A rolling bearing life prediction method based on optimal time spectrum and CNN-ALSTM network

Abstract

This invention discloses a method for predicting the life of rolling bearings based on optimal time-frequency spectrum and CNN-ALSTM network, comprising: Step 1, obtaining the vibration dataset X = {x1, x2, ..., x...} for the entire life cycle of the bearing. n Using the time-frequency energy evaluation index as the fitness function, the snake swarm optimization algorithm is used to optimize the adjustment factor p of the generalized S-transform to obtain the optimal time spectrum of each group of vibration data in the vibration dataset; Step 2, based on the optimal time spectrum obtained in Step 1 for each group of vibration data, the optimal time spectrum dataset S = {S1, S2, ..., S...} for the entire bearing life cycle is generated. n Step 3: Add a dense convolutional layer after the convolutional layers of the CNN network structure, and incorporate the self-attention mechanism into the first layer of the LSTM network to establish a CNN-ALSTM network; input the optimal time-spectrum dataset S into the CNN-ALSTM network for training to obtain the bearing's remaining service life prediction model. This invention can improve the prediction accuracy of rotating machinery bearing components and achieve the goal of accurately predicting the remaining service life of rotating machinery equipment.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing life prediction in rotating machinery, and in particular to a bearing life prediction method based on optimal time spectrum and CNN-ALSTM (Convolutional Neural Network self-attention Long Short-Term Memory, CNN-ALSTM) network. Background Technology

[0002] Rotating machinery, as a crucial component of enterprise production and operation, typically operates continuously at high intensity. Consequently, the service life of rotating machinery bearings often gradually degrades until failure. As a core component of rotating machinery, bearing failure inevitably leads to machine downtime, resulting in economic losses for the enterprise and potentially causing serious production accidents that threaten people's lives and property. Therefore, to reduce the probability of accidents caused by bearing life degradation, it is crucial to shift from reactive, reactive maintenance to proactive, predictive maintenance of rolling bearings. Thus, predicting the lifespan of rolling bearings is of great significance for ensuring production stability, improving product quality, and enhancing economic efficiency for enterprises.

[0003] Traditional rolling bearing life prediction methods typically extract time-domain signal features from input signals after denoising. However, in high-noise environments, fault signals exhibit strong non-stationary characteristics, resulting in weak fault features. The generated periodic pulses are often submerged in background noise and difficult to identify and extract. Consequently, the features extracted using this method are often inaccurate, affecting the prediction accuracy of the prediction network. Therefore, it is impossible to accurately predict the actual remaining life of the rolling bearing.

[0004] The existing methods have two major drawbacks: (1) Since bearing fault signals are often non-stationary signals and have strong noise, it is easy to cause inaccurate signal feature extraction if the fault characteristics of the time domain signal are not enhanced. (2) The data extracted from the time domain signal is limited by the time domain signal itself and cannot intuitively show the fault characteristics. When performing feature extraction, some features are easily lost, which also affects the accuracy of feature extraction and prediction. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the deficiencies in the existing technology by providing a rolling bearing life prediction method based on optimal time spectrum and CNN-ALSTM network, thereby improving the prediction accuracy and speed of rotating machinery bearing components and achieving the goal of accurately predicting the remaining service life of rotating machinery equipment.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] This invention provides a method for predicting the life of rolling bearings based on optimal time-frequency spectrum and CNN-ALSTM network. The method includes the following steps:

[0008] Step 1: Obtain the vibration dataset X = {x1, x2, ..., x} for the entire bearing life cycle. n Using the time-frequency energy evaluation index as the fitness function, the snake swarm optimization algorithm is used to optimize the adjustment factor p of the generalized S-transform, thereby obtaining the optimal time-frequency spectrum of each group of vibration data in the vibration dataset.

[0009] Step 2: Based on the optimal time spectrum obtained from each set of vibration data in Step 1, generate the optimal time spectrum dataset S = {S1, S2, ..., S...} for the entire bearing life cycle. n};

[0010] Step 3: Add a dense convolutional layer after the convolutional layer of the CNN network structure, and add the self-attention mechanism to the first layer of the LSTM network to establish the CNN-ALSTM network; feed the optimal time-spectrum dataset S into the CNN-ALSTM network for training to obtain the bearing's remaining service life prediction model.

[0011] Furthermore, step one of the present invention includes:

[0012] Step 1.1: Collect vibration data throughout the bearing's entire life cycle to obtain the original vibration dataset X = {x1, x2, ..., x...} n}, select the first vibration data x1, randomly select the initial value p0 of the adjustment factor p of the generalized S-transform in the range of (0,1), calculate the generalized S-transform of the vibration data x1, and obtain the time spectrum of x1;

[0013] Step 1.2: Use the time-frequency energy evaluation index E(p0) as the fitness function for the snake swarm optimization algorithm to calculate the fitness function at p0.

[0014] Step 1.3: The smaller the fitness function value, the better the time-frequency clustering of the fault components. The fitness function value can be obtained by calculation. The p value corresponding to the minimum value of the fitness function is the optimal p value under the current iteration number.

[0015] Step 1.4: Iterate the snake swarm optimization algorithm until the fitness function value reaches its minimum. Continuously optimize the population position to obtain a new optimal solution. When the optimization result reaches the optimal solution, the optimization ends.

[0016] Step 1.5, finally, output the optimal p value of the generalized S-transform. best The optimal frequency spectrum of vibration data x1 is obtained.

[0017] Furthermore, in step 1.1 of the present invention, the formula for calculating the generalized S-transform of the vibration data x1 to obtain the time spectrum of x1 is specifically as follows:

[0018]

[0019] in, τ is the generalized S-transform time spectrum of the vibration signal x1 at p0; f is the displacement factor; j is the frequency factor; and t is the time of the vibration signal x1.

[0020] Furthermore, in step 1.2 of the present invention, the time-frequency energy evaluation index E(p0) is used as the fitness function for the snake swarm optimization algorithm. The formula for calculating the fitness function at p0 is as follows:

[0021]

[0022] in,

[0023]

[0024]

[0025]

[0026] Where M(p0) is the time-frequency clustering degree at p0; R(p0) is the Renyi entropy at p0; K(p0) is the kurtosis at p0; N is the number of time sampling points; and M is the number of frequency sampling points.

[0027] Furthermore, in step 1.5 of the present invention, the formula for obtaining the optimal frequency spectrum of vibration data x1 is specifically as follows:

[0028] Based on step 1.1, the formula for calculating the generalized S-transform of vibration data x1 is obtained, leading to the optimal time spectrum of vibration data x1.

[0029] Furthermore, the method for generating the optimal time-spectrum dataset for the entire bearing life cycle in step 2 of the present invention is specifically as follows:

[0030] The vibration dataset X = {x1, x2, ..., x3} of the bearing throughout its entire life cycle is sequentially collected. n Each set of vibration data x in} i Perform the calculation in step one to obtain the optimal time spectrum S for each. i :

[0031]

[0032] Among them, Si Let x be the optimal time spectrum of the i-th data set in the vibration dataset X; i is the data number in the vibration dataset X; x i This refers to the i-th data set in the vibration dataset X;

[0033] Construct all the optimal time-spectrum values ​​into a new optimal time-spectrum dataset S = {S1, S2, ..., S} n}

[0034] Furthermore, the method in step 3 of the present invention specifically includes:

[0035] After the convolutional layers of the CNN network structure, a dense convolutional layer with a kernel of 1 is added. All time-spectrum datasets S are trained through the CNN network using two convolutional layers, a dense convolutional layer, and a pooling layer, to obtain the k-dimensional deep feature vector (y1, y2, ..., y...). k );

[0036] Using a k-dimensional depth feature vector (y1, y2, ..., y k The bearing remaining life prediction model is trained using an ALSTM network, with y1, y2, ..., y3 as the input vector and the remaining bearing life as the output vector. To remove k-dimensional deep feature vectors (y1, y2, ..., y3), a k-dimensional deep feature vector (y1, y2, ..., y3) is used. k To improve prediction accuracy, a self-attention mechanism is added to the first layer of the LSTM network to remove invalid features and redundant information from the LSTM, thereby obtaining a k-dimensional deep feature vector (y1, y2, ..., y). k The optimal weight allocation.

[0037] Furthermore, in the life prediction process, the present invention obtains the predicted value of the remaining life of the bearing by sequentially performing steps 1 to 3 on the collected bearing vibration signals.

[0038] The beneficial effects of this invention are:

[0039] 1. It can automatically optimize the spectrum of the generalized S-transform under conditions of high external noise interference. Engineers do not need to manually set and adjust the parameters of the generalized S-transform; the system automatically completes the parameter optimization.

[0040] 2. Establishing an optimal time-spectrum dataset for the entire bearing life cycle to replace the traditional vibration dataset focuses more on fault characteristic information, greatly improving the efficiency and accuracy of bearing life prediction. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0042] Figure 1 This is a flowchart of the present invention;

[0043] Figure 2 The bearing vibration signal containing strong noise is shown in Case 1 of this invention.

[0044] Figure 3 This is the time spectrum when p = 0.5 in Example 1 of this invention;

[0045] Figure 4 This is the time spectrum when p=1 in Case 1 of the present invention;

[0046] Figure 5 This is the optimal time spectrum found using the snake swarm optimization algorithm in Case 1 of this invention;

[0047] Figure 6 This is the time-domain diagram of the 426th group of vibration signals in bearing 3-3 in Case 2 of this invention;

[0048] Figure 7 This is the optimal time spectrum of the 426th group of vibration signals in bearing 3-3 in Case 2 of this invention;

[0049] Figure 8 The bearings 1-2 in Case 2 of this invention adopt the prediction results of this patent;

[0050] Figure 9 The prediction results of bearings 1-2 in Case 2 of this invention using the DRN-BiGRU network;

[0051] Figure 10 The bearings 1-3 in Case 2 of this invention adopt the prediction results of this patent;

[0052] Figure 11 The prediction results of bearings 1-3 in Case 2 of this invention using the DRN-BiGRU network;

[0053] Figure 12 The bearings 2-6 in Case 2 of this invention adopt the prediction results of this patent;

[0054] Figure 13 The prediction results of bearings 2-6 in Case 2 of this invention using the DRN-BiGRU network;

[0055] Figure 14 This is a diagram of the improved CNN-ALSTM network structure of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] like Figure 1As shown, the bearing life prediction method based on optimal time-frequency spectrum and CNN-ALSTM network includes the following steps:

[0058] Step 1: Using the time-frequency energy evaluation index as the fitness function, the adjustment factor p of the generalized S-transform is optimized through the snake swarm optimization algorithm to obtain the optimal time spectrum of the original signal.

[0059] Vibration data throughout the entire life cycle of the bearing is collected to obtain the original vibration dataset X = {x1, x2, ..., x...} n} Select the first vibration data x1, use the time-frequency energy evaluation index M(p) as the fitness function, and use the snake swarm optimization algorithm to optimize the adjustment factor p of the generalized S-transform to obtain the optimal time spectrum of the original signal.

[0060] The specific optimization steps are as follows:

[0061] (1) Randomly select the initial value p0 of the adjustment factor p of the generalized S-transform in the range of (0,1), calculate the generalized S-transform of the vibration data x1 at this time, and obtain the time spectrum of x1.

[0062]

[0063] in,

[0064] —The generalized S-transform of vibration signal x1 at p0;

[0065] τ — Displacement factor;

[0066] f—frequency factor;

[0067] j — Imaginary unit;

[0068] t — the time of vibration signal x1.

[0069] (2) Initialize the parameters of the snake swarm optimization algorithm, such as population size and number of iterations. Use the time-frequency energy evaluation index E(p0) as the fitness function for the snake swarm optimization algorithm to calculate the fitness function at p0.

[0070]

[0071]

[0072]

[0073]

[0074] in,

[0075] M(p0) – Time-frequency clustering at p0;

[0076] R(p0)——Renyi entropy at p0;

[0077] K(p0) — kurtosis at p0;

[0078] N—Number of time sampling points;

[0079] M — Number of frequency sampling points.

[0080] (3) The smaller the fitness function value, the better the time-frequency clustering of the fault components. The fitness function value can be obtained by calculation. The p value corresponding to the minimum fitness function value is the optimal p value under the current iteration number.

[0081] (4) Iterate the snake swarm optimization algorithm, continuously optimizing the population position to obtain a new optimal solution. When the optimization result reaches the optimal solution, the optimization ends. Finally, output the optimal p value of the generalized S-transform. best According to formula (1), the optimal time spectrum of vibration data x1 is obtained.

[0082] Step 2: Collect the vibration data X = {x1, x2, ..., x...} for the entire bearing life cycle. n Based on the calculations in step one, the optimal time spectrum corresponding to each group of vibration signals is obtained sequentially, generating the optimal time spectrum dataset S = {S1, S2, ..., S} for the entire bearing life cycle. n}

[0083] The vibration dataset X = {x1, x2, ..., x3} of the bearing throughout its entire life cycle is sequentially collected. n Each set of vibration data x in} i Perform the calculation in step one to obtain the optimal time spectrum S for each. i .

[0084]

[0085] in,

[0086] S i —The optimal time spectrum of the i-th data set in the vibration dataset X;

[0087] i — Data number in vibration dataset X;

[0088] x i —The i-th data set in the vibration dataset X.

[0089] Construct all the optimal time-spectrum values ​​into a new optimal time-spectrum dataset S = {S1, S2, ..., S} n}

[0090] Step 3: Input the optimal time-spectrum dataset S into the CNN-ALSTM network for training to obtain the bearing's remaining service life prediction model.

[0091] (1) After the convolutional layers of the traditional CNN (Convolutional Neural Network) network structure, add a dense convolutional layer with a kernel of 1. Train all time-spectrum datasets S through two convolutional layers, a dense convolutional layer, and a pooling layer of the CNN network to obtain the k-dimensional deep feature vector (y1, y2, ..., y) of the time-spectrum. k ).

[0092] (2) Using the k-dimensional depth feature vector (y1, y2, ..., y k Using the k-dimensional deep feature vector (y1, y2, ..., y3) as the input vector and the remaining bearing life as the output vector, an ALSTM network is used to train a bearing remaining life prediction model. To remove the k-dimensional deep feature vector (y1, y2, ..., y4), a further step is to... k To improve prediction accuracy, a self-attention mechanism is added to the first layer of the LSTM network to remove invalid features and redundant information from the LSTM, thereby obtaining a k-dimensional deep feature vector (y1, y2, ..., y). k The optimal weight allocation.

[0093] A self-attention mechanism is added to the input of the LSTM network to filter useful features and automatically assign weights, highlighting beneficial features and eliminating useless ones. This solves the problem of feature redundancy in CNN networks. A residual-temporal attention module is added before the fully connected layer of the LSTM network, which leverages the advantages of residuals to address the vanishing gradient problem caused by large amounts of experimental data in LSTM networks and optimizes prediction results. Its structure diagram is shown below. Figure 14 As shown.

[0094] In life prediction, the collected bearing vibration signals are processed sequentially through steps one through three to obtain the predicted value of the remaining bearing life.

[0095] <Example 1>

[0096] In practical engineering applications, the acquired signals often contain extremely strong background noise, and the impact of bearing failures can also be masked, affecting the feature extraction of the signals. Figure 2 The image shows a bearing vibration signal containing strong noise.

[0097] This patented technology is used to optimize the time-frequency spectrum. Based on the time-frequency transformation formula, the corresponding time-frequency spectrum image is obtained. Furthermore, the snake swarm optimization algorithm is used to optimize the adjustment factor p to obtain the optimal value and its corresponding optimal time-frequency spectrum.

[0098] To more intuitively demonstrate the superiority of the optimal time-frequency spectrum, two p values ​​were randomly selected and compared with the time-frequency images obtained from the optimal p value. Figure 3 and Figure 4 The image shows the time spectrum of the signal under different p values. Figure 5 The optimal time spectrum is obtained by the snake swarm optimization algorithm of this patent.

[0099] <Example 2>

[0100] The dataset used in this experiment is the PHM-2012 bearing life dataset released by IEEE in 2012. This dataset consists of 17 sets of bearing data, all representing the entire process of a bearing from start to finish. The sampling frequency is 25.6 kHz, the sampling time is 0.1 s, and each recording contains 2560 sampling points. The dataset includes three different operating conditions: Condition 1 is 4000 N and 1800 r / min, Condition 2 is 4200 N and 1650 r / min, and Condition 3 is 5000 N and 1500 r / min. Figure 6 This is the time-domain diagram of the vibration signal of bearing group 426 in bearing 3-3.

[0101] Based on step one, its optimal time spectrum is obtained. Figure 7 This is the optimal time spectrum of the vibration signal of the 426th group of bearings in bearing 3-3.

[0102] Based on step two, the optimal time spectrum for each bearing throughout its entire life cycle is obtained, and a new time spectrum dataset is constructed.

[0103] According to step three, the new time-spectrum dataset is used as input and the remaining lifetime is used as output to train the remaining lifetime prediction model through the CNN-ALSTM network.

[0104] During training, bearings 1-2, 1-3, and 2-6 were used as the test set, and the remaining bearings were used as the training set. Bearing 1-2 was sampled 871 times with a lifespan of 8710s; bearing 1-3 was sampled 2375 times with a lifespan of 23750s; and bearings 2-6 were sampled 701 times with a lifespan of 7010s. The time-spectrum images of these three bearing sets were labeled, and the bearing lifespan data was organized using a labeling method. The first image was labeled 1, and so on, with the i-th image representing the time-spectrum of the i-th sample. The remaining lifespan of the bearing is l. i The formula is:

[0105]

[0106] in:

[0107] n—Total time spectrum;

[0108] i — the image number.

[0109] The lifetime prediction methods of this patented method and existing DRN-BiGRU network prediction methods were compared. Figure 8 The prediction results of this patent are used for bearings 1-2. Figure 9 The prediction results for bearings 1-2 using the DRN-BiGRU network are shown. Figure 10 The prediction results of this patent are used for bearings 1-3; Figure 11 The prediction results for bearings 1-3 were obtained using the DRN-BiGRU network; Figure 12 The prediction results of this patent are used for bearings 2-6; Figure 13 The prediction results for bearings 2-6 using the DRN-BiGRU network are shown in Table 1. Table 1 compares the lifetime prediction errors of the two models. The smaller the values ​​of the average error, maximum error, and root mean square error, the higher the prediction accuracy. It is clear that the method presented in this paper has higher prediction accuracy.

[0110] Table 1 Comparison of lifetime prediction errors between the two sets of models

[0111]

[0112] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0113] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A rolling bearing life prediction method based on an optimal time-frequency spectrum and a CNN-ALSTM network, characterized in that, The method comprises the following steps: Step 1: Obtain the vibration dataset for the entire bearing life cycle. X ={ x 1, x 2, …, x n Using time-frequency energy evaluation index as the fitness function, the snake swarm optimization algorithm is applied to generalized... S Adjustment factor of transformation p Optimization is performed to obtain the optimal time spectrum for each group of vibration data in the vibration dataset; Step 2, according to step 1, the optimal time-frequency spectrum corresponding to each group of vibration data is obtained in turn, and the optimal time-frequency spectrum data set of the bearing full life cycle is generated S ={ S 1, S 2, …, S n}; Step 3, a dense convolutional layer is added after the convolutional layer of the CNN network structure, and a self-attention mechanism is added to the first layer of the LSTM network to establish a CNN-ALSTM network; the optimal time-frequency spectrum dataset S is sent into the CNN-ALSTM network for training to obtain a bearing residual service life prediction model; The method of step 3 specifically comprises: After the convolutional layer of the CNN network structure, a dense convolutional layer with a convolution kernel of 1 is added, and all the time-frequency spectrum data sets S After the training of the two convolutional layers, the dense convolutional layer and the pooling layer of the CNN network, the time-frequency spectrum k dimensional deep feature vector y 1, y 2, …, y k is obtained. by k 3D deep feature vector ( y 1, y 2, … , y k The bearing remaining life prediction model is trained using an ALSTM network, with the input vector being the bearing's remaining life and the output vector being the bearing's remaining life. To eliminate... k 3D deep feature vector ( y 1, y 2, … , y k By removing invalid features and redundant information from the LSTM network, the accuracy of predictions can be improved. A self-attention mechanism is added to the first layer of the LSTM network to obtain... k 3D deep feature vector ( y 1, y 2, … , y k ) optimal weight allocation; The self-attention mechanism is added to the input end of the LSTM network, which is used for screening favorable features and automatically assigning weights, highlighting favorable features and eliminating useless features; the residual-time attention module is added before the full connection layer of the LSTM network, which is used for solving the gradient disappearance problem caused by large experimental data of the LSTM network by using the advantage of residual, and optimizing the prediction result.

2. The rolling bearing lifetime prediction method based on the optimal time-frequency spectrum and CNN-ALSTM network according to claim 1, characterized in that, The step one comprises: Step 1.1: Collect vibration data throughout the entire life cycle of the bearing to obtain the raw vibration dataset. X ={ x 1, x 2, …, x n Select the first vibration data. x 1. Randomly select a generalized range within (0,1). S Adjustment factor of transformation p initial value p 0, calculate vibration data x 1 in a broad sense S Transformation, to obtain x The time spectrum of 1; Step 1.2, the time-frequency energy evaluation index E p 0) as the fitness function of the optimization algorithm of the snake optimization algorithm, the fitness function at time 0 is calculated p 0​ Step 1.3, when the smaller the fitness function value indicates the better the time-frequency aggregation of the fault component, the fitness function value can be obtained by calculation, and the minimum value of the fitness function corresponds to the optimal value at the current iteration p number p ; Step 1.4, iteration is carried out on the snake swarm optimization algorithm, so that the fitness function value reaches the minimum value, the population position is continuously optimized, a new optimal solution is obtained, and the optimization is ended when the optimization result reaches the optimal solution; Step 1.5, Finally, the optimal p value of the generalized S-transform is output p best , resulting in vibration data x 1 Optimal time-frequency spectrum.

3. The rolling bearing lifetime prediction method based on the optimal time-frequency spectrum and CNN-ALSTM network according to claim 2, characterized in that, In step 1.1, the vibration data is calculated x 1 is generalized S The formula of the time-frequency spectrum of 1 is specifically x 1 is generalized wherein is the vibration signal x 1 p 0 S transformed time-frequency spectrum; τ is the displacement factor; f is the frequency factor; j is the imaginary unit; t is the vibration signal x 1 4. The rolling bearing lifetime prediction method based on the optimal time-frequency spectrum and CNN-ALSTM network according to claim 3, characterized in that, In step 1.2, the time-frequency energy evaluation index E p 0) as a fitness function of the snake optimization algorithm, the fitness function at time t is calculated p 0, the formula of the fitness function is specifically:​ Wherein, wherein, M ( p 0) is the Renyi entropy at 0; p 0; R ( p p 0; K ( p 0; p 0; N is the number of time samples; M is the number of frequency samples.​ 5. The rolling bearing lifetime prediction method based on the optimal time-frequency spectrum and CNN-ALSTM network according to claim 4, characterized in that, In step 1.5, vibration data is obtained x 1The formula of the optimal time-frequency spectrum is as follows: According to the step 1.1, the vibration data is calculated x 1.1, the vibration data is calculated S 1.1, the vibration data is calculated x 1.1, the vibration data is calculated x 1.1, the vibration data is calculated 1.1, the vibration data is calculated 6. The rolling bearing lifetime prediction method based on the optimal time-frequency spectrum and CNN-ALSTM network according to claim 5, characterized in that, The method for generating the optimal time-frequency spectrum data set of the bearing full life cycle in step 2 is specifically: each set of vibration data in the vibration data set for the full life cycle of the bearing X = x 1, x 2, …, x n} are sequentially subjected to the calculation of step one to obtain respective optimal time-frequency spectra x i S i :​ wherein, S i for a vibration data set X first i group of data; i for a vibration data set X data number in; x i for a vibration data set X first i group of data; constructing all the optimal time-frequency spectra into a new optimal time-frequency spectrum dataset S ={ S 1, S 2, …, S n}。 7. The rolling bearing lifetime prediction method based on the optimal time-frequency spectrum and CNN-ALSTM network according to claim 1, characterized in that, When the life is predicted, the collected bearing vibration signal is sequentially operated according to the operation of steps 1 to 3, and the prediction value of the remaining life of the bearing can be obtained.

Images