Rolling bearing full-life evolution monitoring method based on noise reduction and dynamic activation

Abstract

The invention discloses a method for monitoring full-life evolution of a rolling bearing based on noise reduction and dynamic activation. The method comprises the following steps: firstly, processing vibration signals in different fault modes from overall and local angles by using a parallel self-adaptive noise reduction network module; performing feature fusion on the processed data; inputting the data subjected to feature fusion into a convolutional neural network to extract deep features; and finally, obtaining a monitoring result of the whole life evolution process of the rolling bearing through an optimal output activation function obtained through model training. According to the method, the signal-to-noise ratio of the input monitoring signal is improved through the parallel self-adaptive noise reduction network module, the learnable function is used for replacing a fixed activation function as a new activation function in the output module, dynamic activation of the output result is achieved, the learning ability and modeling ability of a monitoring model are improved, and the real-time performance of the monitoring model is improved. And the accuracy of the whole-life evolution monitoring process of the rolling bearing is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment condition monitoring, and particularly relates to a full-life evolution monitoring method for rolling bearings based on noise reduction and dynamic activation. Background Art

[0002] With the continuous development of science and technology, more and more rotating machinery is applied in fields such as industrial production, wind power generation, and transportation infrastructure. Since most of these mechanical equipment work in relatively extreme environments, the harsh environment will accelerate the failure and degradation of component materials. Once a failure occurs, it may trigger major safety accidents and cause unnecessary economic losses. One of the core components of rotating machinery is the rolling bearing, and the stability of its working state plays a crucial role in the operation of rotating machinery. Therefore, the condition monitoring and fault diagnosis of rolling bearings are one of the key tasks for the reliable operation and maintenance of rotating machinery.

[0003] The prerequisite for using the traditional model-based condition monitoring method (Luo M, Guo Y, Su Z, et al. Defect quantification evaluation of a rolling element bearing based on physical modelling and instantaneous vibration energy investigation [J]. Journal of Sound and Vibration, 2025, 600.) is that a highly reliable mathematical model of the target system can be established. This method starts from the mechanism level of the target system's fault evolution and can essentially achieve the condition monitoring of the target system; however, the mechanical system where the rolling bearing works is often a complex dynamic system, and it is difficult to establish a highly reliable mathematical model. At the same time, due to the differences in the inherent characteristics of different mechanical systems, only limited prior knowledge can be obtained, which leads to certain limitations in the flexibility and applicability of the model-based condition monitoring method.

[0004] The data-driven condition monitoring method (Hu C, Liu Z, Xiao X, et al. A degradation evaluation method with the convolutional neural network for the cyclic symmetry rolling bearing [J]. IOP Publishing Ltd, 2024.) can realize the monitoring of the target system state directly from the data level without relying on the accurate mathematical model of the target system through the analysis and learning of the historical operation data of the target system, and has strong adaptability and flexibility. However, there are still the following problems: First, the monitored data obtained has non-stationarity, and its randomly time-varying statistical characteristics will prevent the full extraction of fault feature information, resulting in some fault information being ignored or omitted, and the model cannot effectively capture the fault mode characteristics. Second, most models use a fixed network layer structure to learn the input data during the training process. However, the monitored data obtained in actual industrial production often comes from relatively complex non-linear systems. This fixed training strategy will cause the model to be unable to effectively model the different fault mode characteristics extracted, resulting in a decrease in the accuracy of the monitoring model. Finally, the full-life evolution monitoring of rolling bearings helps to identify potential faults in advance, optimize the maintenance strategy in time, and extend the service life of rolling bearings. However, the existing monitoring and diagnosis models only focus on the fault mode classification at a specific moment and do not pay attention to the damage evolution process of rolling bearings in the whole life cycle, making it difficult to meet the needs of actual industrial production. Summary of the Invention

[0005] In order to overcome the above shortcomings of the prior art, the purpose of the present invention is to provide a full-life evolution monitoring method for rolling bearings based on noise reduction and dynamic activation, which can effectively monitor the full-life evolution process of rolling bearings and improve the accuracy of the full-life evolution monitoring process of rolling bearings.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A full-life evolution monitoring method for rolling bearings based on noise reduction and dynamic activation. First, use a parallel adaptive noise reduction network module to process the vibration signals under different fault modes from the overall and local perspectives; then fuse the processed data; then input the data after feature fusion into a convolutional neural network to extract deep features; finally, obtain the monitoring results of the full-life evolution process of rolling bearings through the optimal output activation function obtained by model training.

[0008] A full-life evolution monitoring method for rolling bearings based on noise reduction and dynamic activation, comprising the following steps:

[0009] Step 1: Obtain the monitored vibration data sets of the rolling bearing under different fault modes where \(f = 1, 2, \ldots, F\), \(F\) is the number of different fault modes in the data set; \(Z\) is the number of vibration data samples included under different fault modes; is the sampling frequency of the monitored data under different fault modes; the monitored vibration signal data sets of the rolling bearing under different fault modes consist of the time - series measurement values of one or more vibration sensors of the same type, where the sampling frequency remains constant;

[0010] Step 2: Use a parallel adaptive dynamic noise - reduction network module to pre - process the input monitored signal from both the overall and local perspectives; use pooling operations and 1×1 convolution operations to extract the overall periodic features and local pulse features of the monitored signal respectively, and construct an adaptive threshold based on the extracted features; the extracted overall periodic features reflect the arrangement structure of the monitored signal in the time dimension, and the local pulse features reflect the characteristics of the fault mode; for the adaptive threshold \(\tau\) when extracting the overall periodic features g It is expressed as:

[0011]

[0012] where \(x\) is the input monitored data, \(N\) is the number of samples, \(C\) is the number of channels, \(L\) is the length of the monitored data, \(GAP(\cdot)\) is the global average pooling operation, \(W\) 1 and \(W\) 2 are the weights of the 1×1 convolution kernels used to extract the overall periodic features, \(\delta(C)\) is the ReLU activation function, and \(\sigma(\cdot)\) is the Sigmod activation function;

[0013] Similarly, for the adaptive threshold \(\tau\) when extracting the local pulse features m It is expressed as:

[0014]

[0015] where \(MAP(\cdot)\) is the max - pooling operation, \(W\) 3 and \(W\) 4 are the weights of the 1×1 convolution kernels used to extract the local pulse features;

[0016] Drawing on the idea of residual connection and processing the input monitored signal according to the calculated different adaptive thresholds, the output results of different branches of the parallel network are expressed as:

[0017]

[0018] where \(OUT\) i is the output result of different branches of the parallel network, \(\tau\) iare different adaptive thresholds;

[0019] Step 3: Construct a feature fusion layer based on the channel attention mechanism to fuse the output results of different branches of the parallel network; the output results of different branches are merged along the feature dimension to obtain a new input tensor θ. The channel attention weights are obtained by using a fully connected layer and an activation function, and a downsampling operation is used. Finally, the feature fusion result is expressed as:

[0020]

[0021] where θ n,c is the mean value of the c-th channel of the n-th tensor sample at all time steps, W 5 and are the weights of the fully connected layer, K is the size of the convolution kernel used in the downsampling operation, and p is the padding size of the convolution operation;

[0022] Step 4: Use a convolutional neural network to extract the deep features of signals in different fault modes, and construct three custom convolutional units. Each of the three custom convolutional units includes a convolutional layer, a batch normalization layer, a non-linear activation layer, and a pooling layer. The use of the batch normalization layer regularizes the data into the same interval, and its output result is expressed as:

[0023]

[0024] where x b is a batch of data, γ is the scaling parameter, β is the offset parameter, μ is the mean value of all data in the batch, σ 2 is the variance of all data in the batch, and ∈ is a fixed extremely small real number;

[0025] When optimizing the feature extraction ability of the network, a cross-entropy loss function and a triplet center loss function are introduced. The cross-entropy loss function is expressed as:

[0026]

[0027] where label f is the true label of the input monitoring signal, is the predicted probability of the model for the fault mode of the input monitoring signal;

[0028] The triplet center loss function uses a Gaussian kernel function to map the data to the Reproducing Kernel Hilbert Space (RKHS), and expresses the distance relationship between data in a non-linear way; assume that the data set a = {a 1 , a 2 , …, a m}, the process of using the Gaussian kernel function to represent the distance relationship between data is expressed as:

[0029]

[0030] where α is the bandwidth of the Gaussian kernel, and ||·|| is the Euclidean distance calculation function; for a specific sample a m obtain the center of its belonging category and the centers of non - same categories use the above - mentioned Gaussian kernel function to calculate the distance D between the sample a and the same - category center in the RKHS space m and the distance D between the sample a and the minimum different - category center intra ; based on this, the triplet center loss function is expressed as: inter ; based on this, the triplet center loss function is expressed as:

[0031]

[0032] where N b is the number of samples included in a certain batch, and margin is the boundary parameter to ensure that the distance between non - same - class samples is greater than the distance between same - class samples;

[0033] Combining the cross - entropy loss function and the triplet center loss function, the loss function of the network is expressed as:

[0034]

[0035] where is the cross - entropy loss function, is the triplet center loss function, λ is a real number in the range of [0, 1], and the optimization objective of the network is to minimize the loss function

[0036] Step 5, construct a model output module that replaces the fixed activation function with a learnable function. The KAN network based on the Kolmogorov - Arnold theorem consists of an input layer, a hidden layer, and an output layer. This network enables each neuron to independently learn the optimal output activation function. The operation process of the KAN network is specifically expressed as:

[0037]

[0038] where h is a multivariate continuous function representing the mapping relationship between the input e 1 , e 2 , …, e m and the output result, ψ pq is the feature transformation function of the hidden layer, c i is the coefficient optimized during model training, B i(·) is a B-spline basis function defined on the grid, and g q is the output of the hidden layer, and φ q is the non-linear activation function of the output layer;

[0039] The model output module uses the same loss function as the convolutional neural network in step 4, and the optimization also proceeds in the direction of improving the discriminant accuracy of the model, minimizing the distance within the same class and maximizing the distance between different classes;

[0040] Step 6, calculate the probabilities of the monitoring data for different fault modes, and normalize the output results h(e 1 , e 2 , …, e m ) of the model output module constructed in step 5 to between [0, 1] as the probability values of the monitoring data for different fault modes, which are specifically expressed as:

[0041]

[0042] Step 7, output the monitoring results of the whole life evolution process of the rolling bearing. After successively completing the above steps 1 to 6, a trained intelligent monitoring model is obtained. Divide the whole life cycle data of the rolling bearing according to its corresponding sampling strategy and input it into the intelligent monitoring model in turn, then the monitoring results of the whole life evolution process are obtained.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] The present invention proposes a method for monitoring the whole life evolution of a rolling bearing based on noise reduction and dynamic activation. The signal-to-noise ratio of the input monitoring signal is improved through a parallel adaptive noise reduction network module. The convolutional neural network is used to extract the deep features of the monitoring signal under different fault modes, and a learnable function is used instead of a fixed activation function as a new activation function in the model output module, which improves the learning ability and modeling ability of the monitoring model. The method of the present invention can effectively monitor the whole life evolution process of the rolling bearing and improve the accuracy of the whole life evolution monitoring process of the rolling bearing. Brief Description of the Drawings

[0045] Figure 1 is the flowchart of the method of the embodiment of the present invention.

[0046] Figure 2 is the whole life monitoring data of the rolling bearing in the embodiment of the present invention.

[0047] Figure 3 is the effective value curve of the whole life monitoring data of the rolling bearing in the embodiment of the present invention.

[0048] Figure 4 is the output result of the intelligent monitoring model of the present invention for the whole life data of the rolling bearing.

[0049] Figure 5 This is a schematic diagram for testing the disassembly and assembly of a rolling bearing in an embodiment of the present invention. Detailed implementation manners

[0050] The present invention will be further described below in conjunction with embodiments and drawings.

[0051] Referring to Figure 1 , a full-life evolution monitoring method for rolling bearings based on noise reduction and dynamic activation includes the following steps:

[0052] Step 1, obtain the monitored vibration data sets of the rolling bearing under different fault modes where f = 1, 2, …, F, and F is the number of different fault modes included in the data set; Z is the number of vibration data samples included under different fault modes; is the sampling frequency of the monitored data under different fault modes; the monitored vibration signal data sets of the rolling bearing under different fault modes are composed of the time-series measurement values of one or more vibration sensors of the same type, where the sampling frequency remains constant;

[0053] Step 2, use a parallel adaptive dynamic noise reduction network module to preprocess the input monitored signal from both the overall and local perspectives; use pooling operations and 1×1 convolution operations to extract the overall periodic features and local pulse features of the monitored signal respectively, and construct an adaptive threshold based on the extracted features; the extracted overall periodic features reflect the arrangement structure of the monitored signal in the time dimension, and the local pulse features reflect the characteristics of the fault mode; for the adaptive threshold τ when extracting the overall periodic features g It is expressed as:

[0054]

[0055] where x is the input monitored data, N is the number of samples, C is the number of channels, L is the length of the monitored data, GAP(·) is the global average pooling operation, W 1 and W 2 are the weights of the 1×1 convolution kernels used to extract the overall periodic features, δ(·) is the ReLU activation function, and σ(·) is the Sigmod activation function;

[0056] Similarly, for the adaptive threshold τ when extracting the local pulse features m It is expressed as:

[0057]

[0058] where MAP(·) is the maximum pooling operation, W 3 and W 4The weights of the 1×1 convolutional kernel used to extract local pulse features;

[0059] Drawing on the idea of residual connection and processing the input monitoring signal according to different calculated adaptive thresholds, the output results of different branches of the parallel network are expressed as:

[0060]

[0061] where, OUT i is the output result of different branches of the parallel network, and τ i are different adaptive thresholds;

[0062] Step 3, construct a feature fusion layer based on the channel attention mechanism to fuse the output results of different branches of the parallel network; the output results of different branches are merged along the feature dimension to obtain a new input tensor θ. The channel attention weights are obtained by using a fully connected layer and an activation function, and a downsampling operation is used. Finally, the obtained feature fusion result is expressed as:

[0063]

[0064] where, θ n,c is the mean value of the c-th channel of the n-th tensor sample at all time steps, W 5 and are the weights of the fully connected layer, K is the size of the convolutional kernel used in the downsampling operation, and p is the padding size of the convolutional operation;

[0065] Step 4, use a convolutional neural network to extract the deep features of signals in different fault modes, and construct three custom convolutional units. Each of the three custom convolutional units includes a convolutional layer, a batch normalization layer, a non-linear activation layer, and a pooling layer. The use of the batch normalization layer regularizes the data into the same interval, and its output result is expressed as:

[0066]

[0067] where, x b is a batch of data, γ is the scaling parameter, β is the offset parameter, μ is the mean value of all data in the batch, σ 2 is the variance of all data in the batch, and ∈ is a fixed extremely small real number;

[0068] When optimizing the feature extraction ability of the network, a cross-entropy loss function and a triplet center loss function are introduced. The cross-entropy loss function is expressed as:

[0069]

[0070] Among them, label f is the true label of the input monitoring signal, and is the predicted probability of the model for the fault mode of the input monitoring signal;

[0071] The triplet center loss function uses a Gaussian kernel function to map data to a Reproducing Kernel Hilbert Space (RKHS), expressing the distance relationship between data in a non-linear way; assuming that the data set a = {a 1 , a 2 , Ω, a m} in the original space, the process of using the Gaussian kernel function to represent the distance relationship between data is expressed as:

[0072]

[0073] Among them, α is the bandwidth of the Gaussian kernel, and ||·|| is the Euclidean distance calculation function; for a specific sample a m obtain its class center and the non-homogeneous class center use the above Gaussian kernel function to calculate the distance D m from the sample a to the homogeneous class center in the RKHS space intrea and the distance D inter from the sample a to the minimum non-homogeneous class center; based on this, the triplet center loss function is expressed as:

[0074]

[0075] Among them, N b is the number of samples included in a certain batch, and margin is the boundary parameter to ensure that the distance between non-homogeneous samples is greater than the distance between homogeneous samples;

[0076] Combining the cross-entropy loss function and the triplet center loss function, the loss function of the network is expressed as:

[0077]

[0078] Among them, is the cross-entropy loss function, is the triplet center loss function, λ is a real number in the range of [0, 1], and the optimization objective of the network is to minimize the loss function

[0079] Step 5, construct a model output module that replaces the fixed activation function with a learnable function. The KAN network based on the Kolmogorov-Arnold theorem consists of an input layer, a hidden layer, and an output layer. This network enables each neuron to independently learn the optimal output activation function. The operation process of the KAN network is specifically expressed as:

[0080]

[0081] ψ pq (e t )=∑ i c i B i (e t )

[0082]

[0083] Among them, h is a multivariate continuous function representing the mapping relationship between the inputs e 1 , e 2 , …, e m and the output result, ψ pq is the feature transformation function of the hidden layer, c i is the coefficient optimized during model training, B i (·) is the B-spline basis function defined on the grid, g q is the output of the hidden layer, φ q is the non-linear activation function of the output layer;

[0084] The model output module uses the same loss function as the convolutional neural network in step 4, and the optimization also proceeds in the direction of improving the model discrimination accuracy, minimizing the distance between the same class and maximizing the distance between different classes;

[0085] Step 6, calculate the probabilities of the monitoring data for different fault modes, and normalize the output result h(e 1 , e 2 , …, e m ) of the model output module constructed in step 5 to between [0, 1] as the probability values of the monitoring data for different fault modes, which are specifically expressed as:

[0086]

[0087] Step 7, output the monitoring results of the full life evolution process of the rolling bearing. After successively completing the above steps 1 to 6, the trained intelligent monitoring model is obtained. The full life cycle data of the rolling bearing is divided according to its corresponding sampling strategy and sequentially input into the intelligent monitoring model, and then the monitoring results of the full life evolution process are obtained.

[0088] The effectiveness of the method of the present invention is verified through an embodiment. This embodiment is derived from a full life cycle test platform for rolling bearings, and vibration signals in four typical modes of rolling bearing without fault, inner ring fault, outer ring fault, and rolling element fault are selected to train the monitoring model; the full life vibration data of the rolling bearing obtained from the full life cycle test platform of the rolling bearing is as Figure 2As shown; the effective value curve of the full-life vibration data can effectively reflect the degradation trend of the rolling bearing. The effective value curve of the calculation example is as Figure 3 shown. Based on the 3σ principle, the starting degradation time of the rolling bearing is determined to be 75 min, which is represented by the dots on the curve in Figure 3 ; dividing the full-life monitoring data according to the sampling strategy used and sequentially inputting it into the monitoring model according to the time sequence can obtain the output monitoring result of the model as Figure 4 shown. It can be observed from Figure 4 that the health state of the rolling bearing changed around 75 min, which is consistent with the conclusion reflected by the effective value curve of the full-life monitoring data of the rolling bearing. At the same time, the trend presented by the output monitoring result is consistent with the degradation trend of the rolling bearing; disassembling and inspecting the test rolling bearing, the result is as Figure 5 shown. It can be observed that obvious spalling has occurred on the outer ring of the rolling bearing, which is the same as the final manifestation of the outer ring fault in the model output, proving the effectiveness of the method of the present invention in practical applications.

[0089] It should be noted that without departing from the concept of the present invention, the adjustments and deformations made to the method of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A rolling bearing full life evolution monitoring method based on noise reduction and dynamic activation, characterized in that: Firstly, the parallel adaptive noise reduction network module is used to process the vibration signals under different fault modes from the overall and local perspectives; then the processed data is subjected to feature fusion; and then the feature fused data is input into the convolutional neural network to extract deep features; Finally, the optimal output activation function obtained through model training is used to obtain the monitoring results of the rolling bearing's full life evolution process.

2. According to claim 1, a rolling bearing full life evolution monitoring method based on noise reduction and dynamic activation is characterized in that: The following steps are involved: Step 1: Obtain monitoring vibration data sets of rolling bearings under different failure modes Where, f = 1, 2, …, F, F is the number of different fault modes contained in the data set; Z is the number of vibration data samples contained in different fault modes; The sampling frequency of monitoring data under different fault modes; the monitoring vibration signal data set of rolling bearing under different fault modes It consists of time-series measurements from one or more vibration sensors of the same type, where the sampling frequency is kept constant; In the second step, a parallel adaptive dynamic noise reduction network module is used to preprocess the input monitoring signal from the overall and local perspectives; the overall periodic features and local pulse features of the monitoring signal are extracted by pooling operation and 1×1 convolution operation respectively, and an adaptive threshold is constructed based on the extracted features; the extracted overall periodic features reflect the arrangement structure of the monitoring signal in the time dimension, and the local pulse features reflect the characterization of the fault mode; for the adaptive threshold τ when extracting the overall periodic features g It is expressed as: Where x is the input monitoring data, N is the number of samples, C is the number of channels, L is the monitoring data length, GAP(·) is the global average pooling operation, W1 and W2 are the weights of the 1×1 convolution kernel used to extract the overall periodic features, δ(·) is the ReLU activation function, and σ(·) is the Sigmoid activation function; Similarly, for the adaptive threshold τ for extracting local pulse features m It is expressed as: Among them, MAP(·) is the maximum pooling operation, W3 and W4 are the weights of the 1×1 convolution kernel used to extract local pulse features; Drawing on the idea of ​​residual connection and processing the input monitoring signal according to the calculated different adaptive thresholds, the output results of different branches of the parallel network are expressed as: Among them, OUT i is the output result of different branches of the parallel network, τ i are different adaptive thresholds; Step 3: Construct a feature fusion layer based on the channel attention mechanism to fuse the output results of different branches of the parallel network; merge the output results of different branches along the feature dimension to obtain a new input tensor θ, obtain the channel attention weight by using the fully connected layer and activation function, use the downsampling operation, and finally obtain the feature fusion result. It is expressed as: Among them, θ n,c is the mean of the cth channel of the nth tensor sample at all time steps, W5 and is the weight of the fully connected layer, K is the size of the convolution kernel used in the downsampling operation, and p is the padding size of the convolution operation; In the fourth step, a convolutional neural network is used to extract the deep features of different fault mode signals and three custom convolution units are constructed. The three custom convolution units include a convolution layer, a batch normalization layer, a nonlinear activation layer and a pooling layer. The batch normalization layer is used to normalize the data to the same interval. The output result is It is expressed as: Among them, x b is a batch of data, γ is the scaling parameter, β is the offset parameter, μ is the mean of all data in the batch, σ 2 is the variance of all data in the batch, ∈ is a fixed minimal real number; When optimizing the feature extraction capability of the network, the cross entropy loss function and the ternary center loss function are introduced. The cross entropy loss function is expressed as: Among them, label f is the true label of the input monitoring signal, is the predicted probability of the model for the failure mode of the input monitoring signal; The ternary center loss function uses a Gaussian kernel function to map the data to the Regenerated Hilbert Space (RKHS) and expresses the distance relationship between the data in a nonlinear way; assuming that the data set a in the original space is a={a1, a2,…, a m }, the process of using Gaussian kernel function to represent the distance relationship between data is expressed as: Among them, α is the bandwidth of the Gaussian kernel, ||·|| is the Euclidean distance calculation function; for a specific sample a m Get the category center to which it belongs With non-similar centers The Gaussian kernel function above is used to calculate the sample a in the RKHS space. m The center distance D of the same type intra The distance D from the minimum heterogeneous center inter ; Based on this, the triple center loss function is expressed as: Among them, N b is the number of samples included in a batch, and margin is the boundary parameter that ensures that the distance between non-same samples is greater than the distance between same samples; Comprehensive cross entropy loss function and ternary center loss function, network loss function It is expressed as: in, is the cross entropy loss function, is the ternary center loss function, λ is a real number in the range [0,1], and the optimization goal of the network is to minimize the loss function Step 5: Construct a model output module that uses a learnable function instead of a fixed activation function. The KAN network based on the Kolmogorov-Arnold theorem consists of an input layer, a hidden layer, and an output layer. The network enables each neuron to independently learn the optimal output activation function. The operation process of the KAN network is specifically expressed as follows: ψ pq (e t )=S i c i B i (e t ) Where h is a multivariable continuous function representing the input e1, e2, …, e m The mapping relationship between the output result and pq is the feature transformation function of the hidden layer, c i is the coefficient optimized during model training, B i (·) is the B-spline basis function defined on the grid, g q is the output of the hidden layer, φ q is the nonlinear activation function of the output layer; The model output module uses the same loss function as the convolutional neural network in step 4, and the optimization is also carried out in the direction of improving the model's discrimination accuracy, minimizing the same-class distance and maximizing the heterogeneous distance; Step 6: Calculate the probability of different failure modes for the monitoring data and convert the output result h (e1, e2, …, e m ) is normalized to [0,1] as the probability value of monitoring data for different failure modes, which is specifically expressed as: Step 7: Output the monitoring results of the rolling bearing's full life evolution process. After completing steps 1 to 6 above, a trained intelligent monitoring model is obtained. The rolling bearing's full life cycle data is divided according to the corresponding sampling strategy and input into the intelligent monitoring model in sequence to obtain the monitoring results of the full life evolution process.

Images