A method, apparatus, and electronic device for predicting continuous learning performance degradation based on stability-plasticity balance

Abstract

This invention discloses a method, apparatus, and electronic device for predicting performance degradation based on a stability-plasticity balance in continuous learning. The method includes: acquiring a dynamically updated device performance status dataset in real time under different operating task scenarios, based on the device's operating task scenarios; constructing a performance degradation prediction DP-GRUC model based on a convolutional gated network, and training the DP-GRUC model using the CLSP-EWC method for continuous learning performance degradation prediction based on elastic weight consolidation and the device performance status dataset to obtain a trained DP-GRUC model; and predicting the performance degradation of the device under test in the target task scenario based on the trained DP-GRUC model.

Description

Technical Field

[0001] This invention relates to the field of performance degradation prediction technology, and in particular to a method, apparatus and electronic device for predicting continuous learning performance degradation based on a stability-plasticity balance. Background Technology

[0002] The ongoing trend of industrial digitalization has increased the complexity of industrial equipment systems, leading to high complexity and low predictability of failures. Equipment performance degradation is influenced by three key factors: initial degradation, common degradation, and differential degradation.

[0003] Common degradation factors make the performance degradation process of equipment predictable. However, the initial degradation factors determine that different devices have different degradation starting points, and the differential degradation factors determine the differences in degradation characteristics between individual devices. Therefore, how can the equipment performance degradation prediction model consolidate common degradation characteristics in continuously added prediction tasks to achieve a balance between the stability of the model in historical tasks and the plasticity of new tasks? Summary of the Invention

[0004] This invention provides a method, apparatus, and electronic device for predicting continuous learning performance degradation based on a stability-plasticity balance, in order to solve the technical problem of how to consolidate common degradation features in continuously emerging prediction tasks.

[0005] This invention provides a continuous learning performance degradation prediction method based on a stability-plasticity balance, comprising: acquiring a dynamically updated device performance status dataset under different operating task scenarios in real time according to the device's operating task scenarios; constructing a performance degradation prediction DP-GRUC model based on a convolutional gated network, and training the DP-GRUC model using the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance status dataset to obtain a trained DP-GRUC model; and predicting the performance degradation of the device under test under the target task scenario based on the trained DP-GRUC model.

[0006] Preferably, the DP-GRUC model consists of a history encoder module, a future encoder module, and a HI trajectory decoder module.

[0007] Preferably, the equipment operation task scenario includes any of the following: new monitoring data, new operating conditions, new individual equipment, and new equipment model.

[0008] Preferably, the DP-GRUC model is trained using the CLSP-EWC method for continuous learning performance degradation prediction based on elastic weight consolidation and the device performance status dataset to obtain the trained DP-GRUC model, which includes: initializing the DP-GRUC model with randomly initialized model weights; inputting the device performance status dataset under any task scenario into the initialized DP-GRUC model to fit the historical health index HI trajectory and the future health index HI trajectory, thereby obtaining and saving the DP-GRUC model after the first training under that task scenario.

[0009] Preferably, the process of inputting the device performance status dataset under any task scenario into the initial DP-GRUC model to fit the historical health index HI trajectory gradient and the future health index HI trajectory gradient includes: dividing the device performance status dataset under any task scenario into several time windows HI trajectories using a sliding window method, and setting the first half of the data within the sliding window as the historical HI trajectory and the second half as its 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 the historical encoder for historical encoding update, and outputting the historical HI trajectory code; simultaneously inputting the future HI trajectory gradient into the future encoder for future encoding update, and outputting the future HI trajectory code; and combining the output historical HI trajectory code and the future HI trajectory code to form a degradation feature association key value.

[0010] Preferably, calculating the gradients of the historical HI trajectory and the future HI trajectory based on the historical HI trajectory and the future HI trajectory respectively includes:

[0011]

[0012] in, Represents the gradient of the historical HI trajectory; The fraction represents the gradient of the future HI trajectory; This represents the device performance state of device i at time t; This represents the device performance state of device i at time t-1; Indicates the current moment.

[0013] Preferably, training the DP-GRUC model using the CLSP-EWC method for continuous learning performance degradation prediction based on elastic weight consolidation and the device performance status dataset to obtain the trained DP-GRUC model further includes: when any task scenario triggers a DP-GRUC model update, obtaining the device performance status dataset of the newly added task scenario and the DP-GRUC model trained on the previous task; inputting the device performance status dataset of the newly added task scenario into the DP-GRUC model trained on the previous task, and calculating the training loss function of elastic weight consolidation and the current task based on the CLSP-EWC method for continuous learning performance degradation prediction based on elastic weight consolidation; training the DP-GRUC model trained on the previous task by minimizing the training loss function using gradient descent to obtain the DP-GRUC model trained on the current task.

[0014] Preferably, the training loss function is minimized as follows:

[0015] In the formula, Represents the training loss function; This represents the training loss function for task B. Represents the elastic coefficient; Indicates model parameters, This represents the optimal model parameters for model A. Indicates parameter weights; Represents the model function.

[0016] This invention also provides a continuous learning performance degradation prediction device based on stability-plasticity balance, comprising: an acquisition module for acquiring, in real time, a dynamically updated device performance status dataset under different operating task scenarios; a construction and training module for constructing a performance degradation prediction DP-GRUC model based on a convolutional gated network, and training the DP-GRUC model using the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance status dataset to obtain a trained DP-GRUC model; and a prediction module for predicting the performance degradation of the device under test in a target task scenario based on the trained DP-GRUC model.

[0017] This invention also provides an electronic device, including 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, it implements the steps of the continuous learning performance degradation prediction method based on stability-plasticity balance.

[0018] The beneficial effects of this invention are as follows: (1) To address the problem of performance degradation prediction for complex system equipment, a performance degradation prediction model based on convolutional gating network is proposed. The DP-GRUC model mines performance degradation features through convolutional layers and fits them as degradation feature association keys through a gating neural network. Based on the decoder, the mapping relationship between historical HI trajectory and future HI trajectory is fitted to achieve accurate prediction of the equipment's performance at future moments.

[0019] (2) To meet the continuous learning needs of equipment performance degradation prediction, a degradation feature memory model based on elastic weight consolidation was designed based on the DP-GRUC model. This model realizes the memory and consolidation of historical performance degradation features by constraining the learning direction of the neural network during the training process, while ensuring the learning ability of new prediction tasks. In the process of dynamic model update, the stability-plasticity balance in the performance degradation prediction task is achieved.

[0020] (3) Based on the proposed performance degradation prediction model based on convolutional gating network and the degradation feature memory model based on elastic weight consolidation, a set of elastic weight consolidation continuous learning process for performance degradation prediction task was designed. The process designed different task addition scenarios, and based on the general classification task continuous learning index, APRMSE and BTRMSE evaluation indexes were designed for performance degradation prediction task. Attached Figure Description

[0021] Figure 1 This is a flowchart of a continuous learning performance degradation prediction method based on stability-plasticity balance provided by the present invention; Figure 2 This is a flowchart of the continuous learning performance degradation prediction method based on elastic weight consolidation provided in the embodiments of the present invention; Figure 3 This is a flowchart of the elastic weight consolidation continuous learning process for performance degradation prediction tasks provided in this embodiment of the invention. Figure 4 This is a schematic diagram of the task accuracy matrix provided in an embodiment of the present invention; Figure 5 is a flowchart of the DP-GRUC performance degradation prediction model structure and training method provided in the embodiment of the present invention; Figure 6 This is a schematic diagram of elastic weight consolidation training provided in an embodiment of the present invention; Figure 7 is a schematic diagram of the Laplace approximate probability density provided in an embodiment of the present invention. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no inherent meaning. Therefore, "module," "part," or "unit" may be used interchangeably.

[0023] Figure 1 This is a flowchart of a continuous learning performance degradation prediction method based on stability-plasticity balance provided by the present invention, as shown below. Figure 1 As shown, the process includes: Step S1: Based on the device's operating task scenario, obtain the dynamically updated device performance status dataset under different operating task scenarios in real time; Step S2: Construct a performance degradation prediction DP-GRUC model based on a convolutional gated network, and train the DP-GRUC model using the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance status dataset to obtain the trained DP-GRUC model; Step S3: Based on the trained DP-GRUC model, predict the performance degradation of the device under test in the target task scenario.

[0024] In this embodiment of the invention, the DP-GRUC model consists of a historical encoder module, a future encoder module, and an HI trajectory decoder module.

[0025] In this embodiment of the invention, the device operation task scenario includes any of the following: adding monitoring data, adding operating conditions, adding individual devices, and adding device models.

[0026] In this embodiment of the invention, the DP-GRUC model is trained using the CLSP-EWC method for continuous learning performance degradation prediction based on elastic weight consolidation and the device performance status dataset to obtain the trained DP-GRUC model. This process includes: initializing the DP-GRUC model with randomly initialized model weights; inputting the device performance status dataset for any task scenario into the initialized DP-GRUC model to fit the historical health index HI trajectory and the future health index HI trajectory, thereby obtaining and saving the DP-GRUC model after the first training in that task scenario.

[0027] In this embodiment of the invention, inputting the device performance status dataset under any task scenario into the initialization DP-GRUC model to fit the historical health index HI trajectory gradient and the future health index HI trajectory gradient includes: dividing the device performance status dataset under any task scenario into several time windows HI trajectories using a sliding window segmentation method, and setting the first half of the data within the sliding window as the historical HI trajectory and the second half as its 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 the historical encoder for historical encoding update, and outputting the historical HI trajectory code; simultaneously inputting the future HI trajectory gradient into the future encoder for future encoding update, and outputting the future HI trajectory code; combining the output historical HI trajectory code and the future HI trajectory code to form a degradation feature association key value.

[0028] In this embodiment of the invention, calculating the gradient of the historical HI trajectory and the gradient of the future HI trajectory based on the historical HI trajectory and the future HI trajectory respectively includes:

[0029]

[0030] in, Represents the gradient of the historical HI trajectory; The fraction represents the gradient of the future HI trajectory; This represents the device performance state of device i at time t; This represents the device performance state of device i at time t-1; Indicates the current moment.

[0031] In this embodiment of the invention, training the DP-GRUC model using the Continuous Learning Performance Degradation Prediction (CLSP-EWC) method based on elastic weight consolidation and the device performance status dataset to obtain the trained DP-GRUC model further includes: when any task scenario triggers a DP-GRUC model update, obtaining the device performance status dataset of the newly added task scenario and the DP-GRUC model trained on the previous task; inputting the device performance status dataset of the newly added task scenario into the DP-GRUC model trained on the previous task, and calculating the training loss function of elastic weight consolidation and the current task based on the Continuous Learning Performance Degradation Prediction (CLSP-EWC) method based on elastic weight consolidation; training the DP-GRUC model trained on the previous task by minimizing the training loss function using gradient descent to obtain the DP-GRUC model trained on the current task.

[0032] In this embodiment of the invention, the training loss function is minimized as follows:

[0033] In the formula, Represents the training loss function; This represents the training loss function for task B. Represents the elastic coefficient; Indicates model parameters, This represents the optimal model parameters for model A. Indicates parameter weights; Represents the model function.

[0034] This invention also provides a continuous learning performance degradation prediction device based on stability-plasticity balance, comprising: an acquisition module for acquiring, in real time, a dynamically updated device performance status dataset under different operating task scenarios; a construction and training module for constructing a performance degradation prediction DP-GRUC model based on a convolutional gated network, and training the DP-GRUC model using the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance status dataset to obtain a trained DP-GRUC model; and a prediction module for predicting the performance degradation of the device under test in a target task scenario based on the trained DP-GRUC model.

[0035] This invention also provides an electronic device, including 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, it implements the steps of the continuous learning performance degradation prediction method based on stability-plasticity balance.

[0036] Prior to the continuous learning performance degradation prediction method based on stability-plasticity balance, this embodiment of the invention further includes a performance status evaluation method based on a spatiotemporal attention mechanism. Specifically, this method includes: converting equipment operation monitoring performance degradation data into three-dimensional time series data using a sliding window reconstruction method for detection data; graphically representing the performance monitoring sensors 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 evaluating the performance status of the equipment to be evaluated based on the trained STAGCN model.

[0037] In this embodiment of the invention, the STAGCN model includes a spatiotemporal attention model and an attention map 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.

[0038] In this embodiment of the invention, the process of inputting the three-dimensional time series data and the adjacency matrix of the graph neural network into the STAGCN model for training to obtain a trained STAGCN model includes: inputting the three-dimensional time series data into the temporal attention machine of the spatiotemporal attention model to calculate the temporal attention matrix; inputting the temporal attention matrix and the three-dimensional time series data into the spatial attention machine of the spatiotemporal attention model to calculate the spatial attention matrix; multiplying the spatial attention matrix with the adjacency matrix of the graph neural network to construct an attention graph convolutional model, and inputting the three-dimensional time series data into the attention graph convolutional model to obtain the device's health assessment index; calculating the loss value between the health assessment index and the assessment label, and obtaining the trained STAGCN model based on the loss value.

[0039] In this embodiment of the invention, the time attention matrix includes:

[0040] In the formula, E is the time attention matrix. For the third three-dimensional time series data r The input of a spatiotemporal network block, , , , , These are all parameters to be learned in the time attention matrix. This is the activation function.

[0041] In this embodiment of the invention, the spatial attention matrix includes:

[0042] in, S For spatial attention matrix, For the third three-dimensional time series data r The input of a spatiotemporal network block, , , , , These are all parameters to be learned in the spatial attention matrix. This is the activation function.

[0043] In this embodiment of the invention, the performance monitoring sensor is graphically represented as a graph neural network adjacency matrix according to the relationship between the sensor's physical layout and the device structure, including:

[0044] in, Indicates the first The monitoring parameter and the first The edge weight values ​​between the monitoring parameters q The correlation coefficient of sensors at the same location is 0 < q <1, m and n These represent different data collection points.

[0045] In this embodiment of the invention, the sensor is of various types, and one or more different types of sensors are installed at various locations on the device structure; The number of sensor acquisition channels is:

[0046] in, N Indicates the number of acquisition channels; Indicates the first m Number of sensors at each data collection point; M Indicates the number of sensor locations.

[0047] In this embodiment of the invention, the method of converting equipment operation monitoring performance degradation data into three-dimensional time series data using the detection data sliding window reconstruction method includes: using a sliding time window of fixed length and fixed step size to divide the data in each start-up and shutdown cycle of the entire life cycle from operation to failure into small multi-dimensional time series, and stacking the multi-dimensional time series into three-dimensional time series data.

[0048] Addressing the practical needs of industrial scenarios and focusing on the continuous learning objective of equipment performance degradation prediction, this invention proposes a continuous learning of state prediction based on elastic weight consolidation (CLSP-EWC) method inspired by synaptic plasticity mechanisms. The method architecture is as follows: Figure 2 As shown, this 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 HI (History, Development, and Construction) through a two-level encoder structure. During the training and update process, the degradation feature memory model achieves the memorization and consolidation of historical task information by constraining the model's learning direction, thus achieving a balance between the model's stability in historical tasks and its plasticity in new tasks.

[0049] like Figure 2As shown, the performance degradation prediction model in the CLSP-EWC method consists of two encoders (historical and future) and a feature decoder. The encoders take historical and future HI data as input, learn through neural networks to mine degradation features, and combine them to generate key values ​​for these degradation features. The decoder takes these key values ​​as input and fits the mapping relationship between historical and future degradation features through decoding. During task update model retraining, 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 to the historical task. This converges the gradient fitting direction of the model within the superposition space of the historical and new tasks. By combining the performance degradation prediction model and the degradation feature memory model, continuous learning performance degradation prediction is achieved for new task scenarios on different datasets.

[0050] 1.1 Continuous Learning Performance Degradation Prediction Process Based on Elastic Weight Consolidation Based on the CLSP-EWC method designed in this invention, and considering the characteristics of equipment performance degradation, such as Figure 3 As shown, this invention specifically designs an elastic weight consolidation continuous learning process for performance degradation prediction tasks. For four scenarios in actual industrial equipment deployment—new monitoring data, new operating conditions, new individual equipment, and new equipment models—a performance degradation prediction task triggering mechanism is designed. Simultaneously, considering the characteristics of equipment performance degradation evolution, a performance degradation prediction model based on a convolutional gated network (Degradation Prediction based on GRU with Convolutional, DP-GRUC) is specifically designed. Based on the DP-GRUC model structure, an elastic weight consolidation degradation feature memory model dynamic update method is designed. Finally, based on a general evaluation index for continuous learning in classification tasks, this invention, combined with the characteristics of performance degradation trend prediction tasks, specifically designs a continuous learning performance degradation prediction evaluation index to evaluate the ability of the CLSP-EWC method designed in this invention to overcome catastrophic forgetting, i.e., the continuous learning level.

[0051] (11) Task scenario attribute division In the actual use of equipment, performance degradation prediction is typically a non-real-time task within the health management process. That is, it is usually triggered by fixed needs such as monitoring data anomalies or periodic maintenance. Therefore, this invention addresses the equipment operation scenarios in practical industrial applications by designing a corresponding prediction task update triggering mechanism. Furthermore, it sets differentiated elasticity coefficients to accommodate the differences in performance degradation characteristics across different scenarios, thus adapting to the differences in task updates.

[0052] 111) Time Series Expansion Task In the industrial sector, equipment typically has a long service life, with most industrial equipment in use for several years or even decades. During this operation, the performance status (HI) data continuously increases. After the equipment is put into operation, due to factors such as component wear, a specific failure mode may occur in a subsystem, and the equipment's health status may transition to a certain type of abnormal state. Therefore, this invention designs a method based on performance status (HI) data, setting a quantitative performance degradation prediction threshold. When the equipment's HI falls below a certain threshold, a prediction model update is triggered.

[0053] 112) Expansion of the working condition domain During actual operation, the equipment's operating conditions change due to environmental and usage scenarios. Since different operating conditions cause significantly different levels of damage to the equipment's subsystems, they greatly influence the equipment's performance degradation characteristics. Based on this scenario, the equipment's operating environment and operating condition parameters are monitored to identify and determine its operating status. When the equipment enters a new operating condition, a predictive model update is triggered.

[0054] 113) Unit Quantity Expansion Task Due to individual differences such as manufacturing errors, equipment is affected by varying degradation factors during performance degradation, resulting in different individual equipment units exhibiting different personalized performance degradation characteristics. Therefore, in a group of equipment operating scenarios, when a new unit enters the group and generates certain performance status (HI) data, a prediction model update is triggered.

[0055] 114) Degradation Feature Expansion Task In the industrial sector, new models of similar equipment are typically designed, developed, and manufactured within a certain cycle. In most industries, it is common for both old and new models of equipment to be in service simultaneously. Due to differences in equipment structure, formulation, and other design aspects, the performance degradation characteristics of the equipment will also differ accordingly. Therefore, when a new model of equipment is put into operation in a group of equipment, and after the new model of equipment operates and generates certain performance status (HI) data, a predictive model update is triggered.

[0056] (12) Prediction model construction and degradation feature memory For actual equipment operating scenarios, a device performance degradation prediction model is constructed based on the DP-GRUC model structure, using the performance degradation prediction task update triggering mechanism described above. The model parameters are trained and updated using an elastic weight reinforcement feature memory method tailored to the performance degradation prediction task. The specific training process is shown in Table 1. The CLSP-EWC method is used to iterate the DP-GRUC model and update the model weights along the temporal dimension, outputting the corresponding prediction model at different times. PT .

[0057] Table 1

[0058] (13) Evaluation index for continuous learning of performance degradation prediction task Since the widespread and successful applications of continuous learning methods are mainly concentrated in classification tasks, this invention, in order to evaluate the ability of the CLSP-EWC method proposed in this paper to overcome catastrophic forgetting in a task, references widely used evaluation metrics in the field of continuous learning and improves them according to the characteristics of performance degradation prediction tasks, designing an evaluation metric for continuous learning methods oriented towards performance degradation prediction tasks. In the evaluation process of continuous learning in classification tasks, the metric calculation is mainly based on the task accuracy matrix, the structure of which is as follows: Figure 4 As shown. The task accuracy matrix is ​​formed by arranging the model's accuracy for each task after training in chronological order, according to the order in which the tasks occurred. Indicates the first i After training on the data from the [number] task, the model [is in the] [number]th task. j Accuracy in each task.

[0059] 1.2 Performance Degradation Prediction Model Based on Convolutional Gated Networks Equipment performance degradation is a continuous and dynamic time-series process that begins when the equipment is put into operation and ends when it fails and reaches the end of its lifespan. The HI trajectory represents the entire process of equipment performance change, and the length of the HI trajectory represents the length of the equipment's operational life. Throughout the entire equipment's operational lifespan, the performance degradation process has the following two characteristics: (21) Continuity of the degradation process: The performance change caused by a certain start-stop cycle of the equipment is based on the performance state after the end of the previous work cycle. That is, the performance degradation process is a time-sequential evolution. The change in the performance state of the equipment is not only related to the operating conditions of the equipment and the individual condition of the equipment, but also to the operating cycle time of the equipment.

[0060] (22) Differences in degradation characteristics: For equipment of the same model and batch, due to differences in initial manufacturing and assembly errors, as well as differences in operating environment and working conditions, the degradation characteristics of the equipment have certain differences. That is, in the same equipment performance degradation characteristic space, different individuals still have individual differences. Therefore, equipment performance degradation prediction needs to mine common degradation characteristics based on historical data and combine them with the individual degradation characteristics of the equipment to be predicted to output the prediction results.

[0061] Predicting equipment performance degradation allows for early assessment of the equipment's future performance status during operation, providing strong support for operational decisions. This is based on a dynamically updated equipment performance status dataset. Addressing two key characteristics of equipment performance degradation and evolution, the DP-GRUC model comprises a history encoder module, a future encoder module, and a HI trajectory decoder module. The model architecture is as follows: Figure 5 As shown.

[0062] Based on the continuous nature of equipment degradation, this invention designs an HI trajectory encoder based on convolutional neural networks and gated neural networks, extracting equipment degradation features by fitting the HI trajectory gradient. and The data, and the gradient of the HI trajectory, are defined as follows: (1.0)

[0063] (1.1)

[0064] and The gradients of the equipment performance degradation trajectory (HI) at historical and future times represent the relative change in equipment performance before and after a start-stop cycle. Based on and The calculation process for historical degradation feature coding and future degradation feature coding is as follows: (1.2) (1.3) In the formula, and Encoders representing historical and future degradation characteristics respectively and Learnable parameters of the GRU layer in a mid-gated neural network and the Conv1d layer in a one-dimensional convolutional network.

[0065] Based on the encoding results, the historical and future degradation characteristics at a certain point in time are compared. Defined as a pair of degradation feature-related key values. Subsequently, the device performance state at the current runtime. With historical degradation gradient It is fed into the encoder to generate the future HI trajectory. The system trains a performance degradation prediction model using historical data from different devices, learns the key-value mapping relationship of degradation features based on the encoder model, and generates the future degradation gradient based on the HI trajectory information of the device to be predicted using the decoder. The system then derives the device performance state value at the future time from the current performance state, thus achieving device performance degradation prediction.

[0066] 1.3 Degenerate Feature Memory Model Based on Elastic Weight Consolidation The DP-GRUC model can be used to predict device performance degradation. However, the fitting results of a single training cycle of the model converge to the training dataset D. M The degradation features contained within the dataset D, as equipment continues to operate and industrial demands evolve, will lead to the addition of new dataset D. M+1 As the number of learners increases, a task conflict arises between consolidating historical learning capabilities in the model and updating with new datasets. This problem is known in biological neuroscience as the "stability-plasticity" balance. In response to repeated external stimuli, the morphology of connections between neuronal synapses changes, and the reconnection process aims for optimal learning cost. The task architecture of continuous learning prohibits further training on the dataset of a given task after a fixed amount of training on each task. Therefore, achieving continuous learning requires restricting the direction of gradient descent in model weight iterations to ensure the model's ability to complete the preceding task A. Based on this goal, inspired by the biological mechanism of synaptic plasticity, this invention constructs an elastic weight-consolidated feature memory model for performance degradation prediction tasks. In DP-GRUC artificial neural networks, neurons influence each other through connection weights. The core of elastic weight consolidation is to treat the weight protection applied to neuronal synaptic connections as a spring connection, whose stiffness is proportional to the importance of the preceding task's neuronal connection. This invention focuses on the design and research of an elastic weight consolidation method for the continuous learning task of predicting the performance degradation of complex devices based on the DP-GRUC model.

[0067] (131) Design of elastic weight consolidation method for DP-GRUC model The DP-GRUC model neural network consists of multiple layers of linear mapping layers and various types of nonlinear mapping layers stacked together. Fitting learning is a process of continuously adjusting the model weight set based on the deviation of the linear mapping to optimize the model's predictive performance. θ Represents the parameters of an artificial neural network layer. θ This indicates the optimal parameter configuration for the task. θ During configuration, the same performance effect can be achieved with multiple configuration combinations; this theory forms the basis of elastic weight consolidation. This invention develops its method design based on this principle, addressing the issue of over-parameterization affecting the solution configuration for subsequent task B. There is a very high probability that the solution configuration is the same as that of the preceding task A. Approaching. Therefore, as Figure 7 As shown, when learning subsequent task B, EWC restricts the parameters to a range of... The constraint is placed within the low-error region of task A, thus protecting the performance of task A. This constraint is achieved through a quadratic penalty method, so the resetting of neuron weights can be defined as the process of a rigid spring anchoring the weight parameters to the configuration region of the preceding task. The spring stiffness varies among the parameters; it is greater for parameters that are important to the performance of the preceding task A, and less so for those that are less important.

[0068] Different retraining model parameter change paths, such as Figure 6 As shown, after the model has learned the preceding task A, the parameters are located in space. If the model training gradient is adjusted solely based on the performance on task B, the parameter path will be as shown by the red line. The model maximizes its learning ability for task B but destroys the knowledge learned in task A, leading to knowledge loss. If task A constrains the model iteration process, as shown by the green line in the figure, the model's ability to learn task B is severely insufficient due to the overly strict constraints. Therefore, by calculating the importance of different model weights to task A, determining the direction of the parameter gradient change path, and finding a gradient path for task B within the low-loss space of parameter A, effective learning of task B can be achieved while preserving the capabilities of task A.

[0069] To demonstrate the rationality of retraining constraints on deep neural network models and determine the importance of different neuron weights to the task, the model training process is optimized probabilistically, i.e., given a dataset... Determine the optimal importance under the given conditions. Based on the prior probability of the parameters. Calculate conditional probability Based on Bayesian rules: (1.4) Since maximizing the objective function is the same as maximizing the logarithm, the calculation... The logarithmic process is as follows: (1.5) In the formula, Given dataset The logarithm of the probability. During network model training, The objective function to be optimized on the log-likelihood function is: (1.6) Assume the dataset corresponding to the task consists of two independent parts, and task A and task B are defined as follows: and Therefore, the formula is rearranged as follows: (1.7) In the formula, the left side of the equation still describes the given dataset. The posterior probability of the parameters, the right side of the equation depends only on the loss function of the subsequent task B. Therefore, the posterior distribution on the right-hand side of the equation... The model must contain knowledge information about the preceding task A. This knowledge information includes the importance of the model parameters to task A, and this importance information is a key attribute for achieving EWC.

[0070] (132) Approximate solution of the posterior distribution of the preceding task In formula (1.7), Let be the posterior probability distribution of the preceding task A, containing parameter information for the network model fitting task A. However, since it cannot be expressed in an interpretable and explicit form, calculating the posterior probability that perfectly matches the actual data is very difficult. Therefore, we consider calculating its approximation using the Karplacian method.

[0071] First, exist Second Taylor expansion (1.8)

[0072] because Peak value, tangent slope Formula (1.8) can be written as: (1.9) Based on formula (1.6), formula (1.9) can be written as: (2.0) in, .Will Written as Formula (2.0) can be expressed as the standard expression for the normal distribution: (2.1)

[0073] Therefore, the posterior probability distribution of the preceding task A can be obtained as follows: (2.2) That is, such as Figure 7 As shown, the normal distribution approximation of a continuous probability density distribution is found based on the Laplace approximation method. Assume... To smooth the probability distribution, and the peak is located at the maximum point ( Near the mean, it can be used. With variance Normal distribution fitting .

[0074] (133) Calculation of the importance of Fisher information matrix parameters like Figure 7As shown, the Laplace approximation method can be used to... It approximates a normal distribution, and its mean is 1 / 2. The variance of the normal distribution is Due to the Fisher information matrix It can be calculated using the first derivative. This property can be used to avoid calculating the Hessian matrix in formula (1.9).

[0075] (2.3)

[0076] Substituting formula (2.0) into formula (1.7), we get: (2.4)

[0077] In the formula, Represents all constant terms. The introduced hyperparameter is used to balance the network model's ability to learn task B and remember its predecessor task A.

[0078] Substituting formula (2.3) into the equation, we get: (2.5) Therefore, the elastic weight consolidation training of the network model can be expressed as formula (2.5). Wherein, the Fisher information matrix... It contains information about the importance of model parameters to task A. It has three key properties: Equivalent to the second derivative when the loss is close to its minimum; It can be calculated using only the first derivative, therefore the calculation process is relatively simple; It is a positive semi-definite matrix.

[0079] The Laplace calculation is similar to the expectation propagation process, where each subtask is treated as a posterior factor; therefore, the loss function minimized by EWC is: (2.6) In the formula, Let the training loss function be for task B. As the elasticity coefficient, the relative importance of task B and its predecessor task A were configured, and through... Relative importance labeling was performed on each parameter. This enabled the training of the EWC model.

[0080] When the model performs the third task C, EWC uses the sum of the two penalty terms from tasks A and B as a secondary penalty term to bring the network parameters closer to the common parameter region. Similarly, for future prediction tasks, the model maintains its memory of historical tasks through historical task penalty terms, thereby enabling the DP-GRUC model to learn continuously.

[0081] Specifically, prior to the continuous learning performance degradation prediction method based on the stability-plasticity balance, this invention also includes quantitatively characterizing its HI (Health Index) through a performance status assessment method. In this context, a condition monitoring network for complex equipment systems is established. G In total N There are [number] monitoring points, and the monitoring point records are as follows: Recorded at each start-stop cycle monitoring point of the equipment L One measurement value, .by Indicates in t At that moment, the n Among the _ status monitoring points, the _th l Each condition monitoring measurement value. Therefore, with Indicates in t At that moment, the n All status monitoring measurements of each node; express t All status monitoring and measurement values ​​at all times; express All condition monitoring measurements within the time series. Based on the above assumptions, the objective function for equipment condition monitoring is: ,in, ,by Characterization t Quantitative value of device performance status at any given time.

[0082] To address the aforementioned problem description and model fitting objectives, and considering the engineering characteristics of equipment operation and data acquisition during performance status assessment, this invention proposes a performance status assessment method based on Spatial-Temporal Attention based on Graph Convolutional Networks (STAGCN). First, considering the varying durations of single start-stop cycles for equipment, a three-dimensional monitoring data sliding window reconstruction method is designed during data preprocessing to convert equipment operation monitoring data into training and testing data for the performance status assessment model. Second, a graphical representation method for performance monitoring sensors is designed to address the relationship between the physical layout of equipment sensors and the equipment structure. The equipment physical structure design information is input into the STAGCN model via a graph neural network adjacency matrix. Based on this, the STAGCN model is trained using the input data and the adjacency matrix to output the performance status assessment results.

[0083] In engineering applications, due to differences in the sampling frequency, physical dimensions, and sampling quality among different sensors, equipment operation monitoring data cannot usually be directly used for data-driven performance status assessment. Therefore, it is necessary to preprocess the data first, and the processed data is then used as input to the performance status assessment model.

[0084] Because the start-up and shutdown durations of most devices are not consistent, this leads to different... t Monitoring data collected in real time X t Corresponding length L To address this issue of data variability, this invention, based on conventional preprocessing methods such as resampling and normalization, designs a monitoring data sliding window reconstruction method tailored to the characteristics of equipment performance status assessment. This method uses a fixed-length sliding window... Win A sliding time window with a step size of 1 divides the data from each start-stop cycle throughout the entire training lifecycle into small multidimensional time series, and stacks them into a three-dimensional time series. This ensures that the data contained in each sliding window is dimensionally regular input data, i.e., it... X t Convert to In addition, the health status corresponding to each sliding window data point is used as an evaluation label. The original work cycle data of varying lengths are transformed into a regular three-dimensional data structure through a data sliding window reconstruction method, providing standardized data input for the performance status evaluation model.

[0085] In complex systems, operational monitoring sensors are installed at various locations within the equipment structure; some locations may house one or more sensors of different types. This is the basis for establishing a condition monitoring network for complex equipment systems. GIn the middle, the number of sensor locations is M ,but:

[0086] in, M m For the first m Number of sensors at each data collection point N This represents the number of acquisition channels. Sensors are numbered according to their physical layout. Data collected from the same sampling point exhibits a certain correlation; therefore, the performance monitoring sensors are graphically represented as an adjacency matrix as shown in the formula:

[0087] in, w i,j Indicates the first i The monitoring parameter and the first j The edge weight values ​​between the monitoring parameters q The correlation coefficient of sensors at the same location is 0 < q <1, m and n These represent different data collection points.

[0088] The spatiotemporal attention graph convolutional model is the core of the equipment performance status assessment method proposed in this invention. It identifies the time and space dimensions of performance degradation information in the equipment status monitoring data step by step through a stacked model structure, and fits the equipment performance status degradation features through a graph neural network. Based on the feature decoder, the degradation features are fitted into a one-dimensional vector, and finally the large amount of performance status information covered in the monitoring data is transformed into a quantitative health index HI.

[0089] The STAGCN model consists of two parts: a spatiotemporal attention model and an attention graph convolutional model. The spatiotemporal attention model takes multi-dimensional monitoring parameters of the device as input, outputs a temporal attention matrix through a temporal attention machine, and passes the degradation feature density information of the temporal dimension to a spatial attention machine. The spatial attention machine performs multi-dimensional matrix transformations on the high-dimensional input data and outputs a spatial attention matrix to represent the differences in degradation information density between different monitoring parameters. The attention graph convolutional model converts the physical location of the sensors of the device under evaluation into an adjacency matrix of a graph convolutional network through graphical representation. The adjacency matrix in the graph neural network incorporates the layout information of the device status monitoring sensors, and combines the adjacency matrix with the spatial attention matrix to mine device performance status degradation features through graph convolutional encoding. Finally, based on the performance status degradation features, a linear decoder fits the degradation features into a health index HI, thereby achieving a quantitative representation of the device performance status.

[0090] A device performance degradation assessment model is constructed based on the model structure, and the model is trained using training data. The algorithm process is shown in Table 2.

[0091] Table 2

[0092]

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that it can still be used to modify the technical solutions described in the foregoing embodiments, or to make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting continuous learning performance degradation based on a stability-plasticity balance, characterized in that, include: Based on the equipment operation task scenario, acquire the dynamically updated equipment performance status dataset in real time under different operation task scenarios; the equipment operation task scenario includes any of the following: new monitoring data, new operating conditions, new individual equipment, and new equipment model. A performance degradation prediction DP-GRUC model based on a convolutional gated network is constructed, and the DP-GRUC model is trained using the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance state dataset to obtain the trained DP-GRUC model. The process includes: when any task scenario triggers a DP-GRUC model update, obtaining the device performance state dataset of the newly added task scenario and the DP-GRUC model trained on the previous task; inputting the device performance state dataset of the newly added task scenario into the DP-GRUC model trained on the previous task, and calculating the training loss function of elastic weight consolidation and the current task based on the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation; training the DP-GRUC model trained on the previous task using gradient descent to minimize the training loss function, thus obtaining the DP-GRUC model trained on the current task; wherein, the minimized training loss function is: In the formula, Represents the training loss function; This represents the training loss function for task B. Represents the elasticity coefficient; Indicates model parameters, This represents the optimal model parameters for model A. Indicates parameter weights; Represents the model function; Based on the trained DP-GRUC model, the performance degradation prediction of the device under test in the target task scenario is performed; The DP-GRUC model consists of a history encoder module, a future encoder module, and an HI trajectory decoder module. Before acquiring the dynamically updated device performance status dataset under different operational task scenarios based on the device's operational task scenarios, the process also includes: Using the sliding window reconstruction method of detection data, the equipment operation monitoring performance degradation data is converted into three-dimensional time series data including monitoring dimension time series data, time dimension time series data and operation dimension time series data; Based on the relationship between the physical layout of the sensors and the structure of the equipment, the performance monitoring sensors are graphically represented as an adjacency matrix of a graph neural network; A STAGCN model based on spatiotemporal attention graph convolution, comprising a spatiotemporal attention model and an attention graph convolution model, is constructed. The 3D time-series data and the adjacency matrix of the graph neural network are input into the STAGCN model for training to obtain a trained STAGCN model. The process includes: inputting the 3D time-series data into the temporal attention machine of the spatiotemporal attention model to calculate the temporal attention matrix; inputting the temporal attention matrix and the 3D time-series data into the spatial attention machine of the spatiotemporal attention model to calculate the spatial attention matrix; multiplying the spatial attention matrix by the adjacency matrix of the graph neural network to construct an attention graph convolution model, and inputting the 3D time-series data into the attention graph convolution model to obtain the device's health assessment index; calculating the loss value between the health assessment index and the assessment label, and obtaining the trained STAGCN model based on the loss value. The performance status of the device to be evaluated is assessed based on the trained STAGCN model. The time attention matrix includes: In the formula, E is the time attention matrix. For the third three-dimensional time series data r The input of a spatiotemporal network block, It is the first in the three-dimensional time series data r Length of the time dimension in the layer For the third three-dimensional time series data r The number of channels for layer input data, These are all parameters to be learned in the time attention matrix. For activation functions; The spatial attention matrix includes: in, S For spatial attention matrix, For the third three-dimensional time series data r The input of a spatiotemporal network block, It is the first in the three-dimensional time series data r Length of the time dimension in the layer For the third three-dimensional time series data r The number of channels for layer input data, These are all parameters to be learned in the spatial attention matrix. This is the activation function.

2. The method according to claim 1, characterized in that, The DP-GRUC model is trained using the CLSP-EWC method for continuous learning performance degradation prediction based on elastic weight consolidation and the device performance state dataset, resulting in the following trained DP-GRUC model: The DP-GRUC model is initialized with randomly initialized model weights; Input the device performance status dataset for any task scenario into the initial DP-GRUC model to fit the historical health index HI trajectory and the future health index HI trajectory, and obtain and save the DP-GRUC model after the first training in that task scenario.

3. The method according to claim 2, characterized in that, Input the device performance status dataset under any task scenario into the initial DP-GRUC model to fit the historical health index HI trajectory gradient and the future health index HI trajectory gradient, including: The device performance status dataset under any task scenario is divided into several time windows HI trajectories by using a sliding window segmentation method. The first half of the data in the sliding window is set as the historical HI trajectory, and the second half is set as its corresponding future HI trajectory. Calculate the gradient of the historical HI trajectory and the gradient of the future HI trajectory 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, the historical HI trajectory gradient is input into the historical encoder for historical encoding update, and the output is the historical HI trajectory code; at the same time, the future HI trajectory gradient is input into the future encoder for future encoding update, and the output is the future HI trajectory code. The output historical HI trajectory code and the future HI trajectory code are combined to form a degradation feature association key value.

4. The method according to claim 3, characterized in that, Calculating the gradients of the historical HI trajectory and the future HI trajectory based on the historical HI trajectory includes: in, Represents the gradient of the historical HI trajectory; The fraction represents the gradient of the future HI trajectory; This represents the device performance state of device i at time t; This represents the device performance state of device i at time t-1; Indicates the current moment.

5. A device for predicting continuous learning performance degradation based on a stability-plasticity balance, the device being used to implement the method as described in any one of claims 1-4, characterized in that, include: The acquisition module is used to acquire, in real time, a dynamically updated dataset of device performance status under different device operation task scenarios; the device operation task scenarios include any of the following: new monitoring data, new operating conditions, new individual devices, and new device models; A construction and training module is used to construct a performance degradation prediction DP-GRUC model based on a convolutional gating network, and to train the DP-GRUC model using the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation and the device performance status dataset, resulting in a trained DP-GRUC model. The module includes: when any task scenario triggers a DP-GRUC model update, acquiring the device performance status dataset of the newly added task scenario and the DP-GRUC model trained on the previous task; inputting the device performance status dataset of the newly added task scenario into the DP-GRUC model trained on the previous task, and calculating the training loss function of elastic weight consolidation and the current task based on the continuous learning performance degradation prediction CLSP-EWC method based on elastic weight consolidation; and training the DP-GRUC model trained on the previous task using gradient descent to minimize the training loss function, resulting in the DP-GRUC model trained on the current task. The minimized training loss function is: In the formula, Represents the training loss function; This represents the training loss function for task B. Represents the elasticity coefficient; Indicates model parameters, This represents the optimal model parameters for model A. Indicates parameter weights; Represents the model function; The prediction module is used to predict the performance degradation of the device under test in the target task scenario based on the trained DP-GRUC model. The DP-GRUC model consists of a history encoder module, a future encoder module, and a HI trajectory decoder module.

6. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor; when executed by the processor, the computer program implements the steps of the continuous learning performance degradation prediction method based on stability-plasticity balance as described in any one of claims 1-4.

Images