Bearing Health State Evaluation Method Based on Parallel Encoding-Decoding Observer

Abstract

The present invention discloses a method for evaluating the health state of a bearing based on a parallel encoding-decoding observer. A sample set A is generated based on a number of vibration signal samples in the normal working state of the bearing under preset working conditions, and a number of samples are selected from the sample set A to form a training sample set B. A parallel encoding-decoding observer including a position encoding module, an encoding input module, an encoder, a decoding input module, a decoder, and a time step truncation module is constructed and trained using the training sample set B. The samples in the sample set A are input into the trained observer to obtain predicted signals. The predicted vibration signal sequence is subtracted from the true vibration signal sequence to obtain a corresponding residual signal sequence, and signal features are extracted therefrom as a health benchmark. For the moment to be evaluated, a residual signal is obtained through the vibration signal and the predicted signal, and signal features are extracted, and an evaluation is performed based on the health benchmark to obtain a corresponding health score. The present invention can effectively improve the accuracy of bearing health state evaluation.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing health status, and more specifically, relates to a bearing health status assessment method based on a parallel encoding-decoding observer. Background Art

[0002] Bearings are widely used in mechanical and electrical equipment. Their core functions include supporting rotating components, bearing dynamic loads, ensuring smooth movement and precise positioning among components of the equipment, and are key components in many industrial applications. With the progress of industrial technology and the increasing complexity of the production environment, the health status of bearings has a direct impact on the overall reliability and safety of the equipment. At the same time, the industrial community's requirements for the safety of mechanical systems continue to increase. Early fault detection of rotating machinery can reduce the risk of damage, thereby avoiding high emergency repair costs and improving the operating reliability of the equipment. Therefore, accurately assessing the health status of bearings has become a research focus in recent years.

[0003] Traditional bearing health assessment methods mainly rely on physical and empirical models or simple pattern analysis, but such methods still face some challenges when dealing with nonlinear and complex systems. For example, when the dynamic behavior of the system is affected by multiple factors, the response characteristics of the system change under different operating conditions, and those signals containing a large amount of noise or other interference factors, these methods often have difficulty accurately identifying and analyzing the true health status of the bearings, resulting in inaccurate assessment results or increasing the risk of misjudgment.

[0004] Data-driven health assessment methods rely on data collected from systems or equipment, and usually use statistical, machine learning or deep learning techniques to analyze and interpret this data. This method is particularly applicable when physical models are too complex or incomplete. It can automatically extract features from a large amount of monitoring data and then use these features for health assessment or fault prediction. Statistic-based methods usually require a large amount of data and computing power. Machine learning-based methods usually rely on manual feature extraction and can only obtain shallow features, which limits their application in complex scenarios.

[0005] The health assessment method based on data-driven deep learning relies on the powerful analytical ability of neural networks and can fully exploit the deep features in data. In particular, convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely applied in health assessment based on deep learning. For example, a hierarchical gated recurrent unit network is used to predict the health status of rolling bearings, an automatic bearing fault diagnosis scheme using two-dimensional representation of wavelet packet transform and CNN, and a deep joint CNN based on attention mechanism for bearing performance degradation assessment, etc. These methods essentially directly extract features from signal data to obtain desired fault patterns, health indicators, etc. However, there are problems such as large redundancy of input information and various noises in modeling the original vibration signals of bearings. Another approach is to extract features based on the residuals of bearing vibration signals, that is, use a neural network observer to learn the temporal mechanism of the signal, construct an estimate of the current signal, and obtain the residual signal of the feature to be extracted by taking the difference between the estimated value and the true value. Therefore, how to design an efficient signal observer and a reasonable estimated signal evaluation index (loss function) is the core issue of such methods. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for evaluating the health status of bearings based on a parallel encoding-decoding observer. An encoder is constructed based on parallel long short-term memory networks (LSTMs) and temporal convolutional networks (TCNs) to more accurately extract the features of bearing vibration signals, thereby improving the accuracy of bearing health status assessment.

[0007] To achieve the above-mentioned invention purpose, the method for evaluating the health status of bearings based on a parallel encoding-decoding observer of the present invention includes the following steps:

[0008] S1: Obtain a number of vibration signal samples of the bearing in the normal working state under preset working conditions. Each vibration signal sample contains vibration signals at S consecutive acquisition times. Use a sliding window of length L to intercept D vibration signal sub-samples, where D = S - L + 1, L = L S +L P , L S represents the preset known vibration signal length, and L P represents the preset predicted vibration signal length, L S > L P ; The first L S vibration signals in each vibration signal sub-sample form an input sequence The last L P vibration signals form a label output sequence Thus, a sample set A is obtained; several training samples are selected from the sample set A to form a training sample set B;

[0009] S2: Construct a parallel encoding-decoding observer, including an input sequence preprocessing module, a position encoding module, an encoding input module, an encoder, a decoding input module, a decoder, and a time step truncation module, where:

[0010] The input sequence preprocessing module is used to perform high-dimensional mapping on the input sequence to obtain an input matrix H represents a preset dimension;

[0011] The position encoding module is used to generate a position encoding and send it to the encoding input module and the decoding input module;

[0012] The encoding input module is used to superimpose the input sequence X and the position encoding P to obtain an encoded input sequence X′ and send it to the encoder;

[0013] The encoder is used to extract a feature F from the encoded input sequence X′ and send it to the decoding input module; The encoder includes a long short-term memory network, a temporal convolutional network, and a feature fusion module. The long short-term memory network is used to extract a feature from the encoded input sequence X′ [[ID=I3]]and send it to the feature fusion module, where L P <N ≤ L S , the temporal convolutional network is used to extract a feature from the encoded input sequence X′ and send it to the feature fusion module. The feature fusion module is used to superimpose the feature f LSTM and the feature f TCN to obtain a feature F;

[0014] The decoding input module is used to superimpose the feature F, the first N rows of the input sequence X, and the first N rows of the position encoding P to obtain an encoded feature sequence F′ and send it to the decoder;

[0015] The decoder is used to perform decoding processing on the encoded feature sequence F′ to obtain a decoded feature sequence and send it to the time step truncation module;

[0016] The time step truncation module is used to truncate the first L P values from the decoded feature sequence F″ as the predicted output sequence

[0017] S3: Use the training sample set B obtained in step S1 to train the parallel encoding-decoding observer constructed in step S2 to obtain a trained parallel encoding-decoding observer;

[0018] S4: Input each input sequence in the training sample set A into the parallel encoding-decoding observer trained in step S3 to obtain a predicted output sequence; for the D predicted output sequences corresponding to each vibration signal sample, the first predicted signal in each predicted output sequence is used to form a predicted vibration signal sequence in chronological order, and the vibration signal at the corresponding moment is extracted from the vibration signal sample to form a real vibration signal sequence, and the predicted vibration signal sequence and the real vibration signal sequence are subtracted to obtain a residual signal sequence corresponding to the vibration signal sample; set K signal statistical features according to actual needs, calculate K signal statistical feature values of each residual signal sequence, average the signal statistical feature values of all residual signal sequences, and use the vector composed of the average values of the K signal statistical features as the health benchmark

[0019] S5: Obtain and record vibration signals in real time during the entire life cycle of the bearing; when a health assessment is required, obtain the D real vibration signals x closest to the time to be assessed. d , d=1,2,…,D; for each real vibration signal x d , get its previous L S The vibration signals at each acquisition moment constitute the input sequence Input it into the parallel encoder-decoder observer trained in step S3 to obtain the predicted output sequence, from which the vibration signal x is obtained. d The corresponding prediction signal y d ; The predicted signal sequence (y1,y2,…,y D ) and the real vibration signal sequence (x1, x2, ..., x D ) to obtain the residual signal, calculate the K signal statistical features corresponding to the residual signal sequence and form a feature vector

[0020] S6: Calculate the similarity Sim between the feature vector f and the healthy reference F;

[0021] S7: Use a preset normalization function to map the similarity Sim to the range of (0, 1), and use the obtained value as the health score of the bearing at the time to be evaluated.

[0022] The bearing health state evaluation method based on a parallel encoding-decoding observer of the present invention obtains a number of vibration signal samples under the normal working state of the bearing under preset working conditions to form a sample set A, and selects a number of samples from the sample set A to form a training sample set B. A parallel encoding-decoding observer including a position encoding module, an encoding input module, an encoder, a decoding input module, a decoder, and a time step truncation module is constructed and trained using the training sample set B. The samples in the sample set A are input into the trained parallel encoding-decoding observer to obtain a predicted signal. The predicted vibration signal sequence is subtracted from the true vibration signal sequence to obtain a corresponding residual signal sequence, and signal features are extracted therefrom as a health benchmark. For the moment to be evaluated, a residual signal is obtained through the vibration signal and the predicted signal, and signal features are extracted, and an evaluation is performed based on the health benchmark to obtain a corresponding health score.

[0023] The present invention has the following beneficial effects:

[0024] 1) The present invention comprehensively considers problems such as large redundancy of input information and various noises existing in the modeling of the original vibration signal, extracts features using the residual signal of the vibration signal, and the parallel encoding-decoding observer combines the advantages of LSTM in dealing with long-term dependence problems, the better performance and longer memory ability of the temporal convolutional network in multiple sequence learning tasks, and the characteristic that the self-attention mechanism can directly capture long-distance dependence relationships.

[0025] 2) The present invention designs a multi-objective loss function, which not only includes the mean square error on the time-domain signal, but also considers the spectrum constraint and the smoothness constraint, making the observed value of the signal more reasonable.

[0026] 3) Compared with other common neural network-sequence prediction models, the parallel encoding-decoding observer designed by the present invention can better obtain the temporal mechanism of the bearing vibration signal, and takes into account the forward calculation speed of the observer network, providing the possibility for deploying this method on some edge devices. Description of the Drawings

[0027] Figure 1 is a flowchart of the specific implementation of the bearing health state evaluation method based on the parallel encoding-decoding observer of the present invention;

[0028] Figure 2 is a structural diagram of the parallel encoding-decoding observer in the present invention;

[0029] Figure 3 is a normalized function curve graph under different values;

[0030] Figure 4 is a time-domain and frequency-domain comparison graph before and after removing high-frequency noise by a Butterworth low-pass filter in this embodiment;

[0031] Figure 5 It is a comparison chart of the loss function curves of bearing 1-1 in various observer models;

[0032] Figure 6 It is a comparison chart of the loss function curves of bearing 1-2 in various observer models;

[0033] Figure 7 It is a comparison chart of the loss function curves of bearing 1-3 in various observer models;

[0034] Figure 8 It is a comparison chart of the loss function curves of bearing 2-1 in various observer models;

[0035] Figure 9 It is a comparison chart of the loss function curves of bearing 2-2 in various observer models;

[0036] Figure 10 It is a comparison chart of the loss function curves of bearing 2-3 in various observer models;

[0037] Figure 11 It is a comparison chart of the loss function curves of bearing 3-1 in various observer models;

[0038] Figure 12 It is a comparison chart of the loss function curves of bearing 3-2 in various observer models;

[0039] Figure 13 It is a real-time health assessment curve graph of three bearings obtained by using different neural network observers;

[0040] Figure 14 It is the real-time health assessment curve of the bearing under different working conditions in this embodiment. Detailed implementation manner

[0041] The following describes the detailed implementation manner of the present invention with reference to the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0042] Embodiment

[0043] Figure 1 It is a flowchart of the detailed implementation manner of the bearing health state assessment method based on the parallel encoding-decoding observer of the present invention. As Figure 1 shown, the specific steps of the bearing health state assessment method based on the parallel encoding-decoding observer of the present invention include:

[0044] S101: Obtain the normal vibration signal samples of the bearing:

[0045] Obtain a number of vibration signal samples in the normal working state of the bearing under preset working conditions. Each vibration signal sample contains vibration signals at S consecutive acquisition moments. Use a sliding window with a length of L to intercept D vibration signal sub-samples, where D = S - L + 1, and L = L S + L P , L S represents the preset known vibration signal length, and L P represents the preset predicted vibration signal length, and L S > L P . For each vibration signal sub-sample, the first L S vibration signals form the input sequence The last L P vibration signals form the label output sequence Thus, the sample set A is obtained. Select a number of training samples from the sample set A to form the training sample set B.

[0046] The vibration signals can use offline data or online data, where the offline data is obtained in advance, and the online signals are collected in real time by sensors. In this embodiment, the vibration signals of the bearing include horizontal vibration signals and vertical vibration signals, which are obtained by using a horizontal vibration acceleration sensor and a vertical acceleration sensor respectively. In this embodiment, first continuously obtain vibration signals at a number of moments, and use a sliding window to intercept a sequence with a length of L each time, so as to obtain vibration signal samples. Note that cross-file interception is not allowed during interception, that is, ensure that the vibration signals in the sequence are equidistant in time.

[0047] In practical applications, in order to reduce interference, a Butterworth low-pass filter can be used to remove high-frequency noise from each vibration signal sample.

[0048] S102: Construct a parallel encoder-decoder observer:

[0049] In order to realize the estimation of the bearing vibration signal, a parallel encoder-decoder observer based on an encoder-decoder is constructed in the present invention. Figure 2 is the structural diagram of the parallel encoder-decoder observer in the present invention. As Figure 2 shown, the main structure of the parallel encoder-decoder observer in the present invention includes an input sequence preprocessing module, a position encoding module, an encoding input module, an encoder, a decoding input module, a decoder, and a time step interception module. Next, each module will be described in detail.

[0050] The input sequence preprocessing module is used to perform high-dimensional mapping on the input sequence to obtain the input matrix H represents the preset dimension. In this way, stronger signal characterization can be obtained.

[0051] The position encoding module is used to generate position encoding and send it to the encoding input module and the decoding input module. Using position encoding can introduce sequence order information. In this embodiment, the elements in the position encoding P are generated by the following formula:

[0052]

[0053] where h = 1, 2, …, H.

[0054] The encoding input module is used to superimpose the input sequence X with the position encoding P to obtain the encoded input sequence X′ and send it to the encoder.

[0055] The encoder is used to extract the feature F from the encoded input sequence X′ and send it to the decoding input module. The encoder includes a long short-term memory network (LSTM), a temporal convolutional network (TCN), and a feature fusion module. The long short-term memory network is used to extract the feature from the encoded input sequence X′ and send it to the feature fusion module, where L P <N≤L S , the temporal convolutional network is used to extract the feature from the encoded input sequence X′ and send it to the feature fusion module. The feature fusion module is used to superimpose the feature f LSTM and the feature f TCN to obtain the feature F. In the present invention, LSTM is used to solve the defect of the standard latent variable neural network model in dealing with long-term dependence problems. TCN is a convolutional neural network for sequence modeling, which has better performance and longer memory ability in multiple sequence learning tasks.

[0056] The decoding input module is used to superimpose the feature F, the first N rows of the input sequence X, and the first N rows of the position encoding P to obtain the encoded feature sequence F′ and send it to the decoder.

[0057] The decoder is used to perform decoding processing on the encoded feature sequence F′ to obtain the decoded feature sequence and send it to the time step truncation module. In the decoder module of the present invention, a self-attention module is adopted. By calculating the correlation between any two positions in the sequence, long-distance dependence relationships can be directly captured, and long sequences can be better modeled. The self-attention module is a commonly used neural network module, and its specific process will not be elaborated here.

[0058] The time step truncation module is used to truncate the first L P values from the decoded feature sequence F″ as the predicted output sequence

[0059] S103: Train the parallel encoding-decoding observer:

[0060] Using the training sample set B obtained in step S101 to train the parallel encoding-decoding observer constructed in step S102 to obtain a trained parallel encoding-decoding observer.

[0061] To improve the training effect, the present invention adopts a multi-objective loss function when training the parallel encoding-decoding observer, which is composed of the weighted sum of a regression loss, a spectral loss, and a smoothness loss. The specific calculation formula is as follows:

[0062] L MT = λ1L R + λ2L SP + λ3L SM

[0063] Among them, L MT represents the multi-objective loss function. L R represents the regression loss, which is used to measure the difference between the prediction result of the regression model and the true value. In this embodiment, the regression loss adopts the mean square error MSE of the predicted output sequence and the label output sequence.

[0064] L SP represents the spectral loss, and its calculation method is as follows: The input sequence and the predicted output sequence are used to form a new vibration signal sub-sample V pred , and the Fourier transform is respectively performed on the original vibration signal sub-sample V orig and the vibration signal sub-sample V pred to obtain spectral signals, and the mean square error of the two spectral signals is calculated as the spectral loss.

[0065] L SM represents the smoothness loss, and the average value of the second-order difference of the vibration signal sub-sample V pred is calculated as the smoothness loss.

[0066] λ1, λ2, λ3 represent the weighting factors corresponding to the regression loss L R , the spectral loss L SP , and the smoothness loss L SM . Due to the high-frequency characteristics of the vibration signal, the weighting factor λ3 of the smoothness loss L SM should preferably be the smallest compared to other weighting factors, that is, λ3 < λ1 & λ3 < λ2, so that the model does not overly focus on the smoothness of the prediction sequence.

[0067] S104: Extract the health benchmark:

[0068] Input each input sequence in the training sample set A into the parallel encoding-decoding observer trained in step S103 to obtain a predicted output sequence. For the D predicted output sequences corresponding to each vibration signal sample, form a predicted vibration signal sequence by arranging the first predicted signal in each predicted output sequence in chronological order, extract the vibration signal at the corresponding moment from the vibration signal sample to form a true vibration signal sequence, and subtract the predicted vibration signal sequence from the true vibration signal sequence to obtain the residual signal sequence corresponding to this vibration signal sample. Set K signal statistical features according to actual needs, calculate the K signal statistical feature values of each residual signal sequence, average the respective signal statistical feature values of all residual signal sequences, and use the vector composed of the K signal statistical feature averages as the health benchmark In this embodiment, the signal features include the average absolute value, peak value, and root mean square value of the residual signal.

[0069] S105: Obtain the vibration signal features during the whole-life operation:

[0070] During the operation of the bearing throughout its life cycle, obtain the vibration signal in real time and record it. When health assessment is required, obtain the D true vibration signals x closest to the moment to be evaluated d , d = 1, 2, …, D. For each true vibration signal x d , obtain the vibration signals at its previous L S acquisition moments to form an input sequence Input it into the parallel encoding-decoding observer trained in step S103 to obtain a predicted output sequence, and obtain the predicted signal y corresponding to the vibration signal x from it d . Subtract the predicted signal sequence (y1, y2, …, y d ) from the true vibration signal sequence (x1, x2, …, x D ) to obtain a residual signal, calculate the K signal statistical features corresponding to this residual signal sequence and form a feature vector D

[0071] S106: Calculate the feature similarity:

[0072] Calculate the similarity Sim between the feature vector f and the health benchmark F. In this embodiment, the similarity Sim uses the Mahalanobis distance between the signal feature value vector f and the health benchmark vector F.

[0073] S107: Calculate the health score:

[0074] Use a preset normalization function to map the similarity Sim to the range of (0, 1), and use the obtained value as the health score of the bearing at the acquisition moment to be evaluated.

[0075] In this embodiment, the expression of the normalization function is:​

[0076]

[0077] Among them, p represents a preset parameter for changing the steepness of the function. Figure 3 is the normalized function curve graph under different p values. As Figure 3 shown, the smaller the p value, the steeper the normalized function, the faster the convergence, and the p value can be set according to the actual situation in practical applications.

[0078] Embodiment

[0079] To better illustrate the technical solution and technical effect of the present invention, a specific example is adopted to analyze and explain the working process and technical effect of the present invention. This embodiment is based on the XJTU-SY Bearing Datasets collected by the team of Professor Lei Yaguo from Xi'an Jiaotong University.

[0080] For data acquisition, a special test platform for bearing accelerated life test is used. This platform consists of an AC motor, a speed controller, a bearing loading system, etc., and can carry out full life cycle tests of rolling bearings or sliding bearings under different working conditions. The main adjustable parameters include speed and radial force. The radial force is loaded onto the bearing housing by a hydraulic system, and the speed is set by the motor controller. When collecting data, 3 different working conditions are designed, and 5 bearings are tested for each working condition. Table 1 is the data table of the speed and radial force of each working condition in this embodiment.

[0081]

[0082] Table 1

[0083] To collect the bearing vibration signals, PCB 352C33 single-axis acceleration sensors are installed in the horizontal and vertical directions of the bearing respectively. The DT9837 portable dynamic signal acquisition system is adopted, the sampling frequency is set to 25.6 kHz, and 1.28 seconds are sampled every 1 minute. The storage format of the vibration signal is a CSV file, where the first column is the horizontal signal, the second column is the vertical signal, and the file name is numbered in sequence and contains the signals of the full life cycle of the bearing. In this embodiment, several typical working conditions and datasets under fault modes in the original dataset are selected for performance comparison of the observer and health assessment under full-cycle operation. Table 2 is the information table of the datasets under typical working conditions in this embodiment.

[0084]

[0085]

[0086] Table 2

[0087] The main software platforms used in this embodiment are: Python 3.10, PyTorch 1.13, CUDA 12.2; the main hardware platforms are CPU: Intel i7 10750H, GPU: Nvidia GTX 1650Ti 4G.

[0088] Figure 4 It is the time-domain and frequency-domain comparison diagram of the bearing signal before and after removing high-frequency noise by the Butterworth low-pass filter in this embodiment. As Figure 4 shown, it is the time-domain and frequency-domain comparison diagram of the horizontal vibration signal of bearing 1-1 in the 1.csv file, where the cut-off frequency is 10 kHz and the filter order is 4.

[0089] For the purpose of comparing technical effects, in this embodiment, various types of observer models, including Generalized Regression Neural Network (GRNN), LSTM, TCN, and the Parallel Encoder-Decoder (PED) in the present invention, are trained respectively on 8 groups of bearing data sets with different working conditions and fault modes. Due to the differences in the ability of various observers to learn sequence patterns, the trained models will have different degrees of modeling effects on the time-series characteristics of bearing signals. By comparing and analyzing these trained observer models on the test set, the effects of different observers in extracting signal time-series characteristics and constructing residuals can be evaluated.

[0090] In addition, in this embodiment, the data in the normal working state in the early stage of the bearing full life cycle signal is used as the training set, and the test set is the data in other normal working states that are not included in the training set. And it is ensured that the signal data sets with the same bearing number are input into each observer model simultaneously.

[0091] Figure 5 It is the comparison diagram of the loss function curves of bearing 1-1 in various observer models. Figure 6 It is the comparison diagram of the loss function curves of bearing 1-2 in various observer models. Figure 7 It is the comparison diagram of the loss function curves of bearing 1-3 in various observer models. Figure 8 It is the comparison diagram of the loss function curves of bearing 2-1 in various observer models. Figure 9 It is the comparison diagram of the loss function curves of bearing 2-2 in various observer models. Figure 10 It is the comparison diagram of the loss function curves of bearing 2-3 in various observer models. Figure 11 It is the comparison diagram of the loss function curves of bearing 3-1 in various observer models. Figure 12 It is the comparison diagram of the loss function curves of bearing 3-2 in various observer models. As Figures 5 to 12As shown, the displayed losses include the regression losses of the horizontal and vertical signals, and the spectral loss. Since GRNN does not require training, it is not listed. During training, the Adam optimizer is used, training for 100 epochs with a batch size of 256. From Figures 5 to 12 It can be seen that the loss function of the PED of the present invention drops the fastest and has the smallest convergence value, while LSTM and TCN each have a relatively small convergence value for bearing data with different numbers. In addition, there is an unstable phenomenon in the change of the loss value during the training of LSTM. This is because LSTM transmits historical information through hidden states and has limited long-term dependence ability, and cannot accurately remember long-distance information. Therefore, PED with a relatively stable loss curve change and a small convergence value is a better choice.

[0092] After the observer training is completed, the entire time period of the data set (full-life operation data) is used to simulate the online vibration signal and fed into the observer. Table 3 is the comparison table of bearing performance of the horizontal vibration signal in different observers in this embodiment. Table 4 is the comparison table of bearing performance of the vertical vibration signal in different observers in this embodiment.

[0093]

[0094] Table 3

[0095]

[0096] Table 4

[0097] The first column in Table 3 and Table 4 represents the bearing number, and the defined performance indicators are as follows:

[0098] (1) MSE: Mean Square Error. It is used to measure the error between the true value and the predicted value of the bearing vibration sequence, and its value is the average of the squares of the differences between the predicted value and the true value.

[0099] (2) SL: Spectral Loss. It is used to measure the difference between the predicted signal sequence and the true signal sequence in the frequency domain, and its value is the mean square error of the Fourier transform sequences of the entire predicted signal (the predicted time steps are concatenated into one whole) and the true signal.

[0100] (3) SFC: Forward Computation Speed. It is used to measure the speed at which the observer generates observation values, and its value is approximately equal to the number of samples inferred per second. The specific expression is

[0101]

[0102] where 32768 is the number of data points in 1 CSV file, and the test set uses the sample sequence generated by 1 CSV file. For simplicity, the total number of sample sequences not being 32768 due to the sample length is ignored. T cIndicates the time consumed for inferring these samples. The unit of SFC is k samples / s.

[0103] As can be seen from Table 3 and Table 4, the error between the predicted signal and the true signal output by the PED observer of the present invention is the smallest in both the frequency domain and the time domain, indicating that it can best learn the internal timing mechanism of the signal. Since GRNN is a model similar to a radial basis network and has no learnable parameters inside, it is necessary to calculate the distance norm between all training samples during the inference stage. Therefore, its SFC is the smallest, and it is difficult to meet the requirements of real-time monitoring of the system when deploying the health assessment algorithm to some edge devices. While taking accuracy into account, PED has a relatively fast SFC, approximately 6 to 7 times that of GRNN. Compared with simple networks such as LSTM and TCN, the loss of forward calculation speed is acceptable. Therefore, using PED as an observer takes into account both the accuracy and real-time performance of signal prediction.

[0104] Figure 13 is a real-time health assessment curve graph of the health of three bearings obtained by using different neural network observers. As Figure 13 shown, the faults demonstrated in this embodiment include inner ring & outer ring faults, outer ring faults, and inner ring faults. The health value ranges from 0 to 1 after normalization. During the construction of the curve, all other parameters remain the same except for the observer type. From Figure 13 it can be seen that for bearings 1-3, the evaluation curves corresponding to the three types of observers are relatively close, and only the health value corresponding to the PED observer is lower. For bearings 2-3, the evaluation method corresponding to the PED observer has a lower tolerance for faults. Especially at the end of the curve, it can detect faults earlier and provide early warnings, while the curves corresponding to other observers still maintain a relatively high health value at this time. For bearings 3-1, compared with the curve corresponding to LSTM, there are larger spikes, and the beginning of the curve corresponding to the PED observer is smoother, that is, it has a more stable baseline. In summary, the health assessment curve generated by the health assessment method using the PED observer of the present invention has a more stable baseline, earlier warning signs, and lower tolerance for faults.

[0105] Figure 14 is a real-time health assessment curve graph of the bearing health under different working conditions in this embodiment. From Figure 14From the changing trend of the curve, it can be seen that when an inner race fault occurs, the health of the bearing will drop sharply. When an outer race fault and a cage fault occur, the health of the bearing shows a gradual downward trend. This is in line with the characteristics of different faults. The inner race fault has a greater impact and will cause a drastic deterioration of the health state, while the outer race and cage faults have a more gradual impact on the health state. This evaluation curve can intuitively reflect the health degree of the bearing at different time periods during its entire life cycle, verifying that the method in this paper can effectively use the similarity of residual signals to evaluate the running state of the bearing and providing support for the fault prediction and condition detection of the bearing.

[0106] Although the above description of the illustrative specific embodiments of the present invention is provided for the understanding of those skilled in the art of the present technology, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

Claims

1. A bearing health state evaluation method based on a parallel encoding-decoding observer, characterized in that It includes the following steps: S1: Obtain a number of vibration signal samples in the normal working state of the bearing under the preset working conditions. Each vibration signal sample contains vibration signals at S consecutive acquisition moments. Use a sliding window with a length of L to intercept D vibration signal sub-samples, where D = S - L + 1, and L = L S + L P , L S represents the preset known vibration signal length, and L P represents the preset predicted vibration signal length, and L S > L P ; For each vibration signal sub-sample, the first L S vibration signals form the input sequence The last L P vibration signals form the label output sequence Thus, the sample set A is obtained; Select a number of training samples from the sample set A to form the training sample set B; S2: Construct a parallel encoding-decoding observer, including an input sequence preprocessing module, a position encoding module, an encoding input module, an encoder, a decoding input module, a decoder, and a time step truncation module, where: The input sequence preprocessing module is used to perform high-dimensional mapping on the input sequence to obtain an input matrix where H represents a preset dimension; The position encoding module is used to generate position encoding and send it to the encoding input module and the decoding input module; The encoding input module is used to superimpose the input sequence X and the position encoding P to obtain an encoded input sequence X' and send it to the encoder; The encoder is used to extract the feature F from the encoded input sequence X' and send it to the decoding input module. The encoder includes a long short-term memory network, a temporal convolutional network, and a feature fusion module. The long short-term memory network is used to extract the feature from the encoded input sequence X'. and send it to the feature fusion module, where L P <N≤L S , the temporal convolutional network is used to extract the feature from the encoded input sequence X'. and send it to the feature fusion module. The feature fusion module is used to stack the feature f LSTM and the feature f TCN to obtain the feature F; The decoding input module is used to superimpose the feature F, the first N rows of the input sequence X, and the first N rows of the position encoding P to obtain an encoded feature sequence F' and send it to the decoder; The decoder is used to decode the encoded feature sequence F′ to obtain the decoded feature sequence and send it to the time step truncation module; The time step truncation module is used to truncate the first L values from the decoded feature sequence F″ as the predicted output sequence P ​ S3: Use the training sample set B obtained in step S1 to train the parallel encoding-decoding observer constructed in step S2 to obtain a trained parallel encoding-decoding observer; S4: Input each input sequence in the training sample set A into the parallel encoding-decoding observer trained in step S3 to obtain a predicted output sequence; for the D predicted output sequences corresponding to each vibration signal sample, form a predicted vibration signal sequence by arranging the first predicted signal in each predicted output sequence in chronological order, extract the vibration signals at the corresponding moments from the vibration signal sample to form a true vibration signal sequence, and subtract the predicted vibration signal sequence from the true vibration signal sequence to obtain the residual signal sequence corresponding to the vibration signal sample; set K signal statistical features according to actual needs, calculate the K signal statistical feature values of each residual signal sequence, average the various signal statistical feature values of all residual signal sequences, and use the vector composed of the K signal statistical feature averages as the health benchmark S5: Obtain and record vibration signals in real time during the operation of the bearing throughout its life cycle; when health assessment is required, obtain the D true vibration signals x closest to the moment to be evaluated d , where d = 1, 2, …, D; for each true vibration signal x d , obtain the vibration signals at its previous L S acquisition moments to form an input sequence Input it into the parallel encoding-decoding observer trained in step S3 to obtain a predicted output sequence, and obtain the predicted signal y corresponding to the vibration signal x d from it d ; Take the difference between the predicted signal sequence (y1, y2, …, y D ) and the true vibration signal sequence (x1, x2, …, x D ) to obtain a residual signal, calculate the K signal statistical features corresponding to this residual signal sequence and form a feature vector S6: Calculate the similarity Sim between the feature vector f and the healthy benchmark F; S7: Use a preset normalization function to map the similarity Sim to the range of (0,1), and use the obtained value as the health score of the bearing at the moment to be evaluated.

2. The bearing health state evaluation method according to claim 1, wherein In step S1, a Butterworth low-pass filter is used to remove high-frequency noise from each vibration signal sample.

3. The bearing health state evaluation method according to claim 1, characterized in that, The position encoding in step S2 has its elements generated by the following formula: where \(t = 1, 2, \ldots, L\) S and \(h = 1, 2, \ldots, H\).

4. The bearing health state evaluation method according to claim 1, characterized in that, In step S2, the decoder module uses a self-attention module.

5. The bearing health state evaluation method according to claim 1, wherein In step S3, a multi-objective loss function is used when training the parallel encoding-decoding observer, and the calculation formula is as follows: L MT = λ1L R + λ2L SP + λ3L SM Among them, L MT represents the multi-objective loss function, and L R represents the regression loss. L SP represents the spectral loss, and its calculation method is as follows: a new vibration signal subsample V pred is formed by the input sequence and the predicted output sequence. Fourier transforms are respectively performed on the original vibration signal subsample V orig and the vibration signal subsample V pred to obtain spectral signals, and the mean square error of the two spectral signals is calculated as the spectral loss; L SM represents the smoothness loss, and the average value of the second-order difference of the vibration signal subsample V pred is calculated as the smoothness loss; λ1, λ2, and λ3 represent the weighted factors corresponding to the regression loss L R , the spectral loss L SP , and the smoothness loss L SM .

6. The bearing health state evaluation method according to claim 5, wherein, The smoothness loss L SM has a weighting factor λ3 that satisfies λ3 < λ1 & λ3 < λ2.

7. The bearing health state evaluation method according to claim 1, wherein, In step S4, the signal features include the average absolute value, peak value, and root mean square value of the residual signal.

8. The bearing health state evaluation method according to claim 1, wherein, In step S6, the similarity Sim is the Mahalanobis distance between the signal feature value vector f and the healthy benchmark vector F.

9. The bearing health state evaluation method according to claim 1, wherein The expression of the normalization function in step S7 is: Where p represents a preset parameter used to change the steepness of the function.

Images