Bearing performance monitoring and fault early warning system

Abstract

The invention, which belongs to the technical field of mechanical state monitoring and fault diagnosis, discloses a bearing performance monitoring and fault early warning system comprising a multi-dimensional sensing module, an edge processing module, a cloud intelligent analysis module and an early warning interaction module which are in communication connection in sequence. The multi-dimensional sensing module is integrated with a vibration sensor, an acoustic sensor, an oil quality sensor and a temperature sensor; the edge processing module performs continuous wavelet transform time-frequency conversion and lightweight CNN preliminary diagnosis on the acquired data; the cloud intelligent analysis module generates an adversarial network enhanced sample, and outputs a diagnosis result through the attention fusion network and the double-flow CNN; and the early warning interaction module triggers third-level early warning according to the diagnosis result and pushes maintenance suggestions. According to the bearing performance monitoring and fault early warning system, multi-dimensional sensing, complex working condition adaptation and small sample diagnosis capabilities are considered, and full-life-cycle performance monitoring and accurate fault early warning of the bearing are realized.

Description

Technical Field

[0001] This invention relates to the field of mechanical condition monitoring and fault diagnosis technology, and in particular to a bearing performance monitoring and fault early warning system. Background Technology

[0002] As a core load-bearing component of rotating machinery, the health of bearings directly determines the safety and reliability of equipment operation. Statistics show that approximately 30% of mechanical failures originate from bearing failure. Current bearing monitoring technology has gone through three development stages: manual inspection, single-parameter sensing + traditional signal analysis, and AI-assisted inspection. However, many technical bottlenecks still exist. At the sensing level, high-end vibration sensors largely rely on imports, while domestically produced sensors lack stability under high-temperature conditions above 150°C. Furthermore, most systems only monitor a single vibration or temperature parameter, making it difficult to comprehensively reflect the bearing's operating status. At the algorithm level, traditional Fourier transforms and wavelet analyses rely on expert experience and have weak generalization capabilities. Existing machine learning methods (such as SVM) have an accuracy rate of only 85-90%, while deep learning models face overfitting problems due to the scarcity of labeled fault data in industrial scenarios. At the system architecture level, edge computing devices generally suffer from a contradiction between computing power and power consumption, and pure cloud processing is prone to data transmission delays, making it difficult to meet real-time early warning requirements.

[0003] While existing multi-sensor fusion systems (such as diagnostic systems based on MS-1DCNN and DS evidence theory) have improved accuracy, they have not solved the model generalization problem in small sample scenarios and lack monitoring of key parameters such as oil quality. Therefore, there is still room for improvement in the timeliness and industrial applicability of early warning systems. Thus, developing a bearing monitoring and early warning system that combines multi-dimensional perception, adaptation to complex operating conditions, and small sample diagnostic capabilities has become an urgent need in the field of predictive maintenance of industrial equipment. Summary of the Invention

[0004] The purpose of this invention is to provide a bearing performance monitoring and fault early warning system to solve the problems of existing technologies, such as single sensing dimension, poor stability under high temperature conditions, low diagnostic accuracy with small samples, and insufficient real-time early warning.

[0005] To achieve the above objectives, this invention provides a bearing performance monitoring and fault early warning system, comprising a multi-dimensional sensing module, an edge processing module, a cloud-based intelligent analysis module, and an early warning interaction module connected in sequence. The multi-dimensional sensing module integrates a vibration sensor, an acoustic sensor, an oil quality sensor, and a temperature sensor, and uses ceramic encapsulation to achieve stable operation in environments ranging from -50℃ to 200℃. The edge processing module performs continuous wavelet transform (CWT) time-frequency conversion and lightweight CNN preliminary diagnosis on the collected data. The cloud-based intelligent analysis module enhances the samples using an improved QP-Div generative adversarial network (CWT-QP-GAN), and outputs diagnostic results via an attention fusion network and a dual-stream CNN. The early warning interaction module triggers a three-level early warning based on the diagnostic results and pushes maintenance suggestions.

[0006] Preferably, the vibration sensor has a measurement range of 0.1-10kHz and an accuracy of ±0.1g; the acoustic sensor has a response frequency of 20Hz-20kHz; and the oil quality sensor detects parameters including viscosity, contamination level, and moisture content, with a sampling frequency of 1Hz and a detection accuracy of viscosity ±0.5cst. The standardized formula for data acquisition from the multi-dimensional sensing module is as follows: ; in, This represents the standardized signal; This represents the raw time-domain signal (such as vibration acceleration or temperature value) collected by the sensor. This represents the mean of the original time-domain signal; It represents the standard deviation of the original time-domain signal.

[0007] Preferably, the edge processing module uses an ARM Cortex-A72 processor, supports 5G / industrial Ethernet communication, and has a built-in data acquisition unit, preprocessing unit and lightweight diagnostic unit. The data acquisition unit has a sampling frequency of 20kHz and can simultaneously acquire standardized analog signals output by multi-dimensional sensing modules.

[0008] Preferably, the preprocessing unit uses continuous wavelet transform (CWT) to convert the one-dimensional vibration / acoustic signal into a 128×128 pixel two-dimensional time-frequency image. The continuous wavelet transform formula is as follows: ; in, These represent wavelet coefficients (pixel values ​​that make up the time-frequency image; the larger the absolute value of the coefficient, the stronger the signal energy at the corresponding time-frequency point and the more obvious the fault characteristics). The scaling factor (which controls the scaling of the wavelet and corresponds to the frequency resolution) represents the frequency resolution. The smaller the value, the higher the frequency. This represents the translation factor (which controls the position of the wavelet and corresponds to the time resolution). Represents the complex conjugate of the mother wavelet; t Indicates the sampling time point of the signal; The lightweight diagnostic unit embeds a pre-trained shallow CNN model with ≤5 million parameters. It performs fault screening on the pre-processed time-frequency images and transmits only suspected fault data with a confidence level >60% to the cloud-based intelligent analysis module.

[0009] Preferably, the cloud-based intelligent analysis module includes a data augmentation submodule, a multi-feature fusion submodule, an intelligent diagnosis submodule, and a model iteration submodule; The data augmentation submodule uses a modified QP-Div generative adversarial network CWT-QP-GAN to augment small sample time-frequency images uploaded by the edge processing module, generating augmented samples with a structural similarity (SSIM) ≥ 93% with the real samples.

[0010] Preferably, the generator of the CWT-QP-GAN adopts a U-Net network structure, which includes 5 downsampling layers and 5 upsampling layers. Each downsampling layer uses a convolution operation with a stride of 2, and each upsampling layer uses a transposed convolution operation. The discriminator uses a 3-layer convolutional network with a kernel size of 3×3 and the activation function is LeakyReLU with a slope of 0.2. The penalty coefficient of the QP-Div loss function is set to 0.01, which can suppress the mode collapse problem of generated samples. The number of data-augmented samples can be expanded to 5-10 times that of the original samples.

[0011] Preferably, the multi-feature fusion submodule constructs an attention mechanism fusion network, dynamically adjusting the weight coefficients of vibration time-frequency features, acoustic features, and oil parameter features according to the bearing fault type, wherein the vibration feature weight is 0.6, the oil feature weight is 0.3, and the acoustic feature weight is 0.1 in fatigue pitting faults; The attention mechanism fusion network includes channel attention and spatial attention; the channel attention generates weight coefficients for each feature channel through global average pooling and fully connected layers; the spatial attention generates a weight matrix for the feature space by concatenating max pooling and average pooling and combining it with convolutional layers; the output feature dimension of the attention mechanism fusion network is 256-dimensional, which supports direct matching with the input layer of a two-stream CNN. The formula for calculating channel attention weights is as follows: ; in, Indicates the channel attention branch weight; Represents the Sigmoid function; FC (·) indicates a fully connected layer; BN (·) indicates batch normalization; GAP (·) indicates global average pooling; F This represents the input multi-source feature map; The formula for calculating spatial attention weights is as follows: ; in, Indicates spatial attention weights; Conv (·) represents a 1×1 convolution; Concat (·) indicates channel splicing; GMP (·) indicates global max pooling; The final fusion feature formula is as follows: ; in, This represents element-wise multiplication, which applies channel weights and spatial weights to the original features to enhance fault feature regions. This represents the fused features, adapted to dual-stream CNN input.

[0012] Preferably, the intelligent diagnostic submodule adopts a dual-stream CNN architecture. The first stream input is the fused features output by the multi-feature fusion submodule, and the second stream input is the temperature time series data uploaded by the edge processing module. The model outputs the bearing health index (value range 0-100) and fault type, including fatigue pitting, plastic deformation, lubrication failure, outer ring wear, and inner ring wear. The first stream of the dual-stream CNN consists of 4 convolutional layers, 2 pooling layers, and 1 fully connected layer, with convolutional kernel sizes of 3×3, 3×3, 5×5, and 5×5, respectively. The pooling layers use max pooling with a stride of 2. The second stream consists of 3 convolutional layers, 2 pooling layers, and 1 fully connected layer, with all convolutional kernel sizes of 3×3. The pooling layers use average pooling with a stride of 2. The output layer of the dual-stream CNN uses the Softmax activation function to output the probability distribution of health index and fault type. The loss function for model training is cross-entropy loss, the number of iterations is 200 rounds, the batch size is set to 32, the initial learning rate is 0.001, and the learning rate is adjusted using a cosine annealing strategy.

[0013] Preferably, the model iteration submodule adopts a federated learning mechanism, receiving model update parameters uploaded by edge processing modules deployed on multiple edge nodes (factories). Each edge node trains a model based on local data, and the update parameter formula is as follows: ; in, Indicates the first i The edge node s+1 round of local model parameters; Indicates the first s Round global model parameters; Indicates the learning rate; Represents the gradient of the loss function; Indicates the first i Local datasets of edge nodes; Achieve cross-condition iterative optimization of the diagnostic model without sharing the original data.

[0014] Preferably, the early warning interaction module includes a multi-level early warning unit and a user interaction unit; the multi-level early warning unit sets three threshold levels based on the health index output by the cloud intelligent analysis module: health level (80-100), early warning level (50-79), and fault level (<50), which respectively trigger a green light display, an audible and visual alarm (decibel ≥85dB), and a device shutdown control signal. The user interaction unit supports access from both web and mobile devices, displays real-time bearing operating parameter curves, fault location maps, and health index trends, and provides functions for historical data backtracking (storage period ≥ 1 year), generating fault cause analysis reports, and pushing maintenance suggestions. The maintenance suggestions include bearing replacement cycle, lubricant model recommendations, and installation and commissioning precautions.

[0015] Therefore, the present invention employs the above-mentioned bearing performance monitoring and fault early warning system, and the technical effects are as follows: The sensing module adopts high-temperature resistant ceramic packaging, which can work stably in environments ranging from -50℃ to 200℃, and is suitable for various industrial scenarios such as wind power, metallurgy, and machine tools, thus solving the problem of insufficient high-temperature stability of domestic sensors. We construct a full-link technical solution of "CWT time-frequency conversion → QP-GAN data augmentation → attention fusion → dual-stream CNN diagnosis", which is different from the existing MS-1DCNN+DS evidence theory fusion method. It solves the gradient vanishing problem of traditional GAN, realizes high-precision diagnosis in small sample scenarios, and breaks through the limitation of insufficient small sample learning in existing technologies.

[0016] 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

[0017] Figure 1 This is a diagram showing the components of a bearing performance monitoring and fault early warning system according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] 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.

[0020] Example 1 like Figure 1 As shown, the present invention provides a bearing performance monitoring and fault early warning system, including a multi-dimensional sensing module, an edge processing module, a cloud intelligent analysis module and an early warning interaction module that are connected in sequence.

[0021] To make the technical solution of the present invention clearer, the following detailed description is provided in conjunction with specific application scenarios: I. System Deployment Taking the monitoring of rolling mill bearings in a metallurgical enterprise as an example, a multi-dimensional sensing module (ceramic encapsulation, adapted to a 180℃ working environment) is deployed on the outer ring of the main bearing of the rolling mill. The edge processing module is deployed in the equipment control cabinet using an industrial-grade gateway and communicates with the cloud platform through a 5G network. The early warning interaction module is connected to the enterprise's CMMS system.

[0022] II. Hardware Configuration Multi-dimensional sensing module: Uses domestic high-temperature resistant vibration sensor (model: HVT-200), acoustic sensor (AST-100), and oil quality sensor (OIL-300), with a packaging temperature range of -50℃ to 200℃; Edge processing module: It adopts an ARM Cortex-A72 processor (1.5GHz), integrates 16GB of storage, and supports a 5G module (model: ME909s-821). Cloud platform: Based on Huawei Cloud Industrial Internet Platform, GPU servers (NVIDIA A10) are deployed for deep learning model training; Early warning interaction module: industrial touch screen (10 inches), audible and visual alarm (≥85dB), mobile APP (supports Android / iOS).

[0023] III. Parameter Configuration Sensing module: vibration sensor with a sampling frequency of 20kHz, oil quality sensor with a sampling frequency of 1Hz, and temperature sensor for real-time monitoring; Edge model: Shallow CNN with 4.5 million parameters and a confidence threshold of 60%; Cloud-based model: CWT-QP-GAN penalty coefficient 0.01, dual-stream CNN iterations for 200 rounds, learning rate 0.001; Warning thresholds: Healthy 80-100, Warning 50-79, Fault <50.

[0024] IV. Software Implementation Edge software: Developed based on the Linux system, using C language to implement data acquisition and shallow CNN inference, with a latency of <100ms; Cloud-based software: The CWT-QP-GAN and dual-stream CNN models are built using the Python TensorFlow framework, with a batch size of 32 and 200 iterations. Interactive software: The web application is developed using Vue.js, while the mobile application uses the Flutter framework, supporting real-time data refresh and historical data query.

[0025] V. Operational Results Within 6 months of system operation, it successfully issued warnings for 4 bearing failures: Two fatigue pitting failures: The warning was triggered 72 hours in advance, and the failure location was verified to be consistent with the diagnosis results after disassembly. First lubrication failure: Based on abnormal oil moisture content and vibration characteristics, a lubricant replacement recommendation is sent to prevent bearing sintering; One outer ring wear failure: triggered a shutdown signal, reducing equipment damage losses by approximately 2 million yuan.

[0026] Statistics show that the mean time between failures (MTBF) of the equipment has increased by 40%, and the annualized maintenance cost has decreased by 45%, verifying the practicality and reliability of the system.

[0027] Therefore, the present invention adopts the above-mentioned bearing performance monitoring and fault early warning system, which takes into account multi-dimensional perception, complex working condition adaptation, small sample diagnostic capability, and realizes bearing full life cycle performance monitoring and accurate fault early warning.

[0028] 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 bearing performance monitoring and fault early warning system, characterized in that: The system includes a multi-dimensional sensing module, an edge processing module, a cloud-based intelligent analysis module, and an early warning interaction module, all connected in sequence. The multi-dimensional sensing module integrates a vibration sensor, an acoustic sensor, an oil quality sensor, and a temperature sensor. The edge processing module performs continuous wavelet transform (CWT) time-frequency conversion and lightweight CNN preliminary diagnosis on the collected data. The cloud-based intelligent analysis module enhances the samples using an improved QP-Div generative adversarial network (CWT-QP-GAN), and outputs the diagnostic results via an attention fusion network and a dual-stream CNN. The early warning interaction module triggers a level-three early warning based on the diagnostic results and pushes maintenance suggestions.

2. The bearing performance monitoring and fault early warning system according to claim 1, characterized in that: The vibration sensor has a measurement range of 0.1-10kHz and an accuracy of ±0.1g; the acoustic sensor has a response frequency of 20Hz-20kHz; the oil quality sensor detects parameters including viscosity, contamination level, and moisture content, with a sampling frequency of 1Hz and a detection accuracy of viscosity ±0.5cst. The standardized formula for data acquisition from the multi-dimensional sensing module is as follows: ； in, This represents the standardized signal; This represents the raw time-domain signal acquired by the sensor; This represents the mean of the original time-domain signal; It represents the standard deviation of the original time-domain signal.

3. The bearing performance monitoring and fault early warning system according to claim 2, characterized in that: The edge processing module uses an ARM Cortex-A72 processor and supports 5G / industrial Ethernet communication. The edge processing module has a built-in data acquisition unit, a preprocessing unit, and a lightweight diagnostic unit. The data acquisition unit has a sampling frequency of 20kHz and synchronously acquires standardized analog signals output by the multi-dimensional sensing module.

4. The bearing performance monitoring and fault early warning system according to claim 3, characterized in that: The preprocessing unit uses continuous wavelet transform (CWT) to convert the one-dimensional vibration / acoustic signal into a 128×128 pixel two-dimensional time-frequency image. The CWT conversion formula is as follows: ； in, Represents wavelet coefficients; Indicates the scale factor; Indicates the translation factor; Represents the complex conjugate of the mother wavelet; t Indicates the sampling time point of the signal; The lightweight diagnostic unit embeds a pre-trained shallow CNN model to perform fault screening on the pre-processed time-frequency images, and only transmits suspected fault data with a confidence level >60% to the cloud-based intelligent analysis module.

5. The bearing performance monitoring and fault early warning system according to claim 4, characterized in that: The cloud-based intelligent analysis module includes a data augmentation submodule, a multi-feature fusion submodule, an intelligent diagnosis submodule, and a model iteration submodule. The data augmentation submodule uses a modified QP-Div generative adversarial network CWT-QP-GAN to augment small sample time-frequency images uploaded by the edge processing module, generating augmented samples with a structural similarity of ≥93% with real samples.

6. The bearing performance monitoring and fault early warning system according to claim 5, characterized in that: The generator of the CWT-QP-GAN adopts the U-Net network structure, which includes 5 downsampling layers and 5 upsampling layers. Each downsampling layer uses a convolution operation with a stride of 2, and each upsampling layer uses a transposed convolution operation. The discriminator of the CWT-QP-GAN uses a 3-layer convolutional network with a kernel size of 3×3 and the activation function is LeakyReLU with a slope of 0.

2.

7. The bearing performance monitoring and fault early warning system according to claim 6, characterized in that: The multi-feature fusion submodule constructs an attention mechanism fusion network, which dynamically adjusts the weight coefficients of vibration time-frequency features, acoustic features, and oil parameter features according to the bearing fault type. In the case of fatigue pitting fault, the weight of vibration features is 0.6, the weight of oil features is 0.3, and the weight of acoustic features is 0.

1. The attention mechanism fusion network includes channel attention and spatial attention; the channel attention generates weight coefficients for each feature channel through global average pooling and fully connected layers; the spatial attention generates a weight matrix for the feature space by concatenating max pooling and average pooling and combining it with convolutional layers; the output feature dimension of the attention mechanism fusion network is 256-dimensional, which supports direct matching with the input layer of a two-stream CNN. The formula for calculating channel attention weights is as follows: ； in, Indicates the channel attention branch weight; Represents the Sigmoid function; FC (·) indicates a fully connected layer; BN (·) indicates batch normalization; GAP (·) indicates global average pooling; F This represents the input multi-source feature map; The formula for calculating spatial attention weights is as follows: ； in, Indicates spatial attention weights; Conv (·) represents a 1×1 convolution; Concat (·) indicates channel splicing; GMP (·) indicates global max pooling; The final fusion feature formula is as follows: ； in, Indicates element-wise multiplication; This represents the fused features, adapted to dual-stream CNN input.

8. The bearing performance monitoring and fault early warning system according to claim 7, characterized in that: The intelligent diagnostic submodule adopts a dual-stream CNN architecture. The first stream input is the fused features output by the multi-feature fusion submodule, and the second stream input is the temperature time series data uploaded by the edge processing module. The model outputs the bearing health index and fault type, which includes fatigue pitting, plastic deformation, lubrication failure, outer ring wear, and inner ring wear. The first stream of the dual-stream CNN consists of 4 convolutional layers, 2 pooling layers, and 1 fully connected layer, with convolutional kernel sizes of 3×3, 3×3, 5×5, and 5×5, respectively. The pooling layers use max pooling with a stride of 2. The second stream consists of 3 convolutional layers, 2 pooling layers, and 1 fully connected layer, with all convolutional kernel sizes of 3×3. The pooling layers use average pooling with a stride of 2. The output layer of the dual-stream CNN uses the Softmax activation function to output the probability distribution of health index and fault type, and the loss function for model training is cross-entropy loss.

9. The bearing performance monitoring and fault early warning system according to claim 8, characterized in that: The model iteration submodule adopts a federated learning mechanism, receiving model update parameters uploaded by edge processing modules deployed on multiple edge nodes. Each edge node trains its model based on local data, and the update parameter formula is as follows: ； in, Indicates the first i The edge node s +1 round of local model parameters; Indicates the first s Round global model parameters; Indicates the learning rate; Represents the gradient of the loss function; Indicates the first i Local datasets for edge nodes.

10. A bearing performance monitoring and fault early warning system according to claim 9, characterized in that: The early warning interaction module includes a multi-level early warning unit and a user interaction unit; the multi-level early warning unit sets three thresholds based on the health index output by the cloud intelligent analysis module: health level, early warning level, and fault level, which respectively trigger a green light display, an audible and visual alarm, and a device shutdown control signal. The user interaction unit supports access from both web and mobile devices, displays real-time bearing operating parameter curves, fault location maps, and health index trends, and provides functions for historical data backtracking, fault cause analysis report generation, and maintenance suggestion push. The maintenance suggestions include bearing replacement cycle, lubricant model recommendations, and installation and commissioning precautions.

Images