Continuous learning performance degradation prediction method and device and electronic equipment

Abstract

The invention discloses a continuous learning performance degradation prediction method and device and electronic equipment, and the method comprises the steps: constructing a performance degradation prediction sub-model, and constructing a performance degradation prediction model network based on multi-model joint memory enhancement continuous learning MCL-MAN based on the performance degradation prediction sub-model; and obtaining performance state degradation data sets of different batches and different models of equipment in different task scenes, and dynamically updating the MCL-MAN-based performance degradation prediction model networking by using the performance state degradation data sets of different batches and different models of equipment in different task scenes. The dynamically updated performance degradation prediction model networking based on the MCL-MAN is obtained; and acquiring a historical health index HI trajectory gradient of the to-be-predicted equipment, and predicting a future performance degradation trajectory of the to-be-predicted equipment by inputting the historical health index HI trajectory gradient of the to-be-predicted equipment into the dynamically updated MCL-MAN-based performance degradation prediction model network.

Description

Technical Field

[0001] The present invention relates to the technical field of performance degradation prediction, and in particular to a continuous learning performance degradation prediction method, device and electronic equipment. Background Art

[0002] The performance degradation prediction model of artificial neural networks is limited by the network structure and the number of neurons. The memory capacity of a single prediction model has a certain upper limit, and its memory capacity for features cannot be continuously expanded. In addition, for complex system equipment, it is composed of multiple subsystems or components that operate in combination, and the performance degradation characteristics of each component have a certain degree of coupling. For group-degraded product equipment with different failed components and failure modes, a single performance degradation prediction model has strong limitations in the task of fitting the degradation trajectory of multiple performance indicators of complex system equipment. In macroscopic terms, a single artificial neural network model is difficult to fully simulate the complex learning and memory process in the mammalian brain.

[0003] In practice, complex system equipment is often composed of multiple subsystems or components. The subsystems have relatively independent structures and failure propagation mechanisms. The subsystems influence each other through certain pathways, forming failure propagation network links. Limited by the model structure and the number of neurons, a single independent performance degradation network has an upper limit in capturing the complex, long-term, and dynamic degradation laws of group equipment structures. Summary of the invention

[0004] The present invention provides a method, device and electronic device for predicting continuous learning performance degradation, so as to solve the technical problem of how to realize continuous learning performance degradation prediction of complex equipment.

[0005] An embodiment of the present invention provides a continuous learning performance degradation prediction method, including: constructing a performance degradation prediction sub-model, and based on the performance degradation prediction sub-model, constructing a performance degradation prediction model network based on multi-model joint memory enhanced continuous learning MCL-MAN; obtaining performance status degradation data sets of different batches and different models of equipment under different task scenarios, and dynamically updating the performance degradation prediction model network based on MCL-MAN using the performance status degradation data sets of different batches and different models of equipment under the different task scenarios to obtain a dynamically updated performance degradation prediction model network based on MCL-MAN; obtaining a historical health index HI trajectory gradient of the device to be predicted, and predicting the future performance degradation trajectory of the device to be predicted by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated performance degradation prediction model network based on MCL-MAN.

[0006] Preferably, by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated MCL-MAN-based performance degradation prediction model network, predicting the future performance degradation trajectory of the device to be predicted includes: obtaining a historical feature key value by inputting the historical health index HI trajectory gradient of the device to be predicted into the encoder in the dynamically updated MCL-MAN-based performance degradation prediction model network; generating a query key value by using the historical feature key value; based on the query key value, predicting the future performance degradation trajectory of the device to be predicted through iterative key value addressing, numerical value reading, and HI trajectory prediction set update query steps in sequence.

[0007] Preferably, based on the query key value, the query steps of iterative key value addressing, numerical value reading, and HI trajectory prediction set updating are sequentially performed to predict the future performance degradation trajectory of the device to be predicted, including: calculating the addressing probability based on the query key value; calculating the skip-step attention using the addressing probability as the reading weight; generating the health indicator HI trajectory through the skip-step attention and the historical feature key value; in each iteration, combining the query key value of the previous task with the skip-step attention as the query key value of the current task; repeating the above steps until the memory query demand ends, using different query key values ​​to query each skip-step attention, and predicting the future performance degradation trajectory of the device to be predicted.

[0008] Preferably, the MCL-MAN-based performance degradation prediction model networking is based on the dynamic update of each performance degradation prediction sub-model to achieve continuous learning performance degradation prediction of complex equipment, wherein the dynamic update of each performance degradation prediction sub-model includes a pre-training stage and a continuous learning stage.

[0009] Preferably, the pre-training stage includes dynamic updating of the encoder, dynamic updating of the controller and dynamic updating of the prediction model.

[0010] Preferably, the dynamic update of the encoder includes: dividing the historical data into several time window HI trajectories by sliding window cutting, 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, and inputting the historical HI trajectory gradient into the historical encoder for historical coding update, and outputting the historical HI trajectory coding; at the same time, inputting the future HI trajectory gradient into the 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 associated key-value pair, and writing them into an external memory module based on a controller; decoding the degradation feature associated key-value pair by a decoder, taking the current performance status indicator as the starting point, and extrapolating the future HI trajectory, so as to realize the dynamic update of the encoder.

[0011] Preferably, the controller dynamically updates including: inputting the historical HI trajectory gradient into the updated encoder module to generate the historical HI trajectory code, comparing it with the historical memory code in the external memory module, and obtaining the historical memory code reading weight; arranging the reading weights in order and assigning the sorting result to the historical memory code index, combining and pairing the future memory code with the historical HI trajectory code according to the index value to form a degradation feature associated key-value pair, decoding through the updated decoder, and outputting the predicted HI trajectory; combining all the predicted HI trajectories into a predicted HI trajectory set, and calculating the cosine distance between the predicted HI trajectory set and the future HI trajectory set in turn, calculating the output error tolerance rate through the error tolerance, and fitting it as a write probability based on the linear layer; using the write probability to find the differentiated degradation feature associated key-value pair, and write the differentiated degradation feature associated key-value pair into the external memory module.

[0012] Preferably, the dynamic update of the prediction model includes: inputting the historical HI trajectory gradient into the updated encoder module to generate the historical HI trajectory code, and inputting the historical HI trajectory code, the historical memory code, and the future memory code into the linear layer respectively to generate the memory query key value indicator; obtaining the addressing result by inputting the historical degradation trajectory code query key value and the historical memory code key value into the softmax layer, and comparing the addressing result with the future memory code to generate the memory query attention result; constructing the degradation feature associated key value pair by combining and pairing the historical degradation trajectory code with the attention result, and inputting the degradation feature associated key value pair into the decoder for trajectory decoding to generate the future HI trajectory, so as to realize the dynamic update of the performance degradation prediction model.

[0013] An embodiment of the present invention also provides a continuous learning performance degradation prediction device, including: a construction module, used to construct a performance degradation prediction sub-model, and based on the performance degradation prediction sub-model, construct a performance degradation prediction model network based on multi-model joint memory enhanced continuous learning MCL-MAN; a training module, used to obtain performance status degradation data sets of different batches and different models of equipment under different task scenarios, and use the performance status degradation data sets of different batches and different models of equipment under the different task scenarios to dynamically update the performance degradation prediction model network based on MCL-MAN to obtain a dynamically updated performance degradation prediction model network based on MCL-MAN; a prediction module, used to obtain the historical health index HI trajectory gradient of the device to be predicted, and predict the future performance degradation trajectory of the device to be predicted by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated performance degradation prediction model network based on MCL-MAN.

[0014] An embodiment of the present 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, the steps of the continuous learning performance degradation prediction method are implemented.

[0015] The beneficial effects of the present invention are: constructing a networked model cross-linking structure of multiple models, and constructing a model network based on an independent performance degradation prediction neural network model to adapt to the dynamically changing new performance degradation prediction tasks and fit the many-to-many complex mapping relationship between HI trajectories; for the performance degradation prediction task, the complementary learning system theory in biological continuous learning is introduced into the artificial neural network continuous learning process, and the hippocampus-cerebral cortex memory model mechanism is introduced into the dynamic update process of the performance degradation prediction model to solve the limitation of the model structure size on the continuous learning ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a multi-model joint memory-enhanced continuous learning performance degradation prediction method provided by an embodiment of the present invention; Figure 2 is a flow chart of a method for predicting memory enhancement continuous learning performance degradation provided by an embodiment of the present invention; Figure 3 is a flowchart of the dynamic update of the encoder of the MCL-MAN method provided by an embodiment of the present invention; Figure 4 is a flow chart of dynamic update of the MCL-MAN method controller provided by an embodiment of the present invention; Figure 5 It is a flow chart of dynamic updating of the performance degradation prediction model of the MCL-MAN method provided by an embodiment of the present invention; Figure 6 It is a flow chart of a continuous learning performance degradation prediction method provided by the present invention. DETAILED DESCRIPTION

[0017] 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.

[0018] Figure 6 is a flow chart of a method for predicting continuous learning performance degradation provided by the present invention, such as Figure 6 As shown, it includes: step S1: constructing a performance degradation prediction sub-model, and based on the performance degradation prediction sub-model, constructing a performance degradation prediction model network based on multi-model joint memory enhancement continuous learning MCL-MAN; step S2: obtaining performance status degradation data sets of different batches and different models of equipment under different task scenarios, and using the performance status degradation data sets of different batches and different models of equipment under the different task scenarios to dynamically update the performance degradation prediction model network based on MCL-MAN, to obtain a dynamically updated performance degradation prediction model network based on MCL-MAN; step S3: obtaining the historical health index HI trajectory gradient of the device to be predicted, and predicting the future performance degradation trajectory of the device to be predicted by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated performance degradation prediction model network based on MCL-MAN.

[0019] In an embodiment of the present invention, by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated MCL-MAN-based performance degradation prediction model network, predicting the future performance degradation trajectory of the device to be predicted includes: obtaining a historical feature key value by inputting the historical health index HI trajectory gradient of the device to be predicted into the encoder in the dynamically updated MCL-MAN-based performance degradation prediction model network; generating a query key value by using the historical feature key value; based on the query key value, predicting the future performance degradation trajectory of the device to be predicted through iterative key value addressing, numerical value reading, and HI trajectory prediction set update query steps in sequence.

[0020] In an embodiment of the present invention, based on the query key value, the future performance degradation trajectory of the device to be predicted is predicted by iterative key value addressing, numerical value reading, and HI trajectory prediction set update query steps in sequence, including: calculating the addressing probability based on the query key value; calculating the skip-step attention using the addressing probability as the reading weight; generating a health indicator HI trajectory through the skip-step attention and the historical feature key value; in each iteration, combining the query key value of the previous task with the skip-step attention as the query key value of the current task; repeating the above steps until the memory query demand ends, using different query key values ​​to query each skip-step attention, and predicting the future performance degradation trajectory of the device to be predicted.

[0021] In an embodiment of the present invention, the MCL-MAN-based performance degradation prediction model networking is based on the dynamic update of each performance degradation prediction sub-model to achieve continuous learning performance degradation prediction of complex equipment, wherein the dynamic update of each performance degradation prediction sub-model includes a pre-training stage and a continuous learning stage.

[0022] In an embodiment of the present invention, the pre-training stage includes dynamic updating of an encoder, dynamic updating of a controller and dynamic updating of a prediction model.

[0023] In an embodiment of the present invention, the dynamic update of the encoder includes: dividing the historical data into a number 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 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, and inputting the historical HI trajectory gradient into the historical encoder for historical coding update, and outputting the historical HI trajectory coding; at the same time, inputting the future HI trajectory gradient into the 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 associated key-value pair, and writing them into an external memory module based on a controller; decoding the degradation feature associated key-value pair by a decoder, and extrapolating the future HI trajectory with the current performance status indicator as the starting point, thereby realizing the dynamic update of the encoder.

[0024] In an embodiment of the present invention, the controller dynamically updates including: inputting the historical HI trajectory gradient into the updated encoder module to generate the historical HI trajectory code, comparing it with the historical memory code in the external memory module, and obtaining the historical memory code reading weight; arranging the reading weights in order and assigning the sorting result as the historical memory code index, combining and pairing the future memory code with the historical HI trajectory code according to the index value to form a degradation feature associated key-value pair, decoding through the updated decoder, and outputting the predicted HI trajectory; combining all the predicted HI trajectories into a predicted HI trajectory set, and calculating the cosine distance between the predicted HI trajectory set and the future HI trajectory set in turn, calculating the output error tolerance rate through the error tolerance, and fitting it as a write probability based on the linear layer; using the write probability to find the differentiated degradation feature associated key-value pair, and write the differentiated degradation feature associated key-value pair into the external memory module.

[0025] In an embodiment of the present invention, the dynamic update of the prediction model includes: inputting the historical HI trajectory gradient into the updated encoder module to generate the historical HI trajectory code, and inputting the historical HI trajectory code, the historical memory code, and the future memory code into the linear layer respectively to generate a memory query key value indicator; by inputting the historical degradation trajectory code query key value and the historical memory code key value into the softmax layer to obtain the addressing result, and comparing the addressing result with the future memory code to generate the memory query attention result; by combining and pairing the historical degradation trajectory code with the attention result to construct a degradation feature associated key value pair, and inputting the degradation feature associated key value pair into the decoder for trajectory decoding, generating the future HI trajectory, so as to realize the dynamic update of the performance degradation prediction model.

[0026] An embodiment of the present invention also provides a continuous learning performance degradation prediction device, including: a construction module, used to construct a performance degradation prediction sub-model, and based on the performance degradation prediction sub-model, construct a performance degradation prediction model network based on multi-model joint memory enhanced continuous learning MCL-MAN; a training module, used to obtain performance status degradation data sets of different batches and different models of equipment under different task scenarios, and use the performance status degradation data sets of different batches and different models of equipment under the different task scenarios to dynamically update the performance degradation prediction model network based on MCL-MAN to obtain a dynamically updated performance degradation prediction model network based on MCL-MAN; a prediction module, used to obtain the historical health index HI trajectory gradient of the device to be predicted, and predict the future performance degradation trajectory of the device to be predicted by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated performance degradation prediction model network based on MCL-MAN.

[0027] An embodiment of the present 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, the steps of the continuous learning performance degradation prediction method are implemented.

[0028] In practice, complex system equipment is often composed of multiple subsystems or components. The subsystems have relatively independent structures and failure propagation mechanisms. The subsystems influence each other through certain pathways, forming a failure propagation network link. That is, ,in Represents the subsystem performance degradation health indicator, is the number of health indicators of subsystem performance degradation. Limited by the model structure and the number of neurons, a single independent performance degradation network has an upper limit in capturing the complex, long-term, and dynamic degradation laws of group equipment structures. To solve this problem, based on the characteristics of brain partition learning in biological complementary learning systems, this paper designs a multi-model joint continuous learning based memory augmented continuous learning performance degradation prediction method (Multi-model Joint Continual Learningbased Memory Augmented Networks, MCL-MAN). Its evolution process is as follows Figure 1 shown.

[0029] like Figure 1 As shown in the figure, the multi-model joint network architecture designed in this paper is composed of sub-models with the same network structure and differentiated neuron hyperparameters. A single sub-model is used to simulate the sub-brain regions in the brain and fit the one-dimensional HI index in complex system equipment. As different batches and models of equipment are put into operation and performance degradation data is collected, the failure mode is continuously updated, and the sub-models of each brain region are linked to each other to form a performance degradation prediction model network. As the training data is added, different brain regions are selectively activated according to the failure of each subsystem of the new data, realizing the dynamic joint update of the performance degradation prediction network.

[0030] The MCL-MAN method realizes joint learning of multiple brain regions by jointly training and updating through forming hyperparameter links between sub-models. The core of the MCL-MAN method is to fit the HI trajectory of complex equipment based on each sub-model. Based on the multi-model joint network architecture, the dynamic update method of each sub-model is inspired by the "hippocampus-cerebral cortex" long and short-term memory model in the complementary learning system, and the performance degradation prediction task is simulated as a skill to be learned. A memory-enhanced continuous learning method for performance degradation prediction is designed. Its structure is as follows: Figure 2This method is the core of the dynamic update of the MCL-MAN method. By storing and reading and writing the key-value pairs associated with the degradation features, the storage and playback of key memories are realized, so that the performance degradation prediction model can maintain sufficient memory capacity in the continuously increasing tasks.

[0031] The MCL-MAN method proposed in this paper simulates the information learning and memory process between the "hippocampus-cerebral cortex" and designs a sparse experience replay process to solve the catastrophic forgetting problem of the performance degradation prediction model in dynamic updating. The specific process is as follows: (31) Hippocampus-Cortex Learning Process The hippocampal learning process is reflected in the MCL-MAN method as a model pre-training process. First, the encoder and decoder of the performance degradation memory extraction model are pre-trained in the newly added task Task1. The historical performance degradation HI data is fed into the two encoders, and the outputs of the two encoders are connected and input into the decoder to generate future HI trajectories. Then, the training controller decides which samples to write into the memory in Task 1. The training controller selects typical memory keys from the current task, simulating the process of information flowing from the hippocampus to the cerebral cortex. Through the pre-training stage, the HI trajectory in the current task is encoded as a degradation feature association key and written into the external memory module M1.

[0032] (32) Rapid eye movement memory replay process The memory replay process is reflected in the MCL-MAN method as a sparse experience replay continuous learning phase. In the new task Task-B, the memory extraction module in the pre-training phase is used to construct an external memory module M2. Its function is to avoid forgetting the performance degradation feature information learned in Task-A. In addition, an additional core set database is designed to store sparse samples of previous tasks (accounting for less than 1% of the historical data volume). When the current sequence order of the task exceeds 1, that is, when the existing historical information is memorized and the prediction model is dynamically updated, sparse experience replay is used to obtain the memory and partially replay some historical samples in the memory to prevent catastrophic forgetting of historical task information when training the prediction module. During the dynamic update of the model, the memory extraction module is used to further update the external memory module to M2 from the old memory M1. That is, M2 contains the historical degradation information memory in M1 and the newly added memory learned in Task2. Finally, the prediction module is trained based on the external memory M2 to achieve performance degradation prediction.

[0033] like Figure 1 As shown in Figure 2, the MCL-MAN method realizes the continuous learning performance degradation prediction of complex equipment through the joint dynamic update of multiple sub-models. Figure 2As shown in the figure, the dynamic update of the performance degradation prediction sub-model consists of two stages: pre-training and continuous learning. The core models in the two stages are the performance degradation memory extraction model based on the memory enhancement network and the continuous learning performance degradation prediction model based on the external memory storage. Next, the method of the dynamic update process of the multi-sub-model joint memory enhancement network and the two core models is introduced.

[0034] 3.1. Dynamic update process of multi-model joint memory enhancement network In actual industrial production, when complex system equipment is put into operation, the equipment is in a healthy operating state at the beginning of operation. The performance status health index HI of the equipment and its components is fitted through the performance status monitoring parameters. As the operation time of individual equipment or equipment batches increases, the performance status health index HI continues to change, forming a performance status historical data set When the performance degradation prediction task is started, first, based on the performance degradation prediction model, the performance degradation prediction model is initialized by random weights, and the degradation model is copied to Finally, the performance degradation prediction model is networked. During the model network training phase, the HI gradient is calculated. ,when Exceeding the activation threshold When the corresponding performance degradation prediction sub-model The model is set to a trainable state, and all trainable state sub-models are used to jointly participate in the equipment performance degradation prediction model training process.

[0035] After that, based on the task scenario division method, the performance status data is updated in different new data scenarios and a performance status degradation data set is formed. . In the data set The corresponding performance degradation prediction model dynamically updates Task m. When a device object performs the first performance degradation prediction, that is, when Task 1 is completed, based on the training data Training performance degradation prediction network. After that, whenever a new performance degradation prediction task appears, the MCL-MAN method dynamic update process is started. In combination with the performance degradation characteristics of complex system equipment, the present invention, based on the MCL-MAN method, specifically designs a performance degradation prediction process based on continuous learning of multi-model joint memory-enhanced network, which is sequentially encoder dynamic update, controller dynamic update and memory writing, and prediction model dynamic update.

[0036] (311) Encoder dynamic update The encoder dynamic update process is as follows Figure 3 As shown in the figure, the encoder structure follows the DP-GRUC model structure. In the dynamic update process of the encoder, firstly for the HI trajectory The historical data is divided into several time window HI trajectories by sliding window cutting, and the first half of the data in the sliding window is set as the historical HI trajectory, and the second half is the corresponding future HI trajectory. The gradients of the historical and future HI trajectories are calculated respectively. and . The historical HI trajectory gradient With future HI trajectory gradient Enter the history encoder separately With the future encoder In the process, historical coding updates and future coding updates are performed, and the output historical HI trajectory coding and future HI trajectory coding are combined into a degenerate feature associated key-value pair , based on the controller Write to the external memory module. After that, the decoder Decoding is performed, taking the current performance status indicator as the starting point, and extrapolating the future HI trajectory to achieve dynamic update of the encoder.

[0037] (312) Controller dynamic update and memory writing The controller dynamic update and memory writing process is as follows Figure 4 As shown, the historical HI trajectory gradient is input into the updated encoder module to generate the historical HI trajectory code, which is compared with the historical memory code in the external memory module to obtain the historical memory code read weight. After that, the read weights are arranged in order and the sorting result is assigned as the historical memory code index. The future memory code is paired with the historical HI trajectory code according to the index value to form a degradation feature associated key-value pair, which can be decoded by the updated decoder to output the predicted HI trajectory. All predicted HI trajectories are combined into a predicted HI trajectory set, and the cosine distance is calculated with the future HI trajectory in turn. The output error tolerance rate is calculated by error tolerance and fitted as a write probability based on the linear layer. In this process, when the new data is significantly different from the historical memory code, the distance value between the predicted HI trajectory and the future HI trajectory increases, indicating that the new data contains degradation feature information that is not included in the historical memory. In this scenario, the error tolerance rate will increase, thereby increasing the write probability, and the key-value pairs with differentiated degradation features will be written into the memory.

[0038] (313) Dynamic update of prediction model The dynamic update process of the prediction model is as follows Figure 5As shown. First, the historical HI trajectory gradient is input into the updated encoder module to generate the historical HI trajectory encoding. Then the historical HI trajectory encoding, historical memory encoding, and future memory encoding are respectively input into the linear layer to generate the memory query key value indicator. The addressing result is obtained by inputting the historical degradation trajectory encoding query key value and the historical memory encoding key value into the softmax layer, and then comparing it with the future memory encoding to generate the memory query attention result. , encoding the historical degradation trajectory and the attention result The combined pairing constructs the degradation feature associated key-value pairs, which are then input into the decoder for trajectory decoding, and finally generate future HI trajectories to achieve dynamic update of the performance degradation prediction model.

[0039] 3.2 Performance-Degraded Memory Extraction Model Based on Memory-Augmented Network The core of the performance degradation memory extraction module is the memory extraction controller, which has the function of controlling the reading and writing of memory key-value pairs in the external memory module. Based on the autoencoder composed of the encoder and decoder in the performance degradation prediction module, the input monitoring data is transcoded into prior knowledge. The main functions of the controller are divided into two parts: reading the degradation feature associated key-value pair memory samples in the external memory module and writing the memory key-value pair samples generated in the new task into the external memory.

[0040] (321) Degenerate feature association key-value pair memory reading The read operation searches for similar degradation features from the external storage module based on the historical HI trajectories that have been monitored, that is, the key-value pairs associated with the degradation features, and searches for the maximum possible HI trajectory of the device to be predicted in the existing memory. After the HI trajectory of the device to be predicted has a certain time series length, the monitored HI trajectory gradient Input History Encoder Generate key value , and measure its similarity with all key values ​​in the external memory by cosine distance.

[0041]

[0042] in, represents the L2 norm, Represents the similarity score of all memory locations. The higher the similarity score, the closer the device is to the memory location. The more similar the current performance degradation characteristics are. Based on the similarity score, the memory key-value pair to be read is selected and the current key-value pair is Embed the generated future key values ​​separately Combined into key-value pairs ,in The key-value pairs composed of are input into the decoder to generate HI tracks.

[0043]

[0044] in Representation decoder The trainable parameters in the module, represents the fully connected layer in the decoder.

[0045] (322) Degenerate feature association key-value pair memory writing The write operation is to determine whether the newly added task has generated degenerate features that are significantly different from the memory key-value pairs stored in the external memory, and add these new features into the memory to ensure the diversity and low redundancy of the storage memory, thereby improving the HI trajectory representation capability of the storage memory. The write operation mainly reconstructs the error With write probability Parameter control. Reconstruction error Determine whether the network prediction result is close to the benchmark value, and write the probability By comparing with the threshold, it is decided whether to write the current sample into the memory.

[0046] Generated based on memory reading process HI tracks On this basis, the reconstruction error of the generated trajectory is calculated using formula (2.3), and the trajectory with the minimum error is selected.

[0047] In the formula, is the indicator function, indicating that if the HI trajectory is predicted The distance to the true trajectory is less than the threshold , then the function output is 1. represents the time step of the future trajectory.

[0048] The controller consists of a linear layer and a sigmoid activation function layer. As shown in formula (2.5), the minimum reconstruction error is input into the controller to calculate the write probability . Train the controller to adaptively adjust the write probability , and through the loss function defined in formula (2.6) Update its parameters. When e→1, the prediction effect of the representation network is poor, so the write probability needs to be increased. .

[0049]

[0050] In the formula, Represents a linear layer The trainable parameter weights of Represents the sigmoid activation function.

[0051] 3.3. Continuous learning performance degradation prediction model based on external memory storage In the MCL-MAN method designed by the present invention, the prior knowledge learned in the historical prediction task is stored in the external memory module through the performance degradation memory extraction module. In the iterative process of the prediction architecture, the prior knowledge is defined as a key-value memory of the fitting relationship between the historical performance state of the device and the future performance state label. The memory is shared by the memory extraction module and the performance degradation prediction module. Unlike traditional neural Turing machines, the memory in the external memory designed in this paper is continuously updated with the iteration of the prediction task. In this structure, the amount of knowledge that can be stored in the external memory continues to increase, which can continuously improve the performance of the prediction model and the scalability of the task. After training iterations, the controller writes the typical memory key-value storage into the external memory module to minimize storage redundancy.

[0052] On this basis, the performance degradation prediction module queries relevant information from the external memory module through a multi-hop attention mechanism and predicts the future performance status of the device under the current task scenario. This mechanism is used to iteratively address and read task-related knowledge information from memory. This process is similar to the process of the human hippocampus learning to acquire new knowledge and further practicing during the rapid eye movement period. The future HI trajectory is generated by decoding the addressing results of each hop. In the memory extraction module, the pre-trained encoder is obtained , and outputs the embedding vector based on this. The read memory is converted through the multi-hop attention mechanism and combined with the embedding vector as the input of the decoder. The decoder composed of GRU and FC models inputs the future HI trajectory.

[0053] After pre-training performance degradation memory extraction with historical training data, the encoder The device HI trajectory fitting capability is available. Historical HI trajectory gradient Input encoder to generate historical feature key values Based on this, as shown in formula (2.7), using the linear function Generates query key values.

[0054]

[0055] In the formula, It is the first key value queried. Based on this, performance degradation prediction is achieved through four steps: iterative key value addressing, value reading, HI trajectory prediction and update query.

[0056] (331) Key-value addressing Addressing probability The calculation process is as follows:

[0057] In the formula, is the addressing probability, is the key-value mapping linear layer, It is the softmax normalization layer.

[0058] (332) Numerical reading The calculated addressing probability is used as the read weight to calculate the skip attention.

[0059] In the formula, is a linear layer for numerical mapping, is the number of memory key-value pairs in the memory.

[0060] (333) HI trajectory prediction First, the historical HI trajectory is embedded into the vector Attention results connected and then fed into the decoder Finally, the HI trajectory is generated ,

[0061] In the formula, For the decoder Module trainable parameters, is a fully connected layer. Represents a concatenation operation.

[0062] (334) Update query In each iteration, the previous query key value Attention results Combined together, it becomes the key value updated in the next layer ,

[0063] After that, repeat the above four steps until the memory query demand ends, using different query key values For each jump Then the performance degradation trajectory of the device is predicted through the future performance degradation trajectory generated by the decoder.

[0064] Before the continuous learning performance degradation prediction method, the embodiment of the present invention also includes: the implementation steps of the continuous learning performance degradation prediction method based on stability-plasticity balance, which include: according to the equipment operation task scenario, real-time acquisition of the equipment 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 equipment performance status data set to obtain a trained DP-GRUC model; based on the trained DP-GRUC model, predicting the performance degradation of the equipment to be tested in the target task scenario.

[0065] 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.

[0066] 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.

[0067] 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.

[0068] 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.

[0069] 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:

[0070] 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.

[0071] 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.

[0072] In an embodiment of the present invention, the training loss function is minimized as follows:

[0073] 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, represents parameter weight; Represents a model function.

[0074] Before the continuous learning performance degradation prediction method based on stability-plasticity balance, the embodiment of the present invention also includes a performance status evaluation method based on the spatiotemporal attention mechanism, which specifically includes: using the detection data sliding window reconstruction method to convert the equipment operation monitoring performance degradation data into three-dimensional time series data; 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; 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; based on the trained STAGCN model, the performance status of the equipment to be evaluated is evaluated.

[0075] 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.

[0076] 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.

[0077] In an embodiment of the present invention, the temporal attention matrix includes: Where E is the temporal attention matrix, is the input of the rth spatiotemporal network block in the three-dimensional time series data, is the length of the time dimension in the rth layer of the three-dimensional time series data, is the number of channels of the r-th layer input data in the three-dimensional time series data, These are the parameters of the temporal attention matrix to be learned. is the activation function.

[0078] In an embodiment of the present invention, the spatial attention matrix includes: Among them, S is the spatial attention matrix, is the input of the rth spatiotemporal network block in the three-dimensional time series data, is the length of the time dimension in the rth layer of the three-dimensional time series data, is the number of channels of the r-th layer input data in the three-dimensional time series data, These are the parameters of the spatial attention matrix to be learned. is the activation function.

[0079] 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: 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 , and They represent different collection points respectively.

[0080] 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; The number of sensor acquisition channels is:

[0081] in, Indicates the number of acquisition channels; Indicates The number of sensors at each collection point; Indicates the number of sensor locations.

[0082] 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.

[0083] In summary, the present invention aims to solve the problem that under the requirements of differential performance degradation prediction tasks, the memory and consolidation capabilities of degradation feature information in performance degradation prediction tasks are limited due to the limitations of the size of the artificial neural network structure and the number of neurons, thereby affecting the ability to continue learning. The present invention is inspired by the working principle of learning and memory in the biological mammalian brain, combined with the "hippocampus-cerebral cortex" theory in the complementary learning system, and overcomes the limitations of the learning and memory capabilities of artificial neural networks through the structure of external memory storage. In the task of predicting performance degradation with timing and differences of complex system equipment, a multi-model joint performance degradation prediction architecture is designed. This architecture simulates the sub-brain regions in the brain based on the performance degradation prediction sub-model, links the sub-models to each other, and forms a performance degradation prediction model network. Based on the HI degradation rate, the corresponding brain area is activated to achieve the joint prediction of HI indicators of multiple components of complex system equipment. In addition, in order to solve the limitation of learning and memory upper limit of artificial neural network performance degradation prediction model on continuous learning ability, based on the "hippocampus-cerebral cortex" theory, a performance degradation memory extraction model based on memory enhancement network and a continuous learning performance degradation prediction model based on external memory storage were designed on the basis of multi-model joint architecture. By storing and replaying historical degradation feature information through external memory, continuous learning of differentiated performance degradation prediction tasks is realized, which greatly improves the continuous learning performance of the performance degradation prediction model.

[0084] 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 those 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 method for predicting performance degradation of continuous learning, characterized in that: include: Constructing a performance degradation prediction sub-model, and based on the performance degradation prediction sub-model, constructing a performance degradation prediction model network based on multi-model joint memory enhanced continuous learning MCL-MAN; Acquire performance degradation data sets of different batches and different models of equipment under different task scenarios, and dynamically update the MCL-MAN-based performance degradation prediction model network using the performance degradation data sets of different batches and different models of equipment under different task scenarios to obtain a dynamically updated MCL-MAN-based performance degradation prediction model network; The historical health index HI trajectory gradient of the device to be predicted is obtained, and the future performance degradation trajectory of the device to be predicted is predicted by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated MCL-MAN-based performance degradation prediction model network.

2. The method according to claim 1, characterized in that By inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated MCL-MAN-based performance degradation prediction model network, predicting the future performance degradation trajectory of the device to be predicted includes: The historical feature key value is obtained by inputting the historical health index HI trajectory gradient of the device to be predicted into the encoder in the performance degradation prediction model network based on MCL-MAN after dynamic update; Generate a query key value using the historical feature key value; Based on the query key value, the future performance degradation trajectory of the device to be predicted is predicted by iterating key value addressing, value reading, and HI trajectory prediction set update query steps in sequence.

3. The method according to claim 2, characterized in that Based on the query key value, predicting the future performance degradation trajectory of the device to be predicted includes: Based on the query key value, calculating the addressing probability; Using the addressing probability as a reading weight, calculating the skip attention; Generate a health indicator HI trajectory through the step-skipping attention and the historical feature key value; In each iteration, the query key of the previous task is combined with the skip-step attention as the query key of the current task; Repeat the above steps until the memory query demand ends, use different query key values ​​to query each jump attention, and predict the future performance degradation trajectory of the device to be predicted.

4. The method according to claim 1, characterized in that: The MCL-MAN-based performance degradation prediction model networking realizes continuous learning performance degradation prediction of complex equipment based on dynamic update of each performance degradation prediction sub-model, wherein the dynamic update of each performance degradation prediction sub-model includes a pre-training stage and a continuous learning stage.

5. The method according to claim 4, characterized in that The pre-training stage includes dynamic updating of the encoder, dynamic updating of the controller and dynamic updating of the prediction model.

6. The method according to claim 5, characterized in that The encoder dynamic update includes: The historical data is divided into several time window HI trajectories through sliding window cutting, and 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; Based on the historical HI trajectory and the future HI trajectory, the historical HI trajectory gradient and the future HI trajectory gradient are respectively calculated, and the historical HI trajectory gradient is input into the historical encoder for historical coding update, and the historical HI trajectory coding is output; at the same time, the future HI trajectory gradient is input into the future encoder for future coding update, and the future HI trajectory coding is output; The output historical HI trajectory codes and future HI trajectory codes are combined into degradation feature associated key-value pairs, which are written into an external memory module based on the controller; The degradation feature associated key-value pairs are decoded by a decoder, and the future HI trajectory is extrapolated with the current performance status indicator as the starting point, so as to realize the dynamic update of the encoder.

7. The method according to claim 6, characterized in that The controller dynamic update includes: Input the historical HI trajectory gradient into the updated encoder module to generate the historical HI trajectory code, compare it with the historical memory code in the external memory module, and obtain the historical memory code reading weight; Arrange the read weights in order and assign the sorting result as the historical memory encoding index. Pair the future memory encoding with the historical HI trajectory encoding according to the index value to form a degenerate feature associated key-value pair, decode it through the updated decoder, and output the predicted HI trajectory; Combine all predicted HI trajectories into a predicted HI trajectory set, and calculate the cosine distance between the predicted HI trajectory set and the future HI trajectory set in turn, calculate the output error tolerance rate through the error tolerance, and fit it to the write probability based on the linear layer; The write probability is used to find out the key-value pairs associated with the differentiated degradation characteristics, and the key-value pairs associated with the differentiated degradation characteristics are written into the external memory module.

8. The method according to claim 7, characterized in that The prediction model dynamic update includes: Input the historical HI trajectory gradient into the updated encoder module to generate the historical HI trajectory code, and input the historical HI trajectory code, historical memory code, and future memory code into the linear layer to generate the memory query key value indicator; By inputting the historical degradation trajectory encoding query key value and the historical memory encoding key value into the softmax layer to obtain the addressing result, and comparing the addressing result with the future memory encoding to generate the memory query attention result; By pairing the historical degradation trajectory encoding with the attention result combination, a degradation feature associated key-value pair is constructed, and the degradation feature associated key-value pair is input into the decoder for trajectory decoding to generate future HI trajectories, so as to realize the dynamic update of the performance degradation prediction model.

9. A continuous learning performance degradation prediction device, characterized in that: include: A construction module is used to construct a performance degradation prediction sub-model, and based on the performance degradation prediction sub-model, a performance degradation prediction model network based on multi-model joint memory enhanced continuous learning MCL-MAN is constructed; A training module, used to obtain performance degradation data sets of different batches and different models of equipment under different mission scenarios, and dynamically update the MCL-MAN-based performance degradation prediction model network using the performance degradation data sets of different batches and different models of equipment under different mission scenarios to obtain a dynamically updated MCL-MAN-based performance degradation prediction model network; The prediction module is used to obtain the historical health index HI trajectory gradient of the device to be predicted, and predict the future performance degradation trajectory of the device to be predicted by inputting the historical health index HI trajectory gradient of the device to be predicted into the dynamically updated MCL-MAN-based performance degradation prediction model network.

10. 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 continuous learning performance degradation prediction method described in any one of claims 1 to 8 are implemented.

Images