A bearing fault prediction method based on fuzzy cognitive map

Abstract

The application provides a bearing fault prediction method based on a fuzzy cognitive map, and comprises the following steps: S1, a bearing fault prediction model is constructed; S2, obtained fault data is normalized, and each variable of bearing data is taken as a node and input into STFCM, so that the space-time characteristics between bearing data are learned; S3, the bearing state value obtained by STFCM is input into LSTM through a residual structure; S4, the previous hidden state of LSTM is input into TFCM; and S5, an output prediction result is obtained through three full connection layers. Bearings are usually used in complex environments, so that the bearings are prone to faults and the fault signs are not obvious in the early stage of the faults. The application brings potential safety threats and immeasurable economic losses to the industry. At present, the fault diagnosis method of rail transit is difficult to process a large amount of high-dimensional data, and the application can better capture the high-frequency changes of signals and diagnose the faults of rail transit.

Description

TECHNICAL FIELD

[0001] The application provides a bearing fault prediction method based on a fuzzy cognitive map, and belongs to the technical field of bearing fault diagnosis.

BACKGROUND

[0002] Bearings can be the most common and critical mechanical components in various machines, and are widely used in urban transportation systems and industrial applications. However, bearing failures frequently pose a threat to system safety, and therefore, bearing fault prediction has always been a hot research topic. Existing fault diagnosis methods include methods based on analytical models, methods based on signal processing, and methods based on knowledge, etc., and these methods all have certain limitations.

[0003] Fuzzy Cognitive Maps (FCM) is a way of knowledge graph representation, which can connect the trend changes between events through the arc connection between concepts. FCM is composed of nodes and edges connecting the nodes, and is a directed graph. FCM emphasizes the connection between concepts, and the strength of the connection is represented by the weight size. Due to the advantages of good interpretability, reasoning ability, and intuitive representation of knowledge, FCM has been applied to many fields such as military decision-making, power prediction, etc. However, the prediction accuracy is limited. In order to improve the accuracy in time series prediction, many scholars have proposed different algorithms to learn FCM, such as balanced differential learning algorithm, real genetic coding algorithm, etc. With the wide application of machine learning, algorithms such as ridge regression and support vector machine are also used to learn FCM.

[0004] The problem of using FCM to realize long-term prediction is developing rapidly, and sparse auto-encoders are used to extract features from time series to construct FCM. In order to further improve the prediction accuracy, many scholars consider adding other factors to extend the fuzzy cognitive map. After adding the time factor, the extended model has high-order fuzzy cognitive map and dynamic fuzzy cognitive map. Wavelet transform is used to decompose the time series to construct HFCM for prediction. Some scholars also use intuitionistic fuzzy cognitive maps to simulate the uncertain relationship between systems, and combine variational mode decomposition with high-order intuitionistic fuzzy cognitive maps for time series prediction.

[0005] FCM has been widely used in time series forecasting and has achieved good results. Given that it is impossible to directly construct FCM for univariate time series, many scholars have proposed some construction ideas. For example, the fuzzy C-means clustering algorithm is used to cluster the initial univariate time series into several categories to construct FCM. In order to capture time series information of different frequencies, the empirical wavelet transform is combined with HFCM, and the original sequence and the decomposed sequence are used for modeling. The FCM is learned using a method based on support vector machines, which has the advantage of reducing the influence of outliers. For multivariate time series, many scholars have conducted some explorations in different fields. They input the power and other recorded values ​​into the FCM and use the multi-step gradient method, real-coded genetic algorithm, and structure optimization genetic algorithm to learn FCM to obtain prediction results. Some people have proposed a multivariate fast prediction model that combines elastic net and HFCM to speed up prediction and capture changing trends.

[0006] Existing bearing fault diagnosis methods include analytical model-based methods, signal processing-based methods, and knowledge-based methods. The analytical model-based method is the earliest fault detection method and is a mathematical diagnostic method for system samples. The signal processing-based method processes the detected data in a simple manner and is divided into frequency domain analysis and wavelet analysis. Knowledge-based methods include many methods, mainly expert system diagnosis methods, genetic algorithm diagnosis methods, fault tree diagnosis methods, and neural network diagnosis methods, and are more easily adapted to different working conditions.

[0007] Using time series to predict future system states is widely used across various industries. However, achieving high-precision, interpretable, and universal predictions for long, non-stationary time series remains a pressing challenge, resulting in low accuracy in bearing fault prediction. Improvements to FCM have been relatively successful in predicting short, stationary time series, but suffer from poor accuracy for long, non-stationary time series. Deep learning can improve the accuracy of time series predictions, but due to its inherent black box nature, it cannot explain relationships between variables. Furthermore, increasing network depth in deep learning can lead to gradient explosion and vanishing gradients during long-term backpropagation.

[0008] Existing methods offer low diagnostic accuracy, weak generalization, and difficulty processing high-dimensional data. Analytical model-based methods suffer from low accuracy due to limitations in sample quality and capacity, inherent model limitations, the presence of noise, and increasing system complexity. Expert system fault diagnosis methods suffer from knowledge bases constrained by the limitations and subjectivity of expert knowledge, and the assignment of weights in inference methods is highly subjective. Neural network diagnostic methods also face significant challenges with algorithm convergence and generalization. [Summary of the invention]

[0009] The time series of real scenes are mostly non-stationary, especially in bearing failure, and it is difficult for basic prediction models to achieve high-precision prediction. Ordinary models are difficult to achieve high-precision prediction, and some people use deep learning models to improve the prediction accuracy, but they do not have explainability. The prediction accuracy for different bearings is low, and the generalization ability is weak. How to achieve high-precision, explainability and universality for bearing non-stationary time series prediction is still a difficult problem to be solved. There are more track train fault signal collection data, and non-stationarity exists, so it is difficult for early models to capture the trend, therefore, a better fault diagnosis method is needed.

[0010] The present application combines the explainability of FCM with the advantages of deep learning in prediction, and constructs a new bearing fault prediction model with explainability, called Deep Attention Fuzzy Cognitive Maps (DAFCM), for the multivariate time series of bearings, which is suitable for different types of bearings. The model is composed of Spatio-Temporal Fuzzy Cognitive Maps (STFCM), Long short-term memory neural network (LSTM), Temporal Fuzzy Cognitive Maps (TFCM) and residual structure. The use of STFCM can capture the relationship between bearing variables, consider the correlation between nodes and the correlation of each node time. The addition of LSTM and TFCM can improve the difficulty of basic FCM in predicting non-stationary time series of bearings. TFCM can consider the time correlation of bearing sequence, so that the model has dynamic characteristics in the time dimension. At the same time, the residual structure is added to solve the problem of gradient explosion and disappearance in long-term back propagation.

[0011] The present application is a DAFCM model based on FCM framework, which can predict bearing failure by combining deep learning knowledge. The improved attention mechanism is used to construct STFCM for the first time, which is used to enhance the performance of long time series prediction. The residual structure is introduced into the FCM framework for the first time, which solves the problem of gradient explosion and disappearance caused by the increase of neural network depth in DAFCM, and further improves the prediction accuracy. LSTM is introduced into the FCM framework for the first time, and TFCM is constructed by using the improved time attention, which is used to capture the trend of bearing data and improve the prediction accuracy of non-stationary long time series. The proposed DAFCM is applied to rail transit fault diagnosis, which better solves the non-stationarity of collected signals and improves the diagnosis ability.

[0012] The input of DAFCM is the monitoring data of the bearing, and the output is the prediction of the bearing failure target. The constructed DAFCM flow chart is as follows: Figure 1 As shown, Figure 2 It is the DAFCM framework diagram. Assume that N variable bearing data is A=[A 1,t ,A 2,t ,A 3,t ,,A N,t ], A i,t represents the state value of the i-th bearing variable at time t. For an FCM with N nodes, each variable is used as a node to construct the FCM. During data preprocessing, the initial data must be normalized to ensure that any outliers are addressed, speeding up the gradient descent algorithm and improving prediction accuracy.

[0013] The technical solution of the present invention is now described in detail as follows, comprising the following steps:

[0014] Step 1: Input the bearing variable data into the Spatio-Temporal Fuzzy Cognitive Maps (STFCM) to learn the relationship between each bearing variable and its own time dimension. The obtained bearing variable values ​​are input into the residual structure.

[0015] Step 2: The state value obtained by STFCM in step 1 and the initial state value are linearly transformed to form a residual structure to solve the problem of gradient explosion and gradient vanishing caused by the increase in network depth of Deep Attention Fuzzy Cognitive Maps (DAFCM). The bearing state value updated by the residual structure is input into the Long Short-Term Memory Neural Network (LSTM);

[0016] Step 3: Input the bearing state value updated in step 2 into the LSTM to capture the global context of the bearing data, and input the previous hidden state of the LSTM into the Temporal Fuzzy Cognitive Maps (TFCM);

[0017] Step 4: Use TFCM to capture the temporal correlation of the hidden state output from LSTM to improve the prediction ability of the bearing non-stationary time series. After obtaining the new bearing state value, it is input into the next residual structure;

[0018] Step 5: The bearing state value output by TFCM and the initial state value undergo linear changes to form a residual structure, and then pass through three fully connected layers to obtain the final output state value.

[0019] In step one, the present invention further comprises:

[0020] In order to better adaptively capture the relationship between the target value and each external feature as well as the changing trend of the time dimension, we propose a STFCM based on an improved attention mechanism, such as Figure 3 As shown, the node status value starts from ①( Figure 2 As shown in Figure 2, the graph is input into STFCM, and weight values ​​are assigned to different nodes at different time points. The graph is linked to the time dimension, and the correlation between nodes is calculated first to learn the node relationship in the graph. The specific formulas (1) and (2) are as follows:

[0021]

[0022]

[0023] in, W e is a learnable parameter, B e is the bias term, α t is the weight value, f is the transfer function that maps the activation value to a specific range, is the scoring function of node j, is the scoring function of node i, A t is the bearing condition value, is the attention distribution value, i is the current bearing variable value, and N is the total bearing data variable. The state value update formula (3) of the bearing variable is as follows:

[0024]

[0025] Then, the updated node sequence (3) is input into the improved attention mechanism, as shown in formula (4). In STFCM, the state value of a node is associated with its state value in the previous period of time. The single-time layer STFCM is expanded to multiple-time layer STFCM to learn the relationship between different graphs. Here, the attention mechanism with improved attention scoring function is used to enhance the prediction of long-term time series. The specific formulas (4), (5), and (6) are as follows:

[0026]

[0027]

[0028]

[0029] Among them, W f1 、W f2 、W f3 、 is a learnable parameter, B f1 、B f2 is the bias term, βt is the weight value, represents the hidden state and cell state of the transition-gated LSTM, f is the transfer function that maps the activation value to a specific range, and s t The bearing status value after STFCM update, z t is the scoring function of the bearing data at time t, z i is the scoring function of the bearing data time i, T is the total time value of learning bearing data, A l ' is the updated state value of the bearing variable. Here we only consider the information of the time step earlier than the node to realize the correlation relationship. Therefore, after obtaining the weight value, the state value s of the node at time t can be obtained. t .

[0030] In step 2, the present invention further comprises:

[0031] Based on the concept of residual networks, instead of using functions, they reference hidden states. Some researchers have introduced residual structures to construct neural networks with a depth of 152 layers, achieving relatively good results on datasets. In deep neural networks, residual structures combined with linear learning can address the problems of exploding and vanishing gradients, enabling better prediction of bearing faults.

[0032] In view of the good results achieved by introducing residual structure into image processing and time series prediction, the residual structure is introduced into DAFCM. The residual structure is a calculation module, and the residual information is increased through linear transformation to accelerate the convergence speed. The initial node state value is input from ② to the residual structure ( Figure 2 The specific formula (7) is as follows:

[0033]

[0034] in, is the weight of the linear transformation, B q is the bias term, A t is the bearing condition value, is the bearing state value after the residual structure, s t Bearing status value after STFCM update.

[0035] Formula (7) does not increase computational complexity. The residual structure combines the updated node state with the initial node state through linear transformation, which does not increase the training error and makes the loss function approximate to a state mapping. Obviously, the addition of node history state to STFCM does not increase nonlinear transformation. Linear transformation makes it more sensitive to sudden changes in node history state and more robust.

[0036] In step three, the present invention further comprises:

[0037] The basic FCM is difficult to capture the long-term dependency of the time series in the prediction of non-stationary long time series of bearings, resulting in low prediction accuracy. LSTM can capture the long-term dependency of the bearing time series. Therefore, LSTM is introduced to solve this problem, such as Figure 2 As shown, the historical state of the bearing is input into the LSTM, and the time state of each LSTM at the moment is Access to the memory cell is controlled by three gates: the forget gate Input Gate Output Gate They are used to control forgotten, input, and retained information respectively. The specific formulas (8), (9), and (10) are as follows:

[0038]

[0039]

[0040]

[0041] Among them, W c 、U c 、b c are the parameters to be learned in the control gate used, and is the previous hidden state, is the node state, σ(·) and tanh(·) are the sigmoid function and the hyperbolic tangent function respectively, and is the element multiplication. is the input gate, To update the cell status.

[0042] Because LSTM can capture the global node context of bearing data, its hidden state is obtained from LSTM. This node state is transformed and fed into the temporal attention mechanism to capture the dynamic temporal correlation of bearing data. To ensure consistency across the temporal dimension, the hidden state of a certain time step is aggregated as the temporary state value of the bearing state.

[0043] In step 4, the present invention further comprises:

[0044] In the case of bearing failure, unexpected situations may occur, causing sudden changes in the state value at that time. If the current node state is not considered to be related to the previous state, the prediction accuracy will be reduced. Therefore, it is necessary to consider the nonlinear changes between them, solve the non-stationary prediction problem in the bearing time series, and further enhance its prediction ability. The previous hidden state obtained by LSTM is spliced ​​and input into the TFCM based on the improved attention mechanism we proposed, such as Figure 4 As shown, the specific formulas (11), (12), and (13) are as follows:

[0045]

[0046]

[0047]

[0048] Among them, W m1 、W m2 、 is a learnable parameter, B m1 、B m2 is the bias term, λ t is the weight value, is the node state value, f is the transfer function that maps the activation value to a specific range, is the previous hidden state of LSTM, v t is the scoring function of the bearing data at time t, v i is the scoring function of the bearing data time i, and T is the total time value of learning bearing data.

[0049] In step five, the present invention further comprises:

[0050] Since the corresponding network depth is increased, the above residual structure is introduced here, and the initial bearing state value is input into the residual structure from ③ ( Figure 2 The specific formula (14) is as follows:

[0051]

[0052] in, is the weight of the linear transformation, B k is the bias term, A t is the bearing condition value, A t ' is the bearing state value after the residual structure. Finally, the final output is obtained through the three-layer fully connected neural network like Figure 2 shown.

[0053] DAFCM uses an encoder-decoder architecture, directly converting node A i,t The historical state is input into the encoder, and the output of the encoder is obtained through STFCM learning, residual structure and LSTM. The output of the encoder With the node history state A t Then input it into the decoder, which is based on the TFCM mechanism and residual structure, and input it into the output layer to get the final result

[0054] Finally, we use stochastic gradient descent (SGD) and Adam optimizer to train DAFCM. By minimizing the mean square error between the predicted value and the true value, DAFCM can be trained end-to-end via backpropagation. The specific formula (15) is as follows:

[0055]

[0056] Among them, Θ represents all parameters, Q is the number of prediction cycles, i is the i-th cycle, is the predicted bearing condition value, A i is the actual bearing status value.

[0057] The beneficial effects brought about by the technical solution of the present invention are:

[0058] The use of deep learning knowledge can better deal with the problems of basic FCM in large-scale non-stationary time series prediction. Introducing deep learning knowledge into the FCM framework is effective for nonlinear prediction, and the DAFCM model is also more robust in different fields. The basic FCM has good interpretability for linear relationships, but has relatively large problems in nonlinear systems. The proposed DAFCM model can consider multiple factors for interpretability analysis of nonlinear systems, and can provide interpretability that is different from other models. Bearings usually work in complex environments, which makes them prone to failure and the signs of failure are not obvious in the early stages of failure. It brings potential safety threats and immeasurable economic losses to the industry. The current methods of rail transit fault diagnosis are difficult to handle large amounts of high-dimensional data. The present invention can better capture the high-frequency changes of signals and diagnose rail transit faults.

Brief Description of the Drawings

[0059] Figure 1 It is the DAFCM flow chart.

[0060] Figure 2 This is the DAFCM framework diagram.

[0061] Figure 3 This is the STFCM framework diagram.

[0062] Figure 4 This is the TFCM framework diagram.

[0063] Figure 5 It is the acceleration data of the driving end.

[0064] Figure 6 It is the acceleration data at the fan end.

[0065] Figure 7 is the base acceleration data. [Specific implementation method]

[0066] The present invention will be further described below based on specific examples. Figure 1-4 As shown, the present invention provides a method for high-precision, explainable fault prediction for various bearing types. This method utilizes a publicly available dataset from Case Western Reserve University in the United States, comprising four different types of samples: normal, outer ring damaged, inner ring damaged, and rolling element damaged. The bearings to be tested support the motor shaft; the drive-end bearing is an SKF6205, sampled at 12 kHz and 48 kHz; the fan-end bearing is an SKF6203, sampled at 12 kHz. The bearing damage is single-point damage generated by electrospark machining. SKF bearings are tested for damage with diameters of 0.1778, 0.3556, and 0.5334 mm; NTN bearings are tested for damage with diameters of 0.7112 and 1.016 mm. Because the location of outer ring damage is relatively fixed, the position of the damage point relative to the bearing load zone has a direct impact on the vibration response of the motor / bearing system. To quantify this impact, damage points were placed at 3, 6, and 12 o'clock on the outer rings of the drive-end and fan-end bearings, respectively. An accelerometer is placed above the bearing housing on both the fan and drive ends of the motor to collect the vibration acceleration signal of the faulty bearing. The vibration signal is acquired by a 16-channel data logger. Power and speed are measured using a torque sensor / decoder. The three data variables input into the prediction model are: DE (drive end accelerometer data); FE (fan end accelerometer data); and BA (base accelerometer data (normal).

[0067] S1. Build a bearing fault prediction model, construct a deep neural network for the proposed model, and debug the parameters.

[0068] S2. The data variables of the input model are DE, FE and BA. Here we only give the data of the inner ring of a bearing. The specific data is as follows Figure 5-7 As shown in the figure, the obtained fault data is normalized, and each variable of the bearing data is input into the STFCM as a node to learn the spatiotemporal features between the bearing data. The updated node sequence is obtained by learning the spatial features of the data through formulas (1) and (2), as shown in formula (3).

[0069] S3. The updated node sequence is obtained by learning the time characteristics of the data through formulas (4), (5), and (6), which can focus on the influence of a certain important time node instead of evenly distributing weights to all time nodes.

[0070] S4. The bearing state value obtained by the STFCM is input into the residual structure formula (7) to obtain the bearing state updated again. According to the idea of the residual network, the residual structure does not use a function, but refers to a hidden state, and solves the problems of gradient explosion and disappearance.

[0071] S5. The bearing state updated again obtained by the residual structure is input into the LSTM as shown in formulas (8), (9) and (10) to capture the long-term dependence of the bearing time sequence.

[0072] S6. The previous hidden state of the LSTM is input into the TFCM as shown in formulas (11), (12) and (13) to solve the non-stationary prediction problem in the bearing time sequence and further enhance the prediction ability.

[0073] S7. The bearing state output by the TFCM is input into the residual structure formula (14) together with the initial bearing state value.

[0074] S8. The bearing state value of the residual structure formula (14) is input into a three-layer fully connected layer to obtain an output prediction result.

[0075] Compared with other methods, the following table shows:

[0076] method MAE RMSE AR 0.0707 0.0904 ARIMA 0.0896 0.1132 ES 0.1272 0.1508 VAR 0.0712 0.0912 MLP 0.0708 0.0909 FCM 0.1039 0.1201 WFCM 0.0442 0.0566 KFCM 0.0493 0.0621 LSTM 0.0486 0.0607 Proposed method 0.0372 0.0474

[0077] Wherein MAE (mean absolute error) and RMSE (root mean square error) are common evaluation indexes for prediction problems, AR is an autoregressive model, ARIMA is an autoregressive integrated moving average model, ES is an exponential smoothing model, VAR is a vector autoregressive model, MLP is a multilayer perceptron, FCM is a basic fuzzy cognitive map, WFCM is a wavelet transform and fuzzy cognitive map, KFCM is a kernel mapping and fuzzy cognitive map, and LSTM is a long short-term memory neural network. T is the number of prediction cycles, x(t) and

[0078]

[0079]

[0080] It can be seen that the MAE and RMSE of the present application are the smallest, which proves the effectiveness of the method.​

Claims

1. A bearing fault prediction method based on fuzzy cognitive map, characterized in that: The steps include: Step 1: Input the bearing variable data into the spatiotemporal fuzzy cognitive map (STFCM) to learn the relationship between the bearing variables and their respective time dimensions. The obtained bearing variable values ​​are input into the residual structure. Step 2: The state value obtained by STFCM in step 1 and the initial state value are linearly transformed to form a residual structure to solve the problem of gradient explosion and gradient vanishing caused by the increase in the depth of the deep attention fuzzy cognitive map DAFCM network. The bearing state value updated by the residual structure is input into the long short-term memory neural network LSTM; Step 3: Input the bearing state value updated in step 2 into LSTM to capture the global context information of the bearing data, and input the previous hidden state of LSTM into the temporal fuzzy cognitive map TFCM; Step 4: Use TFCM to capture the temporal correlation of the hidden state output from LSTM to improve the prediction ability of the bearing non-stationary time series, and obtain the new bearing state value before inputting it into the next residual structure; Step 5: The bearing state value output by TFCM and the initial state value undergo linear transformation to form a residual structure, which is then passed through three fully connected layers to obtain the final output state value; In step 2, the residual structure is introduced into DAFCM; STFCM is a calculation module, and the residual information is increased through linear transformation to accelerate the convergence speed; the initial node state value is input into the residual structure, as follows: in, is the weight of the linear transformation, B q is the bias term, A t is the bearing condition value, is the bearing state value after the residual structure, s t Bearing status value after STFCM update.

2. The bearing fault prediction method based on fuzzy cognitive map according to claim 1, characterized in that: In step 1, the node state values ​​are input into STFCM, and weight values ​​are assigned to different nodes at different time points. In conjunction with the time dimension, the correlation between nodes is calculated first, and the node relationship is learned. The details are as follows: in, W e is a learnable parameter, B e is the bias term, f is the transfer function that maps the activation value to a specific range, is the scoring function of node j, is the scoring function of node i, A t is the bearing condition value, is the attention distribution value, i is the current bearing variable value, and N is the total bearing data variable; the state value update formula (3) of the bearing variable is as follows:

3. The bearing fault prediction method based on fuzzy cognitive map according to claim 2, characterized in that: Input the updated node sequence formula (3) into the improved attention mechanism; In STFCM, the state value of a node is associated with its state value in the previous period of time. The single-time layer STFCM is expanded to multiple-time layer STFCM. The attention mechanism with improved attention scoring function is used to enhance the prediction of long-term time series. The details are as follows: Among them, W f1 、W f2 、W f3 、 is a learnable parameter, B f1 、B f2 is the bias term, β t is the weight value, represents the hidden state and cell state of the transition-gated LSTM, f is the transfer function that maps the activation value to a specific range, and s t The bearing status value after STFCM update, z t is the scoring function of the bearing data at time t, z i is the scoring function of the bearing data time i, T is the total time value of learning bearing data, A l ' is the updated bearing variable state value; after obtaining the weight value, the state value s of the node at time t can be obtained t .

4. The bearing fault prediction method based on fuzzy cognitive map according to claim 1, characterized in that: In step 3, the historical state of the bearing is input into the LSTM, and the time state of each LSTM at the moment is Access to the memory cell is controlled by three gates: the forget gate f t c , input gate Output Gate They are used to control forgotten, input, and retained information respectively; the details are as follows: Among them, W cT 、U cT 、b c are the parameters to be learned in the control gate used, and is the previous hidden state, is the node state, σ(·) and tanh(·) are the sigmoid function and the hyperbolic tangent function respectively, ⊙ is the element-wise multiplication, is the input gate, To update the cell status.

5. The bearing fault prediction method based on fuzzy cognitive map according to claim 1, characterized in that: In step 4, in the case of a bearing failure, an unexpected situation may occur, causing a sudden change in the state value at that time. If the current node state is not associated with the previous state, the prediction accuracy will be reduced. Therefore, it is necessary to consider the nonlinear changes between them and solve the non-stationary prediction problem in the bearing time series to further enhance its prediction ability. The previous hidden state obtained by LSTM is spliced ​​and input into TFCM as follows: Among them, W m1 、W m2 、 is a learnable parameter, B m1 、B m2 is the bias term, λ r is the weight value, is the node state value, f is the transfer function that maps the activation value to a specific range, is the previous hidden state of LSTM, v t is the scoring function of the bearing data at time t, v i is the scoring function of the bearing data time i, and T is the total time value of learning bearing data.

6. The bearing fault prediction method based on fuzzy cognitive map according to claim 1, characterized in that: In step 5, the initial bearing condition values ​​are input into the residual structure as follows: in, is the weight of the linear transformation, B k is the bias term, A t is the bearing condition value, A t ' is the bearing state value after the residual structure; is the node state value, and finally the final output is obtained through a three-layer fully connected neural network DAFCM uses an encoder-decoder architecture, directly converting node A i,t The historical state is input into the encoder, and the output of the encoder is obtained through STFCM learning, residual structure and LSTM. The output of the encoder With the node history state A t Then input it into the decoder, which is based on the TFCM mechanism and residual structure, and input it into the output layer to get the final result Finally, DAFCM is trained using stochastic gradient descent (SGD) and the Adam optimizer. DAFCM is trained end-to-end via backpropagation by minimizing the mean squared error between the predicted value and the true value, as follows: Among them, Θ represents all parameters, Q is the number of prediction cycles, i is the i-th cycle, is the predicted bearing condition value, A i is the actual bearing status value.

Images