotemporal attention model takes the multi-dimensional monitoring parameters of the equipment as input, outputs the time attention matrix through the time attention machine, and passes the degradation feature density information of the time dimension to the spatial attention machine. The high-dimensional input data is transformed into a multi-dimensional matrix through the spatial attention machine, and the spatial attention matrix is ​​output to characterize the difference in degradation information density between different monitoring parameters. The attention graph convolution model converts the physical position of the sensor of the device to be evaluated into a graph convolution network adjacency matrix through graphical expression, introduces the layout information of the device status monitoring sensor through the adjacency matrix in the graph neural network, and combines the adjacency matrix with the spatial attention matrix to mine the degradation characteristics of the device performance status through graph convolution coding. Finally, based on the performance status degradation characteristics, the degradation characteristics are fitted into the health index HI through the linear decoder. In order to achieve quantitative characterization of the performance status of the equipment.

[0071] based on Figure 3 The model structure in the middle builds the equipment performance degradation assessment model, and uses the training data to train the model. The process is shown in the algorithm in Table 1.

[0072] Table 1 Algorithm performance status evaluation method training framework based on STAGCN model

[0073]

[0074] In the actual context of increasing equipment complexity and increasing monitoring signal parameters, this paper designs a performance status assessment method based on a spatiotemporal attention graph convolution model to solve the needs of accurate identification and positioning of degradation information and efficient characterization, inspired by the human visual attention mechanism. This method identifies the performance degradation information of the time and space dimensions in the equipment status monitoring data step by step through model stacking, and fits the equipment performance status degradation characteristics through graph neural networks, converting a large amount of performance status information covered in the monitoring data into quantitative health indicators, thus achieving efficient and accurate performance status assessment.

[0075] Spatiotemporal attention model based on channel attention mechanism

[0076] In order to identify the spatial and temporal correlation of complex equipment system status monitoring signals, the present invention proposes a spatiotemporal attention model based on the channel attention mechanism on the basis of the performance status evaluation deep learning model, targeting the multi-dimensional and time-series related characteristics of input data in the equipment performance status evaluation process. This module contains two attention attributes, namely, spatial attention attribute and temporal attention attribute.

[0077] In the time dimension, there is a certain correlation between the performance status of the device in different time periods, and the correlation of the device in different stages of operation is different. Therefore, the present invention also introduces an attention mechanism in the time dimension to adaptively assign different importance to data at different time locations.

[0078] (0.3)

[0079] (0.4)

[0080] In the formula, For the r The input of the spatiotemporal network block, For the r The number of channels of the layer input data, when r =1, , , , , These are the parameters of the attention matrix to be learned. Temporal attention matrix E From different time positions t The input data determines E Medium Element Indicates time i With time j Finally, the time matrix is ​​normalized by the softmax function E , get the normalized time matrix ,Will In the input data, the time dimension is dynamically adjusted by merging relevant information:

[0081] (0.9)

[0082] in is the input matrix after time dynamics adjustment, which will be used as the input of the spatial attention machine.

[0083] In the spatial dimension, due to the wide distribution and variety of complex equipment status monitoring nodes, there are certain correlations and differences between different status monitoring nodes. In order to effectively identify the differences in data from different status monitoring sensors and highlight important degradation information, thereby improving the accuracy and generalization ability of the performance status evaluation and prediction network, the present invention uses the attention mechanism to adaptively capture the correlation between nodes in the spatial dimension of the graph neural network. Taking a certain moment as an example, the spatial attention matrix is ​​defined as:

[0084] (0.5)

[0085] (0.6)

[0086] in For the r The input of the spatiotemporal network block, For the r The number of channels of the layer input data. It is r The length of the time dimension in the layer, when r =1, . , , , These are the parameters of the attention matrix to be learned. is the activation function. The spatial attention matrix is ​​dynamically calculated based on the current input of the network layer. S . S Medium Element Representation Node i With Node j The correlation strength between them. Use the softmax function to transform the spatial attention matrix S Each element in the graph ensures that the sum of the attention weights of the nodes is 1. During the graph convolution, the spatial attention matrix Adjust the adjacency matrix A The weights of nodes are calculated to achieve dynamic adjustment of relevant weights between nodes.

[0087] Performance status evaluation model based on attention graph convolution

[0088] In the present invention, complex system status monitoring is regarded as a graph structure, and the characteristics of each equipment performance status monitoring node can be regarded as a signal in the graph. In order to make full use of the topological characteristics of complex system status monitoring, in each working period of equipment startup and operation, the graph convolutional network is used as a signal processing module to mine the information correlation of complex system equipment status monitoring data in the spatial dimension. The graph convolutional neural network converts the graph signal into an algebraic form to analyze the topological properties of the performance status monitoring data graph.

[0089] by t Taking the state monitoring data at the moment as an example, the monitoring signal in the whole figure is recorded as , the graph transformation of the monitoring signal is defined as According to the properties of the Laplace matrix, U is an orthogonal matrix. Therefore, the corresponding inverse Fourier transform is Graph convolution is a convolution operation that replaces the classical convolution operator with a diagonal linear operator in the Fourier domain. Based on this property, graph G Signals in the network x The kernel filtering process is as follows:

[0090] (0.7)

[0091] in, is the graph convolution operation. The convolution operation based on the graph signal is equal to the product of the frequency domain signal after the graph Fourier transform, so the above formula can be transformed into and x are transformed into the frequency domain respectively, and then the transformed results are multiplied, and then the inverse Fourier transform is performed to obtain the final result of the convolution operation.

[0092] However, when there are many nodes to monitor the signal and the scale of the graph is large, the computational cost of directly decomposing the Laplace matrix is ​​high. Therefore, the Chebyshev polynomial is used to approximate the above equation, and the process is as follows:

[0093] (0.8)

[0094] Parameters in the formula is the polynomial coefficient vector. , is the maximum eigenvalue of the Laplace matrix. Recursively defined according to Chebyshev polynomials ,in, , The formula is solved by using the approximate expansion of Chebyshev polynomials, that is, by the convolution kernel Extract the image K The node is the center, and the surrounding area is 0 to K-1 The information of the order neighborhood.

[0095] In order to dynamically adjust the spatial correlation between state monitoring nodes, the spatial attention matrix Adjustment The weight distribution of each node in . Calculation ,in is the Hadamard product. Therefore, the above formula is converted into:

[0096] (0.9)

[0097] Based on the above formula, this definition can be extended to multi-channel graph signals. For example, taking the current time point state evaluation graph neural network as an example, the input is , for each time node t The corresponding input data fragment X t Perform graph convolution filtering transformation and calculate ,in, is the convolution kernel parameter. Therefore, in the attention graph convolution layer, each node is composed of 0~K-1 The neighboring node information is updated.

[0098] After the graph convolution operation identifies the relevant information between each node in the graph neural network from the spatial dimension, the STAGCN model superimposes a time-standard convolution layer in the time dimension to update the node by merging the information in adjacent time segments. r Taking layer calculation as an example, the convolution calculation is as follows:

[0099] (1.0)

[0100] in, represents the standard convolution operation, is the parameter of the time-dimensional convolution kernel, and ReLU is the linear rectification activation function (Rectified Linear Unit, ReLU).

[0101] The STAGCN model, which consists of a spatiotemporal attention model and an attention graph convolution model, mines the performance state degradation features in the monitoring data. Multiple spatiotemporal blocks are superimposed to further extract dynamic spatiotemporal correlations in a larger range. Finally, the degradation features are decoded by the decoder and a fully connected layer is superimposed to ensure that the output between the models has the same dimension and shape as the evaluation target label, and the output HI is output.

[0102] The embodiment of the present invention further includes, after the performance status evaluation method based on the spatiotemporal attention mechanism: the implementation steps of the continuous learning performance degradation prediction method based on elastic weight consolidation, which specifically include: according to the device operation task scenario, real-time acquisition of the device performance status data set that is dynamically updated in different operation task scenarios; constructing a performance degradation prediction DP-GRUC model based on a convolutional gated network, and training the DP-GRUC model through the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance status data set to obtain a trained DP-GRUC model; based on the trained DP-GRUC model, predicting the performance degradation of the device under test in the target task scenario.

[0103] In the embodiment of the present invention, the DP-GRUC model is composed of a history encoder module, a future encoder module and a HI trajectory decoder module.

[0104] In an embodiment of the present invention, the equipment operation task scenario includes any of the following: newly added monitoring data, newly added operating conditions, newly added equipment individuals and newly added equipment models.

[0105] In an embodiment of the present invention, the DP-GRUC model is trained by the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the equipment performance status data set to obtain the trained DP-GRUC model, including: initializing the DP-GRUC model by randomly initializing the model weights; inputting the equipment performance status data set under any task scenario of the equipment into the initialized DP-GRUC model to fit the historical health indicator HI trajectory and the future health indicator HI trajectory, and obtaining and saving the DP-GRUC model after the first training under the task scenario.

[0106] In an embodiment of the present invention, inputting a data set of equipment performance status in any task scenario of the equipment into an initialized DP-GRUC model to fit a historical health indicator HI trajectory gradient and a future health indicator HI trajectory gradient includes: dividing the data set of equipment performance status in any task scenario of the equipment into a plurality of time window HI trajectories by a sliding window cutting method, and setting the first half of the data in the sliding window as the historical HI trajectory, and setting the second half as the corresponding future HI trajectory; calculating the historical HI trajectory gradient and the future HI trajectory gradient based on the historical HI trajectory and the future HI trajectory respectively; based on the historical HI trajectory gradient and the future HI trajectory gradient, inputting the historical HI trajectory gradient into a historical encoder for historical coding update, and outputting the historical HI trajectory coding; at the same time, inputting the future HI trajectory gradient into a future encoder for future coding update, and outputting the future HI trajectory coding; combining the output historical HI trajectory coding and the future HI trajectory coding into a degradation feature association key value.

[0107] In the embodiment of the present invention, respectively calculating the historical HI trajectory gradient and the future HI trajectory gradient based on the historical HI trajectory and the future HI trajectory includes:

[0108] (1.1)

[0109] (1.2)

[0110] in, represents the historical HI trajectory gradient; points represent the future HI trajectory gradient; Indicates the device performance status of device i at time t; Indicates the device performance status of device i at time t-1; Indicates the current moment.

[0111] In an embodiment of the present invention, the DP-GRUC model is trained by the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the equipment performance status data set to obtain the trained DP-GRUC model, which also includes: when any task scenario triggers the update of the DP-GRUC model, the equipment performance status data set of the newly added task scenario of the task scenario and the DP-GRUC model after the training of the previous task are obtained; by inputting the equipment performance status data set of the newly added task scenario of the task scenario into the DP-GRUC model after the training of the previous task, and calculating the training loss function of the elastic weight consolidation and the current task based on the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation; the DP-GRUC model after the training of the previous task is trained by minimizing the training loss function by the gradient descent method to obtain the DP-GRUC model after the training of the current task.

[0112] In an embodiment of the present invention, the training loss function is minimized as follows:

[0113] (1.3)

[0114] In the formula, represents the training loss function; represents the training loss function of task B, represents the elastic modulus; represents the model parameters, represents the optimal model parameters of model A, and W represents the parameter weight; Represents a model function.

[0115] In order to adapt to more diverse equipment performance degradation tasks, learn and memorize more differential degradation features, and improve the robustness of the prediction method, the prediction model of the embodiment of the present invention needs to have the ability to learn the newly added differential degradation features. This feature is the plasticity of the prediction model. The joint learning of common degradation and differential degradation features puts forward the requirement of "stability-plasticity balance" for the performance degradation prediction model. In this scenario, the equipment performance degradation prediction task is mathematically expressed as follows: During the full life operation of the equipment, assuming that the amount of running equipment data at the current moment is N, and the current moment is The health degradation process of the device numbered i is expressed as ,in The goal of performance degradation prediction is to build a data-driven prediction model ,pass Predicting the health of your equipment at future times ,in . Train the prediction model using historical data , to achieve the maximum posterior probability:

[0116] (1.4)

[0117] in, Represents the prediction model weight. The existing performance degradation prediction methods are mainly one-time learning and testing tasks for a single data set. During the model learning process, the training and test data sets are fixed. In the task scenario where continuous learning is the goal, the training data and performance degradation prediction tasks are continuously updated over time. Therefore, the training and test data are a series of data sets consisting of multiple data sets, which are defined as , where D is the dataset set, M is the number of tasks, , N is the number of equipment individuals in the data set D.

[0118] Aiming at the actual needs in industrial scenarios and the continuous learning goal of equipment performance degradation prediction tasks, this paper proposes a continuous learning of state prediction based on elastic weight consolidation (CLSP-EWC) method inspired by the synaptic plasticity mechanism. The method architecture is as follows: Figure 4 As shown in the figure. The method consists of a performance degradation prediction model based on a convolutional gating network and a degradation feature memory model based on elastic weight consolidation. The performance degradation prediction model fits the mapping relationship between historical and future HIs through a two-stage encoder structure. During the training and updating process, the degradation feature memory model realizes the memory and consolidation of historical task information by constraining the model learning direction, achieving a balance between the model's historical task stability and new task plasticity.

[0119] like Figure 4 As shown in the figure, the performance degradation prediction model in the CLSP-EWC method is composed of two encoders, one for history and one for the future, and one feature decoder. The encoder takes the historical and future HI data of the device as input, mines the degradation features through neural network learning, and combines them to generate the key values ​​of the degradation features. The encoder takes the key values ​​of the degradation features as input, and decodes to fit the mapping relationship between the historical and future degradation features. During the retraining process of the task update model, the degradation feature memory model in the CLSP-EWC method flexibly limits the iteration intensity of each neuron by identifying the contribution of each neuron in the model to the historical tasks. The gradient fitting direction of the model is converged in the superposition space of the historical tasks and the newly added tasks. Through the combination of the performance degradation prediction model and the degradation feature memory model, continuous learning performance degradation prediction in the scenarios of newly added tasks for different data sets is realized.

[0120] Continuous learning performance degradation prediction process based on elastic weight consolidation

[0121] Based on the CLSP-EWC method designed by the present invention, combined with the characteristics of equipment performance degradation, such as Figure 5 As shown, the present invention specifically designs an elastic weight consolidation continuous learning process for performance degradation prediction tasks. A performance degradation prediction task trigger mechanism is designed for four scenarios: new monitoring data, new operating conditions, new equipment individuals, and new equipment models in actual industrial equipment. At the same time, in view of the evolution characteristics of equipment performance degradation, a performance degradation prediction model based on a convolutional gated network (Degradation Prediction based on GRU withConvolutional, DP-GRUC) is specifically designed. Based on the DP-GRUC model structure, a dynamic update method for the elastic weight consolidation degradation feature memory model is designed. Finally, based on the general evaluation index of continuous learning for classification tasks, the present invention combines the characteristics of performance degradation trend prediction tasks to specifically design a continuous learning performance degradation prediction evaluation index to evaluate the ability of the CLSP-EWC method designed by the present invention to overcome catastrophic forgetting, that is, the continuous learning level.

[0122] Task scenario attribute division

[0123] In the process of the equipment being put into actual use, performance degradation prediction is usually a non-real-time task in the health management link. That is, it is usually triggered under fixed requirements such as abnormal monitoring data and regular maintenance. Therefore, the present invention is oriented to the equipment operation scenarios in actual industrial applications, and designs a corresponding prediction task update trigger mechanism. At the same time, differentiated elasticity coefficients are set according to the differences in performance degradation characteristics in different scenarios to adapt to the task differences between different task updates.

[0124] ①Time series expansion task

[0125] In the industrial field, the equipment is usually used for a long time. Most industrial equipment has been in use for several years or decades. During the operation of the equipment, the performance status HI data continues to grow. During the use of the equipment after it is put into operation, due to component wear and other conditions, a certain subsystem of the equipment has a specific failure mode, and the health status of the equipment may transition to a certain type of abnormal state. Therefore, the present invention is designed to set a quantitative performance degradation prediction threshold based on the performance status HI data. When the equipment HI is lower than a certain threshold, a prediction model update is triggered.

[0126] ②Expanding the capacity of working conditions

[0127] During the actual operation of the equipment, the operating conditions of the equipment will change due to the influence of the environment and usage scenarios. Since different operating conditions have very different damage to the equipment subsystems, the operating conditions will greatly affect the performance degradation characteristics of the equipment. Based on this scenario, the equipment operating conditions are identified and judged by monitoring the equipment operating environment and operating condition parameters. When the equipment enters a new operating condition, a prediction model update is triggered.

[0128] ③Unit quantity expansion task

[0129] Due to individual differences such as factory manufacturing errors, the equipment is affected by differential degradation factors during the performance degradation process, and different equipment units have different individual performance degradation characteristics. Therefore, in the group equipment operation scenario, when a new unit appears in the group and is put into operation and generates certain performance status HI data, a prediction model update is triggered.

[0130] ④Degraded feature expansion task

[0131] In the industrial field, the same type of equipment is usually designed, developed and produced in a certain cycle. In most industries, it is common for new and old models of equipment to be in service at the same time. Due to differences in the structure, formula and other design of the equipment, the performance degradation characteristics of the equipment will also be different. Therefore, when a new model of equipment is put into operation in a group equipment service scenario, after the new model of equipment is running and generates certain performance status HI data, a prediction model update is triggered.

[0132] Prediction model construction and degradation feature memory

[0133] According to the actual equipment operation scenario conditions, the performance degradation prediction task update trigger mechanism mentioned above is used to build the equipment performance degradation prediction model based on the DP-GRUC model structure, and the elastic weight consolidation feature memory method for the performance degradation prediction task is used to train and update the model parameters. The specific training process is shown in the algorithm in Table 2. The CLSP-EWC method is used to iterate the DP-GRUC model in the time series dimension, update the model weights, and output the corresponding prediction model P at different times T. T .

[0134] Table 2 Algorithm CLSP-EWC training framework based on DP-GRUC model

[0135]

[0136] Continuous learning evaluation indicators for performance degradation prediction tasks

[0137] Since the widespread and successful application of continuous learning methods is mainly concentrated in classification tasks, in order to evaluate the ability of the CLSP-EWC method proposed in this invention to overcome catastrophic forgetting of tasks, this invention refers to the evaluation indicators that are widely used in the field of continuous learning, and improves them according to the characteristics of performance degradation prediction tasks, and designs evaluation indicators for continuous learning methods for performance degradation prediction tasks. In the process of continuous learning evaluation for classification tasks, the indicator calculation is mainly based on the task accuracy matrix, whose structure is as follows: Figure 6 As shown in the figure. According to the order in which the tasks appear, the accuracy of the model for each task after each task training is arranged in order to form the task accuracy matrix. i,j Indicates i After the task data is trained, the model j The accuracy in each task.

[0138] Performance degradation prediction model based on convolutional gating network

[0139] Equipment performance degradation is a continuous dynamic time series process, which starts from the time of commissioning until the failure occurs and reaches the end of life. The HI trajectory during this period represents the entire process of equipment performance change, and the length of the HI trajectory represents the length of the equipment's operating life. During the entire life cycle of equipment operation, the performance degradation process has the following two characteristics:

[0140] Continuity of degradation process: The performance change caused by a certain work start-stop cycle of the equipment is based on the performance state after the previous work cycle. That is, the performance degradation process is a process of sequential evolution. The change of equipment performance state is not only related to the equipment operating conditions and individual equipment conditions, but also to the operating cycle time of the equipment.

[0141] Differences in degradation characteristics: For the same model and batch of equipment, the degradation characteristics of the equipment have certain differences due to differences in initial manufacturing and assembly errors and differences in operating environment, working conditions and other conditions. That is, in the same equipment performance degradation feature space, different individuals still have personalized differences. Therefore, equipment performance degradation prediction needs to mine common degradation characteristics based on historical data, and output prediction results in combination with personalized degradation characteristics of the equipment to be predicted.

[0142] Predicting equipment performance degradation can determine the performance status of the equipment in the future working process in advance, providing strong support for operation decision-making. Based on dynamically updated equipment performance status data set , Aiming at the two characteristics of the equipment performance degradation evolution process, the DP-GRUC model consists of a history encoder module, a future encoder module and a HI trajectory decoder module. The model architecture is as follows Figure 7 shown.

[0143] Based on the continuity of the equipment degradation process, the present invention designs an HI trajectory encoder based on convolutional neural network and gated neural network, and extracts equipment degradation features by fitting the HI trajectory gradient. and Data, define the HI trajectory gradient as follows:

[0144]

[0145] and The HI trajectory gradients of equipment performance degradation in the past and future are characterized respectively, which characterize the relative performance degradation change of the equipment before and after a start-stop cycle. and , the calculation process of historical degradation feature coding and future degradation feature coding is as follows:

[0146]

[0147] In the formula, and Represents the historical and future degradation feature encoders respectively and Learnable parameters of the GRU layer and the Conv1d layer of the 1D convolutional network.

[0148] Based on the encoding results, the historical and future degradation characteristics at a certain moment are is defined as a pair of degradation feature associated key values. With historical degradation gradient is fed into the encoder to generate future HI trajectories The performance degradation prediction model is trained through the historical data of different equipment operations, the key-value mapping relationship of degradation features is learned based on the encoder model, and the future degradation gradient is generated based on the decoder's comprehensive HI trajectory information of the device to be predicted. The performance state value of the device at the future moment is deduced from the current performance state to achieve device performance degradation prediction.

[0149] Degraded feature memory model based on elastic weight consolidation

[0150] The DP-GRUC model can be used to predict device performance degradation. However, the fitting results of the model in a single training process converge to the training data set. The degradation characteristics contained in the data set are updated with the continuous operation of the equipment and the update of industrial needs. With the increase of , there is a task conflict between the consolidation of historical capabilities learned in the model and the update of the newly added data set. This problem is called the "stability-plasticity" balance in biological neuroscience. In response to repeated external stimulation, the connection morphology between neuronal synapses will change, and the reconnection process aims at the optimal learning cost. The task architecture of continuous learning is that after a fixed number of trainings on each task, no further training is allowed on the data set of the task. Therefore, to achieve continuous learning, it is necessary to limit the direction of the iterative gradient descent of the model weights to ensure the model's ability to complete the previous task A. Based on this goal, inspired by the biological synaptic plasticity mechanism, the present invention constructs an elastic weight consolidation feature memory model for the performance degradation prediction task. The neurons of the DP-GRUC artificial neural network influence each other through the connection weights. The core of the elastic weight consolidation is to regard the weight protection applied to the neuronal synaptic connection as a spring link, and its stiffness is proportional to the importance of the neuronal connection of the previous task. The present invention uses the complex equipment performance degradation prediction task based on the DP-GRUC model as a continuous learning task to carry out the design and research of the elastic weight consolidation method.

[0151] Design of elastic weight consolidation method for DP-GRUC model

[0152] The DP-GRUC model neural network consists of multiple layers of linear mapping layers and multiple types of nonlinear mapping layers. Fitting learning is a process of continuously adjusting the model weight set based on the linear mapping deviation to optimize the model prediction performance. θ represents the artificial neural network layer parameters, Indicates the optimal parameter configuration for the task. θ During the configuration process, the same performance effect can match multiple configuration combinations. This theory is the basis for the consolidation of elastic weights. The present invention develops a method design based on this. Due to over-parameterization, the solution configuration of the subsequent task B There is a high probability that the solution configuration is the same as the previous task A Therefore, if Figure 8 As shown in Figure 2, when learning the subsequent task B, EWC constrains the parameters to The performance of task A is protected by keeping the task in the low error region of task A centered on . This constraint is implemented by the quadratic penalty method, so the reset of neuron weights can be defined as the process of a rigid spring anchoring the weight parameters to the configuration region of the previous task. The spring stiffness varies between parameters. For parameters that are important to the performance in the previous task A, the stiffness is greater, and vice versa.

[0153] Different retraining model parameter change paths are as follows Figure 8 As shown, after the model learns the previous task A, the parameters are located in the space , if the model training gradient is adjusted only according to the effect of task B, the parameter path will be as shown by the red line. The model maximizes the learning ability of task B, but destroys the knowledge learned in task A, causing knowledge forgetting. If the model iteration process is constrained by task A, as shown by the green line in the figure, due to the overly strict restrictions, the model's ability to learn task B is seriously insufficient. Therefore, by calculating the importance of different model weights to task A, determining the direction of the parameter gradient change path, and finding the gradient path for task B in the low-loss space of parameter A, it is possible to achieve effective learning of task B while retaining the ability of task A.

[0154] In order to demonstrate the rationality of the retraining constraints of the deep neural network model and determine the importance of different neuron weights to the task, the model training process is optimized by probability, that is, for a given data set Determine the optimal importance under the condition of Calculating conditional probabilities , based on Bayes' rule

[0155] (1.0)

[0156] Since the objective function to be maximized is the same as the logarithm, the calculation The logarithmic process is as follows:

[0157] (1.1)

[0158] In the formula, Given a data set The probability logarithm of . During the network model training process, is the objective function to be optimized on the log-likelihood function:

[0159] (1.2)

[0160] Assume that the dataset corresponding to the task is divided into two independent parts, and task A and task B are defined as and , so rearrange the formula to:

[0161] (1.3)

[0162] In the formula, the left side of the equation still describes the given data set The parameter posterior probability of the equation only depends on the loss function of the subsequent task B. Therefore, the posterior distribution on the right side of the equation is It must contain the knowledge information of the previous task A. This knowledge information includes the importance of model parameters to task A, which is the key attribute to realize EWC.

[0163] Approximate solution of the posterior distribution of the previous task

[0164] In formula (1.3), is the posterior probability distribution of the previous task A, which contains the parameter information of the network model fitting task A. However, since it cannot be expressed in an interpretable display form, it is very difficult to calculate the posterior probability that fully conforms to the actual data. Therefore, consider calculating its approximate value through the Karplus method.

[0165] First, exist Secondary Taylor expansion

[0166] (1.4)

[0167] because is the peak value, the tangent slope , formula (2.7) can be written as:

[0168] (1.5)

[0169] Based on formula (1.2), formula (1.5) can be written as:

[0170] (1.6)

[0171] in, .Will Written as , formula (1.6) can be expressed as the standard expression of normal distribution:

[0172] (1.7)

[0173] Thus, the posterior probability distribution of the previous task A can be obtained as:

[0174] (1.8)

[0175] That is, Fig. 9 As shown, the normal distribution approximation of the continuous probability density distribution is found based on the Laplace approximation method. Assume is a smooth probability distribution with a peak at the maximum value Nearby, that is, the average With variance Normal distribution fitting of .

[0176] Fisher information matrix parameter importance calculation

[0177] like Fig. 9 As shown, the Laplace approximation method can be used to The approximate fit is a normal distribution, and the mean of the normal distribution is , the variance of the normal distribution is Since the Fisher information matrix can be calculated by the first-order derivative. This property can be used to avoid calculating the Hessian matrix in formula (1.5).

[0178] (1.9)

[0179] Substituting formula (1.5) into formula (1.3), we have: (2.0)

[0180] In the formula, represents all constant terms, It is a hyperparameter introduced to weigh the ability of the network model to learn task B and remember the previous task A.

[0181] Substituting formula (1.9) into the equation, we have:

[0182] (2.1)

[0183] Therefore, the network model elastic weight consolidation training can be expressed as formula (2.1). Where the Fisher information matrix It contains the importance information of model parameters to task A. There are three key properties:

[0184] ① Equivalent to the second-order derivative when the loss is close to the minimum value;

[0185] ② It can be calculated only from the first-order derivative, so the calculation process is relatively simple;

[0186] ③ is a positive semidefinite matrix.

[0187] The Laplace calculation is similar to the expectation propagation process, where each subtask is regarded as a posterior factor, so the loss function minimized by EWC is:

[0188] (2.2)

[0189] In the formula, is the training loss function for task B, The elasticity coefficient is configured with the relative importance of task B and the previous task A, and Relative importance marking is performed on each parameter, thereby realizing EWC model training.

[0190] When the model performs the third task C, EWC uses the sum of the two penalty terms of task A and task B as the secondary penalty term to bring the network parameters close to the common parameter area. Similarly, for future prediction tasks, the historical task penalty term is used to maintain the model's ability to remember historical tasks, thereby achieving continuous learning of the DP-GRUC model.

[0191] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the present invention is not limited thereby. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention shall be within the scope of the present invention.

Claims

1. A performance status evaluation method based on spatiotemporal attention mechanism, characterized in that: include: The equipment operation monitoring performance degradation data is converted into three-dimensional time series data using the detection data sliding window reconstruction method; According to the relationship between the physical layout of sensors and the structure of equipment, the performance monitoring sensors are graphically expressed as a graph neural network adjacency matrix; Constructing a STAGCN model based on spatiotemporal attention graph convolution, and inputting the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; Performing performance status evaluation on the device to be evaluated based on the trained STAGCN model; The STAGCN model includes a spatiotemporal attention model and an attention graph convolution model; the three-dimensional time series data includes monitoring dimension time series data, time dimension time series data and operation dimension time series data; The three-dimensional time series data and the graph neural network adjacency matrix are input into the STAGCN model for training, and the trained STAGCN model is obtained, including: Inputting the three-dimensional time series data into the temporal attention machine in the spatiotemporal attention model to calculate the temporal attention matrix; Input the temporal attention matrix and the three-dimensional time series data into the spatial attention machine in the spatiotemporal attention model to calculate the spatial attention matrix; Multiplying the spatial attention matrix with the graph neural network adjacency matrix to construct an attention graph convolution model, and inputting the three-dimensional time series data into the attention graph convolution model to obtain an evaluation health index of the device; Calculate the loss value of the evaluation health index and the evaluation label, and obtain the trained STAGCN model based on the loss value; Among them, the temporal attention matrix includes: ； Where E is the temporal attention matrix, is the first r The input of the spatiotemporal network block, , , , These are the parameters of the temporal attention matrix to be learned. is the activation function; Among them, the spatial attention matrix includes: ； in, S is the spatial attention matrix, is the first r The input of the spatiotemporal network block, , , , , These are the parameters of the spatial attention matrix to be learned. is the activation function.

2. The method according to claim 1, characterized in that According to the relationship between the physical layout of the sensor and the device structure, the performance monitoring sensor is graphically expressed as a graph neural network adjacency matrix including: ； in, Indicates The monitoring parameters and The edge weights between monitoring parameters are is the correlation coefficient of sensors at the same point, and its range is , m and n They represent different collection points respectively.

3. The method according to claim 2, characterized in that The sensors are of various types, and one or more sensors of different types are installed at various positions of the equipment structure; The number of sensor acquisition channels is: ； in, Indicates the number of acquisition channels; Indicates m The number of sensors at each collection point; Indicates the number of sensor locations.

4. The method according to claim 1, characterized in that: Using the detection data sliding window reconstruction method, the equipment operation monitoring performance degradation data is converted into three-dimensional time series data including: Using a sliding time window with a fixed length and a fixed step size, the data of each start-up and shutdown working cycle in the entire life cycle from operation to failure is divided into small multidimensional time series, and the multidimensional time series are stacked into three-dimensional time series data.

5. A performance status evaluation device based on spatiotemporal attention mechanism, characterized in that: include: A data preprocessing module is used to convert the equipment operation monitoring performance degradation data into three-dimensional time series data by using a detection data sliding window reconstruction method; A graphical expression module is used to graphically express the performance monitoring sensor as a graph neural network adjacency matrix according to the relationship between the sensor physical layout and the device structure; A construction and training module is used to construct a STAGCN model based on spatiotemporal attention graph convolution, and input the three-dimensional time series data and the graph neural network adjacency matrix into the STAGCN model for training to obtain a trained STAGCN model; An evaluation module, used for performing a performance status evaluation on the device to be evaluated based on the trained STAGCN model; The STAGCN model includes a spatiotemporal attention model and an attention graph convolution model; the three-dimensional time series data includes monitoring dimension time series data, time dimension time series data and operation dimension time series data; Among them, the construction and training module is specifically used to input the three-dimensional time series data into the time attention machine in the spatiotemporal attention model to calculate the time attention matrix; input the time attention matrix and the three-dimensional time series data into the space attention machine in the spatiotemporal attention model to calculate the space attention matrix; multiply the space attention matrix with the graph neural network adjacency matrix to construct an attention graph convolution model, and input the three-dimensional time series data into the attention graph convolution model to obtain the evaluation health index of the device; calculate the loss value of the evaluation health index and the evaluation label, and obtain the trained STAGCN model based on the loss value; Among them, the temporal attention matrix includes: ； Where E is the temporal attention matrix, is the first r The input of the spatiotemporal network block, , , , These are the parameters of the temporal attention matrix to be learned. is the activation function; Among them, the spatial attention matrix includes: ； in, S is the spatial attention matrix, is the first r The input of the spatiotemporal network block, , , , , These are the parameters of the spatial attention matrix to be learned. is the activation function.

6. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the performance status evaluation method based on the spatiotemporal attention mechanism described in any one of claims 1 to 4 are implemented.

Images