Bearing life prediction method and device based on text features and convolutional neural network

Abstract

The present application provides a bearing life prediction method and device based on text features and convolutional neural networks. The method includes: obtaining time domain vibration data of a rolling bearing to be predicted; performing noise reduction processing on the time domain vibration data; obtaining frequency domain data by Fourier transforming the noise-reduced time domain vibration data; extracting frequency domain features of the frequency domain data, and encoding the frequency domain data to obtain text features; inputting the frequency domain features and text features into a pre-trained neural network model to predict the predicted life of the rolling bearing to be predicted. The method provided by the present application uses wavelet decomposition and reconstruction methods for noise reduction, avoiding the problem of threshold selection; the text features used in the present application are used as new degradation indicators, avoiding the shortcomings of incomplete representation of traditional degradation indicators; and the attention mechanism is applied to the feature channel to enhance the feature association capability in the convolutional neural network.

Description

Technical Field

[0001] The present application relates to the technical field of bearing life prediction methods, and in particular to a bearing life prediction method and device based on text features and convolutional neural networks. Background Art

[0002] Bearings are widely used in modern machinery. Based on the friction properties of their components, they are categorized as rolling bearings and plain bearings, with rolling bearings being the most widely used. Rolling bearings primarily support rotating shafts and components attached to them. Their operating condition significantly impacts the safe operation of equipment. To minimize factory maintenance costs, predicting the life of rolling bearings is crucial.

[0003] Traditional methods typically use mathematical models to predict the remaining life of bearings. However, given the complex operating conditions and diverse components, traditional model-building approaches are inefficient and rely on expert knowledge. Neural networks have developed rapidly in recent years, with the emergence of various convolutional, recurrent, and attention-based networks. These networks have shown significant success in image and natural language processing, but they are difficult to directly apply to the problem of predicting the remaining life of rolling bearings.

[0004] New neural networks face the following challenges in predicting the life of rolling bearings: First, vibration signals contain numerous noise factors, necessitating prior noise reduction. While noise reduction methods based on fixed-value filtering are effective, they often require data and experience to determine the optimal threshold. Second, traditional degradation metrics such as kurtosis, skewness, margin, and shock factor struggle to comprehensively characterize degradation trends at different stages. For example, margin and kurtosis are sensitive to signals but lack stability, while waveform metrics offer greater stability but lower sensitivity. Third, different bearing components have varying preferences for degradation metrics. While convolutional networks are well-suited for signal processing, their parameter-sharing strategy across feature channels makes them insensitive to feature correlations. This inability to accurately predict feature correlations contributes to the current low accuracy of rolling bearing remaining life prediction. Summary of the Invention

[0005] This application provides a bearing life prediction method and device based on text features and convolutional neural networks, which can be used to solve the technical problem of low accuracy in predicting the remaining life of rolling bearings.

[0006] In a first aspect, the present application provides a bearing life prediction method based on text features and convolutional neural networks, the method comprising:

[0007] Acquire time domain vibration data of the rolling bearing to be predicted;

[0008] performing noise reduction processing on the time-domain vibration data;

[0009] The noise-reduced time-domain vibration data is subjected to Fourier transformation to obtain frequency-domain data;

[0010] Extracting frequency domain features of the frequency domain data, and encoding the frequency domain data to obtain text features;

[0011] The frequency domain features and the text features are input into a pre-trained neural network model to predict the predicted life of the rolling bearing to be predicted.

[0012] In conjunction with the first aspect, in one implementation of the first aspect, obtaining time-domain vibration data of a rolling bearing includes:

[0013] The horizontal vibration amplitude and the vertical vibration amplitude of the rolling bearing to be predicted are collected respectively; the time domain vibration data includes the horizontal vibration amplitude and the vertical vibration amplitude.

[0014] In conjunction with the first aspect, in one implementation of the first aspect, performing noise reduction processing on the time-domain vibration data includes:

[0015] The time domain vibration data is subjected to noise reduction processing by using wavelet decomposition.

[0016] In conjunction with the first aspect, in one implementation of the first aspect, performing noise reduction processing on the time-domain vibration data includes:

[0017] A sparse autoencoder is used to perform noise reduction on the time domain vibration data.

[0018] In conjunction with the first aspect, in one implementation of the first aspect, extracting the frequency domain features of the frequency domain data includes:

[0019] Extracting frequency domain features of the frequency domain data, and performing smoothing transformation and cumulative transformation;

[0020] The frequency domain characteristics include variance, skewness, kurtosis, maximum value and margin factor;

[0021] The smooth transformation utilizes SG filter function;

[0022] The formula used for cumulative transformation is:

[0023] X=[x1,x2,…,x i ,…,x N ]

[0024]

[0025]

[0026] Among them, X is the horizontal feature and vertical feature of the input data, S i is the cumulative value at time i, C i is the cumulative transformation value at time i.

[0027] In conjunction with the first aspect, in one implementation of the first aspect, encoding the frequency domain data to obtain text features includes:

[0028] The frequency domain data has a total of N samples, each sample D i Contains T time points, each time point is vibration amplitude data of length L;

[0029] The range of the vibration amplitude data is divided into j intervals, and each sample D is counted. i The number of different amplitudes contained in each time point, these j intervals are regarded as j different words w j , get the word frequency matrix M;

[0030] The word frequency feature is constructed by the following method;

[0031] Perform singular value decomposition on the word frequency matrix M, reconstruct the matrix using the largest singular value, and obtain a low-dimensional feature matrix M* as the word frequency feature;

[0032] The word frequency-inverse text frequency feature is constructed by the following method:

[0033] The word frequency matrix M is divided by the sampling frequency to obtain a first intermediate quantity M';

[0034] The first intermediate quantity M' is multiplied by the normalized inverse text frequency to obtain a second intermediate quantity M", and the second intermediate quantity M" is subjected to singular value decomposition. The matrix is reconstructed using the largest singular value, and the obtained low-dimensional feature matrix M^ is used as the word frequency-inverse text frequency feature;

[0035] Among them, word frequency-inverse text frequency includes word frequency coefficient and normalized inverse text frequency;

[0036] in, Normalized Word frequency - inverse text frequency = word frequency coefficient * normalized inverse text frequency, denominator of normalized inverse text frequency | j:w j ∈D i | contains the word w j Sample D i The number of

[0037] In combination with the first aspect, in one implementation of the first aspect, the neural network model includes 18 convolutional layers, wherein one residual connection and attention module is set for every three forward convolutional layers;

[0038] The neural network model uses an adaptive moment estimation optimizer for parameter optimization with a learning rate of 0.001;

[0039] Batch normalization and ReLU activation functions are used between convolutional layers of the neural network model;

[0040] The last layer of the neural network is a fully connected layer, and the sigmoid activation function is used to normalize the output of the neural network.

[0041] In a second aspect, the present application provides a bearing life prediction device based on text features and convolutional neural networks, the device comprising:

[0042] An acquisition module, used for acquiring time domain vibration data of a rolling bearing to be predicted;

[0043] A noise reduction module is used to perform noise reduction processing on the time domain vibration data.

[0044] a processing module for performing Fourier transform on the noise-reduced time-domain vibration data to obtain frequency-domain data; and extracting frequency-domain features of the frequency-domain data, and encoding the frequency-domain data to obtain text features;

[0045] The prediction module is used to input the frequency domain features and the text features into a pre-trained neural network model to predict the predicted life of the rolling bearing to be predicted.

[0046] In conjunction with the second aspect, in one implementation of the second aspect, the acquisition module is specifically configured to:

[0047] The horizontal vibration amplitude and the vertical vibration amplitude of the rolling bearing to be predicted are collected respectively; the time domain vibration data includes the horizontal vibration amplitude and the vertical vibration amplitude.

[0048] In conjunction with the second aspect, in one possible implementation of the second aspect, the noise reduction module is specifically configured to:

[0049] The time domain vibration data is subjected to noise reduction processing by using wavelet decomposition.

[0050] In conjunction with the second aspect, in one possible implementation of the second aspect, the noise reduction module is specifically configured to:

[0051] A sparse autoencoder is used to perform noise reduction on the time domain vibration data.

[0052] In conjunction with the second aspect, in one implementation of the second aspect, the processing module is specifically configured to:

[0053] Extracting frequency domain features of the frequency domain data, and performing smoothing transformation and cumulative transformation;

[0054] The frequency domain characteristics include variance, skewness, kurtosis, maximum value and margin factor;

[0055] The smooth transformation utilizes SG filter function;

[0056] The formula used for cumulative transformation is:

[0057] X=[x1,x2,…,x i ,…,x N ]

[0058]

[0059]

[0060] Among them, X is the horizontal feature and vertical feature of the input data, S i is the cumulative value at time i, C i is the cumulative transformation value at time i.

[0061] In conjunction with the second aspect, in one implementation of the second aspect, the processing module is further specifically configured to:

[0062] The frequency domain data has a total of N samples, each sample D i Contains T time points, each time point is vibration amplitude data of length L;

[0063] The range of the vibration amplitude data is divided into j intervals, and each sample D is counted. i The number of different amplitudes contained in each time point, these j intervals are regarded as j different words w j , get the word frequency matrix M;

[0064] The word frequency feature is constructed by the following method;

[0065] Perform singular value decomposition on the word frequency matrix M, reconstruct the matrix using the largest singular value, and obtain a low-dimensional feature matrix M* as the word frequency feature;

[0066] The word frequency-inverse text frequency feature is constructed by the following method:

[0067] The word frequency matrix M is divided by the sampling frequency to obtain a first intermediate quantity M';

[0068] The first intermediate quantity M' is multiplied by the normalized inverse text frequency to obtain a second intermediate quantity M", and the second intermediate quantity M" is subjected to singular value decomposition. The matrix is reconstructed using the largest singular value, and the obtained low-dimensional feature matrix M^ is used as the word frequency-inverse text frequency feature;

[0069] Among them, word frequency-inverse text frequency includes word frequency coefficient and normalized inverse text frequency;

[0070] in, Normalized Word frequency - inverse text frequency = word frequency coefficient * normalized inverse text frequency, denominator of normalized inverse text frequency | j:w j ∈D i | contains the word w j Sample D i The number of

[0071] In conjunction with the second aspect, in one implementation of the second aspect, the neural network model includes 18 convolutional layers, wherein one residual connection and attention module is set for every three forward convolutional layers;

[0072] The neural network model uses an adaptive moment estimation optimizer for parameter optimization with a learning rate of 0.001;

[0073] Batch normalization and ReLU activation functions are used between convolutional layers of the neural network model;

[0074] The last layer of the neural network is a fully connected layer, and the sigmoid activation function is used to normalize the output of the neural network.

[0075] The method provided in this application uses wavelet decomposition and reconstruction methods for noise reduction, avoiding the problem of threshold selection; the text features used in this application are used as new degradation indicators, avoiding the shortcomings of incomplete representation of traditional degradation indicators; and the attention mechanism is applied to the feature channel, enhancing the feature association ability in the convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 A schematic diagram of a flow chart of a bearing life prediction method based on text features and convolutional neural networks provided in an embodiment of the present application;

[0077] Figure 2 A schematic diagram of wavelet decomposition provided in an embodiment of the present application;

[0078] Figure 3 A partial feature smoothing effect diagram provided in an embodiment of the present application;

[0079] Figure 4 A partial feature cumulative transformation effect diagram provided in an embodiment of the present application;

[0080] Figure 5 A text feature rendering provided in an embodiment of the present application;

[0081] Figure 6A schematic diagram of an attention mechanism and residual link provided in an embodiment of the present application;

[0082] Figure 7 A flow chart of a bearing life prediction method based on text features and convolutional neural networks provided in an embodiment of the present application. DETAILED DESCRIPTION

[0083] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0084] like Figure 1 The figure shows a flow chart of a bearing life prediction method based on text features and convolutional neural networks provided in an embodiment of the present application. The method provided in the present application includes the following steps:

[0085] Step S101: Acquire time domain vibration data of a rolling bearing to be predicted.

[0086] In the embodiment of the present application, the horizontal vibration amplitude and the vertical vibration amplitude of the rolling bearing to be predicted are collected separately. The time domain vibration data includes the horizontal vibration amplitude and the vertical vibration amplitude.

[0087] Specifically, vibration data of the rolling bearing is collected using an accelerometer to obtain time-domain vibration data of the rolling bearing. In this application, two accelerometers are used to collect the vibration amplitude in the horizontal direction and the vibration amplitude in the vertical direction, respectively. The accelerometer samples the time-domain vibration data every 10 seconds, with each sampling time lasting 0.1 seconds and a sampling frequency of 25.6 kHz.

[0088] Step S102: performing noise reduction processing on the time domain vibration data.

[0089] In the embodiment of the present application, a variety of methods can be used to reduce the noise of the time domain vibration data. One feasible method is to use wavelet decomposition to reduce the noise of the time domain vibration data.

[0090] Specifically, the Dobesi wavelet, db-2, is used as the wavelet basis to decompose the original data to obtain low-frequency coefficients and high-frequency coefficients, and the low-frequency coefficients are filtered. The high-frequency coefficients are further decomposed and repeated twice. The process is as follows Figure 2 shown. Figure 2 A wavelet decomposition diagram is provided in an embodiment of the present application. The high-frequency coefficients and the filtered low-frequency coefficients are then reconstructed by wavelet to obtain new data.

[0091] Another feasible method is to use a sparse autoencoder to perform noise reduction on time-domain vibration data, which can be selected according to the situation.

[0092] Step S103 : Performing Fourier transform on the noise-reduced time-domain vibration data to obtain frequency-domain data.

[0093] Specifically, the noise-reduced time-domain vibration data is subjected to Fourier transformation to convert the time coordinate into the frequency coordinate to obtain frequency domain data.

[0094] Step S104: extract frequency domain features of the frequency domain data, and encode the frequency domain data to obtain text features.

[0095] In the embodiment of the present application, frequency domain features of frequency domain data are extracted, and smoothing transformation and cumulative transformation are performed.

[0096] Among them, the frequency domain characteristics include variance, skewness, kurtosis, maximum value and margin factor.

[0097] The smoothing transformation uses the least squares smoothing filter, namely the SG filter function. The SG filter window is the data length. This application uses a lower polynomial coefficient to enhance the smoothing effect. The smoothing polynomial coefficient is set to 3. The smoothing effect diagram of some features is as follows: Figure 3 shown.

[0098] The formula used for cumulative transformation is:

[0099] X=[x1,x2,…,x i ,…,x N ]

[0100]

[0101]

[0102] Among them, X is the horizontal feature and vertical feature of the input data, S i is the cumulative value at time i, C i is the cumulative transformation value at time i. The cumulative transformation effect diagram of some features is as follows Figure 4 shown.

[0103] The specific process of encoding frequency domain data to obtain text features is as follows:

[0104] The frequency domain data has a total of N samples, each sample D i Contains T time points, each time point is the vibration amplitude data of length L. Among them, the frequency domain data D dimension is (N, T, L), each sample D i The dimension is (T, L).

[0105] Divide the range of vibration amplitude data into j intervals and count each sample D i The number of different amplitudes contained in each time point, these j intervals are regarded as j different words w j, and get the word frequency matrix M. The dimension of the word frequency matrix M is (T, j).

[0106] The word frequency feature is constructed using the following method.

[0107] Perform singular value decomposition on the word frequency matrix M and reconstruct the matrix using the largest singular value. The resulting low-dimensional feature matrix M* is used as the word frequency feature. The dimension of M* is (T, 1).

[0108] The word frequency-inverse text frequency feature is constructed by the following method:

[0109] The word frequency matrix M is divided by the sampling frequency to obtain the first intermediate quantity M'.

[0110] The first intermediate quantity M' is multiplied by the normalized inverse text frequency to obtain the second intermediate quantity M", and the second intermediate quantity M" is subjected to singular value decomposition. The matrix is reconstructed using the largest singular value, and the obtained low-dimensional feature matrix M^ is used as the word frequency-inverse text frequency (tf-idf) feature.

[0111] The term frequency-inverse document frequency (tf-idf) includes the term frequency coefficient and the normalized inverse document frequency.

[0112] in, Normalized Word frequency - inverse text frequency = word frequency coefficient * normalized inverse text frequency, denominator of normalized inverse text frequency | j:w j ∈D i | contains the word w j Sample D i The number of

[0113] The effect diagram of word frequency feature and word frequency-inverse text frequency (tf-idf) feature is as follows Figure 5 shown.

[0114] The method provided in this application also performs preprocessing before data encoding. This is because, in the case of unprocessed bearing vibration signals, approximately 98% of the amplitude data falls within [0, 2cm], and 2% of the data falls within [2cm, 18cm]. This means that most time series data falls within a very small range, and the number of amplitude data in other ranges is not counted, making degradation trends difficult to characterize. To address this problem, the inverse tangent function is first used to map all data to the range [-1, 1]. Then, an inverse hyperbolic tangent transform is used to reduce the kurtosis of the vibration signal data distribution. After two trigonometric transformations, approximately 75% of the amplitude data is mapped to the range [0, 0.75], and 25% of the data falls between [0.75, 1]. This facilitates the generation of singular vectors with more obvious degradation trends.

[0115] The inverse tangent function expression is:

[0116]

[0117] Where x is the frequency domain data and y is the data after arc tangent normalization.

[0118] The expression of the inverse hyperbolic tangent function is:

[0119]

[0120] Among them, x is the data after arc tangent normalization, and y is the data after inverse hyperbolic tangent normalization.

[0121] Step S105: inputting the frequency domain features and the text features into a pre-trained neural network model to predict the life of the rolling bearing to be predicted.

[0122] Specifically, the input of the neural network model provided in this application is frequency domain features and text features, and the output is the total life of the rolling bearing to be predicted. The predicted life is obtained by subtracting the existing life from the total life of the rolling bearing to be predicted.

[0123] The neural network model provided in this application has been trained multiple times before being officially used.

[0124] In the embodiment of the present application, the neural network model includes 18 convolutional layers, in which one residual connection and attention module are set for every three forward convolutional layers. Specifically, the residual connection is used to alleviate the vanishing gradient conduction, and the attention mechanism is used to enhance the correlation of feature dimensions. The attention mechanism performs global pooling and full connection calculations on each feature dimension, calculates the weight of each feature dimension, and then uses the weight value to multiply and scale each feature dimension. The schematic diagram of the residual connection and attention mechanism is shown in Figure 2. Figure 6 shown.

[0125] The neural network model was optimized using the adaptive moment estimation Adam optimizer with a learning rate of 0.001.

[0126] Batch normalization and linear rectifier unit (ReLU) activation function are used between the convolutional layers of the neural network model to enhance the nonlinear fitting ability and training speed of the convolutional neural network.

[0127] Specifically, ReLU activation function is used between layers to enhance the nonlinear expression ability of the model, and batch normalization is performed before the activation function to improve the model training speed.

[0128] The last layer of the neural network is a fully connected layer, and the output of the neural network is normalized using the sigmoid activation function to normalize the predicted lifespan to between 0 and 1.

[0129] The sigmoid activation function formula is as follows:

[0130]

[0131] Here, x is the neuron input and e is a natural constant.

[0132] The method provided in this application uses the wavelet decomposition-reconstruction method for noise reduction, avoiding the problem of threshold selection; the text features used in the present invention are used as new degradation indicators, avoiding the shortcomings of incomplete representation of traditional degradation indicators; and the attention mechanism is applied to the feature channel, thereby enhancing the feature association ability in the convolutional neural network.

[0133] This application also provides a bearing life prediction device based on text features and convolutional neural networks. For technical details not disclosed in the device provided by this application, please refer to the method provided by this application, which will not be repeated here.

[0134] like Figure 7 As shown in the figure, a bearing life prediction device based on text features and convolutional neural network is provided in an embodiment of the present application. The device provided in the present application includes:

[0135] The acquisition module 701 is used to acquire the time domain vibration data of the rolling bearing to be predicted.

[0136] The noise reduction module 702 is used to perform noise reduction processing on the time domain vibration data.

[0137] The processing module 703 is configured to obtain frequency domain data by Fourier transforming the de-noised time domain vibration data, extract frequency domain features of the frequency domain data, and encode the frequency domain data to obtain text features.

[0138] The prediction module 704 is used to input the frequency domain features and the text features into a pre-trained neural network model to predict the predicted life of the rolling bearing to be predicted.

[0139] Optionally, the acquisition module 701 is specifically configured to:

[0140] The horizontal vibration amplitude and vertical vibration amplitude of the rolling bearing to be predicted are collected separately. The time domain vibration data includes the horizontal vibration amplitude and the vertical vibration amplitude.

[0141] Optionally, the noise reduction module 702 is specifically configured to:

[0142] Wavelet decomposition is used to reduce the noise of time domain vibration data.

[0143] Optionally, the noise reduction module 702 is specifically configured to:

[0144] The sparse autoencoder is used to denoise the time domain vibration data.

[0145] Optionally, the processing module 703 is specifically configured to:

[0146] Extract the frequency domain features of the frequency domain data and perform smoothing and cumulative transformations.

[0147] Among them, the frequency domain characteristics include variance, skewness, kurtosis, maximum value and margin factor.

[0148] The smooth transformation uses the SG filter function.

[0149] The formula used for cumulative transformation is:

[0150] X=[x1,x2,…,x i ,…,x N ]

[0151]

[0152]

[0153] Among them, X is the horizontal feature and vertical feature of the input data, S i is the cumulative value at time i, C i is the cumulative transformation value at time i.

[0154] Optionally, the processing module 703 is further specifically configured to:

[0155] The frequency domain data has a total of N samples, each sample D i Contains T time points, each of which is vibration amplitude data of length L.

[0156] Divide the range of vibration amplitude data into j intervals and count each sample D i The number of different amplitudes contained in each time point, these j intervals are regarded as j different words w j , and get the word frequency matrix M.

[0157] The word frequency feature is constructed using the following method.

[0158] Perform singular value decomposition on the word frequency matrix M, use the largest singular value to reconstruct the matrix, and obtain the low-dimensional feature matrix M* as the word frequency feature.

[0159] The word frequency-inverse text frequency feature is constructed by the following method:

[0160] The word frequency matrix M is divided by the sampling frequency to obtain the first intermediate quantity M'.

[0161] The first intermediate quantity M' is multiplied by the normalized inverse text frequency to obtain the second intermediate quantity M", and the second intermediate quantity M" is subjected to singular value decomposition. The matrix is reconstructed using the largest singular value, and the obtained low-dimensional feature matrix M^ is used as the word frequency-inverse text frequency feature.

[0162] The word frequency-inverse text frequency includes the word frequency coefficient and the normalized inverse text frequency.

[0163] in, Normalized Word frequency - inverse text frequency = word frequency coefficient * normalized inverse text frequency, denominator of normalized inverse text frequency | j:w j ∈D i | contains the word w j Sample D i The number of

[0164] Optionally, the neural network model contains 18 convolutional layers, where one residual connection and attention module is set for every three forward convolutional layers.

[0165] The neural network model was optimized using an adaptive moment estimation optimizer with a learning rate of 0.001.

[0166] Batch normalization and ReLU activation functions are used between convolutional layers of the neural network model.

[0167] The last layer of the neural network is a fully connected layer, and the sigmoid activation function is used to normalize the output of the neural network.

[0168] The method provided in this application uses wavelet decomposition and reconstruction for noise reduction, avoiding the problem of threshold selection. It also uses text features as a new degradation indicator, avoiding the incomplete representation shortcomings of traditional degradation indicators. Furthermore, it applies an attention mechanism to feature channels, enhancing the feature association capabilities of convolutional neural networks.

[0169] Those skilled in the art can clearly understand that the technology in the embodiments of the present application can be implemented by means of software plus the necessary general hardware platform. Based on this understanding, the technical solutions in the embodiments of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application or certain parts of the embodiments.

[0170] In this specification, the same or similar parts between the various embodiments can be referred to each other. In particular, for the service construction device and service loading device embodiments, since they are basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.

[0171] The above-described embodiments of the present application do not constitute a limitation on the scope of protection of the present application.

Claims

1. Bearing life prediction method based on text features and convolutional neural network, characterized by: The method comprises: Acquire time domain vibration data of the rolling bearing to be predicted; performing noise reduction processing on the time-domain vibration data; The noise-reduced time-domain vibration data is subjected to Fourier transformation to obtain frequency-domain data; Extracting frequency domain features of the frequency domain data, and encoding the frequency domain data to obtain text features; Inputting the frequency domain features and the text features into a pre-trained neural network model to predict the predicted life of the rolling bearing to be predicted; Extracting frequency domain features of the frequency domain data includes: Extracting frequency domain features of the frequency domain data, and performing smoothing transformation and cumulative transformation; The frequency domain characteristics include variance, skewness, kurtosis, maximum value and margin factor; The smooth transformation utilizes SG filter function; The formula used for cumulative transformation is: X=[x1,x2,…,x i ,…,x N ] Among them, X is the horizontal feature and vertical feature of the input data, S i is the cumulative value at time i, C i is the cumulative transformation value at time i; Encoding the frequency domain data to obtain text features includes: The frequency domain data has a total of N samples, each sample D i Contains T time points, each time point is vibration amplitude data of length L; The range of the vibration amplitude data is divided into j intervals, and each sample D is counted. i The number of different amplitudes contained in each time point, these j intervals are regarded as j different words w j , get the word frequency matrix M; The word frequency feature is constructed by the following method; Perform singular value decomposition on the word frequency matrix M, reconstruct the matrix using the largest singular value, and obtain a low-dimensional feature matrix M* as the word frequency feature; The word frequency-inverse text frequency feature is constructed by the following method: The word frequency matrix M is divided by the sampling frequency to obtain a first intermediate quantity M'; The first intermediate quantity M' is multiplied by the normalized inverse text frequency to obtain a second intermediate quantity M", and the second intermediate quantity M" is subjected to singular value decomposition. The matrix is reconstructed using the largest singular value, and the obtained low-dimensional feature matrix M^ is used as the word frequency-inverse text frequency feature; Among them, word frequency-inverse text frequency includes word frequency coefficient and normalized inverse text frequency; in, Word frequency - inverse text frequency = word frequency coefficient * normalized inverse text frequency, denominator of normalized inverse text frequency | j:w j ∈D i | contains the word w j Sample D i The number of 2. The method according to claim 1, characterized in that Acquire time-domain vibration data of rolling element bearings, including: The horizontal vibration amplitude and the vertical vibration amplitude of the rolling bearing to be predicted are collected respectively; the time domain vibration data includes the horizontal vibration amplitude and the vertical vibration amplitude.

3. The method according to claim 1, characterized in that The time-domain vibration data is subjected to noise reduction processing, comprising: The time domain vibration data is subjected to noise reduction processing by using wavelet decomposition.

4. The method according to claim 1, wherein The time-domain vibration data is subjected to noise reduction processing, comprising: A sparse autoencoder is used to perform noise reduction on the time domain vibration data.

5. The method according to claim 1, wherein The neural network model contains 18 convolutional layers, where every three forward convolutional layers are equipped with a residual connection and an attention module. The neural network model uses an adaptive moment estimation optimizer for parameter optimization with a learning rate of 0.001; Batch normalization and ReLU activation functions are used between convolutional layers of the neural network model; The last layer of the neural network is a fully connected layer, and the sigmoid activation function is used to normalize the output of the neural network.

6. Bearing life prediction device based on text features and convolutional neural network, characterized in that: The device comprises: An acquisition module, used for acquiring time domain vibration data of a rolling bearing to be predicted; A noise reduction module is used to perform noise reduction processing on the time domain vibration data. a processing module for performing Fourier transform on the noise-reduced time-domain vibration data to obtain frequency-domain data; and extracting frequency-domain features of the frequency-domain data, and encoding the frequency-domain data to obtain text features; A prediction module, configured to input the frequency domain features and the text features into a pre-trained neural network model to predict the predicted life of the rolling bearing to be predicted; The process of extracting the frequency domain features of the frequency domain data by the processing module includes: Extracting frequency domain features of the frequency domain data, and performing smoothing transformation and cumulative transformation; The frequency domain characteristics include variance, skewness, kurtosis, maximum value and margin factor; The smooth transformation utilizes SG filter function; The formula used for cumulative transformation is: X=[x1,x2,…,x i ,…,x N ] Among them, X is the horizontal feature and vertical feature of the input data, S i is the cumulative value at time i, C i is the cumulative transformation value at time i; The processing module extracts and encodes the frequency domain data to obtain text features, including: The frequency domain data has a total of N samples, each sample D i Contains T time points, each time point is vibration amplitude data of length L; The range of the vibration amplitude data is divided into j intervals, and each sample D is counted. i The number of different amplitudes contained in each time point, these j intervals are regarded as j different words w j , get the word frequency matrix M; The word frequency feature is constructed by the following method; Perform singular value decomposition on the word frequency matrix M, reconstruct the matrix using the largest singular value, and obtain a low-dimensional feature matrix M* as the word frequency feature; The word frequency-inverse text frequency feature is constructed by the following method: The word frequency matrix M is divided by the sampling frequency to obtain a first intermediate quantity M'; The first intermediate quantity M' is multiplied by the normalized inverse text frequency to obtain a second intermediate quantity M", and the second intermediate quantity M" is subjected to singular value decomposition. The matrix is reconstructed using the largest singular value, and the obtained low-dimensional feature matrix M^ is used as the word frequency-inverse text frequency feature; Among them, word frequency-inverse text frequency includes word frequency coefficient and normalized inverse text frequency; in, Word frequency - inverse text frequency = word frequency coefficient * normalized inverse text frequency, denominator of normalized inverse text frequency | j:w j ∈D i | contains the word w j Sample D i The number of 7. The device according to claim 6, characterized in that The acquisition module is specifically used for: The horizontal vibration amplitude and the vertical vibration amplitude of the rolling bearing to be predicted are collected respectively; the time domain vibration data includes the horizontal vibration amplitude and the vertical vibration amplitude.

8. The device according to claim 6, characterized in that The noise reduction module is specifically used for: The time domain vibration data is subjected to noise reduction processing by using wavelet decomposition.

Images