Bearing life prediction method based on double-branch feature extraction under long-time sequence

Abstract

The invention provides a bearing life prediction method based on double-branch feature extraction under a long-time sequence, and relates to the technical field of mechanical fault diagnosis and artificial intelligence. The method comprises the following steps: collecting multi-sensor vibration signals for normalization processing, and realizing multi-source information fusion through a Transform network. Then, a parallel double-branch feature extraction network composed of the Bi-TCN and the Bi-ILSTM is constructed; wherein the Bi-TCN excavates a long time sequence dependency relationship in a vibration signal from a front direction and a back direction, and the utilization capability of future information is improved; the Bi-ILSTM learns dynamic changes in a time sequence, and captures a long-term evolution trend. Meanwhile, a feature fusion sparse attention mechanism is designed, and weighted fusion of features is realized according to network output weights. And finally, outputting an RUL prediction result of the bearing through the full connection layer. Experimental verification is carried out on a PHM2012 data set, and the result shows that the method has higher accuracy and lower errors in a bearing RUL prediction task and shows better prediction performance and generalization ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault diagnosis and artificial intelligence, and particularly relates to a bearing life prediction method for extracting dual-branch features under a long time series, which is applicable to the intelligent monitoring of the bearing health state and the prediction of the remaining life in industrial equipment. Background Art

[0002] Rolling bearings are the core components of rotating machinery, and their operating states directly affect the safety and stability of the whole machine. In industrial production, due to being in a variable load and high-noise environment for a long time, bearings are prone to wear and even failure, resulting in equipment failures, increasing maintenance costs and downtime losses.

[0003] With the development of deep learning technology, models such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) have been widely used in bearing RUL prediction. However, the existing methods still have the following deficiencies: The temporal convolutional network (TCN) mostly has a unidirectional structure and only uses historical information, unable to fully capture the features of future time; Although the long short-term memory network (LSTM) can extract temporal correlations, it is difficult to learn long time series features; The multi-sensor data fusion strategy is insufficient, and the complementary advantages of multi-source information cannot be exploited; The feature fusion method is relatively rough, and the importance differences of the output features of different branch networks are not considered. Therefore, there is an urgent need for a new architecture that can efficiently fuse multi-sensor data under small sample conditions, accurately extract bidirectional information of time series, and achieve accurate bearing RUL prediction. Summary of the Invention

[0004] The purpose of the present invention is to provide a bearing remaining useful life (RUL) prediction method and system that can still achieve high accuracy, strong temporal modeling ability, and good feature expressiveness under the input of multi-source vibration signals, so as to overcome the deficiencies of the existing methods in weak temporal dependence modeling ability, single feature fusion method, and high prediction error.

[0005] Aiming at the modeling requirements for the degradation behavior of bearings under long-term complex working conditions, the present invention proposes an end-to-end prediction framework based on parallel Bi-TCN and Bi-ILSTM. By constructing a parallel structure of a bidirectional temporal convolutional network (Bi-TCN) and a bidirectional improved long short-term memory network (Bi-ILSTM), the bidirectional temporal dependence features and dynamic evolution laws of vibration signals are effectively extracted, and a feature fusion sparse attention mechanism is designed to achieve adaptive weighted integration of multi-channel information, thereby improving the prediction accuracy and generalization ability of the model for the remaining life.

[0006] The main content of the present invention includes:

[0007] A bearing life prediction method for extracting dual-branch features under a long time series, comprising the following steps:

[0008] (1) The collected multi-sensor data is normalized and efficiently fused through the Transformer network, which serves as the training set for the dual-branch deep feature extraction model;

[0009] (2) Constructing parallel Bi-TCN and Bi-ILSTM models;

[0010] (3) Parallel Bi-TCN and Bi-ILSTM model parameter setting;

[0011] (4) Divide the multi-sensor fusion data into training bearing data and test bearing data for offline modeling training;

[0012] (5) Input the test data into the trained prediction network model for online prediction;

[0013] (6) The accuracy of the prediction results is evaluated using two error indicators: MAE and RMSE.

[0014] In step (1), the vibration signals collected by multiple sensors are normalized and then efficiently fused through the Transformer network. The normalization is defined as:

[0015]

[0016] Among them, X t is the normalized signal, X is the original signal, and X min is the minimum value of the original signal, X max is the maximum value of the original signal.

[0017] In step (2), a parallel feature extraction network model is constructed: one branch is composed of two consecutive bidirectional improved long short-term memory networks, in which a residual gate is added to improve the long short-term memory network, and the other branch is composed of a wide convolutional layer, a maximum pooling layer, and three consecutive bidirectional temporal convolutional networks. The features extracted by the two branches are fused through a feature fusion attention mechanism, and finally the prediction results are output through a fully connected layer.

[0018] The parameter setting in step (3) should be consistent with reducing the burden of model training while avoiding the grid effect problem of void convolution, and the model training method is multi-batch training.

[0019] In step (4), during the offline modeling phase, the training samples are input into the constructed neural network model for iterative optimization. In each round of training, the model calculates a loss value based on the error between the predicted output and the actual label, and gradually adjusts the network parameters through a backpropagation mechanism. When the number of training rounds i reaches the set maximum number of iterations N, the model completes the training process and the parameters converge stably.

[0020] In step (5), the test data is input into the trained prediction network model for real-time prediction, and evaluation metrics are set to verify the prediction performance of the model. The evaluation metrics are the mean absolute error and the root mean square error.

[0021] In step (6), the prediction performance of the prediction network model is evaluated using MAE and RMSE. The definitions of the two error metrics, MAE and RMSE, are as follows:

[0022]

[0023] Among them, this method uses the mean absolute error (MAE) and the root mean square error (RMSE) to evaluate the accuracy of the prediction results. The error term er t represents the gap between the predicted RUL and the true RUL at the t-th moment, where n represents the total duration of the bearing from the initial state to failure. The smaller the values of the two metrics, the more accurate the prediction results and the better the performance of the model.

[0024] The main innovation points of the present invention include:

[0025] 1. Design of a remaining useful life modeling framework for multi-source vibration data: The present invention combines the bidirectional temporal convolutional network (Bi-TCN) and the bidirectional long short-term memory network (Bi-LSTM) for the first time to construct a parallel feature extraction structure, realizing the simultaneous capture of bidirectional long dependencies and short-term dynamic features in vibration signals, effectively making up for the problems that traditional TCN cannot utilize future information and LSTM has weak modeling ability for long-term dependencies, and improving the modeling ability for complex degradation trends.

[0026] 2. Complementary advantages of parallel network structure fusion: The Bi-TCN branch extracts remote dependency features through a multi-scale dilated convolutional structure, adapting to the periodicity and local oscillation patterns in the degradation trend; the Bi-LSTM branch focuses on modeling the time dependencies and trend changes point-to-point in the sequence. The two branches share the input data, achieving bidirectional modeling and feature enhancement at the structural level, providing richer temporal information for RUL prediction. The following table shows the main structural parameters of the parallel network:

[0027]

[0028] 3. Design of a feature fusion sparse attention mechanism: The present invention introduces a feature fusion sparse attention mechanism, adaptively allocating fusion weights according to the importance of the output features of the Bi-TCN and Bi-LSTM, realizing dynamic feature weighting. This mechanism not only improves the model's attention to key features but also enhances the ability to suppress redundant or weak information, showing better generalization ability and robustness in multi-condition data.

[0029] 4. End-to-End RUL Prediction Model Architecture: The overall structure is an end-to-end prediction model, which eliminates the manual construction of health indicators (HI) and the post-processing fitting process, directly outputs the RUL value, and improves the prediction efficiency and deployment simplicity. In model training, the mean squared error loss function is adopted, and time series sample pairs are constructed through a sliding window mechanism to adapt to the modeling requirements of different life stages.

[0030] Through the above innovative design, the experimental results of the present invention on the publicly available bearing degradation dataset of PHM2012 show that this method exhibits lower root mean squared error (RMSE) and mean absolute error (MAE) under multiple degradation conditions, and has significant advantages compared with traditional CNN, TCN or LSTM structures. Especially in predicting the early degradation stage, it shows higher trend capture ability and stability.

[0031] The bearing RUL prediction method based on the parallel fusion of Bi-TCN and Bi-ILSTM proposed by the present invention integrates long-term and short-term dependence modeling, multi-channel feature enhancement and sparse attention fusion mechanism, and realizes high-precision remaining life prediction under multiple fault conditions without relying on too much prior knowledge and label construction, and has good robustness, transferability and industrial practical value. Brief Description of the Drawings

[0032] Figure 1 : Structure diagram of the parallel Bi-TCN and Bi-ILSTM feature extraction model;

[0033] Figure 2 : Feature fusion diagram of the sparse attention fusion mechanism;

[0034] Figure 3 : RUL prediction result diagram of the test bearings B1_2 and B2_4 in the PHM2012 dataset. Detailed Description of the Invention

[0035] The present invention provides a bearing life prediction method for double-branch feature extraction under long time series, and its specific implementation includes the following key technical details:

[0036] The dataset adopted by the present invention is the publicly available PHM2012 bearing dataset. In the data preprocessing process of the proposed method, data normalization and multi-sensor data fusion operations are performed on the original data. The collected multi-sensor data is normalized. The original vibration signal is collected with a length of 1280 points per group. According to the actual RUL prediction requirements, it is divided into a training set and a test set. The data samples are sliced by the sliding window method, and the sliding step is set to 100 points, and the generated training samples cover different degradation stages. These data are efficiently fused through the Transformer network to serve as the training set of the double-branch deep feature extraction model.

[0037] In terms of the model structure, it is mainly composed of two main branches, namely the Bi-TCN branch and the Bi-ILSTM branch:

[0038] Bi-TCN branch: It consists of a wide convolutional layer, a pooling layer, and three layers of bidirectional temporal convolutional modules. Convolution kernels with different dilation rates (1, 2, 3) are used to perform multi-scale modeling on the input signal to obtain long-distance temporal dependence features in the forward and backward directions, making up for the limitation that the traditional TCN structure can only model historical information.

[0039] Bi-ILSTM branch: It is composed of two layers of stacked bidirectional improved LSTM. The improved LSTM network obtains ILSTM (Res-LSTM) by introducing a residual gate. It contains 16 and 32 gated units respectively, and the Dropout ratio is set to 0.2, focusing on learning the time series correlation and trend features in the vibration signal.

[0040] The feature dimensions finally output by the two branches are both 128-dimensional and are fed into the feature fusion attention mechanism module. According to the feature information extracted from each path, this module calculates dynamic weighting coefficients through a lightweight attention network, performs weighted fusion on the outputs of Bi-TCN and Bi-ILSTM, and realizes the adaptive integration between features.

[0041] The fused features are output through a fully connected layer to obtain the final bearing RUL prediction value. The loss function uses the mean square error (MSE), the optimizer uses Adam, the initial learning rate is set to 0.001, and 50 epochs are used for iterative training. During the training process, data is input in a mini-batch manner, the batch size is 32, and GPU acceleration training is enabled to improve the model fitting efficiency.

[0042] In the model testing stage, vibration signal samples from different working conditions and degradation stages are input, and the model can output the corresponding RUL prediction results in real time. Through experimental evaluation on the PHM2012 bearing public dataset, the model proposed in the present invention has good effects in terms of prediction accuracy and stability, the average RMSE error is significantly reduced, and it shows better response sensitivity especially in the early stage of degradation and near the mutation points. Further experimental results also verify the importance of multi-channel input, and it is found that multi-channel fusion input has lower errors and higher trend capture ability during the whole life prediction process.

[0043] To sum up, the present invention constructs a parallel feature extraction structure of Bi-TCN and Bi-ILSTM, combines the feature fusion attention mechanism and the multi-sensor input strategy, realizes the deep modeling of the vibration signal of the bearing throughout its life cycle and the high-precision remaining life prediction, is applicable to the health monitoring and intelligent maintenance system of industrial equipment, and has good engineering generalizability and industrial application prospects.

Claims

1. A bearing life prediction method for double-branch feature extraction under a long time series, characterized in that It includes the following steps: (1) Normalize the multi-sensor data collected, and achieve efficient fusion of multi-source information through the Transformer network, which is used as the training set for the dual-branch deep feature extraction model; (2) Construct parallel Bi-TCN and Bi-ILSTM models; (3) Set the parameters of the parallel Bi-TCN and Bi-ILSTM models; (4) Divide the multi-sensor fusion data into training bearing data and test bearing data for offline modeling training; (5) Input the test data into the trained prediction network model for online prediction; (6) Evaluate the accuracy of the prediction results through two error metrics, MAE and RMSE.

2. The method according to claim 1, wherein In step (1), the vibration signals collected by the multi-sensor are normalized and then efficiently fused through the Transformer network. The normalization is defined as: Among them, X t is the signal after normalization, X is the original signal, X min is the minimum value of the original signal, X max is the maximum value of the original signal.

3. The method according to claim 1, wherein In step (2), a parallel feature extraction network model is constructed: one branch consists of two consecutive bidirectional improved long short-term memory networks, and a residual gate is added to improve the long short-term memory network. The other branch consists of a wide convolutional layer, a max pooling layer, and three consecutive bidirectional temporal convolutional networks; the features extracted by the two branches are fused through a feature fusion attention mechanism, and finally the prediction result is output through a fully connected layer.

4. The method according to claim 1, characterized in that, In step (3), the parameter setting should meet the requirement of reducing the model training burden while avoiding the grid effect problem of dilated convolution. The model training method is multi-batch training.

5. The method according to claim 1, wherein In step (4) during the offline modeling stage, the training samples are input into the constructed neural network model for iterative optimization; in each round of training, the model calculates the loss value according to the error between the predicted output and the actual label, and gradually adjusts the network parameters through the backpropagation mechanism; when the training round number i reaches the set maximum iteration number N, the model completes the training process and the parameters converge stably.

6. The method according to claim 1, wherein In step (5), the test data is input into the trained prediction network model for real-time prediction, and evaluation metrics are set to verify the prediction performance of the model. The evaluation metrics are the mean absolute error and the root mean square error.

7. The method according to claim 1, characterized in that, In step (6), the prediction performance of the prediction network model is evaluated using MAE and RMSE. The definitions of these two error metrics, MAE and RMSE, are: Among them, the average absolute error (MAE) and root mean square error (RMSE) are used in this method to evaluate the accuracy of the prediction results; the error term er t represents the gap between the predicted RUL and the true RUL at the t-th moment, where n represents the total duration of the bearing from the initial state to failure; the smaller the values of the two indicators, the more accurate the prediction results and the better the performance of the model.

Images