Bearing remaining life prediction model training method, device, medium and prediction method

Abstract

The present disclosure relates to a method, device, medium and prediction method for training a bearing remaining life prediction model, belonging to the field of artificial intelligence technology. The method includes: acquiring a bearing vibration signal; performing adaptive frequency processing on the bearing vibration signal to obtain a fusion feature; performing exponential smoothing processing on the bearing vibration signal to obtain an exponential smoothing output; performing Haar wavelet feature enhancement processing on the bearing vibration signal to obtain a Haar wavelet feature enhancement output; dividing the bearing vibration signal that has undergone adaptive frequency processing, exponential smoothing processing, and Haar wavelet feature enhancement processing to obtain a training set; training an initial bearing remaining life prediction model using the obtained training set to obtain a target bearing life prediction model. The model trained by the present training method can more accurately predict the remaining life of the bearing.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and particularly relates to a bearing remaining life prediction model training method, device, medium and prediction method. Background Art

[0002] Bearings are vital components in various mechanical equipment and are widely used in many fields such as motors, automobiles, aerospace, and industrial automation. With the long-term operation of mechanical equipment, bearings will experience varying degrees of wear and damage, which directly affects the stability and operating efficiency of the equipment. Therefore, accurately predicting the remaining service life of bearings is of great significance for equipment maintenance, fault warning, and performance optimization. However, at present, there is no way to accurately predict the remaining service life of bearings. Summary of the invention

[0003] The present disclosure proposes a bearing remaining life prediction model training method, device, medium and prediction method to solve the problem that the remaining service life of a bearing cannot be accurately predicted.

[0004] According to a first aspect of the present disclosure, a method for training a bearing remaining life prediction model is provided, the method comprising: acquiring a bearing vibration signal; performing adaptive frequency processing on the bearing vibration signal to obtain a fusion feature; performing exponential smoothing processing on the fusion feature to obtain an exponential smoothing output; performing Haar wavelet feature enhancement processing on the exponential smoothing output to obtain a Haar wavelet feature enhanced output; dividing the Haar wavelet feature enhanced output to obtain a training set; using the obtained training set to train an initial bearing remaining life prediction model to obtain a target bearing life prediction model, wherein the initial bearing remaining life prediction model comprises an encoder and a decoder, and the encoder processes the bearing vibration signal The processing steps include: using a one-dimensional convolution layer to perform a convolution operation on the bearing vibration signal, using the output of the one-dimensional convolution layer as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of the cyclic timing convolution module, and using the output of the cyclic timing convolution module as the input of the bidirectional gated recurrent unit; the processing steps of the decoder on the output of the bidirectional gated recurrent unit include: performing an attention mechanism processing on the output of the bidirectional gated recurrent unit, using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of the unidirectional gated unit, and inputting the output of the unidirectional gated unit into the fully connected layer.

[0005] In some embodiments, the adaptive frequency processing of the bearing vibration signal includes: performing a fast Fourier transform on the bearing vibration signal to obtain a frequency domain; calculating the energy of each frequency domain and performing normalization processing to obtain normalized frequency domain energy; generating an adaptive frequency domain mask based on the normalized frequency domain energy and an adaptive threshold parameter; multiplying the frequency domain with the corresponding complex weight to obtain a weighted frequency domain; multiplying the frequency domain with the adaptive frequency domain mask to obtain a masked frequency domain; performing an inverse fast Fourier transform on the weighted frequency domain and the masked frequency domain to obtain a time domain; extracting time domain features through multiple convolutional layers, and fusing multiple features to obtain fused features.

[0006] In some embodiments, the exponential smoothing processing of the bearing vibration signal includes: determining an exponential smoothing weight and an initial smoothing weight; performing Fourier convolution on the bearing vibration signal according to the exponential smoothing weight to obtain a Fourier convolution output; and calculating the exponential smoothing output according to the Fourier convolution output, the initial smoothing weight, and the bearing vibration signal.

[0007] In some embodiments, the Haar wavelet feature enhancement processing of the bearing vibration signal includes: dimensional expansion of the bearing vibration signal to obtain a change result; two-dimensional convolution and Haar wavelet transform of the change result to obtain a transformation result; two-dimensional convolution and transposition of the transformation result, and Haar wavelet transform to obtain the original feature; removing the third dimension of the original feature to obtain a conversion feature; and performing feature fusion on the second dimension of the conversion feature to obtain a Haar wavelet feature enhanced output.

[0008] In some embodiments, the output of the adaptive frequency processing module is used as the input of a circular timing convolution module, wherein the processing of the input data by the circular timing convolution module includes: normalizing the input data to obtain an affine transformation result; performing fast Fourier transform on the affine transformation result to obtain a circular convolution result; performing a first-layer linear transformation and activation on the circular convolution result to obtain an activation result; performing a second-layer linear transformation on the activation result to obtain a second-layer linear transformation result; and denormalizing the second-layer linear transformation result to obtain a denormalized result.

[0009] In some embodiments, the normalization of the input data to obtain an affine transformation result includes: according to the formula: , get the mean of the input data ,in, represents the input data sequence, Indicates the length of the data sequence, Indicates the serial number; according to the formula: , and get the standard deviation ,in, Represents a constant; according to the formula: , the normalized result is obtained ; According to the formula: , the affine transformation result is obtained , where and both represent learnable parameters.

[0010] In some embodiments, performing a fast Fourier transform on the affine transformation result to obtain a circular convolution result includes: According to the formula: , the circular convolution result is obtained , where represents the fast Fourier transform, represents the inverse fast Fourier transform, represents the convolution kernel, represents element-wise multiplication; performing a first-layer linear transformation and activation on the circular convolution result to obtain an activation result includes: According to the formula: , the activation result is obtained , where represents the weight matrix of the fully connected layer, represents the bias term, represents the activation function; performing a second-layer linear transformation on the activation result to obtain a second-layer linear transformation result includes: According to the formula: , the second-layer linear transformation result is obtained , where represents the weight matrix of the fully connected layer, represents the bias term; performing denormalization on the second-layer linear transformation result to obtain a denormalized result includes: According to the formula: , the denormalized result is obtained .

[0011] According to a second aspect of the present disclosure, a bearing remaining life prediction model training device is provided, comprising: an acquisition module for acquiring a bearing vibration signal; an adaptive frequency processing module for performing adaptive frequency processing on the bearing vibration signal to obtain a fusion feature; an exponential smoothing processing module for performing exponential smoothing processing on the fusion feature to obtain an exponential smoothing output; a Haar wavelet feature enhancement processing module for performing Haar wavelet feature enhancement processing on the exponential smoothing output to obtain a Haar wavelet feature enhanced output; a division module for dividing the Haar wavelet feature enhanced output to obtain a training set; a training module for training an initial bearing remaining life prediction model using the obtained training set to obtain a target bearing life prediction model, wherein the initial bearing remaining life prediction model The remaining life prediction model includes an encoder and a decoder. The encoder processes the bearing vibration signal, including: using a one-dimensional convolution layer to perform a convolution operation on the bearing vibration signal, using the output of the one-dimensional convolution layer as the input of an adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of a cyclic timing convolution module, and using the output of the cyclic timing convolution module as the input of a bidirectional gated recurrent unit; the decoder processes the output of the bidirectional gated recurrent unit, including: performing an attention mechanism on the output of the bidirectional gated recurrent unit, using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of a unidirectional gated unit, and inputting the output of the unidirectional gated unit into a fully connected layer.

[0012] According to a third aspect of the present disclosure, a method for predicting the remaining life of a bearing is provided, the method comprising: acquiring a real-time bearing vibration signal; inputting the real-time bearing vibration signal into a bearing remaining life prediction model trained using the bearing remaining life prediction model training method as described above, to predict the remaining life of the bearing.

[0013] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the instructions are executed by a processor, the bearing remaining life prediction model training method or the bearing remaining life prediction method as described above is implemented.

[0014] By adopting the above technical solution, the embodiments of the present disclosure can achieve beneficial technical effects: the present disclosure adopts an adaptive frequency processing method, utilizes Fourier analysis, enhances feature representation, and captures long-term and short-term interactions of bearing vibration information, while reducing noise through adaptive thresholds.

[0015] The present disclosure adopts an exponential smoothing method, and combines fast Fourier transform and exponential weighted smoothing, which can effectively smooth the input signal, reduce the impact of short-term fluctuations, and retain long-term trends.

[0016] The present invention adopts a Haar wavelet feature enhancement processing method, which uses the Haar wavelet to decompose the vibration signal into different frequency sub-bands, extract low-frequency trends and high-frequency fault information respectively, and enhance the ability to capture key features of bearing vibration signals after reducing feature resolution, thereby improving the accuracy and robustness of model prediction.

[0017] The present invention adopts cyclic temporal convolution and reduces the interference of the external environment on data distribution through dynamic de-mean normalization, thereby improving the robustness and generalization ability of the model.

[0018] The present invention adds an adaptive frequency method to both the encoder and decoder stages, thereby improving the model's ability to decode complex time series and further deepening and strengthening the feature mining of vibration signals.

[0019] As mentioned above, first, when pre-processing the historical bearing vibration signal, the present invention sequentially undergoes adaptive frequency processing, exponential smoothing processing, and Haar wavelet feature enhancement processing, thereby enhancing the characteristic representation of the bearing vibration signal, and being able to capture the long-term and short-term interactions of the vibration signal, while removing noise; at the same time, through smoothing processing, the impact of short-term vibrations is reduced; finally, the bearing vibration signal is decomposed into different frequency sub-bands using Haar wavelet feature enhancement, and low-frequency trends and high-frequency fault information are extracted respectively, and the feature resolution is reduced to enhance the ability to capture the key features of the bearing vibration signal. Since the obtained bearing vibration signal pays more attention to the key features, the prediction accuracy of the model can be improved by training the model using these training sets.

[0020] In the convolutional neural network, the encoder combines a one-dimensional convolution layer, an adaptive frequency processing module, a recurrent temporal convolution module, and a bidirectional gated recurrent unit. The encoder can weight and filter the frequency domain representation of the bearing vibration signal, and can effectively focus on the important information within a specific frequency range. By calculating the frequency domain energy of the bearing vibration signal and normalizing it, the model can automatically adjust which frequency components contribute more, thereby achieving dynamic frequency domain enhancement. The decoder combines the attention mechanism and the unidirectional gating unit to calculate the attention weight between the current hidden state and the encoder output to obtain the context vector. Before the unidirectional gating unit input, the adaptive frequency processing module is used to fuse and compress the attention output and embedded features, which helps to reduce redundancy and enhance the model's ability to focus on important features. The input is then spliced ​​with the context vector, and the output of the next time step is generated through the unidirectional gating unit. The improvements in the encoder and decoder enable the model itself to have more accurate prediction capabilities, providing support for bearing maintenance, fault warning, and performance optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings, which constitute a part of the specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0022] The present disclosure can be more clearly understood from the following detailed description with reference to the accompanying drawings.

[0023] Figure 1 is a flowchart illustrating a method for training a bearing remaining life prediction model according to some embodiments of the present disclosure.

[0024] Figure 2 is a flowchart (in its entirety) illustrating a method for training a bearing remaining life prediction model according to some embodiments of the present disclosure.

[0025] Figure 3 is a flow chart showing a method for adaptive frequency processing of bearing vibration signals according to some embodiments of the present disclosure.

[0026] Figure 4 is a flow chart showing a method for exponential smoothing of a bearing vibration signal according to some embodiments of the present disclosure.

[0027] Figure 5 is a flow chart showing a method for enhancing the Haar wavelet features of a bearing vibration signal according to some embodiments of the present disclosure.

[0028] Figure 6 is a flowchart illustrating a method for processing cyclic sequential convolution of input data according to some embodiments of the present disclosure.

[0029] Figure 7 is a schematic diagram showing the prediction results of the bearing 1_3 task in the XJTU bearing dataset according to some embodiments of the present disclosure.

[0030] Figure 8 is a schematic diagram showing the prediction results of the bearing 2_3 task in the XJTU bearing dataset according to some embodiments of the present disclosure.

[0031] Fig. 9 is a block diagram showing a bearing remaining life model training device according to some embodiments of the present disclosure.

[0032] Fig.10 is a block diagram showing a bearing remaining life model training device according to some other embodiments of the present disclosure.

[0033] Fig.11 is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure.

[0034] Fig.12 is a flow chart showing a method for predicting the remaining life of a bearing according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0035] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0036] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0037] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0038] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0039] In all examples shown and discussed herein, any specific values ​​should be understood as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0040] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0041] At present, bearings are vital components in various mechanical equipment and are widely used in many fields such as motors, automobiles, aerospace, and industrial automation. With the long-term operation of mechanical equipment, bearings will experience varying degrees of wear and damage, which directly affects the stability and operating efficiency of the equipment. Therefore, accurately predicting the remaining service life of bearings is of great significance for equipment maintenance, fault warning, and performance optimization. However, at present, there is no way to accurately predict the remaining service life of bearings.

[0042] In view of this, the present disclosure proposes a bearing remaining life prediction model training method, device, medium and prediction method. The present disclosure adopts an adaptive frequency processing method, uses Fourier analysis to enhance feature representation, and captures long-term and short-term interactions of bearing vibration information, while reducing noise through adaptive thresholds. The present disclosure adopts an exponential smoothing processing method, and combines fast Fourier transform and exponential weighted smoothing to effectively smooth the input signal, reduce the impact of short-term toggling, and retain long-term trends. The present disclosure adopts a Haar wavelet feature enhancement processing method, uses Haar wavelets to decompose the vibration signal into different frequency sub-bands, extracts low-frequency trends and high-frequency fault information respectively, and reduces the feature resolution to enhance the ability to capture key features of bearing vibration signals, thereby improving the accuracy and robustness of model prediction. The present disclosure adopts cyclic time series convolution, and reduces the interference of the external environment on data distribution through dynamic mean normalization, thereby improving the robustness and generalization ability of the model. The present disclosure adds an adaptive frequency method to both the encoder and decoder stages, which improves the model's ability to decode complex time series and deepens and strengthens the feature mining of vibration signals. The technology disclosed in this disclosure can more accurately predict the remaining service life of bearings, providing support for equipment maintenance, fault warning and performance optimization.

[0043] First of all, it should be noted that convolutional neural network is a type of feedforward neural network with deep structure and convolution operation. It extracts local features of input data through convolution operation, forms complex feature representation through multi-layer convolution and pooling operation, and finally performs classification or regression tasks through fully connected layer.

[0044] Convolutional neural networks have the following advantages in vibration signal processing: capturing local features. By setting the appropriate convolution kernel size and step size, vibration modes in different frequency ranges can be extracted, thereby better understanding the local structure of the signal; parameter sharing can greatly reduce the number of network parameters; translation invariance is achieved through convolution operations, which can better handle translation changes in signals; low-level convolution layers extract simple local features, and high-level convolution layers extract more abstract and complex features, thereby better representing the structure of vibration signals; CNN can better adapt to different types of vibration signals and has better generalization capabilities.

[0045] Figure 1 FIG. 1 is a flow chart showing a method for training a bearing remaining life prediction model according to some embodiments of the present disclosure. Figure 1 As shown, the bearing remaining life prediction model training method includes steps S110 to S160.

[0046] In step S110, a bearing vibration signal is acquired.

[0047] In step S120, adaptive frequency processing is performed on the bearing vibration signal to obtain a fusion feature.

[0048] In some embodiments, the adaptive frequency processing of the bearing vibration signal includes: performing a fast Fourier transform on the bearing vibration signal to obtain a frequency domain; calculating the energy of each frequency domain and performing normalization processing to obtain normalized frequency domain energy; generating an adaptive frequency domain mask based on the normalized frequency domain energy and an adaptive threshold parameter; multiplying the frequency domain with the corresponding complex weight to obtain a weighted frequency domain; multiplying the frequency domain with the adaptive frequency domain mask to obtain a masked frequency domain; performing an inverse fast Fourier transform on the weighted frequency domain and the masked frequency domain to obtain a time domain; extracting time domain features through multiple convolutional layers, and fusing multiple features to obtain fused features.

[0049] like Figure 3 As shown in the figure, the adaptive frequency processing is performed by the adaptive frequency module (AE): first, the input signal Perform a fast Fourier transform to convert it from the time domain to the frequency domain, and we get .Right now: , where is the input signal, is the fast Fourier transform function, is the output after the fast Fourier transform.

[0050] After fast Fourier transform, the energy of each frequency domain component is calculated And normalize it to get , to balance the energy differences of different frequency components. That is: , , where is the output after fast Fourier transformation, Represents the energy of the signal in the frequency domain, and calculates the square of the modulus length of each frequency component. is the calculated frequency domain energy. is the energy matrix Taking the median along the dimension of each batch, we get the median energy of each batch. A small constant to avoid division by zero errors, initially . It is the normalized frequency domain energy, which indicates the relative energy of each frequency component.

[0051] Then based on the normalized energy and adaptive threshold parameters , generating an adaptive frequency domain mask , which is used to selectively retain important frequency components and discard unimportant frequency components. That is: , where is the normalized frequency domain energy, is the adaptive threshold parameter, which is a learnable parameter. is the generated binary mask, It is a function used to determine whether a value is positive or negative.

[0052] Then, the signal and the corresponding complex weight Multiply to get the weighted frequency domain signal , and in the adaptive mask The masked frequency domain signal is generated under the action of Further adjust the frequency domain signal. That is: , , where is the output after fast Fourier transformation, is the weight matrix, is the generated binary mask, is the weighted frequency domain signal, is the masked frequency domain signal.

[0053] Finally, the processed frequency domain signal is converted back to the time domain through inverse fast Fourier transform to obtain the output signal .Right now: , where is the weighted frequency domain signal, is the frequency domain signal after masking, is the inverse fast Fourier function, It is the time domain signal after inverse fast Fourier transform.

[0054] On this basis, the time domain features are further extracted through the convolution layer, and the outputs of different convolution paths are combined to obtain the final feature representation through fusion. , complete the entire signal processing and feature extraction process. That is: , , , , where It is the time domain signal after inverse fast Fourier transform representing the convolution operation. , and Represents convolution kernels of different sizes, is a one-dimensional convolution function, is the result of convolution kernel 1, After the activation function After the signal, increase its nonlinearity, is the result of convolution kernel 2, After the activation function After the signal, increase its nonlinearity, and is the signal after feature fusion, It is the feature after the third layer of convolution processing.

[0055] The adaptive frequency masking method is used to weight the bearing vibration signal in the frequency domain, which can effectively remove the high-frequency noise components and retain the low-frequency information that is crucial to the judgment of the bearing health status.

[0056] In step S130, exponential smoothing is performed on the fused features to obtain an exponential smoothing output.

[0057] In some embodiments, the exponential smoothing processing of the fusion features includes: determining an exponential smoothing weight and an initial smoothing weight; performing Fourier convolution on the bearing vibration signal according to the exponential smoothing weight to obtain a Fourier convolution output; and calculating the exponential smoothing output according to the Fourier convolution output, the initial smoothing weight, and the bearing vibration signal.

[0058] like Figure 4 As shown, exponential smoothing is handled by the exponential smoothing module (SM): first, the exponential weighting coefficients are calculated for each time step. , the weight controls the smoothness of the input signal over the time step. The coefficient of the exponential decay Determines the relative importance of new information and historical information. , , in the formula, is the smoothing factor, Between 0-1, used to control the degree of smoothing. is at the time step The smoothing weight of is the initial weight of exponential smoothing, is the total number of time steps.

[0059] Then the input signal is Fourier convolved, and the convolution kernel is the calculated exponential smoothing weight. That is: , where is the input signal, is at the time step The smoothing weight of is the output of the convolution Fourier, The convolution Fourier function formula is as follows:

[0060] , , , where is the fast Fourier transform, is the input signal, is the incoming convolution kernel, is to convert the signal into frequency domain, The convolution kernel is converted into the frequency domain, and then the point multiplication in the frequency domain is performed. is the complex conjugate of the convolution sum in the frequency domain, is the result of frequency domain convolution, is the inverse fast Fourier function, convert it to the time domain, and apply a rolling operation , ensure that the convolution results are aligned, and get the result of Fourier convolution .

[0061] After convolution, the initial weights are combined for initialization to control the initial value of smoothing and obtain the output of the exponential smoothing module. is the output of the exponential smoothing module.

[0062] Exponential smoothing algorithm is used to smooth the vibration signal, thereby reducing the impact of short-term fluctuations on life prediction and improving data stability.

[0063] In step S140, Haar wavelet feature enhancement processing is performed on the exponential smoothing output to obtain a Haar wavelet feature enhanced output.

[0064] In some embodiments, the exponential smoothing output is subjected to Haar wavelet feature enhancement processing, including: dimensional expansion of the bearing vibration signal to obtain a change result; two-dimensional convolution and Haar wavelet transform of the change result to obtain a transformation result; two-dimensional convolution and transposition of the transformation result, and Haar wavelet transform to obtain the original feature; removing the third dimension of the original feature to obtain a conversion feature; and performing feature fusion on the second dimension of the conversion feature to obtain a Haar wavelet feature enhanced output.

[0065] like Figure 5 As shown in the figure, the Haar wavelet feature enhancement processing is processed by the Haar wavelet feature enhancement module (HE): first, the dimension of the input feature is expanded to provide more feature information channels, and in accordance with the module's requirements for the input state, and then the input feature is subjected to the Haar filter through the forward Haar wavelet transform to obtain the result of reducing the resolution. That is: , , where Indicates adding input features The third dimension, Represents the specified dimension, Indicates the result after the change. represents a two-dimensional convolution operation, Represents the Haar filter, which divides the input into four parts: similar part, horizontal part, vertical part, and diagonal part. It is the result of reducing the resolution by Haar wavelet transform.

[0066] After that, the inverse Haar wavelet transform is performed to restore the features. The output tensor will be reshaped according to the order of the Haar wavelet transform. Then the inverse convolution operation will restore the image from the transformed low-resolution space to the original feature space. Finally, the newly added third dimension is removed, and the feature dimension is restored through feature fusion to obtain the output of the Haar wavelet feature enhancement module. That is: , , where It is the result of reducing the resolution after Haar transform. represents the operation of two-dimensional convolution and transposition, represents the Haar filter, is the original feature recovered by the inverse transform, Yes Function of the third dimension, Represents the specified dimension, is the feature fusion function, Indicates that after removing the third dimension Feature fusion is performed on the second dimension to restore the reduced dimension features. It is the output of the Haar wavelet feature enhancement module.

[0067] The Haar wavelet transform is used to separate and focus on the frequency components in the bearing vibration signal that are related to fault development and life prediction, thereby improving the accuracy and robustness of the model prediction.

[0068] In step S150, the Haar wavelet feature enhancement output is divided to obtain a training set.

[0069] In step S160, the obtained training set is used to train the initial bearing remaining life prediction model to obtain the target bearing life prediction model, wherein the initial bearing remaining life prediction model includes an encoder and a decoder, and the encoder processes the bearing vibration signal, including: using a one-dimensional convolution layer to perform a convolution operation on the bearing vibration signal, using the output of the one-dimensional convolution layer as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of the cyclic timing convolution module, and using the output of the cyclic timing convolution module as the input of the bidirectional gated loop unit; the decoder processes the output of the bidirectional gated loop unit, including: performing an attention mechanism on the output of the bidirectional gated loop unit, using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of the unidirectional gated unit, and inputting the output of the unidirectional gated unit into the fully connected layer.

[0070] In some embodiments, a cyclic timing convolution module is introduced into the initial bearing remaining life prediction model to reduce the interference of the external environment on the data distribution, wherein the processing of the input data by the cyclic timing convolution module includes: normalizing the input data to obtain an affine transformation result; performing a fast Fourier transform on the affine transformation result to obtain a circular convolution result; performing a first-layer linear transformation and activation on the circular convolution result to obtain an activation result; performing a second-layer linear transformation on the activation result to obtain a second-layer linear transformation result; and denormalizing the second-layer linear transformation result to obtain a denormalized result.

[0071] like Figure 6 As shown, in some embodiments, the normalization of the input data to obtain an affine transformation result includes: according to the formula: , get the mean of the input data ,in, represents the input data sequence, Indicates the length of the data sequence, Indicates the serial number; according to the formula: , and get the standard deviation ,in, Represents a constant; according to the formula: , and get the normalized result ; According to the formula: , and get the affine transformation result ,in, and They all represent learnable parameters.

[0072] The fast Fourier transform is performed on the affine transformation result to obtain a circular convolution result, including:

[0073] According to the formula:

[0074] , , and get the circular convolution result ,in, represents the fast Fourier transform, represents the inverse fast Fourier transform, represents the convolution kernel, represents element-wise multiplication;

[0075] The first-layer linear transformation and activation of the circular convolution result are performed to obtain the activation result, including: according to the formula: , get the activation result ,in, represents the weight matrix of the fully connected layer, represents the bias term, represents an activation function; performing a second-layer linear transformation on the activation result to obtain a second-layer linear transformation result, including: according to the formula: , and get the second layer linear transformation result ,in, represents the weight matrix of the fully connected layer, represents the bias term; the denormalization of the second layer linear transformation result to obtain the denormalized result includes: according to the formula: , and get the denormalized result .

[0076] Specifically, we first expand the dimension of the input features to provide more feature information channels, and in line with the module's requirements for the input state, and then use the forward Haar wavelet transform to perform a Haar filter on the input features to obtain a result with reduced resolution. That is: , , where Indicates adding input features The third dimension, Represents the specified dimension, is the result of the change, represents a two-dimensional convolution operation, Represents the Haar filter, which divides the input into four parts: similar part, horizontal part, vertical part, and diagonal part. It is the result of reducing the resolution by Haar wavelet transform.

[0077] After that, the inverse Haar wavelet transform is performed to restore the features. The output tensor will be reshaped according to the order of the Haar wavelet transform. Then the inverse convolution operation will restore the image from the transformed low-resolution space to the original feature space. Finally, the newly added third dimension is removed, and the feature dimension is restored through feature fusion to obtain the output of the Haar wavelet feature enhancement module. That is: , , where It is the result of reducing the resolution after Haar transform. represents the operation of two-dimensional convolution and transposition, represents the Haar filter, is the original feature recovered by the inverse transform, Yes Function of the third dimension, Represents the specified dimension, is the feature fusion function, Indicates that after removing the third dimension Feature fusion is performed on the second dimension to restore the reduced dimension features. It is the output of the Haar wavelet feature enhancement module.

[0078] In the recurrent temporal convolution module (RC), the input data is first normalized to ensure that it has a suitable scale to improve the stability and efficiency of training. Normalization is done by subtracting the mean and divided by the standard deviation To standardize the data so that the data is distributed in the same range. That is: , , , , where is the input sequence, is the sequence length, Indicates the serial number, Represents a constant to prevent division by zero errors, is the mean of the input data, is the standard deviation, is the result of normalization, is the result of the affine transformation, and are learnable parameters.

[0079] The result after normalization transformation The fast Fourier transform is used to implement circular convolution, and then the fully connected layer is used to perform feature mapping to enhance the features and increase the expressiveness of the model. That is: , , , where is the fast Fourier transform, is the inverse fast Fourier transform, is the result of the affine transformation, is the convolution kernel, is element-wise multiplication, is the result of circular convolution. , is the weight matrix of the fully connected layer, , is the bias term. is the result of the first layer linear transformation and activation, is the activation function, It is the result of the second layer linear transformation.

[0080] Finally, the linearly transformed results are denormalized to restore the data to its original scale. , where is the result of the linear change of the second layer, and is a learnable parameter, is the mean of the input data, is the standard deviation, is the result of denormalizing and restoring the scale of the data.

[0081] like Figure 2 As shown, in some embodiments, the prediction method includes:

[0082] The first step is to extract the vibration signal features. The adaptive frequency module and exponential smoothing module are used for preliminary vibration signal processing to obtain the output , and then calculate the time-frequency domain features of the bearing vibration data, such as the mean, root mean square, kurtosis, skewness, peak-to-peak value, variance, peak factor, pulse factor, margin factor, shape factor, root square amplitude and waveform factor. Then combine these features to obtain the comprehensive vibration characteristics. .Right now: , , where Represents the adaptive frequency module, represents the exponential smoothing module, represents the Haar wavelet feature enhancement module, is the result of applying the adaptive frequency module and the exponential smoothing module. , etc. represent the calculation functions of features such as mean and root mean square. It is a comprehensive vibration characteristic.

[0083] After feature extraction With the initial hidden state The first hidden state is then input into the encoder, compressed by convolution operation, and then input into the adaptive frequency module to enhance feature representation and capture long-term and short-term interactions and reduce noise. Input the bidirectional gating unit together to get the encoder output and the current hidden state .Right now:

[0084] ;

[0085] .

[0086] In the formula is a one-dimensional convolution function, is the comprehensive vibration characteristic, Represents the adaptive frequency module, represents the recurrent temporal convolution module, is the initial hidden state, is the feature fusion function, Indicates feature fusion on the second dimension of the adaptive frequency module output and the cyclic timing convolution module output. It is a two-way gating unit. is the encoder output, is the current hidden state.

[0087] In the decoder stage, the adaptive frequency module is used again to compress and fuse the attention output and contextual information , which helps reduce redundancy and enhances the model's ability to focus on important features. That is: , , where is the hidden state at the current time step, is the hidden state of the previous step, yes function, is an exponential function, is the length of the time series, is the attention output, It is context information.

[0088] Output your attention and contextual information Fusion, together with the hidden state of the current time step, is input into the gating unit to obtain the output , and then and contextual information Connect them together to the linear layer to get the decoder output of the current time step. That is: , , where It is a unidirectional gating unit, the gating unit output and hidden state , is a linear function, It is the output of the decoder, i.e. the prediction result.

[0089] The XJTU bearing dataset is provided by a certain unit and contains complete running to failure data of 15 rolling bearings running under different working conditions. The sampling frequency used for data collection is 25.6kHz. The vibration signals in the X and Y directions of the bearings are sampled, and 32768 samples are recorded every 60 seconds. The information of the dataset and tasks is shown in Table 1. For each, three tests are performed, that is, bearing 1 and bearing 2 are set as training sets, and the others are set as test sets.

[0090] Table 1 XJTU bearing dataset tasks

[0091]

[0092] As shown in Table II, the evaluation index RMSE is compared with the existing methods on the XJTU bearing dataset.

[0093] Table 2

[0094]

[0095] As shown in Table III, the evaluation index MAE is compared with the existing methods on the XJTU bearing dataset.

[0096] Table 3

[0097]

[0098] like Figure 7 and Figure 8 As shown in the figure, the bearing remaining life prediction framework is proposed and verified on Task A and Task B on the XJTU bearing dataset. The results show that the proposed method is superior to the existing methods in terms of root mean square error (RMSE) and mean absolute error (MAE), which proves the effectiveness of the proposed method.

[0099] As mentioned above, the prediction results of the proposed method on the bearing 1_3 and bearing 2_3 tasks of the XJTU bearing dataset are as follows Figure 7 and Figure 8 The task division of the XUTU bearing dataset is shown in Table 1. The RMSE and MAE comparisons of this method and some advanced bearing life prediction methods in recent years in Task A and Task B of the XJTU bearing dataset are shown in Tables 2 and 3.

[0100] Please note that the input feature sequence is processed by the result convolutional neural network and the adaptive frequency domain module and the recurrent time series convolution module, and the frequency domain representation of the input signal is weighted and filtered, which can effectively focus on the important information within a specific frequency range. By calculating the frequency domain energy of the input signal and normalizing it, the model can automatically adjust which frequency components contribute more to the signal, thereby achieving dynamic frequency domain enhancement. The decoder combines the attention mechanism and GRU to calculate the attention weight between the current hidden state and the encoder output to obtain the context vector. Before the GRU input, the adaptive frequency domain module can be used to fuse and compress the attention output and embedded features, which helps to reduce redundancy and enhance the model's ability to focus on important features. The input is then concatenated with the context vector to generate the output of the next time step through the GRU layer.

[0101] Fig. 9 is a block diagram showing a bearing remaining life model training device according to some embodiments of the present disclosure. Fig. 9 As shown, the bearing remaining life model training device 900 includes:

[0102] An acquisition module 910 is configured to acquire a bearing vibration signal;

[0103] The adaptive frequency processing module 920 is configured to perform adaptive frequency processing on the bearing vibration signal to obtain a fusion feature;

[0104] An exponential smoothing processing module 930 is configured to perform exponential smoothing processing on the bearing vibration signal to obtain an exponential smoothing output;

[0105] The Haar wavelet feature enhancement processing module 940 is configured to perform Haar wavelet feature enhancement processing on the bearing vibration signal to obtain a Haar wavelet feature enhancement output;

[0106] A division module 950 is configured to divide the bearing vibration signal after the adaptive frequency processing, the exponential smoothing processing, and the Haar wavelet feature enhancement processing to obtain a training set and a test set;

[0107] The training module 960 is configured to use the obtained training set to train the initial bearing remaining life prediction model to obtain the target bearing life prediction model, wherein the initial bearing remaining life prediction model includes an encoder and a decoder, and the encoder processes the bearing vibration signal, including: using a one-dimensional convolution layer to perform a convolution operation on the bearing vibration signal, using the output of the one-dimensional convolution layer as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of the cyclic timing convolution module, and using the output of the cyclic timing convolution module as the input of the bidirectional gated loop unit; the decoder processes the output of the bidirectional gated loop unit, including: performing an attention mechanism on the output of the bidirectional gated loop unit, using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of the unidirectional gated unit, and inputting the output of the unidirectional gated unit into the fully connected layer.

[0108] In the device of the embodiment of the present disclosure, a bearing remaining life model training device is provided. The present disclosure adopts an adaptive frequency processing method, uses Fourier analysis to enhance feature representation, and captures long-term and short-term interactions of bearing vibration information, while reducing noise through adaptive thresholds. The present disclosure adopts an exponential smoothing processing method, and combines fast Fourier transform and exponential weighted smoothing to effectively smooth the input signal, reduce the impact of short-term toggling, and retain long-term trends. The present disclosure adopts a Haar wavelet feature enhancement processing method, uses Haar wavelets to decompose the vibration signal into different frequency sub-bands, extracts low-frequency trends and high-frequency fault information respectively, and reduces the feature resolution to enhance the ability to capture key features of the bearing vibration signal, thereby improving the accuracy and robustness of model prediction. The present disclosure adopts cyclic time series convolution, and reduces the interference of the external environment on the data distribution through dynamic demean normalization, thereby improving the robustness and generalization ability of the model. The present disclosure adds an adaptive frequency method in both the encoder and decoder stages, improves the model's ability to decode complex time series, and deepens and strengthens the feature mining of vibration signals. The technology disclosed in the present disclosure can more accurately predict the remaining service life of the bearing, and provide support for equipment maintenance, fault warning and performance optimization.

[0109] Fig.10is a block diagram showing a bearing remaining life model training device according to some other embodiments of the present disclosure.

[0110] like Fig.10 As shown, the bearing remaining life model training device 1000 includes a memory 1010; and a processor 1020 coupled to the memory 1010. The memory 1010 is used to store instructions for executing the corresponding embodiment of the bearing remaining life model training method. The processor 1020 is configured to execute the bearing remaining life model training method in any of the embodiments of the present disclosure based on the instructions stored in the memory 1010.

[0111] Fig.11 is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure.

[0112] like Fig.11 As shown, the computer system 1100 may be embodied in the form of a general-purpose computing device. The computer system 1100 includes a memory 1110, a processor 1120, and a bus 1130 that connects various system components.

[0113] The memory 1110 may include, for example, a system memory, a non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs. The system memory may include a volatile storage medium, such as a random access memory (RAM) and / or a cache memory. The non-volatile storage medium may store, for example, instructions for executing at least one of the corresponding embodiments of the bearing remaining life model training method. The non-volatile storage medium includes, but is not limited to, a disk memory, an optical memory, a flash memory, and the like.

[0114] The processor 1120 can be implemented by a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistors, etc. Discrete hardware components. Accordingly, each module such as the acquisition module, the adaptive frequency processing module, the exponential smoothing processing module, the Haar wavelet feature enhancement processing module, the partitioning module, and the training module can be implemented by a central processing unit (CPU) running instructions in a memory that execute corresponding steps, or can be implemented by a dedicated circuit that executes corresponding steps.

[0115] The bus 1130 may use any of a variety of bus architectures, including, but not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, and a Peripheral Component Interconnect (PCI) bus.

[0116] The computer system 1100 may also include an input / output interface 1140, a network interface 1150, a storage interface 1160, etc. These interfaces 1140, 1150, 1160, the memory 1110, and the processor 1120 may be connected via a bus 1130. The input / output interface 1140 may provide a connection interface for input / output devices such as a display, a mouse, and a keyboard. The network interface 1150 may provide a connection interface for various networked devices. The storage interface 1160 may provide a connection interface for external storage devices such as a floppy disk, a USB flash drive, and an SD card.

[0117] Here, various aspects of the present disclosure are described with reference to flowcharts and / or block diagrams of methods, devices, and computer program products according to embodiments of the present disclosure. It should be understood that each frame of the flowchart and / or block diagram and the combination of frames can be implemented by computer-readable program instructions.

[0118] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable device to produce a machine, so that the processor executes the instructions to produce means for implementing the functions specified in one or more blocks in the flowchart and / or block diagram.

[0119] These computer-readable program instructions may also be stored in a computer-readable memory, which cause the computer to work in a specific manner to produce an article of manufacture, including instructions for implementing the functions specified in one or more blocks in the flowchart and / or block diagram.

[0120] The present disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.

[0121] like Fig.12 As shown, an embodiment of the present disclosure also provides a method for predicting remaining life of a bearing, including steps S1210 to S1220.

[0122] In step S1210, a real-time bearing vibration signal is obtained;

[0123] In step S1220, the real-time bearing vibration signal is input into the bearing remaining life prediction model trained by the above-mentioned bearing remaining life prediction model training method to predict the remaining life of the bearing.

[0124] The present disclosure provides a bearing remaining life prediction model training method, device, medium and prediction method. The present disclosure adopts an adaptive frequency processing method, uses Fourier analysis to enhance feature representation, and captures long-term and short-term interactions of bearing vibration information, while reducing noise through adaptive thresholds. The present disclosure adopts an exponential smoothing processing method, and combines fast Fourier transform and exponential weighted smoothing to effectively smooth input signals, reduce the impact of short-term toggling, and retain long-term trends. The present disclosure adopts a Haar wavelet feature enhancement processing method, uses Haar wavelets to decompose vibration signals into different frequency sub-bands, extracts low-frequency trends and high-frequency fault information respectively, and reduces feature resolution to enhance the ability to capture key features of bearing vibration signals, thereby improving the accuracy and robustness of model prediction. The present disclosure adopts cyclic time series convolution, and reduces the interference of the external environment on data distribution through dynamic demean normalization, thereby improving the robustness and generalization ability of the model. The present disclosure adds an adaptive frequency method in both the encoder and decoder stages, improves the model's ability to decode complex time series, and deepens and strengthens the feature mining of vibration signals. The technology disclosed in this disclosure can more accurately predict the remaining service life of bearings, providing support for equipment maintenance, fault warning and performance optimization.

[0125] So far, the bearing remaining life prediction model training method, device, medium and prediction method according to the present disclosure have been described in detail. In order to avoid obscuring the concept of the present disclosure, some details known in the art are not described. Based on the above description, those skilled in the art can fully understand how to implement the technical solution disclosed here.

[0126] Although some specific embodiments of the present disclosure have been described in detail by way of examples, it should be understood by those skilled in the art that the above examples are for illustration only and are not intended to limit the scope of the present disclosure. It should be understood by those skilled in the art that the above embodiments may be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. A method for training a bearing remaining life prediction model, characterized in that: The method comprises: Obtain bearing vibration signal; Adaptive frequency processing is performed on the bearing vibration signal to obtain fusion features; Perform exponential smoothing on the fused features to obtain exponential smoothing output; Performing Haar wavelet feature enhancement processing on the exponential smoothing output to obtain Haar wavelet feature enhanced output; Divide the Haar wavelet feature enhancement output to obtain a training set; The obtained training set is used to train the initial bearing remaining life prediction model to obtain the target bearing life prediction model, wherein the initial bearing remaining life prediction model includes an encoder and a decoder, and the encoder processes the bearing vibration signal in the following steps: performing a convolution operation on the bearing vibration signal using a one-dimensional convolution layer, using the output of the one-dimensional convolution layer as the input of an adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of a cyclic timing convolution module, and using the output of the cyclic timing convolution module as the input of a bidirectional gated recurrent unit; the decoder processes the output of the bidirectional gated recurrent unit in the following steps: performing an attention mechanism on the output of the bidirectional gated recurrent unit, using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of a unidirectional gated unit, and inputting the output of the unidirectional gated unit into a fully connected layer.

2. The method for training a bearing remaining life prediction model according to claim 1, characterized in that: The adaptive frequency processing of the bearing vibration signal comprises: Perform fast Fourier transform on the bearing vibration signal to obtain the frequency domain; Calculate the energy of each frequency domain and perform normalization to obtain normalized frequency domain energy; generating an adaptive frequency domain mask based on normalized frequency domain energy and an adaptive threshold parameter; Multiply the frequency domain with the corresponding complex weight to obtain the weighted frequency domain; Multiply the frequency domain with the adaptive frequency domain mask to obtain the masked frequency domain; Perform inverse fast Fourier transform on the weighted frequency domain and the masked frequency domain to obtain the time domain; The time domain features are extracted through multiple convolutional layers, and multiple features are fused to obtain fused features.

3. The method for training a bearing remaining life prediction model according to claim 1, characterized in that: The exponential smoothing process of the fused features includes: Determine the exponential smoothing weight and the initial smoothing weight; Perform Fourier convolution on the bearing vibration signal according to the exponential smoothing weight to obtain the Fourier convolution output; The exponential smoothing output is calculated based on the Fourier convolution output, the smoothing initial weight, and the bearing vibration signal.

4. The method for training a bearing remaining life prediction model according to claim 1, characterized in that: The Haar wavelet feature enhancement processing is performed on the exponential smoothing output, comprising: Expand the dimension of the bearing vibration signal to obtain the change result; Perform two-dimensional convolution and Haar wavelet transform on the change result to obtain the transformation result; Perform two-dimensional convolution and transposition on the transformation results, and perform Haar wavelet transform to obtain the original features; Remove the third dimension of the original feature to obtain the transformed feature; Feature fusion is performed on the second dimension of the transformed features to obtain the Haar wavelet feature enhanced output.

5. The method for training a bearing remaining life prediction model according to claim 1, characterized in that: The output of the adaptive frequency processing module is used as the input of the cyclic timing convolution module, wherein the processing of the input data by the cyclic timing convolution module includes: Normalize the input data to obtain the affine transformation result; Perform fast Fourier transform on the affine transformation result to obtain the circular convolution result; Perform the first-layer linear transformation and activation on the circular convolution result to obtain the activation result; Perform a second-layer linear transformation on the activation result to obtain a second-layer linear transformation result; The second layer linear transformation result is denormalized to obtain a denormalized result.

6. The method for training a bearing remaining life prediction model according to claim 5, characterized in that: The normalization process of the input data to obtain an affine transformation result includes: According to the formula: , get the mean of the input data ,in, represents the input data sequence, Indicates the length of the data sequence, Indicates the serial number; According to the formula: , and get the standard deviation ,in, represents a constant; According to the formula: , and get the normalized result ; According to the formula: , and get the affine transformation result ,in, and They all represent learnable parameters.

7. The method for training a bearing remaining life prediction model according to claim 6, characterized in that: The fast Fourier transform is performed on the affine transformation result to obtain a circular convolution result, including: According to the formula: , and get the circular convolution result ,in, represents the fast Fourier transform, represents the inverse fast Fourier transform, represents the convolution kernel, represents element-wise multiplication; The first-layer linear transformation and activation are performed on the circular convolution result to obtain the activation result, including: According to the formula: , get the activation result ,in, represents the weight matrix of the fully connected layer, represents the bias term, represents the activation function; The performing a second-layer linear transformation on the activation result to obtain a second-layer linear transformation result includes: According to the formula: , and get the second layer linear transformation result ,in, represents the weight matrix of the fully connected layer, represents the bias term; The denormalizing the second-layer linear transformation result to obtain the denormalized result includes: According to the formula: , and get the denormalized result .

8. A bearing remaining life prediction model training device, characterized in that: include: An acquisition module, used for acquiring bearing vibration signals; An adaptive frequency processing module is used to perform adaptive frequency processing on bearing vibration signals to obtain fusion features; An exponential smoothing processing module is used to perform exponential smoothing on the fused features to obtain exponential smoothing output; A Haar wavelet feature enhancement processing module is used to perform Haar wavelet feature enhancement processing on the exponential smoothing output to obtain a Haar wavelet feature enhanced output; A partitioning module is used to partition the Haar wavelet feature enhancement output to obtain a training set; A training module is used to train an initial bearing remaining life prediction model using the obtained training set to obtain a target bearing life prediction model, wherein the initial bearing remaining life prediction model includes an encoder and a decoder, and the encoder processes the bearing vibration signal, including: using a one-dimensional convolution layer to perform a convolution operation on the bearing vibration signal, using the output of the one-dimensional convolution layer as the input of an adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of a cyclic timing convolution module, and using the output of the cyclic timing convolution module as the input of a bidirectional gated cyclic unit; the decoder processes the output of the bidirectional gated cyclic unit, including: performing an attention mechanism on the output of the bidirectional gated cyclic unit, using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, using the output of the adaptive frequency processing module as the input of a unidirectional gated unit, and inputting the output of the unidirectional gated unit into a fully connected layer.

9. A method for predicting the remaining life of a bearing, characterized in that: The method comprises: Get real-time bearing vibration signal; The real-time bearing vibration signal is input into a bearing remaining life prediction model trained by the bearing remaining life prediction model training method as described in any one of claims 1 to 7 to predict the remaining life of the bearing.

10. A computer-readable storage medium, characterized in that: Computer program instructions are stored thereon, and when the instructions are executed by a processor, the bearing remaining life prediction model training method as described in any one of claims 1 to 7 or the bearing remaining life prediction method as described in claim 9 is implemented.

Images