Degradation prediction method based on random process and deep learning interaction, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of equipment degradation prediction, and particularly discloses a degradation prediction method based on a random process and deep learning interaction, electronic equipment and a storage medium, and the method comprises the following steps: S1, constructing a neural network based on a multi-scale residual module, a bidirectional long-short-term memory network and a cross attention mechanism; s2, constructing a Wiener degradation model with a drift coefficient in normal distribution, wherein the Wiener degradation model is used for reflecting a current degradation state; s3, adopting a maximum likelihood estimation method to obtain a drift coefficient in the Wiener degradation model; s4, obtaining two parameters of drift coefficient normal distribution according to the Bayesian theorem; s5, performing interaction between the Wiener degradation model and the neural network, updating the two parameters in the step S4 by using data output by the neural network, and obtaining a drift coefficient through the step S3; and S6, setting a failure threshold value, and obtaining the residual life by using the neural network.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment degradation prediction, and specifically relates to a degradation prediction method based on random process and deep learning interaction, an electronic device and a storage medium. Background Art

[0002] In key fields such as industry, aerospace and energy, the reliability and life prediction of equipment and systems play a vital role in ensuring safety, reducing maintenance costs and improving operational efficiency.

[0003] In the life prediction process, the Wiener process is used as a degradation modeling tool because it can describe the randomness in degradation behavior; however, in many engineering systems, the degradation rate of the equipment (i.e., the drift parameter) is not constant, but fluctuates according to changes in environmental conditions, operating modes, or internal states. The traditional Wiener process (since the drift parameter is a constant value) is difficult to reflect the degradation path under such fluctuations, that is, it is unable to handle the uncertainty of the drift coefficient.

[0004] In addition, the traditional Wiener process is usually modeled offline based on historical data. However, in application scenarios without real-time monitoring, due to the lack of degradation data collection, this static method cannot update the distribution of the drift coefficient through real-time data, making it difficult to reflect the current degradation state, which is not conducive to the formulation of subsequent maintenance strategies.

[0005] Therefore, the life prediction method based on traditional degradation modeling has difficulty in dealing with the uncertainty of drift coefficient due to changes in environmental conditions, operating modes or internal states. It is also unable to update degradation data in real time in application scenarios without real-time monitoring, making it difficult to reflect the current degradation state of the system, which will affect the reliability assessment and life prediction of the system. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a degradation prediction method, an electronic device and a storage medium based on the interaction of random process and deep learning.

[0007] Based on the above purpose, the present invention is achieved through the following technical solutions:

[0008] A first aspect of the present invention provides a degradation prediction method based on random process and deep learning interaction, comprising the following steps:

[0009] S1. Construct a neural network based on multi-scale residual modules, bidirectional long short-term memory networks and cross-attention mechanisms.

[0010] S2. Construct a Wiener degradation model with a drift coefficient of normal distribution to reflect the current degradation state.

[0011] S3. Use the maximum likelihood estimation method to obtain the drift coefficient in the Wiener degradation model.

[0012] S4. According to Bayes' theorem, two parameters of the normal distribution of the drift coefficient are obtained.

[0013] S5, the interaction between the Wiener degradation model and the neural network, uses the data output by the neural network to update the two parameters in S4, and then obtains the drift coefficient through S3.

[0014] S6. Set the failure threshold and use the neural network to obtain the remaining life.

[0015] According to the above degradation prediction method based on random process and deep learning interaction, preferably, in step S1, the neural network inputs t i The historical degradation value X(t i ) predict the future i+1 The degradation value X(t i+1 );Can be based on t i+1 The degradation value predicted at time t is used as the approximate true degradation value to predict i+2 The degradation value at the moment, and so on; the neural network passes through the multi-scale residual module, the input data is X1 n×1 After transformation, a tensor of shape (B, n, 1) is obtained, where B is the batch size and n is the length of the time series; the steps of input data transformation include: first layer 1D convolution transformation; splicing operation; second layer 1D convolution transformation; splicing and 1x1 convolution; maximum pooling and residual connection.

[0016] The first layer of 1D convolution transformation is achieved through two 1D convolution kernels of different scales, that is, the convolution kernel sizes are 3 and 5 respectively, with padding operation, so that the output length remains unchanged. The output expression is:

[0017]

[0018] in, represents the weights of the two convolution kernels; X1 represents the input data.

[0019] The concatenation operation is to concatenate the two convolution results in the channel dimension, and its expression is:

[0020] C1=Concat(S1,P1).

[0021] The second layer 1D convolution transformation is a second multi-scale convolution operation, and its expression is:

[0022]

[0023] The concatenation and 1x1 convolution concatenate the results of the second convolution in the channel dimension, and then transform the number of channels into the number of input channels through 1x1 convolution. The expression is:

[0024] C2 = Concat(S2, P2);

[0025] S′=Conv1d(C2,W 1×1 ).

[0026] The maximum pooling and residual connection performs a maximum pooling operation on the input data, and its expression is:

[0027]

[0028] Then add the pooled input and the convolutional output to form a residual connection:

[0029]

[0030] Among them, S1 and P1 represent the output results of the first convolution of the input data with a convolution kernel size of 3 and 5 respectively; S2 and P2 represent the output results of the second convolution of the input data with a convolution kernel size of 3 and 5 respectively; C1 and C2 represent the first and second S i and P i (i=1, 2) is the result of concatenation in the channel dimension; X1 pool Represents the input after pooling; S' represents the output result after 1x1 convolution transforms the number of channels into the number of input channels.

[0031] According to the above degradation prediction method based on random process and deep learning interaction, preferably,

[0032] The most commonly used LSTM network is concisely called the vanilla LSTM, and the central idea is that three nonlinear gating units (i.e., input, forget, and output gates) can regulate the flow of information into and out of the unit. LSTM is a variant of recurrent neural network (RNN) that can solve the gradient vanishing problem. For a basic LSTM unit, there are three inputs and two outputs.

[0033]

[0034] Where σ represents the sigmoid function; h t and x t Represent the hidden state and current input respectively; i t , o t and f t Respectively represent input gate, output gate and forget gate; W i , W o and W fRespectively represent the weights of the input gate, output gate and forget gate; b i 、b o and b f Respectively represent the bias of the input gate, output gate and forget gate; C t and Represent memory cells and candidate memory cells respectively.

[0035] In step S1, the neural network adopts a bidirectional long short-term memory network to improve sequence learning by considering both past and future contexts. It contains two hidden layers: one layer processes data forward at each time step, and the other layer processes data backward;

[0036]

[0037] in, and Represent the forward and reverse hidden states respectively; y t Represents the final output; and Represents the forward and reverse weights respectively; b y Represents the bias parameter.

[0038] According to the above degradation prediction method based on the interaction of random process and deep learning, preferably, in step S1, the neural network uses a multi-head attention mechanism to filter important information from the input features and assigns different weights to them according to their importance. The model can focus on information with greater weights; the scaled dot product attention mechanism is used to obtain the corresponding weights of the query and key matrices through dot multiplication, and then the Softmax function is used for normalization. The attention is obtained by weighted summation, and the expressions are:

[0039]

[0040] Among them, Q represents the query matrix; K represents the keyword matrix; V represents the value matrix; A f represents the input matrix; W Q , W K and W V Represent the weight matrices of Q, K and V respectively; d k Represents the dimensions of Q, K, and V.

[0041] The multi-head attention mechanism is based on the combination of several self-attention mechanisms. The self-attention mechanism pays attention to the details of the information according to the target input. The multi-head attention mechanism is used to realize the simultaneous attention to different parameters of multiple sensor signals, and the obtained results are spliced ​​to obtain the final attention. Its expression is:

[0042]

[0043] in, and denote the weight matrices of Q, K and V in the i-th attention head respectively; W denotes the weight matrix of the multi-head attention mechanism.

[0044] The neural network uses a cross-attention mechanism to allow information interaction between two different feature representations; its two features come from the multi-scale residual module (MSRB module) and the bidirectional long short-term memory network-multi-head attention mechanism module (BILSTM-MHA module); the cross-attention obtains two cross results by taking the output of the multi-scale residual module as the query Query, key Key and value Value, and the output of the bidirectional long short-term memory network-multi-head attention mechanism module as the query Query, key Key and value Value, and the two cross results are concatenated and then passed through the linear layer to obtain the final output.

[0045] According to the above degradation prediction method based on random process and deep learning interaction, preferably,

[0046] In step S2, since the initial drift coefficient μ of the Wiener degradation model is unknown, it is assumed to be a normal distribution N(μ0,τ 2 ), the diffusion coefficient σ is a known constant, and the expression of the Wiener degradation model is:

[0047] X(t)=X(0)+μt+σB(t),μ~N(μ0,τ 2 );

[0048] Among them, X(0) represents the initial degradation state; t represents time; X(t) represents the degradation state at time t; μ represents the average rate of degradation, that is, the drift coefficient; σ represents the random fluctuation amplitude, that is, the diffusion coefficient; B(t) is the standard Wiener process, which represents randomness.

[0049] According to the above-mentioned degradation prediction method based on random process and deep learning interaction, preferably, in step S3, according to the m groups of historical degradation data {(t i ,X(t i ))}(where i=1,2,…,m), the historical drift coefficient is obtained by calculation from the historical degradation data; the influence of the error term σB(t) is small and can be ignored; according to the observed values ​​of the historical drift coefficients {μ1,μ2,…,μ m}Get the initial mean μ0 and variance τ of the drift coefficient 2 , the expression is:

[0050]

[0051] Among them, {μ1,μ2,…,μ m} are the observed values ​​of m historical drift coefficients.

[0052] Since at each time point t i The degradation value X(t i ) obeys the normal distribution N(X0+μt i ,σ 2 t i ), therefore, the observed value X(t i ) is:

[0053]

[0054] The likelihood function of the observed value is:

[0055]

[0056] Then, the log-likelihood function is:

[0057]

[0058] Since μ follows a normal distribution N(μ0,τ 2 ), combined with Bayesian theory, the joint log-likelihood function is:

[0059]

[0060] Expand to get:

[0061]

[0062] Then maximize the joint log-likelihood function to get t m The a posteriori estimate of the drift coefficient μ at the moment is expressed as:

[0063]

[0064] Among them, X(t i ) indicates t i The degradation value at the moment; X0 represents the initial degradation level of the system at time t=0; σ represents the random fluctuation amplitude, that is, the diffusion coefficient; μ0 represents the mean of the normal distribution of the drift coefficient; τ 2 Represents the variance of the normal distribution of the drift coefficient.

[0065] According to the above-mentioned degradation prediction method based on random process and deep learning interaction, preferably, in step S4, according to Bayes' theorem, the posterior distribution of the drift coefficient is the product of the likelihood function and the normal distribution of the initial drift coefficient; the distribution of the initial drift parameter and the likelihood function are both in the form of normal distribution, and due to conjugation, the posterior distribution is also a normal distribution; t m The posterior mean μ of the drift coefficient at the moment post and variance τ post 2 The expressions are:

[0066]

[0067] Among them, X(t i ) indicates t i The degradation value at the moment; X0 represents the initial degradation level of the system at time t=0; σ represents the random fluctuation amplitude, that is, the diffusion coefficient; μ0 represents the mean of the normal distribution of the drift coefficient; τ 2 Represents the variance of the normal distribution of the drift coefficient.

[0068] By introducing the normal distribution of the drift coefficient, the randomness and volatility in the degradation process can be more realistically reflected to adapt to various degradation modes, solving the problem of drift coefficient uncertainty that is difficult to handle due to changes in environmental conditions, operating modes or internal states, which is also helpful for the subsequent formulation of maintenance strategies.

[0069] According to the above degradation prediction method based on random process and deep learning interaction, preferably, in step S5, the tth m The historical degradation data at time t is taken as input and the output is m+1 The degradation value at time t m The mean and variance of the posterior distribution of the drift coefficient at time t are taken as m+1 The mean and variance of the prior distribution of the drift coefficient at time, the known diffusion coefficient and the degradation data {(t i ,X(t i ))}, where i = 1, 2, ..., m + 1; get t m+1 The drift coefficient at time.

[0070] t m+1 The mean and variance of the posterior distribution of the drift coefficient at time t are calculated respectively; m+1 The mean and variance of the posterior distribution of the drift coefficient at time t are taken as m+2 The mean and variance of the prior distribution of the drift coefficient at time t, and the known diffusion coefficient and the degradation data updated by the neural network {(t i ,X(t i ))}, where i =

[0071] 1,2,…,m+2), used to update t m+2 The drift coefficient at time.

[0072] t m+2 The mean and variance of the posterior distribution of the drift coefficient at time t are used as m+3 The mean and variance of the prior distribution of the drift coefficient at the moment; repeat the above steps to update the degradation data in the application scenario without real-time monitoring in real time to reflect the problem of the current degradation state of the system.

[0073] In step S6, a failure threshold is set. When the degradation value output by the neural network reaches the failure threshold, the difference between the corresponding failure time and the initial time is the remaining life.

[0074] A second aspect of the present invention provides an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, any step of the degradation prediction method based on the interaction of random process and deep learning as described in the first aspect is implemented.

[0075] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a computer processor, the computer program implements any step in the degradation prediction method based on the interaction of random processes and deep learning as described in the first aspect.

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] 1. The present invention is based on a Wiener degradation model in which the drift coefficient obeys a normal distribution. It solves the problem of uncertainty in the drift coefficient that is difficult to handle due to changes in environmental conditions, operating modes or internal states, and more realistically reflects the randomness and volatility in the equipment degradation process to adapt to various degradation modes.

[0078] 2. The neural network of the present invention is based on a multi-scale residual module, a bidirectional long short-term memory network and a cross-attention mechanism, which can capture the relationship between the previous and subsequent dependencies in the time series, can predict the degradation value after the next cycle based on the degradation value after the previous cycle, and can update the degradation data in real time in application scenarios without real-time monitoring, thereby accurately reflecting the current degradation state of the system.

[0079] 3. The present invention can not only predict the future degradation process through the interaction between the Wiener degradation model and the deep learning neural network, but also update the posterior distribution of the drift coefficient in the Wiener degradation model in real time according to the future degradation prediction value, thereby solving the problem that the degradation data cannot be updated in real time to reflect the current degradation status of the system in the absence of real-time monitoring application scenarios, and is helpful for the subsequent maintenance and management of the system.

[0080] 4. The present invention adopts a neural network based on a multi-scale residual module, a bidirectional long short-term memory network and a cross-attention mechanism, which can solve the problem that degradation data cannot be updated in real time to reflect the current degradation state of the system in application scenarios without real-time monitoring, thereby helping to improve the accuracy of remaining life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a flow chart of the present invention in Example 1.

[0082] Figure 2 is a schematic diagram of the structure of the neural network in Example 1;

[0083] Figure 3 This is a schematic diagram of the existing LSTM structure;

[0084] Figure 4 Schematic diagram of the BiLSTM structure of the present invention in Example 1;

[0085] Figure 5 is a schematic diagram of the life prediction result in Example 1;

[0086] Figure 6-14 It is a schematic diagram comparing the life prediction results in Example 1. DETAILED DESCRIPTION

[0087] The present invention is further described in detail below through specific examples, but the scope of the present invention is not limited thereto.

[0088] Example 1

[0089] A degradation prediction method based on the interaction of random process and deep learning, the flow chart is as follows Figure 1 As shown, a battery is used for case analysis to collect battery capacity decay data, which usually includes battery capacity data that changes over time (or number of charge and discharge cycles); the data can be obtained by regularly measuring the capacity of the battery (such as measuring the actual discharge capacity under specific discharge conditions).

[0090] 1. Experimental data

[0091] This embodiment uses four batteries, B0005, B0006, B0007 and B00018, from the NASA battery data set as samples for experiments. Table 1 is the basic data of the life experiment of these four batteries.

[0092] Table 1-Basic data of life test of four batteries

[0093]

[0094] Since battery 7 has not reached the end of its life after the cycle, that is, the battery capacity has not dropped to 1.4 Ahr, the battery capacity data of batteries 5, 6, and 18 are used as experimental objects in this embodiment.

[0095] 2. Build a neural network

[0096] The neural network is based on the multi-scale residual module (MSRB), the bidirectional long short-term memory network (BiLSTM) and the cross-attention mechanism (Cross-Attention), such as Figure 2 shown.

[0097] This embodiment uses battery capacity data as the input of the neural network and predicts capacity as the output. Specifically, assuming that the kth cycle is the prediction starting point, the designed neural network can predict the predicted capacity value of the k+1th cycle, as shown in the following formula:

[0098]

[0099] The updated value of the battery capacity can be obtained by the following formula:

[0100]

[0101] Among them, q k-L+1 ,q k-L+2 ,…,q k-1 ,q k Respectively represent the battery capacity of the first L cycles;

[0102] They represent v predicted capacity values ​​respectively.

[0103] The specific process of the neural network is as follows: In the multi-scale residual module (MSRB), the input data is X1 n×1 , after transformation, a tensor of shape (B, n, 1) is obtained, where B represents the batch size and n represents the length of the time series.

[0104] The steps for inputting battery capacity data transformation include:

[0105] The first layer of 1D convolution transformation:

[0106] After two 1D convolution kernels (Conv1d) of different scales, that is, the convolution kernel sizes are 3 and 5 respectively, with padding operation, so that the output length remains unchanged, the output is:

[0107]

[0108] in, They represent the weights of the two convolution kernels respectively; X1 represents the input data.

[0109] Splicing operation:

[0110] Concatenate the two convolution results in the channel dimension:

[0111] C1=Concat(S1,P1);

[0112] The second layer of 1D convolution transformation:

[0113] After the second multi-scale convolution operation:

[0114]

[0115]

[0116] Concatenation and 1x1 convolution:

[0117] The results of the second convolution are concatenated in the channel dimension, and then the number of channels is transformed into the number of input channels through 1x1 convolution:

[0118] C2 = Concat(S2, P2);

[0119] S′=Conv1d(C2,W 1×1 );

[0120] Max pooling and residual connection:

[0121] Perform a maximum pooling operation on the input battery capacity data:

[0122]

[0123] Add the pooled input and the convolutional output to form a residual connection:

[0124]

[0125] Among them, S1 and P1 represent the battery capacity output results of the first convolution of the battery capacity input data with a convolution kernel size of 3 and 5 respectively; S2 and P2 represent the battery capacity output results of the second convolution of the battery capacity input data with a convolution kernel size of 3 and 5 respectively; C1 and C2 represent the first and second S i and P i (i=1, 2) is the result of concatenation in the channel dimension; X1 pool is the battery capacity input after pooling; S' is the battery capacity output result after the number of channels is transformed into the number of input channels through 1x1 convolution.

[0126] Bidirectional long short-term memory network (BiLSTM);

[0127] The most commonly used LSTM network is concisely called the vanilla LSTM. The central idea is that three nonlinear gating units (i.e., input, forget, and output gates) can regulate the flow of information into and out of the unit. LSTM is a variant of recurrent neural network (RNN) that can solve the gradient vanishing problem. For a basic LSTM unit, there are three inputs and two outputs, such as Figure 3 shown.

[0128]

[0129] Where σ represents the sigmoid function; h t and x tRepresent the hidden state and current input respectively; i t , o t and f t Respectively represent input gate, output gate and forget gate; W i , W o and W f Respectively represent the weights of the input gate, output gate and forget gate; b i 、b o and b f Respectively represent the bias of the input gate, output gate and forget gate; C t and Represent memory cells and candidate memory cells respectively.

[0130] Unlike traditional LSTM, the BiLSTM in this embodiment improves sequence learning by considering both past and future contexts, and can understand complex relationships more carefully and enhance feature extraction. It contains two hidden layers: one layer processes data forward at each time step, and the other layer processes data backward; its structure is as follows: Figure 4 shown.

[0131]

[0132] in, and Represent the forward and reverse hidden states respectively; y t Represents the final output; and Represents the forward and reverse weights respectively; b y Represents the bias parameter.

[0133] Multi-head attention mechanism;

[0134] The multi-head attention mechanism filters important information from the input features and assigns different weights to them according to their importance; therefore, the model can focus on information with greater weights; this embodiment uses the scaled dot product attention mechanism, and its specific steps are: first, the corresponding weights of the query and key matrices are obtained by dot multiplication, and then normalized using the Softmax function; attention is obtained by weighted summation, and the expressions are:

[0135]

[0136] Among them, Q represents the query matrix; K represents the keyword matrix; V represents the value matrix; A f represents the input matrix; W Q , W K and W V Represent the weight matrices of Q, K and V respectively; d k Represents the dimensions of Q, K, and V.

[0137] The multi-head attention mechanism is based on the combination of several self-attention mechanisms. The self-attention mechanism focuses on the details of the information according to the target input. For multiple sensor signals, the multi-head attention mechanism is used to achieve simultaneous attention to different parameters. Finally, the results are spliced ​​to obtain the final attention. Its expression is:

[0138]

[0139] in, and denote the weight matrices of Q, K and V in the i-th attention head respectively; W denotes the weight matrix of the multi-head attention mechanism.

[0140] Cross-Attention Mechanism;

[0141] Cross-Attention is a mechanism used to allow information interaction between two different feature representations. The two features come from the multi-scale residual module (MSRB module) and the bidirectional long short-term memory network-multi-head attention mechanism module (BILSTM-MHA module). The calculation formula of cross-attention is similar to that of ordinary self-attention, but the input query, key, and value come from different features.

[0142] The output of the multi-scale residual module is used as the query, key and value, and the output of the bidirectional long short-term memory network-multi-head attention mechanism module is used as the query, key and value to obtain two cross results. The two results are concatenated and passed through the linear layer to obtain the final output.

[0143] 3. Wiener degradation model;

[0144] The initial drift coefficient μ of the Wiener degradation model is unknown. Assume that the drift coefficient is normally distributed N(μ0,τ 2 ), the diffusion coefficient σ is a known constant, and its expression is:

[0145] X(t)=X(0)+μt+σB(t),μ~N(μ0,τ 2 );

[0146] Among them, X(0) represents the initial capacity of the battery, which is generally set to 100% or the nominal capacity of the battery; t represents the number of charge and discharge cycles; X(t) represents the battery capacity after the tth cycle; μ represents the average rate of capacity decay, and its unit is the capacity change per cycle; σ represents the random fluctuation amplitude based on the capacity decay after each cycle; B(t) is the standard Wiener process, which represents randomness.

[0147] First, the Wiener degradation model (the error term σB(t) has little effect and can be ignored) is used to fit each independent historical capacity data {(t i ,X(t i ))} (where i = 1, 2, ..., m), and obtain a set of drift coefficient observations {μ1, μ2, ..., μ m}; and then calculate t m The initial mean and variance of the normal distribution of the drift coefficient after the cycle are expressed as:

[0148]

[0149] The calculated initial mean and variance, the known diffusion coefficient, and the cycle number and capacity data {(t i ,X(t i ))}(where i=1,2,…,m) is substituted into the two formulas of the posterior distribution of the drift coefficient, and we get t m The mean and variance of the posterior distribution of the drift coefficient after cycles.

[0150]

[0151] At the same time, t can also be calculated from the above data m The drift coefficient after the cycle is calculated as follows:

[0152]

[0153] 4. Interaction between Wiener Degradation Model and Neural Network

[0154] The t m The historical battery capacity after the tth cycle is input into the neural network, and the output is m+1 The capacity after the first cycle; then t m The mean and variance of the posterior distribution of the drift coefficient after the cycle are taken as t m+1 The mean and variance of the prior distribution of the drift coefficient after cycles, the known diffusion coefficient and capacity data {(t i ,X(t i ))}(where i=1,2,…,m+1) is substituted into the following formula to calculate t m+1 Drift coefficient after cycles.

[0155]

[0156] In addition, t m+1 The mean and variance of the posterior distribution of the drift coefficient after cycles can also be calculated by the following two formulas:

[0157]

[0158] t m+1 The mean and variance of the posterior distribution of the drift coefficient after the cycle are taken as t m+2 The mean and variance of the prior distribution of the drift coefficient after 10 cycles are compared with the known diffusion coefficient and capacity data (t i ,X(t i ))} (where i = 1, 2, ..., m + 2) are used together to update t m+2 Drift coefficient after cycles; t m+2 The mean and variance of the posterior distribution of the drift coefficient after cycles are used as t m+3 The mean and variance of the prior distribution of the drift coefficient after cycles.

[0159] Repeating the above steps can achieve degradation prediction without real-time status monitoring. Due to the continuous updating of battery capacity data, the degradation trend described by the Wiener degradation model is more in line with the actual battery life. In addition, the estimated capacity change for each cycle (i.e., drift coefficient) is helpful for the subsequent formulation of maintenance strategies.

[0160] When the battery capacity decays to 70% of its nominal capacity, it is considered that the battery has reached the end of its life; when the predicted capacity output by the neural network reaches the capacity threshold, a final cycle number t is obtained. c , minus the number of cycles t at the starting point m , which is the remaining life RUL. The calculation formula for the remaining life is:

[0161] RUL=t c -t m .

[0162] 5. Experimental results

[0163] like Figure 5 As shown in FIG. 1 , the neural network of this embodiment performs well in predicting the capacity value. Figure 6-14 As shown, compared with other neural networks, the capacity value predicted by this embodiment and the neural network based on multi-scale residual module, bidirectional long short-term memory network and cross attention mechanism is closer to the actual capacity value and can better reflect the degradation state of the battery.

[0164] The interaction of the above-mentioned statistical model and deep learning methods enables degradation prediction without real-time status monitoring. Due to the continuous updating of battery capacity data, the degradation trend described by the Wiener degradation model is more in line with the actual situation of battery life, and therefore can provide more detailed information for equipment management and maintenance. If the degradation process shows an accelerating trend, it may be necessary to perform preventive maintenance in advance or check the equipment status more frequently. In addition, the degradation rate (i.e., drift coefficient) can be used to decide when to perform preventive maintenance, thereby avoiding unexpected failures or extending the service life of the equipment.

[0165] Example 2

[0166] An electronic device comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, any step in a degradation prediction method based on random process and deep learning interaction as described in Example 1 is implemented.

[0167] The hardware of the electronic device in this embodiment also includes a GPU, a display buffer memory, a RAMD / A converter, and a heat sink that cooperate with the processor; the GPU is responsible for processing the graphic display of the electronic device, providing image rendering and acceleration functions, and using its parallel computing advantages to accelerate the processing of large-scale data-intensive tasks such as deep learning training and scientific computing.

[0168] Further, the degradation prediction method process based on random process and deep learning interaction described in Example 1 can be implemented as a computer software program. For example, this embodiment includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method. In such an embodiment, the computer program can be downloaded and installed from a network, and / or installed from a removable medium. When the computer program is executed by a processor, the above functions defined in the method of the present application are executed.

[0169] Example 3

[0170] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step in a degradation prediction method based on the interaction of a random process and deep learning as described in Example 1.

[0171] The computer-readable medium described in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0172] The computer program code for performing the operation of the present application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages, such as python, C++, and also conventional procedural programming languages ​​or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer, partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., using an Internet service provider to connect through the Internet).

[0173] The computer-readable storage medium of this embodiment can be accelerated by hardware such as a GPU, and the parallel computing advantages of the GPU can be used to accelerate the processing of any step in a degradation prediction method based on the interaction of random processes and deep learning as described in Example 1.

[0174] In summary, the present invention effectively overcomes the deficiencies in the prior art and has a high industrial utilization value. The above embodiments serve to illustrate the substantive content of the present invention, but do not limit the protection scope of the present invention. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and protection scope of the technical solution of the present invention.

[0175] The above embodiments are specific implementation methods of the present invention, but the implementation methods of the present invention are not limited to the above embodiments. Any other combination, change, modification, substitution and simplification that does not exceed the design concept of the present invention shall fall within the protection scope of the present invention.

Claims

1. A degradation prediction method based on the interaction of random process and deep learning, characterized in that: The following steps are involved: S1. Construct a neural network based on multi-scale residual modules, bidirectional long short-term memory networks and cross-attention mechanisms; S2, constructing a Wiener degradation model with a drift coefficient of normal distribution to reflect the current degradation state; S3, using the maximum likelihood estimation method to obtain the drift coefficient in the Wiener degradation model; S4. Obtain two parameters of the normal distribution of the drift coefficient according to Bayes' theorem; S5, the interaction between the Wiener degradation model and the neural network, uses the data output by the neural network to update the two parameters in S4, and then obtains the drift coefficient through S3; S6. Set the failure threshold and use the neural network to obtain the remaining life.

2. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S1, the neural network is fed with input data X1 into the multi-scale residual module. n×1 , the transformation is obtained into a tensor of shape (B, n, 1), where B is the batch size and n is the length of the time series; the steps of input data transformation include: first layer 1D convolution transformation; splicing operation; second layer 1D convolution transformation; splicing and 1x1 convolution; maximum pooling and residual connection; The first layer of 1D convolution transformation is achieved through two 1D convolution kernels of different scales, that is, the convolution kernel sizes are 3 and 5 respectively, with padding operation, so that the output length remains unchanged. The output expression is: in, Represents the weights of two convolution kernels; X1 represents the input data; The concatenation operation is to concatenate the two convolution results in the channel dimension, and its expression is: C1=Concat(S1,P1); The second layer 1D convolution transformation is a second multi-scale convolution operation, and its expression is: The concatenation and 1x1 convolution concatenate the results of the second convolution in the channel dimension, and then transform the number of channels into the number of input channels through 1x1 convolution. The expression is: C2 = Concat(S2, P2); S’=Conv1d(C2,W 1×1 ); The maximum pooling and residual connection performs a maximum pooling operation on the input data, and its expression is: Then add the pooled input and the convolutional output to form a residual connection: Among them, S1 and P1 represent the output results of the first convolution of the input data with a convolution kernel size of 3 and 5 respectively; S2 and P2 represent the output results of the second convolution of the input data with a convolution kernel size of 3 and 5 respectively; C1 and C2 represent the first and second S i and P i (i=1, 2) is the result of concatenation in the channel dimension; X1 pool Represents the input after pooling; S' represents the output result after 1x1 convolution transforms the number of channels into the number of input channels.

3. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S1, the neural network adopts a bidirectional long short-term memory network to improve sequence learning by considering both past and future contexts. It contains two hidden layers: one layer processes data forward at each time step, and the other layer processes data backward; in, and Represent the forward and reverse hidden states respectively; y t Represents the final output; and Represents the forward and reverse weights respectively; b y Represents the bias parameter.

4. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S1, the neural network uses a multi-head attention mechanism to filter important information from the input features and assign different weights to them according to their importance; The scaled dot product attention mechanism is used to obtain the corresponding weights of the query and key matrices through dot multiplication, and the Softmax function is used for normalization. The attention is obtained by weighted summation. The expressions are: Among them, Q represents the query matrix; K represents the keyword matrix; V represents the value matrix; A f represents the input matrix; W Q , W K and W V Represent the weight matrices of Q, K and V respectively; d k Represents the dimensions of Q, K, and V; The multi-head attention mechanism is used to realize the simultaneous attention to different parameters of multiple sensor signals, and the obtained results are spliced ​​to obtain the final attention, which is expressed as: in, and Denote the weight matrices of Q, K, and V in the i-th attention head respectively; W denotes the weight matrix of the multi-head attention mechanism; The neural network uses a cross-attention mechanism to allow information interaction between two different feature representations; its two features come from the multi-scale residual module and the bidirectional long short-term memory network-multi-head attention mechanism module respectively; the cross-attention obtains two cross results by taking the output of the multi-scale residual module as the query Query, key Key and value Value, and the output of the bidirectional long short-term memory network-multi-head attention mechanism module as the query Query, key Key and value Value, and the two cross results are concatenated and then passed through the linear layer to obtain the final output.

5. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S2, since the initial drift coefficient μ is unknown, it is assumed to be a normal distribution N(μ0,τ 2 ), the diffusion coefficient σ is a known constant, and the expression of the Wiener degradation model is: X(t)=X(0)+μt+σB(t),μ~N(μ0,τ 2 ); Among them, X(0) represents the initial degradation state; t represents time; X(t) represents the degradation state at time t; μ represents the average rate of degradation, that is, the drift coefficient; σ represents the random fluctuation amplitude, that is, the diffusion coefficient; B(t) is the standard Wiener process, which represents randomness.

6. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S3, the influence of the error term σB(t) is small and can be ignored. According to the observed values ​​of the historical drift coefficients {μ1,μ2,…,μ m }Get the initial mean μ0 and variance τ of the drift coefficient 2 , the expression is: Among them, {μ1,μ2,…,μ m } is the observed value of m historical drift coefficients; Since at each time point t i The degradation value X(t i ) obeys the normal distribution N(X0+μt i ,σ 2 t i ), therefore, the observed value X(t i ) is: The likelihood function of the observed value is: Then, the log-likelihood function is: Since μ follows a normal distribution N(μ0,τ 2 ), combined with Bayesian theory, the joint log-likelihood function is: Expand to get: Then maximize the joint log-likelihood function to get t m The a posteriori estimate of the drift coefficient μ at time is expressed as: Among them, X(t i ) indicates t i The degradation value at the moment; X0 represents the initial degradation level of the system at time t=0; σ represents the random fluctuation amplitude, that is, the diffusion coefficient; μ0 represents the mean of the normal distribution of the drift coefficient; τ 2 Represents the variance of the normal distribution of the drift coefficient.

7. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S4, according to Bayes' theorem, the posterior distribution of the drift coefficient is the product of the likelihood function and the normal distribution of the initial drift coefficient; the distribution of the initial drift parameter and the likelihood function are both in the form of normal distribution, and due to conjugation, the posterior distribution is also a normal distribution; t m The posterior mean μ of the drift coefficient at the moment post and variance τ post 2 The expressions are: Among them, X(t i ) indicates t i The degradation value at the moment; X0 represents the initial degradation level of the system at time t=0; σ represents the random fluctuation amplitude, that is, the diffusion coefficient; μ0 represents the mean of the normal distribution of the drift coefficient; τ 2 Represents the variance of the normal distribution of the drift coefficient.

8. The degradation prediction method based on random process and deep learning interaction according to claim 1, characterized in that: In step S5, the tth m The historical degradation data at time t is taken as input and the output is m+1 The degradation value at time t m The mean and variance of the posterior distribution of the drift coefficient at time t are taken as m+1 The mean and variance of the prior distribution of the drift coefficient at time, the known diffusion coefficient and the degradation data {(t i ,X(t i ))}, Where i = 1, 2, ..., m + 1; get t m+1 The drift coefficient at the moment; t m+1 The mean and variance of the posterior distribution of the drift coefficient at time t are taken as m+2 The mean and variance of the prior distribution of the drift coefficient at time t, and the known diffusion coefficient and the degradation data updated by the neural network {(t i ,X(t i ))}, where i = 1, 2, ..., m + 2), is used to update t m+2 Drift coefficient at time; t m+2 The mean and variance of the posterior distribution of the drift coefficient at time t are used as m+3 The mean and variance of the prior distribution of the drift coefficient at the moment; repeat the above steps to update the degradation data in the application scenario without real-time monitoring in real time; In step S6, a failure threshold is set. When the degradation value output by the neural network reaches the failure threshold, the difference between the corresponding failure time and the initial time is the remaining life.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, any step in the degradation prediction method based on the interaction of random process and deep learning as described in any one of claims 1-8 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer processor, it implements any step in the degradation prediction method based on random process and deep learning interaction as described in any one of claims 1 to 8.

Images