A bearing life prediction method based on deep learning hybrid model

Abstract

The present invention discloses a bearing life prediction method based on a deep learning hybrid model, which belongs to the technical field of bearing remaining life prediction, and includes the following steps: First, the root mean square value of the vibration signal is extracted as the original data. In order to enhance the ability to extract degradation information, the variational mode decomposition optimized by the Grey Wolf Optimization Algorithm is used to decompose the original data to obtain each subsequence. Then, the hidden information is extracted through a bidirectional time convolutional network, and input into a bidirectional gated recurrent unit combined with an attention mechanism (AM) for prediction. Finally, the overall prediction result is obtained by summing the predicted values ​​of each component. The proposed algorithm is compared with the advanced algorithm in an experiment, which verifies the significant advantages of the prediction model proposed in this article. It can improve the prediction accuracy and model robustness, and provide a new approach for the field of bearing remaining life prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing life prediction, and more specifically to a bearing life prediction method based on a deep learning hybrid model. Background Art

[0002] As one of the most important components in rotating machinery systems, rolling bearings are widely used in various industrial machinery devices, including but not limited to high-speed trains, machine tools and aircraft engines. Due to their complex operating conditions and frequent exposure to harsh external environments, rolling bearings are prone to failure. Such failures may not only cause mechanical equipment to shut down, but also cause significant economic losses and even endanger personnel safety. Therefore, predictive health management (PHM) technology based on fault prediction has received widespread attention. In PHM technology, remaining useful life (RUL) assessment is one of the core contents. RUL refers to the time that a device or component can continue to operate normally in the future. Accurate RUL assessment is of great significance for equipment maintenance, fault warning and resource optimization configuration. Through accurate prediction of RUL, maintenance plans can be made in advance, unexpected downtime can be avoided, resource allocation can be optimized, and thus equipment reliability and operation efficiency can be improved. In recent years, with the development of artificial intelligence and big data technology, PHM technology has made rapid progress, providing more effective solutions for health monitoring and fault diagnosis of key components such as rolling bearings.

[0003] At present, domestic and foreign research institutions and scholars have conducted a lot of research on the prediction of the remaining life of bearings. In general, the prediction of the remaining life mainly adopts the following two methods: model-based prediction method and data-driven method. The working conditions of the actual production line are relatively complex, and the vibration signal of the bearing is more likely to jump, which is difficult to predict using the model. In contrast, the data-driven method can learn and establish a mapping model between the vibration signal and RUL to predict the RUL, and has better recognition ability for complex vibration characteristics. With the vigorous development of sensor technology and monitoring equipment, the amount of data continues to increase, and the data-driven RUL prediction method has become a hot topic of research. The deep learning algorithm is used for modeling. This method trains the model by extracting the time-frequency domain features of the full life data of the bearing, and finally uses the trained deep learning model to predict the remaining life of the bearing. Because the deep learning method has the ability to automatically extract features, it can effectively process complex and high-dimensional data, thereby improving the accuracy and robustness of RUL prediction, and has attracted widespread attention.

[0004] Although data-driven deep learning algorithms have made significant progress in bearing remaining life prediction, they still face some challenges. The degradation process of bearings is extremely complex. When the bearing is in a healthy state, its vibration signal tends to be stable. However, as the operating time increases, the bearing will enter the accelerated degradation stage, at which time its degradation rate may be significantly different from the normal rate, which significantly increases the difficulty of remaining life prediction. Traditional deep learning algorithms mostly focus on positive features, but unidirectional features are difficult to deal with the signal jump problem when the bearing enters the degradation stage, resulting in poor prediction results. In summary, the complexity of bearing remaining life prediction poses a severe challenge to existing deep learning algorithms. We urgently need to develop new algorithms that can flexibly capture the dynamic change characteristics of bearings throughout their life cycle to achieve more accurate and reliable predictions of bearing remaining life. Summary of the invention

[0005] The purpose of the present invention is to provide a reliable bearing life prediction algorithm model. The technical problem to be solved is to provide a bearing life prediction method based on a deep learning hybrid model, which can extract the bidirectional hidden features of the vibration signal and greatly improve the prediction accuracy. In this way, the degradation state of the bearing can be predicted in time, and the components can be processed and maintained in time before a failure occurs, so as to avoid accidents and extend the service life of the bearing.

[0006] To solve the above problems, the technical solution adopted by the present invention is:

[0007] A bearing life prediction method based on a deep learning hybrid model comprises the following steps:

[0008] S1. Obtaining the original degradation time series data collected by the bearing vibration signal sensor and performing data preprocessing; extracting the root mean square value as the original feature data;

[0009] S2, the time-frequency information of bearing degradation extracted by S1 is first input into the variational mode decomposition optimized by the Grey Wolf optimization algorithm to extract the time-frequency domain features of the vibration signal to form an original feature data set;

[0010] S3: Input the feature data decomposed by S2 into the bidirectional temporal convolutional network to extract bidirectional feature information.

[0011] S4. The bidirectional feature information is sent to the bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network can capture the long-term dependencies in the sequence, and output the feature sequence that integrates the global information through the gated recurrent unit units in the forward and backward directions, and use the attention mechanism to further extract key information. Finally, the final output is obtained through a fully connected layer.

[0012] The further improvement is that S1 specifically includes the following steps:

[0013] S11, obtaining the original degradation time series data collected by the bearing vibration signal sensor, and performing data preprocessing, including removing abnormal points in the data and filling in missing points;

[0014] S12. Extract the root mean square value of the vibration signal as the initial feature. The root mean square value is a commonly used signal amplitude measurement indicator, which reflects the energy of the signal. For each sampling point, we calculate the root mean square value of the corresponding vibration signal. The formula is as follows:

[0015]

[0016] Where N is the number of sampling points, x i is the vibration signal amplitude of the i-th sampling point;

[0017] The further improvement is that S2 specifically includes the following steps:

[0018] S21. First, we need to select a suitable fitness function for the Gray Wolf optimization algorithm. For nonlinear complex signals such as bearing vibration signals, we use multi-scale permutation entropy as the fitness function. Multi-scale permutation entropy can effectively detect complex time series signals and is particularly suitable for the analysis of nonlinear signals. The formula of the fitness function is as follows:

[0019]

[0020] Among them, H p The value of (M) indicates the randomness and complexity of the time series, and m is the embedding dimension. The embedding dimension m of multi-scale permutation entropy is an important characteristic parameter. If m is too small, there will be too few permutation patterns and the information contained in the dynamic changes will be lost. If m is too large, the permutation patterns will increase significantly, which will lead to a significant increase in the calculation time of multi-scale permutation entropy.

[0021] S22. First, the extracted RMS value features are used as the input of the Grey Wolf Optimization Algorithm. The dynamic changes of the signal at multiple scales are captured through the multi-scale permutation entropy. By calculating the multi-scale permutation entropy of the input feature signal, we can obtain an evaluation value that reflects the complexity of the signal.

[0022] S23. We take the evaluation value of multi-scale permutation entropy as prey and introduce it into the gray wolf optimization algorithm. During the algorithm iteration process, the wolf pack will continuously update its position according to the position of the prey to approach the optimal solution; at the same time, the prey (i.e., the evaluation value of multi-scale permutation entropy) will also be updated according to the update of the wolf pack to reflect the new optimal feature subset, and the first two feature subsets will be passed into the variational mode decomposition as parameters α and K; the formula for wolf pack update is as follows:

[0023] X i (t+1)=X i (t)-A*D i +A*C*(X p(t) -X i (t))

[0024] Among them, X i (t) represents the position of the ith wolf in the tth generation, A and C represent the convergence factor and exploration factor respectively, and D i represents the distance between the i-th wolf and its prey, X p(t) Indicates the location of prey.

[0025] S24, the α and K obtained in S23 are passed in as the initial parameters of the variational mode decomposition; the core of the variational mode decomposition process lies in the expression and solution of the variational problem; in the variational mode decomposition process, each eigenmode function is iteratively updated, and the center frequency and bandwidth are limited. The determination of the center frequency and bandwidth of the component signal involves finding K mode functions to ensure that the total bandwidth of all decomposed eigenmode functions is the smallest and their sum is equal to the target signal. The formula of variational mode decomposition is as follows:

[0026]

[0027] In the formula, u k is the eigenmode function of the decomposed modal component, ω k is the kth center frequency, δ(t) is the unit impulse function, and * is the convolution symbol;

[0028] The further improvement is that S3 specifically includes the following steps:

[0029] S31. Input the modal components obtained in S15 in S24 into a bidirectional temporal convolutional network. The bidirectional temporal convolutional network is a network structure that combines a temporal convolutional network with a bidirectional processing mechanism. It improves the feature extraction capability of complex time series data by capturing the bidirectional temporal dependencies in the sequence data.

[0030] S32, the input sequence is processed by the forward and backward time convolutional networks respectively, and then the outputs of the two are merged to obtain the final output. In addition, the residual connection is used to prevent the problem of gradient disappearance and gradient explosion. The forward time convolutional network performs convolution operations in the order of time steps from front to back to generate a forward feature sequence, and the backward time convolutional network performs convolution operations in the order of time steps from back to front to generate a backward feature sequence. For a given input sequence X∈R T×D , where T is the sequence length and D is the feature dimension. The outputs of the forward time convolutional network and the backward time convolutional network can be expressed as:

[0031]

[0032] in, and are the feature sequences output by the forward and backward temporal convolutional networks, respectively, and C is the output feature dimension. After obtaining the forward and backward feature sequences, the bidirectional temporal convolutional network merges them to generate the final output. The final output can be expressed as:

[0033]

[0034] The further improvement is that S4 specifically includes the following steps:

[0035] S41, next, the feature sequence F BiTCN It is fed into a bidirectional gated recurrent unit network, which can capture long-term dependencies in the sequence and output a feature sequence that incorporates global information through gated recurrent units in both forward and backward directions.

[0036] S42. In a bidirectional gated recurrent unit network, the hidden state h at each time step t t Determined by the hidden states of the forward and backward gated recurrent units, the forward gated recurrent unit processes the input in the forward order of the sequence to generate a forward hidden state sequence The backward gated recurrent unit processes the input in the reverse order of the sequence to generate a backward hidden state sequence Finally, the hidden state h at each time step is t It is the concatenation or combination of the forward hidden state and the backward hidden state. The formula of the gated recurrent unit is:

[0037]

[0038] S43, output a feature sequence that integrates global information, and use the attention mechanism to further extract key information. The attention mechanism introduces an additional attention weight vector, which can be calculated in real time based on the dynamic characteristics of the input data;

[0039] By applying different degrees of weight to different parts of the input data, the attention mechanism enables the model to pay more attention to information that is critical to the task while ignoring irrelevant information. The weight formula of the attention mechanism is expressed as follows:

[0040] score t,s =tanh(w t h t +w s h s )

[0041]

[0042] The score t,s is the attention probability distribution function; w t , w s is the attention weight matrix; h is the hidden layer vector; a is the attention score; a is the attention layer output value. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the overall framework diagram of the prediction method of the present invention;

[0044] Figure 2 It is a schematic diagram of the gray wolf optimization algorithm in the present invention;

[0045] Figure 3 It is a bidirectional temporal convolutional network structure diagram in the present invention;

[0046] Figure 4 is a schematic diagram of a hybrid model in the present invention; DETAILED DESCRIPTION

[0047] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0048] like Figure 1 As shown, a bearing life prediction method based on a deep learning hybrid model includes the following steps:

[0049] S1. Acquire the original degradation time series data collected by the bearing vibration signal sensor, and perform data preprocessing to obtain complete bearing condition monitoring data; specifically, the following steps are included;

[0050] S11. After collecting the raw data collected by the bearing vibration signal sensor, it is crucial to perform systematic preprocessing on the raw data. First, the data needs to be detected for outliers and all detected outliers should be removed. This step can ensure the accuracy and reliability of subsequent analysis. Next, for missing values ​​in the data, appropriate methods should be used to fill them in to ensure the integrity and continuity of the data. Through these preprocessing steps, more reliable and consistent bearing vibration signal data can be obtained, providing a solid foundation for subsequent analysis and research.

[0051] S2. Input the preprocessed bearing degradation time series data into the variational mode decomposition method optimized by the gray wolf optimization algorithm. The gray wolf optimization algorithm optimizes the parameter settings in the variational mode decomposition by simulating the predation behavior of the gray wolf to achieve the optimal decomposition of the vibration signal. The variational mode decomposition extracts the time-frequency domain characteristics of the signal by decomposing the different modes of the signal. These characteristics not only contain the change information of the vibration signal in the time domain, but also reflect the characteristics in the frequency domain, which can more comprehensively describe the operating status and degradation process of the bearing.

[0052] S21. First, we need to select a suitable fitness function for the Gray Wolf optimization algorithm. For nonlinear complex signals such as bearing vibration signals, we use multi-scale permutation entropy as the fitness function. Multi-scale permutation entropy can effectively detect complex time series signals and is particularly suitable for the analysis of nonlinear signals. The formula of the fitness function is as follows:

[0053]

[0054] Among them, H p The value of (M) indicates the randomness and complexity of the time series, and m is the embedding dimension. The embedding dimension m of multi-scale permutation entropy is an important characteristic parameter. If m is too small, there will be too few permutation patterns and the information contained in the dynamic changes will be lost. If m is too large, the permutation patterns will increase significantly, which will lead to a significant increase in the calculation time of multi-scale permutation entropy.

[0055] S22, taking the evaluation value of multi-scale permutation entropy as prey, we introduce it into the gray wolf optimization algorithm; Figure 2 As shown in the figure, during the algorithm iteration process, the wolf pack will continuously update its position according to the position of the prey to approach the optimal solution; at the same time, the prey (i.e., the multi-scale permutation entropy evaluation value) will also be updated according to the update of the wolf pack to reflect the new optimal feature subset, and the first two feature subsets will be passed into the variational mode decomposition as parameters α and K; the formula for wolf pack update is as follows:

[0056] X i (t+1)=X i (t)-A*D i +A*C*(X p(t) -X i (t))

[0057] Among them, X i (t) represents the position of the ith wolf in the tth generation, A and C represent the convergence factor and exploration factor respectively, and D i represents the distance between the i-th wolf and its prey, X p(t) Indicates the location of prey.

[0058] S23, the α and K obtained in S23 are passed in as the initial parameters of variational mode decomposition; the core of the variational mode decomposition process lies in the expression and solution of the variational problem; in the variational mode decomposition process, each decomposed modal component is iteratively updated, and the center frequency and bandwidth are limited; the determination of the center frequency and bandwidth of the component signal involves finding K modal functions to ensure that the total bandwidth of all decomposed modal components is minimized and their sum is equal to the target signal; the formula of variational mode decomposition is as follows:

[0059]

[0060] In the formula, u k To decompose the modal components, ω k is the kth center frequency, δ(t) is the unit impulse function, and * is the convolution symbol.

[0061] S3, such as Figure 3 As shown, a bidirectional temporal convolutional network is used to extract bidirectional features from the decomposed data; specifically, the following steps are included:

[0062] S31. Input the obtained modal components into a bidirectional temporal convolutional network. The bidirectional temporal convolutional network is a network structure that combines a temporal convolutional network with a bidirectional processing mechanism. It improves the feature extraction capability of complex time series data by capturing the bidirectional temporal dependencies in the sequence data.

[0063] S32, the input sequence is processed by the forward and backward time convolutional networks respectively, and then the outputs of the two are merged to obtain the final output; in addition, residual connection is used to prevent the problem of gradient disappearance and gradient explosion; the forward time convolutional network performs convolution operation in the order of time steps from front to back to generate a forward feature sequence; the backward time convolutional network performs convolution operation in the order of time steps from back to front to generate a backward feature sequence; for a given input sequence X∈R T×D Where T is the sequence length, D is the feature dimension, and the outputs of the forward time convolutional network and the backward time convolutional network can be expressed as:

[0064]

[0065] in, and are the feature sequences output by the forward and backward temporal convolutional networks, respectively, and C is the output feature dimension; after obtaining the forward and backward feature sequences, the bidirectional temporal convolutional network merges them to generate the final output; the final output can be expressed as:

[0066]

[0067] S4, such as Figure 4 As shown in the figure, the extracted bidirectional features are then passed into the bidirectional gated recurrent unit to learn the features, and the attention mechanism is used to make the model focus on important features; specifically, the following steps are included:

[0068] S41, feature sequence F BiTCN It is sent to the bidirectional gated recurrent unit network; the bidirectional gated recurrent unit can capture the long-term dependencies in the sequence, and output a feature sequence that integrates global information through the gated recurrent unit units in both forward and backward directions.

[0069] S42. In the bidirectional gated recurrent unit, the hidden state h at each time step t t It is determined by the hidden states of the forward and backward gated recurrent units; the forward gated recurrent unit processes the input in the forward order of the sequence to generate a forward hidden state sequence h t The backward gated recurrent unit processes the input in the reverse order of the sequence to generate a backward hidden state sequence Finally, the hidden state h at each time step is t It is the concatenation or combination of the forward hidden state and the backward hidden state; the formula of the gated recurrent unit is:

[0070]

[0071] S43, output a feature sequence that integrates global information, and use the attention mechanism to further extract key information; the attention mechanism introduces an additional attention weight vector, which can be calculated in real time based on the dynamic characteristics of the input data;

[0072] By applying different degrees of weight to different parts of the input data, the attention mechanism enables the model to pay more attention to information that is critical to the task while ignoring irrelevant information; the weight formula of the attention mechanism is expressed as follows:

[0073] score t,s =tanh(w t h t +w s h s )

[0074]

[0075] The score t,s is the attention probability distribution function; w t , w s is the attention weight matrix; h is the hidden layer vector; a is the attention score; a is the attention layer output value.

[0076] Example

[0077] A bearing life prediction method based on deep learning hybrid model is used to verify the effectiveness and superiority of the proposed method using the XJTU-SY bearing dataset jointly collected by Xi'an Jiaotong University and Shenyang Blower Group. The bearing accelerated life test platform used in the experiment was designed by the joint laboratory and manufactured by Shengyang Technology. The platform consists of an AC motor, a motor speed controller, a rotating shaft, a support bearing, a hydraulic loading system, and a test bearing. It can carry out accelerated life tests of various rolling bearings or sliding bearings under different working conditions and obtain full life cycle monitoring data of the test bearing. The adjustable working conditions of the test platform mainly include radial force and speed. The radial force is generated by the hydraulic loading system and acts on the bearing seat of the test bearing. The speed is set and adjusted by the speed controller of the AC motor.

[0078] During the test, two PCB352C33 accelerometers installed in the vertical and horizontal directions of the bearing shell were applied to the collected vibration signals, with a sampling frequency of 25.6kHz. The continuous acquisition time of each sample was 1.28s, and the sampling interval between two adjacent samples was 1 minute. This paper selects the bearing data set collected under operating condition 3 to analyze the proposed method, and uses the vibration data of bearings 3_1, 3_2 and 3_4 as test samples.

[0079] Table 1

[0080]

[0081] In this study, we conducted a detailed experimental analysis of the data of three bearings under working condition 3 of the XJTU-SY bearing dataset to verify the effectiveness of the proposed model. According to the full life of the bearing and the data collection once a minute, we assigned corresponding time labels to the vibration signals of the bearings. Considering the differences in the life of each bearing, we normalized the time labels of the bearings and limited their range to between 0 and 1 to characterize the remaining useful life (RUL) of the bearings. In order to avoid the randomness of the training results, the data of two of the bearings were used as training sets in turn to train the model. The last bearing dataset was used as a test set to predict the remaining life.

[0082] As shown in Table 2, the prediction results of this example and other frontier bearing life prediction methods based on deep learning are given.

[0083] Table 2

[0084]

[0085] As can be seen from Table 2, compared with other convolutional neural network variants, this embodiment significantly reduces the mean absolute error and root mean square error of the prediction, thereby improving the life prediction performance. This improvement is mainly attributed to the fact that the method proposed in this embodiment performs bidirectional feature extraction on the bearing vibration signal, successfully suppresses the influence of noise and interference, and enhances the weight of useful information. In addition, the bearing degradation characteristics are learned through the hybrid model, so that the bearing degradation feature information finally obtained is enhanced in both integrity and pertinence. The above results further verify the application value of the present invention in actual bearing life prediction.

[0086] In summary, the present invention effectively extracts bidirectional features from raw data containing a large amount of noise and interference by combining variational mode decomposition optimized by the Gray Wolf Optimization Algorithm with a hybrid model based on a bidirectional temporal convolutional network, significantly improving the prediction accuracy and robustness of the model. This method provides a new approach to the field of bearing remaining life prediction and demonstrates its potential for application under complex working conditions.

Claims

1. A bearing life prediction method based on a deep learning hybrid model, characterized in that: S1. Obtaining the original degradation time series data collected by the bearing vibration signal sensor and performing data preprocessing; extracting the root mean square value as the original feature data; S2, the time-frequency information of bearing degradation extracted by S1 is first input into the variational mode decomposition optimized by the Grey Wolf optimization algorithm to extract the time-frequency domain features of the vibration signal to form an original feature data set; S3, input the feature data decomposed by S2 into the bidirectional temporal convolutional network to extract bidirectional feature information; S4. Send the bidirectional feature information to the bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network can capture the long-term dependencies in the sequence and output the feature sequence that integrates the global information through the gated recurrent unit units in the forward and backward directions. The attention mechanism is used to further extract the key information, and finally the final output is obtained through a fully connected layer. The further improvement is that S1 specifically includes the following steps: S11, obtaining the original degradation time series data collected by the bearing vibration signal sensor, and performing data preprocessing, including removing abnormal points in the data and filling in missing points; S12. Extract the root mean square value of the vibration signal as the initial feature. The root mean square value is a commonly used signal amplitude measurement indicator, which reflects the energy of the signal. For each sampling point, we calculate the root mean square value of the corresponding vibration signal. The formula is as follows: Where N is the number of sampling points, x i is the vibration signal amplitude of the i-th sampling point; The further improvement is that S2 specifically includes the following steps: S21. First, we need to select a suitable fitness function for the Gray Wolf optimization algorithm. For nonlinear complex signals such as bearing vibration signals, we use multi-scale permutation entropy as the fitness function. Multi-scale permutation entropy can effectively detect complex time series signals and is particularly suitable for the analysis of nonlinear signals. The formula of the fitness function is as follows: Among them, H p The value of (M) indicates the randomness and complexity of the time series, and m is the embedding dimension. The embedding dimension m of multi-scale permutation entropy is an important characteristic parameter. If m is too small, there will be too few permutation patterns, and the information contained in the dynamic changes will be lost. If m is too large, the permutation patterns will increase significantly, which will lead to a significant increase in the calculation time of multi-scale permutation entropy. S22. First, the extracted RMS value features are used as the input of the Grey Wolf Optimization Algorithm. The dynamic changes of the signal at multiple scales are captured by multi-scale permutation entropy. By calculating the multi-scale permutation entropy of the input feature signal, we can get an evaluation value, which reflects the complexity of the signal. S23. We take the evaluation value of multi-scale permutation entropy as prey and introduce it into the gray wolf optimization algorithm. During the algorithm iteration process, the wolf pack will continuously update its position according to the position of the prey to approach the optimal solution; at the same time, the prey (i.e., the evaluation value of multi-scale permutation entropy) will also be updated according to the update of the wolf pack to reflect the new optimal feature subset, and the first two feature subsets will be passed into the variational mode decomposition as parameters α and K; the formula for wolf pack update is as follows: X i (t+1)=X i (t)-A*D i +A*C*(X p(t) -X i (t)) Among them, X i (t) represents the position of the ith wolf in the tth generation, A and C represent the convergence factor and exploration factor respectively, and D i represents the distance between the i-th wolf and its prey, X p(t) Indicates the location of prey; S24, the α and K obtained in S23 are passed in as the initial parameters of the variational mode decomposition; the core of the variational mode decomposition process lies in the expression and solution of the variational problem; in the variational mode decomposition process, each eigenmode function is iteratively updated, and the center frequency and bandwidth are limited. The determination of the center frequency and bandwidth of the component signal involves finding K mode functions to ensure that the total bandwidth of all decomposed eigenmode functions is the smallest and their sum is equal to the target signal. The formula of variational mode decomposition is as follows: In the formula, u k is the eigenmode function of the decomposed modal component, ω k is the kth center frequency, δ(t) is the unit impulse function, and * is the convolution symbol; The further improvement is that S3 specifically includes the following steps: S31, inputting the modal component obtained in S15 in S24 into a bidirectional temporal convolutional network, wherein the bidirectional temporal convolutional network is a network structure that combines a temporal convolutional network with a bidirectional processing mechanism, and improves the feature extraction capability of complex time series data by capturing the bidirectional temporal dependency in the sequence data; S32, the input sequence is processed by the forward and backward time convolutional networks respectively, and then the outputs of the two are merged to obtain the final output. In addition, the residual connection is used to prevent the problem of gradient disappearance and gradient explosion. The forward time convolutional network performs convolution operations in the order of time steps from front to back to generate a forward feature sequence, and the backward time convolutional network performs convolution operations in the order of time steps from back to front to generate a backward feature sequence. For a given input sequence X∈R T×D , where T is the sequence length and D is the feature dimension. The outputs of the forward time convolutional network and the backward time convolutional network can be expressed as: in, and are the feature sequences output by the forward and backward temporal convolutional networks, respectively, and C is the output feature dimension. After obtaining the forward and backward feature sequences, the bidirectional temporal convolutional network merges them to generate the final output. The final output can be expressed as: The further improvement is that S4 specifically includes the following steps: S41, next, the feature sequence F BiTCN It is fed into a bidirectional gated recurrent unit network, which can capture long-term dependencies in the sequence and output a feature sequence that incorporates global information through gated recurrent units in both forward and backward directions. S42. In a bidirectional gated recurrent unit network, the hidden state h at each time step t t Determined by the hidden states of the forward and backward gated recurrent units, the forward gated recurrent unit processes the input in the forward order of the sequence to generate a forward hidden state sequence The backward gated recurrent unit processes the input in the reverse order of the sequence to generate a backward hidden state sequence Finally, the hidden state h at each time step is t It is the concatenation or combination of the forward hidden state and the backward hidden state. The formula of the gated recurrent unit is: S43, output a feature sequence that integrates global information, and use the attention mechanism to further extract key information. The attention mechanism introduces an additional attention weight vector, which can be calculated in real time based on the dynamic characteristics of the input data; By applying different degrees of weight to different parts of the input data, the attention mechanism enables the model to pay more attention to information that is critical to the task while ignoring irrelevant information. The weight formula of the attention mechanism is expressed as follows: score t,s =tanh(w t h t +w s h s ) The score t,s is the attention probability distribution function; w t , w s is the attention weight matrix; h is the hidden layer vector; a is the attention score; a is the attention layer output value.

Images