A method for diagnosing motor bearings based on frozen convolution kernels

By combining hierarchical convolutional neural networks with frozen and fine-tunable convolutional kernels, the problem of accurate identification of motor bearing component faults has been solved, achieving efficient component-level fault diagnosis, adapting to industrial small-sample scenarios, and improving operation and maintenance efficiency and identification accuracy.

CN123069153APending Publication Date: 2026-11-10NINGDE NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611268909.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately distinguish independent faults in the five major structural components of motor bearings. Furthermore, they are prone to overfitting under small sample sizes and high-noise conditions, failing to identify early, subtle faults, resulting in poor targeted maintenance and wasted resources.

Method used

A hierarchical convolutional neural network is adopted, which extracts general vibration features by using frozen convolutional kernels and adapts component-specific features by using fine-tunable convolutional kernels. Combined with batch normalization and random deactivation layers, the model is trained using a small sample dataset and a focus loss function is constructed for model optimization to achieve component-level fault identification.

Benefits of technology

It achieves high-precision identification of motor bearing component-level faults, improves operation and maintenance efficiency, reduces model training and inference computation, adapts to small sample scenarios in industrial sites, and improves fault identification accuracy and recall rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069153A_ABST
    Figure CN123069153A_ABST
Patent Text Reader

Abstract

This invention discloses a motor bearing diagnostic method based on frozen convolutional kernels. The method includes: first, acquiring the time-domain vibration signal of the motor bearing during operation; preprocessing it to generate a two-dimensional time-frequency feature map; constructing a hierarchical frozen convolutional kernel neural network model, splitting it into a frozen convolutional kernel module with fixed parameters and a finely adjustable convolutional kernel module with adaptively updated parameters, respectively extracting general vibration and impact features of the bearing and component-specific fault features of the inner ring, outer ring, rolling elements, cage, and seals; introducing a category-weighted focus loss function with a focus factor to balance the imbalance of fault samples among multiple components and the learning differences between easy and difficult samples, using the Adam optimizer to complete small-sample transfer training, and achieving secondary fault verification through a confidence discrimination mechanism. This invention can effectively suppress overfitting during small-sample training, significantly improve fault identification accuracy under strong noise and variable load industrial conditions, and achieve precise operation and maintenance at the bearing component level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method for diagnosing motor bearings based on frozen convolution kernels. Background Technology

[0002] Rolling bearings are core, easily damaged components of industrial motors. Statistics show that over 80% of motor mechanical failures originate from the failure of various bearing structural components. Bearings consist of five key structures: inner ring, outer ring, rolling elements, cage, and seals. The causes and vibration characteristics of failure in these different components are fundamentally different: the inner ring rotates with the shaft, and a failure in this component will produce speed-modulated periodic impacts; the outer ring is fixed to the end cap, and a failure in this component will produce fixed-frequency impact vibrations; rolling element failures will induce irregular alternating pulses; cage wear and deformation will lead to low-frequency turbulent vibrations; and aging or damage to the seals will cause frictional noise and weak amplitude disturbances.

[0003] Existing traditional bearing fault diagnosis technologies mainly rely on spectrum analysis, wavelet transform, and single deep learning classification models, which have significant technical shortcomings: First, traditional methods can only determine the overall bearing fault and cannot accurately distinguish individual faults of the inner ring, outer ring, rolling elements, cage, and seals, making it difficult to locate the failed component and resulting in poor targeted maintenance. Second, existing deep learning diagnostic models mostly adopt full-parameter training or global fine-tuning modes, which are prone to overfitting under the conditions of small sample sizes, strong noise, and variable speed in industrial bearing faults, and cannot capture the early and weak fault features of each component. Third, existing algorithms have not been adapted and optimized for the differentiated fault features of different bearing components, and general features and component-specific features interfere with each other, resulting in extremely low accuracy in identifying minor faults, seal faults, and low-frequency cage faults. Fourth, industrial bearing data suffers from severe sample imbalance, with a large number of normal samples and a scarcity of fault samples for various components. Traditional loss functions can cause the model to bias towards the normal category, missing early component faults.

[0004] Currently, the industry lacks an intelligent diagnostic solution that can specifically adapt to the differentiated fault characteristics of the five major structural components of bearings, adapt to small-sample industrial scenarios, and suppress overfitting, thus failing to achieve accurate fault diagnosis at the bearing component level. Summary of the Invention

[0005] The purpose of this invention is to provide a motor bearing diagnostic method based on frozen convolution kernels, which separates general vibration features from component-specific fault features through a hierarchical convolution kernel mechanism, thereby achieving high-precision fault identification of all structural components of the motor bearing.

[0006] To achieve the above objectives, the present invention provides a method for diagnosing motor bearings based on frozen convolution kernels, comprising the following steps: S1. Collect the original vibration signals of each component of the motor bearing and perform standardized preprocessing on the original vibration signals; S2. Perform short-time Fourier transform on the standardized vibration signal to generate a two-dimensional time-frequency feature map and perform pixel normalization processing. S3. Construct a hierarchical convolutional neural network diagnostic model; S4. Pre-train the basic weights of the hierarchical convolutional neural network diagnostic model, and complete the transfer training using a small sample dataset of actual measurements of multiple components of the target motor bearing. S5. Construct the model training objective function and iteratively optimize the trainable network parameters; S6. Collect the real-time vibration signal of the bearing of the motor under test, input it into the trained hierarchical convolutional neural network diagnostic model, perform secondary fault verification in combination with confidence threshold, and output the final diagnostic results and operation and maintenance suggestions.

[0007] Preferably, in S1, the standardized calculation formula is: ; In the formula, Indicates the first Amplitude values ​​at each vibration sampling point; Indicates the standardized first Amplitude values ​​at each vibration sampling point; This represents the mean value of the original entire vibration signal. This represents the standard deviation of the original entire vibration signal. ; This represents the total number of sampling points for a single vibration signal segment.

[0008] Preferably, in step S2, a short-time Fourier transform is performed on the standardized vibration signal to generate a two-dimensional time-frequency feature map containing time-domain impact characteristics and frequency-domain modulation characteristics. The time-frequency feature map is then subjected to pixel normalization. The short-time Fourier transform formula is as follows: ; In the formula, Indicates time ,frequency The corresponding time-frequency complex amplitude; This represents the length of the sliding window in the short-time Fourier transform. This indicates the Hanning window function at the vibration sampling point. Weight at each location; The imaginary unit; The formula for normalizing the pixels of the image is: ; In the formula, Represents the coordinates of the normalized time-frequency feature map. The corresponding pixel feature value; Representing the coordinates of the time-frequency matrix The corresponding amplitude; Represents the whole-frame time-frequency amplitude matrix; This represents the maximum amplitude of the time-frequency characteristic spectrum. This represents the minimum amplitude of the time-frequency spectrum.

[0009] Preferably, in S3, a hierarchical convolutional neural network diagnostic model for fault identification of multiple structural components of bearings is constructed. The hierarchical convolutional neural network diagnostic model includes, in sequence, an input layer, a frozen convolutional kernel module, a batch normalization layer, a fine-tunable convolutional kernel module, a random deactivation layer, a global pooling layer, a fully connected layer, and a multi-component fault classification output layer. The frozen convolution kernel module is used to extract periodic impact, vibration edge, and noise substrate features of motor bearings, including... Each convolutional layer has a convolution operation expression as follows: ; In the formula, Indicates the frozen convolution kernel module. The input feature map of the convolutional layer is at the pixel location Eigenvalues ​​at; Indicates the frozen convolution kernel module. Pixel positions in the output feature map of the convolutional layer Eigenvalues ​​at; The relative coordinate vector within the convolution kernel; The Euclidean modulus of the convolution kernel relative to its coordinates; Define the receptive field of the convolution kernel; The characteristic spatial attenuation weighting function of bearing vibration is used to simulate the natural attenuation characteristics of bearing vibration signal with the distance of the sensing field. It is a non-linear activation function; Indicates the frozen convolution kernel module. Layer convolution kernel weights; Indicates the frozen convolution kernel module. Layer convolution bias parameters.

[0010] Preferably, in S3, the fine-tunable convolution kernel module is used to adaptively learn the fault modulation characteristics, frequency characteristics, and impact interval characteristics of the bearing inner ring, outer ring, rolling elements, cage, and seals, so as to realize the differentiated identification of multi-component faults. The fine-tunable convolution kernel module includes... Each convolutional layer has a convolution operation expression as follows: ; In the formula, This indicates the fine-tunable convolution kernel module. Layer convolution input feature map pixel position Eigenvalues ​​at; This indicates the fine-tunable convolution kernel module. Pixel positions in the output feature map of the convolutional layer Eigenvalues ​​at; The component failure feature enhancement weighting function; This indicates the fine-tunable convolution kernel module. Layer-specific trainable convolutional kernel weights; This indicates the fine-tunable convolution kernel module. Layer-trainable convolutional bias parameters; It is a non-linear activation function.

[0011] Preferably, in S4, the basic weights of the hierarchical convolutional neural network diagnostic model are pre-trained using a publicly available bearing source domain dataset. All weights and bias parameters of the convolutional kernel module are fixed and frozen, while only the fine-tunable convolutional kernel module and fully connected layer parameters are opened. Transfer training is completed using a small sample dataset of actual measurements of multiple components of the target motor bearing. Among them, the ratio of frozen convolutional kernel layers to the total number of convolutional layers. Satisfying 0.3≤ ≤0.6, proportion The calculation formula is: ; In the formula, To freeze the number of convolutional layers; The number of convolutional layers can be finely adjusted; The transfer learning process uses the Adam optimizer to update the trainable parameters. The parameter update formula is as follows: ; In the formula, for The set of trainable parameters of the model before the next iteration update For the first After the [number]th iteration update The set of trainable parameters for which the iteration step takes effect; Number of training iterations for the model; For the first The first-order momentum corresponding to each iteration; for The second-order momentum corresponding to each iteration; The first-order momentum decay coefficient; It is the second-order momentum decay coefficient; The base learning rate is used to train the model; Here is the Adam numerical stability constant; This represents the gradient of the focus loss function with respect to the trainable parameters in the current iteration. Hadamard is the element-wise product; This is the corrected first-order momentum; This is the corrected second-order momentum.

[0012] Preferably, a batch normalization layer is added between the frozen convolutional kernel module and the fine-tunable convolutional kernel module. The batch normalization formula is as follows: ; In the formula, The convolutional layer outputs the original feature values; This represents the mean of the features in the current training batch. The variance of the features in the current training batch; These are the batch normalized learnable scaling and offset parameters, respectively. To prevent the batch normalization from eliminating the zero constant; A random deactivation layer is added after the fine-tunable convolutional module to suppress overfitting on small samples. The formula for calculating random deactivation is: ; In the formula, The original features are output for the upper convolutional layer; Features after random inactivation processing; Let be a random variable that follows a uniform distribution from 0 to 1; Preset the neuron inactivation probability; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

[0013] Preferably, in S5, a focus loss function containing a focus factor is constructed as the model training objective function, and the trainable network parameters are iteratively optimized. The expression for the focus loss function is: ; In the formula, This represents the total loss value for a single training batch; Indicates the number of samples in a single training batch; This indicates the total number of multiple structural failure categories in the bearing; Indicates the first Category weighting coefficients for component failures; Indicates the fault focus factor; Indicates the first Each sample of real component failure categories has a unique hot label; Indicates the model output of the first The sample belongs to the first Predicted probability of type of failure; Minimizing the expected batch focus loss is taken as the overall iterative optimization objective of the model. The overall training objective expression is: ; In the formula, The overall training target loss for the model; Indicates the training parameter set Perform a minimum solution optimization on top; This is used to calculate the expected mean of a batch of samples.

[0014] Preferably, in step S6, real-time vibration signals of the bearing of the motor under test are acquired, and steps S1-S2 are repeated to generate the time-frequency feature spectrum of the bearing under test. This spectrum is then input into the trained diagnostic model. Based on a multi-component, multi-gradient fault identification system and combined with the vibration fault feature identification mechanism of each bearing structural component, the model outputs the bearing health status, the location of the faulty component, and the corresponding damage degree, with a pre-set confidence threshold. When the model has the highest prediction confidence If the fault characteristics of the bearing component are determined to be masked by noise or not significant, an early warning signal is output and vibration data is re-acquired for secondary diagnosis.

[0015] Preferably, the multi-component, multi-gradient fault identification system identifies fault categories including normal bearing condition, inner ring fault, outer ring fault, rolling element fault, cage fault, and seal fault. A quantitative grading criterion is constructed using a vibration amplitude multiplier coefficient, the expression for which is: ; In the formula, This is the vibration amplitude multiplier for bearing failure; This represents the effective value of the steady-state vibration amplitude of the bearing under test. This is the effective value of the reference vibration amplitude of a normal, fault-free bearing under the same operating conditions. Based on the vibration amplitude multiplier, combined with the three-dimensional judgment index of fault characteristic spectrum significance and noise inundation degree, the faults of each structural component are divided into three levels of damage: mild, moderate, and severe. The judgment rules are as follows: when The fault characteristics of the component are weak in spectrum, the high-frequency harmonics are not significant, the fault characteristics are easily drowned out by industrial noise, the motor bearing only has micro wear, micro pitting, and slight frictional distortion, and there are no structural material peeling or deformation defects, the motor damage level is mild damage. when At that time, the fault spectrum of the component was determined to be clear and stable, with regular periodic impulse modulation characteristics, low noise interference, non-destructive structural defects in the motor bearing, and the motor damage level was moderate. when At that time, the component fault characteristic spectrum was rich in high amplitude and multi-order harmonics, the impact pulse density was high, the motor bearing had serious structural defects, and the motor damage level was severe.

[0016] Therefore, the present invention employs the above-mentioned method for diagnosing motor bearings based on frozen convolution kernels, and the beneficial effects are as follows: (1) This invention breaks through the limitations of traditional models that can only determine overall faults. It can accurately distinguish the independent faults of the five major structural components and identify the light, medium and heavy damage levels of each component. It provides accurate data support for targeted bearing repair and component replacement, greatly improves motor operation and maintenance efficiency, and avoids the waste of resources caused by blindly replacing the entire bearing.

[0017] (2) The present invention freezes the general feature convolution kernel, which greatly reduces the number of trainable parameters of the model, reduces the model fitting capacity, and avoids the model learning invalid features such as industrial noise and load interference under small sample working conditions. At the same time, it retains the pre-trained general vibration feature extraction capability and only fine-tunes and optimizes the bearing component-specific features. Compared with the traditional full fine-tuning and zero-training model, the accuracy of multi-component fault identification under small sample working conditions is improved by 10%~18%, which effectively adapts to the actual scenario where industrial field fault samples are scarce.

[0018] (3) In view of the characteristics of a large number of normal samples and a scarcity of various component failure samples, especially minor failure samples, in industrial sites, this invention introduces a category weight coefficient to increase the loss constraint weight for rare component failures and early minor failures, avoids the model from being biased towards predicting normal categories, and balances the identification accuracy and recall rate of the five major structural components of bearings, adapting to the distribution characteristics of real industrial data.

[0019] (4) This invention only updates the gradient of some convolutional and fully connected layer parameters, which greatly reduces the amount of backpropagation computation, significantly improves the convergence speed of model training iteration, and significantly reduces the training time; the model inference stage has a lightweight structure, does not require high-end computing power support, and can be directly deployed on embedded edge monitoring terminals to meet the 7×24-hour online real-time component-level fault diagnosis requirements of industrial motors.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of an embodiment of a motor bearing diagnosis method based on frozen convolution kernels according to the present invention; Figure 2 This is a comparison chart of the overall performance of three schemes in a multi-component fault diagnosis method based on frozen convolution kernels according to an embodiment of the present invention. Figure 3 This is an embodiment of the present invention, which is a method for diagnosing motor bearings based on frozen convolution kernels. It is a diagram showing the independent identification index of faults in each bearing structural component of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] like Figure 1 As shown, this invention provides a method for diagnosing motor bearings based on frozen convolution kernels, comprising the following steps: S1. Collect the original vibration signals of each component of the motor bearing during operation, and perform standardized preprocessing on the original vibration signals to eliminate the amplitude baseline differences caused by load and speed fluctuations.

[0025] The standardized calculation formula is as follows: ; In the formula, Indicates the first Amplitude values ​​at each vibration sampling point; Indicates the standardized first Amplitude values ​​at each vibration sampling point; This represents the mean value of the original entire vibration signal. This represents the standard deviation of the original entire vibration signal. ; This represents the total number of sampling points for a single vibration signal segment.

[0026] S2. Perform a short-time Fourier transform on the standardized vibration signal to generate a two-dimensional time-frequency feature map containing time-domain impact characteristics and frequency-domain modulation characteristics. Then, perform pixel normalization on the time-frequency feature map. The short-time Fourier transform formula is as follows: ; In the formula, Indicates time ,frequency The corresponding time-frequency complex amplitude; This represents the length of the sliding window in the short-time Fourier transform. This indicates the Hanning window function at the vibration sampling point. Weight at each location; It is the imaginary unit.

[0027] The formula for normalizing the pixels of the image is: ; In the formula, Represents the coordinates of the normalized time-frequency feature map. The corresponding pixel feature value; Representing the coordinates of the time-frequency matrix The corresponding amplitude; Represents the whole-frame time-frequency amplitude matrix; This represents the maximum amplitude of the time-frequency characteristic spectrum. This represents the minimum amplitude of the time-frequency spectrum.

[0028] S3. Construct a hierarchical convolutional neural network diagnostic model for fault identification of multiple structural components in bearings. The hierarchical convolutional neural network diagnostic model includes, in sequence, an input layer, a frozen convolutional kernel module, a batch normalization layer, a fine-tunable convolutional kernel module, a random deactivation layer, a global pooling layer, a fully connected layer, and a multi-component fault classification output layer.

[0029] The frozen convolution kernel module of this invention is used to extract basic features such as periodic impact, vibration edge and noise base features of motor bearings, providing general feature support for subsequent extraction of multi-component specific fault features.

[0030] The frozen convolution kernel module includes Each convolutional layer has a convolution operation expression as follows: ; In the formula, Indicates the frozen convolution kernel module. The input feature map of the convolutional layer is at the pixel location Eigenvalues ​​at; Indicates the frozen convolution kernel module. Pixel positions in the output feature map of the convolutional layer Eigenvalues ​​at; The relative coordinate vector within the convolution kernel; The Euclidean modulus of the convolution kernel relative to its coordinates; Define the receptive field of the convolution kernel; The characteristic spatial attenuation weighting function of bearing vibration is used to simulate the natural attenuation characteristics of bearing vibration signal with the distance of the sensing field. It is a non-linear activation function; Indicates the frozen convolution kernel module. The weights of the convolutional kernels are fixed and not updated throughout the entire training process; Indicates the frozen convolution kernel module. The convolutional layer bias parameters are fixed and not updated throughout the training process.

[0031] The present invention provides a finely adjustable convolution kernel module for targeted adaptive learning of the fault modulation characteristics, frequency characteristics, and impact interval characteristics of the inner ring, outer ring, rolling elements, cage, and seals of bearings, thereby enabling differentiated identification of faults in multiple components.

[0032] The fine-tunable convolution kernel module includes... Each convolutional layer has a convolution operation expression as follows: ; In the formula, This indicates the fine-tunable convolution kernel module. Layer convolution input feature map pixel position Eigenvalues ​​at; This indicates the fine-tunable convolution kernel module. Pixel positions in the output feature map of the convolutional layer Eigenvalues ​​at; This represents a component fault feature enhancement weighting function that specifically strengthens weak local fault features of bearings and suppresses global noise interference. This indicates the fine-tunable convolution kernel module. The trainable convolutional kernel weights of each layer participate in gradient updates during the transfer learning phase. This indicates the fine-tunable convolution kernel module. The trainable convolutional bias parameters of each layer are transferred to gradient updates during the training phase. It is a non-linear activation function.

[0033] Furthermore, this invention adds a batch normalization layer between the frozen convolution kernel module and the fine-tunable convolution kernel module, and the batch normalization formula is as follows: ; In the formula, The convolutional layer outputs the original feature values; This represents the mean of the features in the current training batch. The variance of the features in the current training batch; These are the batch normalized learnable scaling and offset parameters, respectively. Batch normalization is used to eliminate zero minimum constants. By unifying the fault characteristic distribution of multiple components through batch normalization, the convergence speed and identification stability of the model are improved.

[0034] A random deactivation layer is added after the fine-tunable convolutional module to suppress overfitting on small samples. The formula for calculating random deactivation is: ; In the formula, The original features are output for the upper convolutional layer; Features after random inactivation processing; Let be a random variable that follows a uniform distribution from 0 to 1; Preset the neuron inactivation probability; This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise, adapting to training scenarios with small sample fault data of various bearing components.

[0035] S4. The basic weights of the hierarchical convolutional neural network diagnostic model are pre-trained using the publicly available bearing source domain dataset. All weights and bias parameters of the convolution kernel module are fixed and frozen. Only the fine-tunable convolution kernel module and fully connected layer parameters are opened. Transfer training is completed using a small sample dataset of actual measurements of multiple components of the target motor bearing.

[0036] Among them, the ratio of frozen convolutional kernel layers to the total number of convolutional layers. Satisfying 0.3≤ ≤0.6, proportion The calculation formula is: ; In the formula, To freeze the number of convolutional layers; The number of convolutional layers can be finely adjusted.

[0037] This invention sets 0.3≤ A scaling range of ≤0.6 is suitable for balancing the extraction of general and specific features from multiple bearing components, avoiding feature redundancy or feature loss.

[0038] The transfer learning process uses the Adam optimizer to update the trainable parameters. The parameter update formula is as follows: ; In the formula, for The set of trainable parameters of the model before the next iteration update For the first After the [number]th iteration update The set of all trainable parameters for which the iteration step takes effect; Number of training iterations for the model; For the first The first-order momentum corresponding to each iteration is used to accumulate historical gradient information, smooth parameter update fluctuations, and accelerate model convergence. for The second-order momentum corresponding to each iteration is used to accumulate historical gradient squared information and adaptively adjust the parameter update step size. This is the corrected first-order momentum; This is the corrected second-order momentum. It is the first-order momentum decay coefficient, used to control the retention weight of historical first-order momentum and smooth the gradient update trajectory; It is the second-order momentum decay coefficient, used to control the retention weight of historical second-order momentum, so as to achieve adaptive adjustment of the learning rate; The base learning rate is used to train the model and control the update magnitude of network parameters in a single iteration, balancing training convergence speed and parameter stability. This is the Adam numerical stability constant, used to avoid numerical overflow issues where the denominator is zero, and to ensure stability during the training process; The gradient of the focus loss function with respect to the trainable parameters in the current iteration represents the direction and magnitude of the parameter update to be optimized. Hadamard element-wise product represents the individual multiplication of corresponding elements of two matrices, enabling precise pixel-wise statistics of second-order gradient information.

[0039] S5. Construct a focus loss function containing a focus factor as the objective function for model training, iteratively optimize the trainable network parameters, focus on small, difficult-to-distinguish, and early weak fault samples of bearings, balance the imbalance of fault samples of multiple components and the learning difference between easy and difficult samples, and solve the defects of traditional cross-entropy in insufficient fitting of low-frequency weak faults and high false negative rate.

[0040] The expression for the focus loss function is: ; In the formula, This represents the total loss value for a single training batch; Indicates the number of samples in a single training batch; This indicates the total number of multiple structural failure categories in the bearing; Indicates the first Category weighting coefficients for component failures; This represents the fault focusing factor, used to suppress the loss weights of easily learnable normal samples and severely faulty samples, and to focus on difficult-to-learn early mild faults and low-frequency weak fault samples. Indicates the first Each sample of real component failure categories has a unique hot label; Indicates the model output of the first The sample belongs to the first Predicted probability of class-specific faults.

[0041] This invention uses minimizing the expected batch focus loss as the overall iterative optimization objective of the model. The overall training objective expression is: ; In the formula, The overall training target loss for the model; Indicates the training parameter set Perform a minimum solution optimization on top; The expected mean of the batch samples is calculated to balance the global sample loss distribution and ensure the integrity and stability of the model's fitting of multi-component fault features.

[0042] S6. Acquire real-time vibration signals of the bearing of the motor under test. Repeat steps S1-S2 to generate the time-frequency feature spectrum of the bearing under test. Input the spectrum into the trained diagnostic model. Based on the multi-component multi-gradient fault identification system and combined with the vibration fault feature identification mechanism of each bearing structural component, output the bearing health status, fault component location, and corresponding damage degree. Set the confidence threshold. When the model has the highest prediction confidence If the fault characteristics of the bearing component are determined to be masked by noise or not significant, an early warning signal is output and vibration data is re-acquired for secondary diagnosis.

[0043] The multi-component, multi-gradient fault identification system of this invention can identify fault categories including normal bearing condition, inner ring fault, outer ring fault, rolling element fault, cage fault, and seal fault. A quantitative grading criterion is constructed using a vibration amplitude multiplier coefficient, the expression of which is: ; In the formula, This is the vibration amplitude multiplier for bearing failure; This represents the effective value of the steady-state vibration amplitude of the bearing under test. This represents the effective value of the reference vibration amplitude of a normal, fault-free bearing under the same operating conditions.

[0044] This invention classifies structural component failures into three levels of damage: mild, moderate, and severe, based on a vibration amplitude multiplier coefficient, combined with a three-dimensional judgment index of fault characteristic spectrum significance and noise inundation degree. The judgment rules are as follows: when The fault characteristics of the component were determined to be weak, with insignificant high-frequency harmonics. The fault characteristics were easily drowned out by industrial noise. The motor bearings only showed micro-wear, micro-pitting, and slight frictional distortion, with no structural material spalling or deformation defects. The motor damage level was classified as mild damage.

[0045] when At that time, the fault spectrum of the component was determined to be clear and stable, with regular periodic impulse modulation characteristics, low noise interference, non-destructive structural defects in the motor bearing, and the motor damage level was moderate.

[0046] when At that time, the component fault characteristic spectrum was rich in high amplitude and multi-order harmonics, the impact pulse density was high, the motor bearing had serious structural defects, and the motor damage level was severe.

[0047] Based on the above amplitude ratio quantization formula and spectral feature criteria, the model completes the accurate matching and determination of the fault level of each structural component, and realizes the dual refined diagnosis output of accurate positioning of bearing failure components and fault damage degree.

[0048] Example 1: This embodiment uses a comprehensive test bench for simulating motor bearing failures. The test motor is a three-phase asynchronous motor, matched with a mainstream 6205 deep groove ball bearing (including a complete inner ring, outer ring, steel rolling elements, nylon cage, and double-sided rubber seals); the signal sampling frequency is... The motor has a rated operating speed of 1480 r / min and a variable load of 0~80%, simulating an industrial variable operating condition scenario.

[0049] Signal processing parameters: Number of sampling points for a single vibration signal segment The short-time Fourier transform uses a Hanning window with a window length of [missing information]. With a window overlap rate of 50%, a 128×128 pixel standardized time-frequency feature map is finally generated.

[0050] Network structure and training hyperparameters: number of layers in the frozen convolutional kernel module The number of convolution kernel modules can be finely adjusted. Freezing ratio ; batch normalized minimum constant Random inactivation probability =0.25; Adam optimizer parameters , Training learning rate Stability constant Focus loss focus factor Confidence threshold The maximum number of iterations is 120, and an early stopping strategy is enabled to prevent overfitting.

[0051] This embodiment constructs a dataset of 9 types of faults covering the entire bearing structure. The specific categories are: C1 normal condition, C2 minor inner ring fault, C3 severe inner ring fault, C4 minor outer ring fault, C5 severe outer ring fault, C6 rolling element fault, C7 cage wear fault, C8 seal aging fault, and C9 seal damage fault.

[0052] Dataset partitioning: The source domain dataset was pre-trained using the CWRU public bearing general dataset; the target domain dataset was a small sample dataset of multiple components measured on the test bench, with 60 time-frequency images per class in the training set and 30 images per class in the test set, simulating a small sample industrial scenario.

[0053] Comparison schemes: Scheme A is the multi-part diagnostic method for freezing convolutional kernels of this invention; Scheme B is the traditional full-network fine-tuning transfer learning method; Scheme C is the traditional CNN training from scratch diagnostic method. Tables 1 and 2 show the comparison results of the three schemes. Table 1. Comparison of overall multi-component fault diagnosis performance among the three schemes

[0054] Table 2 Independent Fault Identification Indicators for Each Bearing Structural Component of the Invention

[0055] From Table 1 and Figure 2As can be seen, the overall diagnostic accuracy of Scheme A of this invention reaches 95.87%, which is 9.75 percentage points higher than the 86.12% of the traditional full-network fine-tuning scheme B, and 15.42 percentage points higher than the 80.45% of the CNN training from scratch scheme C. In terms of the recognition rate of weak / low-frequency faults such as cage low-frequency faults and weak friction faults of seals, this invention reaches 93.21%, which is 13.65 and 20.88 percentage points higher than the 79.56% of scheme B and the 72.33% of scheme C, respectively, effectively solving the technical problem that traditional models cannot distinguish weak faults at the bearing component level.

[0056] In terms of training efficiency, the single-round training time of this invention is only 1.75 hours, which is 33.2% shorter than the 2.62 hours of scheme B and 45.3% shorter than the 3.20 hours of scheme C. Moreover, there is no obvious overfitting phenomenon, while scheme B shows mild overfitting and scheme C shows severe overfitting.

[0057] From Table 2 and Figure 3 As can be seen, this invention achieves high-precision independent identification of each bearing structural component: rolling element fault accuracy is 96.43% and recall rate is 95.87%, inner ring severe fault accuracy is 97.35% and recall rate is 97.11%, and outer ring severe fault accuracy is 97.92% and recall rate is 98.05%. Even for cage wear faults (accuracy rate 92.89% and recall rate 92.15%) and seal aging faults (accuracy rate 92.35% and recall rate 91.88%) with subtle characteristics, the identification accuracy is maintained above 92%. This fully verifies the technical effect of freezing the convolution kernel to retain general vibration characteristics and finely adjusting the convolution kernel to accurately adapt to the specific fault characteristics of the component, and fully meets the engineering requirements of precise operation and maintenance of industrial motor bearing components.

[0058] Therefore, this invention employs the aforementioned motor bearing diagnostic method based on frozen convolution kernels. By using hierarchical frozen convolution kernels, it achieves layered extraction of general bearing fault features and component-specific features. Combined with a three-dimensional quantitative judgment standard of amplitude ratio formula, spectral saliency, and noise inundation level, it can accurately identify all structural components of the bearing and light, medium, and severe gradient faults. Coupled with a weighted focus loss function and a confidence level secondary verification mechanism, it effectively solves the problems of missed and false detections of weak faults under small industrial samples and high-noise conditions, significantly improving the accuracy and stability of multi-component fault identification and diagnosis, and is suitable for online precision operation and maintenance scenarios of industrial motors.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing motor bearings based on frozen convolution kernels, characterized in that, Includes the following steps: S1. Collect the original vibration signals of each component of the motor bearing and perform standardized preprocessing on the original vibration signals; S2. Perform short-time Fourier transform on the standardized vibration signal to generate a two-dimensional time-frequency feature map and perform pixel normalization processing. S3. Construct a hierarchical convolutional neural network diagnostic model; S4. Pre-train the basic weights of the hierarchical convolutional neural network diagnostic model, and complete the transfer training using a small sample dataset of actual measurements of multiple components of the target motor bearing. S5. Construct the model training objective function and iteratively optimize the trainable network parameters; S6. Collect the real-time vibration signal of the bearing of the motor under test, input it into the trained hierarchical convolutional neural network diagnostic model, perform secondary fault verification in combination with confidence threshold, and output the final diagnostic results and operation and maintenance suggestions.

2. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 1, characterized in that, In S1, the standardized calculation formula is: ; In the formula, Indicates the first Amplitude values ​​at each vibration sampling point; Indicates the standardized first Amplitude values ​​at each vibration sampling point; This represents the mean value of the original entire vibration signal. This represents the standard deviation of the original entire vibration signal. ; This represents the total number of sampling points for a single vibration signal segment.

3. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 2, characterized in that, In S2, a short-time Fourier transform is performed on the standardized vibration signal to generate a two-dimensional time-frequency feature map containing time-domain impact characteristics and frequency-domain modulation characteristics. The time-frequency feature map is then subjected to pixel normalization. The short-time Fourier transform formula is as follows: ; In the formula, Indicates time ,frequency The corresponding time-frequency complex amplitude; This represents the length of the sliding window in the short-time Fourier transform. This indicates the Hanning window function at the vibration sampling point. Weight at each location; The imaginary unit; The formula for normalizing the pixels of the image is: ; In the formula, Represents the coordinates of the normalized time-frequency feature map. The corresponding pixel feature value; Represents the coordinates of the time-frequency matrix The corresponding amplitude; Represents the whole-frame time-frequency amplitude matrix; This represents the maximum amplitude of the time-frequency characteristic spectrum. This represents the minimum amplitude of the time-frequency spectrum.

4. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 3, characterized in that, In S3, a hierarchical convolutional neural network diagnostic model for fault identification of multiple structural components of bearings is constructed. The hierarchical convolutional neural network diagnostic model includes, in sequence, an input layer, a frozen convolutional kernel module, a batch normalization layer, a fine-tunable convolutional kernel module, a random deactivation layer, a global pooling layer, a fully connected layer, and a multi-component fault classification output layer. The frozen convolution kernel module is used to extract periodic impact, vibration edge, and noise substrate features of motor bearings, including... Each convolutional layer has a convolution operation expression as follows: ; In the formula, Indicates the frozen convolution kernel module. The input feature map of the convolutional layer is at the pixel location Eigenvalues ​​at; Indicates the frozen convolution kernel module. Pixel positions in the output feature map of the convolutional layer Eigenvalues ​​at; The relative coordinate vector within the convolution kernel; The Euclidean modulus of the convolution kernel relative to its coordinates; Define the receptive field of the convolution kernel; The characteristic spatial attenuation weighting function of bearing vibration is used to simulate the natural attenuation characteristics of bearing vibration signal with the distance of the sensing field. It is a non-linear activation function; Indicates the frozen convolution kernel module. Layer convolution kernel weights; Indicates the frozen convolution kernel module. Layer convolution bias parameters.

5. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 4, characterized in that, In S3, the fine-tunable convolution kernel module is used to adaptively learn the fault modulation characteristics, frequency characteristics, and impact interval characteristics of the bearing inner ring, outer ring, rolling elements, cage, and seals, so as to realize the differentiated identification of multi-component faults. The fine-tunable convolution kernel module includes... Each convolutional layer has a convolution operation expression as follows: ; In the formula, This indicates the fine-tunable convolution kernel module. Layer convolution input feature map pixel position Eigenvalues ​​at; Indicates that it can be finely adjusted. No. Pixel positions in the output feature map of the convolutional layer Eigenvalues ​​at; The component failure feature enhancement weighting function; This indicates the fine-tunable convolution kernel module. Layer-specific trainable convolutional kernel weights; This indicates the fine-tunable convolution kernel module. Layer-trainable convolutional bias parameters; It is a non-linear activation function.

6. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 5, characterized in that, In S4, the basic weights of the hierarchical convolutional neural network diagnostic model are pre-trained using the public bearing source domain dataset. All weights and bias parameters of the convolution kernel module are fixed and frozen. Only the fine-tunable convolution kernel module and the parameters of the fully connected layer are opened. Transfer training is completed using a small sample dataset of actual measurements of multiple components of the target motor bearing. Among them, the ratio of frozen convolutional kernel layers to the total number of convolutional layers. Satisfying 0.3≤ ≤0.6, proportion The calculation formula is: ; In the formula, To freeze the number of convolutional layers; The number of convolutional layers can be finely adjusted; The transfer learning process uses the Adam optimizer to update the trainable parameters. The parameter update formula is as follows: ; In the formula, for The set of trainable parameters of the model before the next iteration update For the first After the [number]th iteration update The set of trainable parameters for which the iteration step takes effect; Number of training iterations for the model; For the first The first-order momentum corresponding to each iteration; for The second-order momentum corresponding to each iteration; It is the first-order momentum decay coefficient; It is the second-order momentum decay coefficient; The base learning rate is used to train the model; Here is the Adam numerical stability constant; This represents the gradient of the focus loss function with respect to the trainable parameters in the current iteration. Hadamard is the element-wise product; This is the corrected first-order momentum; This is the corrected second-order momentum.

7. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 6, characterized in that, A batch normalization layer is added between the frozen convolutional kernel module and the fine-tunable convolutional kernel module. The batch normalization formula is as follows: ; In the formula, The convolutional layer outputs the original feature values; This represents the mean of the features in the current training batch. The variance of the features in the current training batch; These are the batch normalized learnable scaling and offset parameters, respectively. To prevent the batch normalization from eliminating the zero constant; A random deactivation layer is added after the fine-tunable convolutional module to suppress overfitting on small samples. The formula for calculating random deactivation is: ; In the formula, The original features are output for the upper convolutional layer; Features after random inactivation processing; Let be a random variable that follows a uniform distribution from 0 to 1; The preset neuron inactivation probability; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise.

8. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 7, characterized in that, In S5, a focus loss function containing a focus factor is constructed as the objective function for model training. The trainable network parameters are iteratively optimized. The expression for the focus loss function is: ; In the formula, This represents the total loss value for a single training batch; Indicates the number of samples in a single training batch; This indicates the total number of multiple structural failure categories in the bearing; Indicates the first Category weighting coefficients for component failures; Indicates the fault focus factor; Indicates the first Each sample of real component failure categories has a unique hot label; Indicates the model output of the first The sample belongs to the first Predicted probability of type of failure; Minimizing the expected batch focus loss is taken as the overall iterative optimization objective of the model. The overall training objective expression is: ; In the formula, The overall training target loss for the model; Indicates the training parameter set Perform a minimum solution optimization on top; This is used to calculate the expected mean of a batch of samples.

9. The method for diagnosing motor bearings based on frozen convolution kernels according to claim 8, characterized in that, In step S6, real-time vibration signals of the bearing of the motor under test are acquired. Steps S1-S2 are repeated to generate the time-frequency feature spectrum of the bearing under test. This spectrum is then input into the trained diagnostic model. Based on a multi-component, multi-gradient fault identification system and combined with the vibration fault feature identification mechanism of each bearing structural component, the model outputs the bearing health status, the location of the faulty component, and the corresponding damage degree, with a pre-set confidence threshold. When the model has the highest prediction confidence If the fault characteristics of the bearing component are determined to be masked by noise or not significant, an early warning signal is output and vibration data is re-acquired for secondary diagnosis.

10. A method for diagnosing motor bearings based on frozen convolution kernels according to claim 9, characterized in that, The multi-component, multi-gradient fault identification system identifies fault categories including normal bearing condition, inner ring fault, outer ring fault, rolling element fault, cage fault, and seal fault. A quantitative grading criterion is constructed using a vibration amplitude multiplier coefficient, the expression of which is: ; In the formula, This is the vibration amplitude multiplier for bearing failure; This represents the effective value of the steady-state vibration amplitude of the bearing under test. This is the effective value of the reference vibration amplitude of a normal, fault-free bearing under the same operating conditions. Based on the vibration amplitude multiplier, combined with the three-dimensional judgment index of fault characteristic spectrum significance and noise inundation degree, the faults of each structural component are divided into three levels of damage: mild, moderate, and severe. The judgment rules are as follows: when The fault characteristics of the component are weak in spectrum, the high-frequency harmonics are not significant, the fault characteristics are easily drowned out by industrial noise, the motor bearing only has micro wear, micro pitting, and slight frictional distortion, and there are no structural material peeling or deformation defects, the motor damage level is mild damage. when At that time, the fault spectrum of the component was determined to be clear and stable, with regular periodic impulse modulation characteristics, low noise interference, non-destructive structural defects in the motor bearing, and the motor damage level was moderate. when At that time, the component fault characteristic spectrum was rich in high amplitude and multi-order harmonics, the impact pulse density was high, the motor bearing had serious structural defects, and the motor damage level was severe.