Bearing defect intelligent detection method and system based on deep learning

Abstract

The invention discloses a bearing defect intelligent detection method and system based on deep learning, and relates to the field of bearing defect detection. Multi-modal data such as bearing vibration, acoustics, thermal imaging and the like are acquired by using various sensors, and a four-dimensional feature tensor is constructed through preprocessing such as noise reduction and feature extraction; features are fused through a reconfigurable multi-branch convolutional neural network, and defects are identified and a development trend is predicted in combination with a meta-learning twin network; a decision threshold is optimized by adopting a quantum heuristic algorithm, and multi-level early warning is realized; and continuous evolution of the model is completed through edge-cloud collaboration and federated learning, the functions of data calibration compensation, model dynamic optimization and the like are achieved, and efficient and accurate detection of bearing defects is achieved. The detection time is remarkably shortened, and the positioning precision is high; the novel defect response speed is high, and faults can be predicted in advance; system energy consumption is reduced, model updating is improved, stable operation of equipment is effectively guaranteed, and cost reduction and efficiency improvement of industrial intelligent operation and maintenance are facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing defect detection, and in particular to a bearing defect intelligent detection method and system based on deep learning. Background Art

[0002] In modern industrial production systems, bearings, as core components of mechanical equipment, have a direct impact on equipment safety and production efficiency. Traditional bearing defect detection relies heavily on manual visual inspection, auscultation, or simple instrument measurement. This method is not only inefficient and unable to meet the demands of automated production lines inspecting dozens of pieces per minute, but is also significantly affected by subjective factors such as the inspector's experience and fatigue. As the manufacturing industry transitions toward intelligent and unmanned operations, traditional manual inspection has become a bottleneck restricting production efficiency.

[0003] In recent years, automated detection methods based on sensor technology and computer vision have gradually emerged, but there are still significant technical limitations. Single-modality detection technology (such as relying solely on vibration sensors or infrared thermal imagers) is severely affected by environmental factors. Under complex working conditions such as high temperature, strong electromagnetic fields, and dust, the signal is easily distorted, resulting in a decrease in detection accuracy. For example, the misjudgment rate of vibration detection can be high when the equipment resonates or is disturbed by external vibrations; infrared thermal imaging detection has difficulty in accurately identifying tiny defects when the ambient temperature fluctuates greatly. Although multi-sensor fusion technology has been applied, the existing solutions lack effective spatiotemporal calibration and data fusion strategies. The data from each sensor cannot be accurately synchronized and deeply correlated, and the reliability and stability of the detection results are insufficient.

[0004] In terms of deep learning applications, existing bearing defect detection models face the dual challenges of "data dependence" and "poor generalization." Training high-quality models requires a large number of labeled samples, but it is difficult to collect samples of new defects or rare working conditions, resulting in weak adaptability of the model to new scenarios. At the same time, traditional model updates rely on manual data collection and retraining, which takes weeks or even months and cannot respond to equipment changes in the production process in a timely manner. In addition, existing systems generally lack autonomous optimization capabilities and find it difficult to achieve a dynamic balance between energy consumption, detection speed and accuracy, which limits their large-scale application in industrial scenarios. Therefore, there is an urgent need for a bearing defect detection technology with multimodal deep fusion, adaptive learning and intelligent decision-making capabilities to meet the urgent needs of modern industry for equipment reliability and intelligent operation and maintenance. Summary of the Invention

[0005] The present invention proposes a deep learning-based intelligent bearing defect detection method and system to solve the problems mentioned in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a bearing defect intelligent detection method and system based on deep learning, comprising:

[0007] Acquisition preprocessing steps: deploy triaxial vibration sensors, acoustic sensors, infrared thermal imagers, and hyperspectral cameras, and synchronize time using the IEEE 1588v2 protocol. Wavelet packet transform is used to decompose vibration signals, combined with variational mode decomposition to suppress modal aliasing. Spectral angle mapping is used to remove background noise and construct a four-dimensional feature tensor.

[0008] Feature fusion steps: Design a reconfigurable multi-branch convolutional neural network equipped with a dynamic routing module to adjust the convolution kernel size and expansion rate according to the characteristics of the input data; introduce a cross-modal attention gating mechanism to assign feature weights by calculating mutual information entropy; adopt a progressive feature pyramid structure to fuse multi-scale features;

[0009] Diagnosis and prediction steps: Build a twin network, design a defect evolution prediction module, use a spatiotemporal graph convolutional network to model historical inspection data, and predict defect development trends; combine reinforcement learning algorithms to optimize inspection strategies, and balance inspection accuracy and efficiency through interactive learning with the virtual bearing system;

[0010] Decision-making and early warning steps: Use quantum annealing algorithms to optimize Gaussian mixture model parameters and construct a high-dimensional anomaly detection space; design a multi-threshold decision mechanism to assess defect severity; develop a cross-modal correlation early warning system that triggers an emergency shutdown command when the spatiotemporal matching between vibration anomalies and thermal imaging hotspots exceeds a threshold;

[0011] Collaborative optimization steps: Build a heterogeneous collaborative framework, use model distillation technology to compress parameters on edge nodes, and protect data privacy through secure multi-party computing; introduce evolutionary strategies to adjust the network architecture, and combine digital twin technology to pre-verify the model optimization effect.

[0012] Furthermore, it also includes:

[0013] Calibration and compensation steps: Deploy a laser Doppler vibrometer and establish a spatiotemporal conversion model through cross-correlation analysis; when sensor data is missing, use the implicit neural representation technology of the generative adversarial network to reconstruct the missing signal; develop a sensor health prediction module and use the long short-term memory network to analyze the sensor's historical data.

[0014] Furthermore, it also includes:

[0015] Model evolution steps: Build a real-time monitoring system for model performance and evaluate model uncertainty by calculating the Shannon entropy of test results; trigger an active learning mechanism when the uncertainty exceeds a threshold, and use a Bayesian active learning strategy to screen samples; design a dynamic adjustment algorithm for the model structure and optimize the number of network layers and connection methods based on a genetic algorithm.

[0016] Furthermore, the diagnostic prediction step uses a quantum convolutional neural network to replace the traditional convolutional layer, and uses quantum bits to extract defect features; constructs a defect semantic map, associates defect images, maintenance records and material properties, and jointly infers multi-source information through a graph attention network; develops a sub-pixel positioning algorithm to locate defects based on super-resolution reconstruction technology.

[0017] Furthermore, the decision-making and early warning step uses digital twin technology to construct a digital mirror of the bearing health, generates virtual detection data by simulating different working conditions, and expands the abnormality detection boundary; uses the quantum Monte Carlo method to calculate the probability of defect occurrence, and quantifies the uncertainty as the quantum state probability amplitude; designs a graded early warning response mechanism, and automatically generates a maintenance work order and dispatches drones to deliver spare parts when the remaining life is predicted to be lower than the threshold.

[0018] Furthermore, the collaborative optimization step develops a model interpretability module, using layer-by-layer correlation propagation technology to generate defect feature heat maps; builds a global inspection data blockchain network to share cross-enterprise model parameters through smart contracts; and uses quantum reinforcement learning to optimize the model update strategy.

[0019] Furthermore, we will develop an intelligent defect repair module, generate maintenance robot operation paths based on reinforcement learning, and automatically replace bearing seals and add lubricating oil; build a predictive maintenance knowledge graph, use natural language processing technology to extract expert experience, and trace the cause of the fault; and design an energy consumption optimization system by dynamically adjusting the sensor sampling frequency and algorithm calculation accuracy.

[0020] Furthermore, the following modules are also included:

[0021] Multimodal sensing module: Integrates a MEMS three-axis vibration sensor, a fiber optic acoustic sensor, an uncooled infrared focal plane array, and a hyperspectral imaging module, with a built-in field programmable gate array for real-time noise reduction and feature pre-extraction;

[0022] Reconstruction processing module: Adopting a heterogeneous computing architecture, deploying a dynamic network reconstruction engine, and automatically switching computing resource allocation strategies based on detection tasks; designing a hardware-level security unit to protect the security of detection data and model parameters through a trusted execution environment;

[0023] Decision-making and control module: Build a decision-making engine, integrate industry standards and expert experience; develop a 3D visual diagnosis system, use AR to mark defect locations, and simulate repair plans; integrate industrial Internet of Things interfaces and link PLC and SCADA systems;

[0024] Evolutionary Operations and Maintenance Module: Build a federated learning server cluster to collaboratively train cross-regional models; deploy an automated testing platform to generate virtual test cases through digital twin technology; configure intelligent inspection robots to regularly calibrate system hardware and troubleshoot faults.

[0025] Furthermore, it also includes:

[0026] Defect microscopic analysis module: equipped with a scanning electron microscope and an energy dispersive X-ray spectrometer, which automatically triggers microscopic inspection when suspected material defects are detected;

[0027] Carbon footprint management module: Optimizes algorithm execution strategies by analyzing and detecting process energy consumption data;

[0028] Supply chain collaboration module: When batch defects are detected, quality warnings are automatically sent to suppliers and the emergency dispatch process for spare parts is initiated.

[0029] Furthermore, the multimodal perception module adopts a self-powered design, integrating a vibration energy harvester and a thermoelectric power generation unit to convert energy; the reconstruction processing module supports dynamic loading and unloading of models, and enters a low-power sleep mode during idle periods; the evolutionary operation and maintenance module uses digital thread technology to record information throughout the entire life cycle and trace and reproduce problems.

[0030] Compared with the existing technology, the beneficial effects of the present invention are:

[0031] In terms of inspection efficiency and accuracy, the system optimizes the single-bearing inspection process, enhancing inspection reliability under complex operating conditions, accurately locating defects, and effectively avoiding equipment failures caused by missed or false detections. For example, in high-temperature metallurgical environments, traditional methods are susceptible to heat wave interference and misjudgment. However, this system, relying on multimodal data fusion and a dynamic adaptive network, can accurately identify tiny cracks and wear.

[0032] In terms of intelligence and adaptability, the system incorporates meta-learning and federated learning technologies to accelerate responses to new defects and achieve rapid learning with a small number of samples. Edge-cloud co-evolution mechanisms facilitate automatic model updates, significantly reducing manual intervention. Predictive maintenance features provide early warning of equipment failures and, in conjunction with intelligent decision-making systems, generate optimal maintenance strategies to minimize unplanned downtime.

[0033] Regarding energy consumption and cost control, reinforcement learning optimizes detection strategies and combines them with energy-efficient hardware design to optimize overall system energy consumption. Model distillation technology compresses parameters to adapt to edge device resource constraints, reducing hardware investment. Furthermore, the reliability of sensor fault warnings has been improved, and combined with data reconstruction and self-calibration capabilities, this ensures long-term stable system operation, laying a solid foundation for core technologies that help industrial enterprises reduce costs, increase efficiency, and achieve intelligent upgrades. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a schematic block diagram of a bearing defect intelligent detection method based on deep learning proposed by the present invention;

[0035] Figure 2This is a schematic block diagram of a bearing defect intelligent detection system based on deep learning proposed by the present invention;

[0036] Figure 3 This is a schematic diagram comparing the detection accuracy of multiple scenarios;

[0037] Figure 4 This is a schematic diagram comparing detection time and energy consumption;

[0038] Figure 5 This is a schematic diagram of defect location accuracy comparison;

[0039] Figure 6 Schematic diagram for comparing model optimization iteration efficiency;

[0040] Figure 7 A schematic diagram showing the time-consuming comparison of multimodal data processing;

[0041] Figure 8 This is a schematic diagram comparing the accuracy of sensor fault warning. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0044] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0045] Reference Figures 1 to 8 : A bearing defect intelligent detection method and system based on deep learning, comprising:

[0046] Acquisition preprocessing steps: When a bearing enters the monitoring area, the synchronization trigger module within the sensing terminal generates a timestamp with 500ns accuracy based on the IEEE 1588v2 precision clock protocol, ensuring sub-microsecond synchronized acquisition of multi-sensor data. To reduce signal interference, the system employs multi-stage filtering. The vibration signal is first amplified and anti-aliased (with a cutoff frequency set to 10kHz) by a PCBPiezotronics 4-channel signal conditioner before being input into the edge computing unit. Within the unit, a 5-layer wavelet packet decomposition using the sym8 wavelet basis is performed, decomposing the signal into 32 frequency bands. The energy of each frequency band is extracted as preliminary features. Further processing is then performed using the variational mode decomposition (VMD) algorithm, which adaptively determines the number of decomposed modes, effectively suppressing modal aliasing and separating the different vibration components. After being converted to digital signals by the fiber optic sensor, the acoustic signal is first bandpass filtered (with a passband range of 20Hz-10kHz), and then a short-time Fourier transform (STFT) is used to generate a time-frequency plot. To extract key features, the Mel Frequency Cepstral Coefficient (MFCC) algorithm is used to extract 13-dimensional feature vectors from the time-frequency diagram, and then expanded to 39-dimensional features through first-order and second-order difference operations. After the infrared thermal imaging data is collected, non-uniformity correction (NUC) is first performed using a neural network-based non-uniformity correction algorithm. Assume that the original response of the i-th pixel of the detector is R i , the corrected response is R′ i , constructing a non-uniformity correction model Where N is the total number of pixels, is the ideal response of the j-th pixel learned by the neural network, W i，jThe inter-pixel correction weights (characterizing the compensation relationship between the response differences between pixels i and j) are output by the neural network and learned through training on a large amount of thermal imaging data, achieving high-precision correction. Background suppression is performed using a deep learning model based on the U-Net architecture. This model uses a large number of bearing thermal images containing different backgrounds as training data. It can automatically identify and remove interference from environmental heat sources, highlighting the temperature characteristics of the bearing surface. After hyperspectral image acquisition, radiometric calibration is first performed to convert the image grayscale values ​​into actual spectral radiance. Then, using the spectral angle mapping (SAM) algorithm, the spectral vector of each pixel is compared with a pre-established library of bearing material spectra, the spectral angle is calculated, the bearing surface material area is automatically identified, and the spectral characteristics of the region of interest (ROI) are extracted. Finally, the system aligns the vibration, acoustic, thermal imaging, and hyperspectral data along the time dimension and fuses them into a four-dimensional feature tensor (time × frequency × space × spectrum), providing rich multidimensional information for subsequent analysis.

[0047] Feature fusion step: A reconfigurable multi-branch convolutional neural network (R-CNN) is deployed on the GPU of the edge computing unit. The overall network structure adopts a layered design, consisting of an input layer, a feature extraction layer, a fusion layer, and an output layer. The feature extraction layer consists of three parallel branches, each targeting different modal data. The vibration signal branch adopts a 1D-CNN architecture consisting of eight dynamically routed convolutional layers. Each dynamically routed convolutional layer has a built-in parameter prediction module. Based on the frequency distribution and time domain characteristics of the input signal, this module uses a small neural network to predict the optimal kernel size (dynamically adjustable between 3 and 11) and dilation rate for the current convolutional layer, thereby adaptively extracting vibration features at different scales. The acoustic signal branch uses a 2D-CNN combined with an attention mechanism. First, a series of convolutional and pooling layers are used to extract basic features of the time-frequency graph. Then, a channel attention module (CAM) and a spatial attention module (SAM) are introduced. The channel attention module calculates the importance weight of each channel to highlight frequency components related to defects. The spatial attention module focuses on the spatial location of defects in the time-frequency graph. The combination of the two effectively enhances the ability to extract acoustic defect features. The thermal imaging and hyperspectral branches employ a 3D-CNN architecture, utilizing three-dimensional convolutional kernels to simultaneously extract spatiotemporal and spectral features. A residual connection structure is incorporated into the network to address the vanishing gradient problem in deep network training, ensuring that the complex features of the bearing surface temperature distribution and spectral characteristics can be learned. The cross-modal attention gating mechanism (CAGM) plays a key role in the fusion layer. This mechanism evaluates the contribution of each modality to the final detection result by calculating the mutual information entropy between features from different modalities. Specifically, the features extracted by each branch are first normalized, and then the mutual information entropy between each modal feature is calculated. When localized bearing overheating is detected, the system calculates the mutual information entropy between the thermal imaging modality and other modalities. If the thermal imaging modality is found to have the highest correlation with the final detection result, the weight of the thermal imaging modality is increased to 0.6 (the default weight is 0.25), while the weights of the other modalities are correspondingly reduced. Dynamically assigning feature weights enables effective fusion of multimodal features. The progressive feature pyramid (PFP) architecture utilizes a bidirectional bottom-up and top-down approach. In the bottom-up process, shallow networks preserve pixel-level details through small-sized convolution kernels and high-resolution feature maps. As the number of network layers increases, larger-sized convolution kernels and downsampling operations are used in deeper networks to extract component-level semantic features. The top-down path combines deep semantic information with shallow detail information through upsampling and feature fusion operations, ultimately outputting a feature vector containing multi-scale information, providing a comprehensive feature representation for subsequent defect diagnosis.

[0048] Diagnosis and prediction steps: A cloud server cluster runs a meta-learning-driven Siamese network (Meta-SiameseNet), trained using the Model-Agnostic Meta-Learning (MAML) algorithm. In a few-shot learning scenario, when a new defect emerges (e.g., with only five labeled samples), the system first divides these few samples into a support set and a query set. The support set is used to quickly adjust the network's initial parameters, enabling it to quickly adapt to the new task; the query set is used to evaluate the performance of the adjusted network. Through multiple iterations of training, the network completes model adaptation within 10 minutes, achieving a detection accuracy of 82% (compared to 45% for traditional methods, which typically require retraining large amounts of data and take days or even weeks). The defect evolution prediction module utilizes a spatiotemporal graph convolutional network (ST-GCN), modeling historical bearing inspection data as a spatiotemporal graph structure. The nodes in the graph represent the inspection status at different points in time, each containing a multimodal feature vector at that moment. Edges represent the temporal and spatial relationships between states. By training the ST-GCN network, the feature propagation patterns between nodes are learned, enabling prediction of defect development trends within the next 24 hours. For example, when a slight crack is detected in a bearing, the system can predict the expansion speed and direction of the crack in the future, providing a basis for planning maintenance strategies in advance and reducing unplanned downtime. The reinforcement learning module uses OpenAI Gym as a framework to construct a virtual bearing system as the environment and the detection strategy as the agent action. The state space of the agent contains information such as the real-time operating parameters of the bearing (such as speed, load, temperature), multimodal detection features, and historical detection results, which is recorded as the state vector S; the action space includes operations such as adjusting the sensor sampling frequency, selecting different detection algorithms, and changing the detection cycle, which is recorded as the action set A. The reward function R is designed to comprehensively consider factors such as detection accuracy, energy consumption, and detection time. The formula is R = α·A acc -β·E-γ·T, where A acc is the detection accuracy (ranging from 0 to 1, with values ​​closer to 1 indicating higher accuracy), reflecting positive incentives for accurate detection; E is the energy consumption per detection, reflecting energy costs; T is the detection time, representing time costs; α, β, and γ are weight coefficients (α>0, β, γ≥0, which can be adjusted according to actual needs to ensure that the reward function balances all objectives), used to adjust the impact of different factors on the reward. Through extensive interactive learning with the virtual environment, the intelligent agent selects action A based on state S and dynamically adjusts the detection strategy based on reward R, reducing energy consumption per detection by 30% while maintaining 99% detection accuracy. For example, when a bearing is under low load, the system automatically reduces the sensor sampling frequency and switches to a lightweight detection model to reduce computing resource consumption.

[0049] Decision-making and early warning steps: The D-Wave Advantage 2X quantum computing accelerator card is used to optimize the parameters of the Gaussian mixture model (Q-GMM). Traditional CPU optimization methods often take hours or even days to converge when processing high-dimensional data. Quantum computing, however, leverages the superposition and entanglement properties of qubits to explore multiple solution spaces in parallel, accelerating the convergence of Q-GMM parameters by 20 times compared to traditional methods. A multi-threshold decision-making mechanism, designed based on the principle of quantum state superposition, categorizes defect severity into five probability levels: safe (0-0.2), concern (0.2-0.4), warning (0.4-0.6), emergency (0.6-0.8), and critical (0.8-1). The system determines the probability level of the defect by calculating the probability distribution of the current detection feature vector in the Q-GMM model. When the spatiotemporal match between the abnormal vibration signal characteristics and the thermal imaging hotspot exceeds 0.8, the system immediately triggers an emergency shutdown. Simultaneously, an augmented reality visualization report containing the defect location, type, and severity is sent to maintenance personnel via the 5G network. Operation and maintenance personnel can use a smart terminal equipped with a dedicated APP to scan the bearing using the mobile phone camera to intuitively see the three-dimensional annotation information of the defect, as well as maintenance suggestions and operating steps on the screen, thereby improving maintenance efficiency and accuracy.

[0050] Collaborative optimization steps: The edge node uses model distillation technology to compress the complex deep learning model in the cloud by 10 times before deployment. The specific process is: train a large teacher model in the cloud, and then train a small student model on the edge node. By allowing the student model to learn the output probability distribution of the teacher model, the student model can significantly reduce the model parameters and calculation amount while maintaining a high accuracy rate. The edge node participates in federated learning through the secure multi-party computing (MPC) protocol to ensure data privacy and security. After every 1,000 detections, the edge node uploads the encrypted model parameter update gradient to the cloud. The cloud server uses the TensorFlowFederated framework to aggregate multi-node data and update the global model. The digital twin module is built based on the Unity3D engine. By collecting the operating data and test results of the bearing in real time, a digital mirror image consistent with the actual bearing state is created in a virtual environment. Let the actual bearing operating state vector be S real (including speed, load and other parameters), the digital twin mirror state vector is S virt , build state consistency verification formula Here, ||·||2 is the Euclidean norm, which measures the differences between vectors; Consistency is an indicator of state consistency (ranging from 0 to 1, with the closer to 1, the more consistent the digital image is with the actual state). After optimizing the model, this module was used to simulate bearing operation under different operating conditions in a digital twin environment, generating millions of virtual test cases to pre-verify the optimized model. By comparing virtual test results with actual operating data and evaluating the model optimization effect according to the above formula, the model deployment success rate was increased from 75% to 98%, effectively reducing the risk of offline deployment.

[0051] The present invention further comprises the following steps:

[0052] Calibration and compensation steps: A Polytec CLV-2534 laser Doppler vibrometer is deployed as a reference sensor. This device uses a non-contact measurement method with a velocity resolution of 0.1 mm / s and an acceleration resolution of 0.01 m / s. 2 , the measurement frequency range is 0-50kHz. Multi-sensor spatiotemporal calibration is automatically performed every hour. The calibration process is as follows: First, the laser Doppler vibrometer measures the vibration data of a specific point on the bearing surface as a benchmark; then, the system compares the measurement data of other vibration sensors with the benchmark data, calculates the time delay and spatial offset through cross-correlation analysis, and establishes a spatiotemporal conversion model between multiple sensors. When the data of a certain sensor is missing, the system starts the generative adversarial network (GAN) based on implicit neural representation (INR) to reconstruct the data. The GAN network consists of a generator and a discriminator. The generator takes other modal data and time and space coordinates as input, and generates an estimate of the missing signal through a multi-layer perceptron (MLP); the discriminator is used to judge the authenticity of the generated signal. According to actual measurements, the root mean square error (RMSE) of the vibration signal reconstruction is 0.05g, and the RMSE of the acoustic signal reconstruction is 3dB, which meets the detection requirements. The sensor health prediction module uses an LSTM network to analyze historical sensor data, including signal amplitude, signal-to-noise ratio, temperature, and other characteristics, to predict sensor failures 72 hours in advance with 92% accuracy. When a sensor failure is predicted, the system automatically switches to a backup sensor and issues a maintenance reminder.

[0053] The present invention further comprises the following steps:

[0054] Model Evolution Steps: The model performance monitoring system calculates the Shannon entropy of the test results in real time. Shannon entropy reflects the uncertainty of the test results. When the entropy exceeds 0.8 (indicating high uncertainty), the active learning mechanism is triggered. Based on the Bayesian Active Learning (BALD) strategy, the system selects the most valuable samples from the unlabeled data pool. The specific selection process is as follows: First, the current model performs multiple forward propagations on the unlabeled data to obtain multiple predictions. Then, the entropy of each prediction is calculated, and the sample with the highest entropy is selected as the most valuable sample. Annotation is completed using a human-machine collaborative annotation platform (integrated image and semantic annotation capabilities). This platform supports simultaneous online annotation by multiple people and features automatic annotation suggestions and annotation conflict detection, improving annotation efficiency and accuracy. The dynamic model structure adjustment algorithm uses a genetic algorithm, randomly generating 10 network structure variants in each iteration. Each variant is evaluated on a virtual test set using performance metrics such as accuracy, recall, F1 score, and computational complexity. Through genetic operations such as selection, crossover, and mutation, high-performing variants are retained and poorer ones are eliminated. After multiple rounds of iteration, the optimal structure is ultimately selected. This approach increases the model's response speed to new defects by three times, enabling it to quickly adapt to new problems that arise during the production process.

[0055] In the present invention, the diagnosis and prediction step uses a quantum convolutional neural network (QCNN) to replace the traditional convolutional layer, and uses the superposition characteristics of quantum bits to extract defect features in parallel, which improves the reasoning speed by 40%; constructs a defect semantic map, associates defect images with maintenance records, material properties and other knowledge, and realizes joint reasoning of multi-source information through the graph attention network (GAT); develops a sub-pixel positioning algorithm, and uses deep learning-based super-resolution reconstruction technology to improve the defect positioning accuracy to ±0.05mm.

[0056] In the present invention, the decision-making and early warning steps are combined with digital twin technology to construct a digital mirror of the bearing health, generate virtual detection data by simulating different working conditions, and expand the abnormality detection boundary; the quantum Monte Carlo method is used to calculate the probability of defect occurrence, and the uncertainty is quantified as the quantum state probability amplitude; a hierarchical early warning response mechanism is designed. When the predicted remaining life is less than 24 hours, a maintenance work order is automatically generated and a drone is dispatched to deliver spare parts.

[0057] In the present invention, a collaborative optimization step is used to develop a model interpretability module, and layer-by-layer correlation propagation (LRP) technology is used to generate defect feature heat maps to assist engineers in understanding decision-making logic. A global detection data blockchain network is constructed to achieve secure sharing and incentives of cross-enterprise model parameters through smart contracts. Quantum reinforcement learning is used to optimize the model update strategy, increasing the slope of the model performance improvement curve by 60% under resource-constrained conditions.

[0058] In this invention, an intelligent defect repair module is developed, and the operation path of the maintenance robot is generated based on reinforcement learning, which supports operations such as automatic replacement of bearing seals and filling of lubricating oil; a predictive maintenance knowledge graph is constructed, and natural language processing technology is used to extract expert experience from the maintenance manual to achieve rapid tracing of the cause of the fault; an energy consumption optimization system is designed, which reduces the overall energy consumption of the system by 35% by dynamically adjusting the sensor sampling frequency and algorithm calculation accuracy.

[0059] The present invention also includes the following modules:

[0060] Multimodal perception module: Integrates a MEMS three-axis vibration sensor (sensitivity 0.001g), a fiber optic acoustic sensor (dynamic range 120dB), an uncooled infrared focal plane array (NETD ≤ 25mK), and a hyperspectral imaging module (spectral resolution 5nm); a built-in field programmable gate array (FPGA) enables real-time noise reduction and feature pre-extraction of data, with a data compression ratio of 20:1; supports 5G-MEC (mobile edge computing) communications, with end-to-end latency less than 10ms.

[0061] Reconstruction processing module: Adopts a heterogeneous computing architecture, including GPU (NVIDIA A100), TPU (Tensor Processing Unit) and quantum computing accelerator card; deploys a dynamic network reconstruction engine to automatically switch computing resource allocation strategies according to detection tasks; designs a hardware-level security module to protect the security of detection data and model parameters through a trusted execution environment (TEE).

[0062] Decision-making and control module: Build a knowledge graph-based decision engine, integrating more than 100,000 industry standards and expert experience; develop a 3D visual diagnostic system that supports AR annotation of defect locations and repair plan simulation; integrate the Industrial Internet of Things (IIoT) interface to achieve seamless linkage with PLC and SCADA systems.

[0063] Operation and maintenance evolution module: Build a federated learning server cluster to support cross-regional model collaborative training; deploy a model automation testing platform to generate millions of virtual test cases through digital twin technology; configure intelligent inspection robots to regularly self-calibrate and troubleshoot system hardware.

[0064] The present invention also includes the following modules:

[0065] Defect Microscopic Analysis Module: Equipped with a scanning electron microscope (SEM) and an energy dispersive X-ray spectrometer (EDX), it automatically triggers microscopic inspection when suspected material defects are detected.

[0066] Carbon footprint management module: By analyzing the energy consumption data of the detection process and optimizing the algorithm execution strategy, the carbon emissions of a single detection can be reduced by 20%.

[0067] The supply chain collaboration module automatically sends quality warnings to suppliers and initiates the emergency dispatch process for spare parts when batch defects are detected.

[0068] In the present invention, the multimodal perception module adopts a self-powered design, integrating a vibration energy harvester and a temperature difference power generation module, with an energy conversion efficiency of 15%; the reconstruction processing module supports dynamic loading and unloading of the model, enters a low-power sleep mode during idle periods, and reduces standby power consumption by 85%; the operation and maintenance evolution module adopts digital thread technology to fully record the entire life cycle information from data collection to model optimization, supporting rapid tracing and reproduction of problems.

[0069] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A bearing defect intelligent detection method based on deep learning, characterized in that: The following steps are involved: Data collection and preprocessing steps: deploy triaxial vibration sensors, acoustic sensors, infrared thermal imagers, and hyperspectral cameras, and synchronize time using the IEEE 1588v2 protocol; Wavelet packet transform is used to decompose vibration signals, and variational mode decomposition is combined to suppress modal aliasing. Spectral angle mapping algorithm is used to remove background noise and construct a four-dimensional feature tensor. Feature fusion steps: Design a reconfigurable multi-branch convolutional neural network equipped with a dynamic routing module to adjust the convolution kernel size and dilation rate according to the characteristics of the input data; introduce a cross-modal attention gating mechanism to assign feature weights by calculating mutual information entropy; Adopting a progressive feature pyramid structure to fuse multi-scale features; Diagnosis and prediction steps: Build a twin network, design a defect evolution prediction module, use a spatiotemporal graph convolutional network to model historical inspection data, and predict defect development trends; Combined with reinforcement learning algorithms to optimize detection strategies, through interactive learning with the virtual bearing system, balance detection accuracy and efficiency; Decision-making and early warning steps: Use quantum annealing algorithms to optimize Gaussian mixture model parameters and construct a high-dimensional anomaly detection space; design a multi-threshold decision mechanism to assess defect severity; develop a cross-modal correlation early warning system that triggers an emergency shutdown command when the spatiotemporal matching between vibration anomalies and thermal imaging hotspots exceeds a threshold; Collaborative optimization steps: Build a heterogeneous collaborative framework, use model distillation technology to compress parameters on edge nodes, and protect data privacy through secure multi-party computing; introduce evolutionary strategies to adjust the network architecture, and combine digital twin technology to pre-verify the model optimization effect.

2. The bearing defect intelligent detection method based on deep learning according to claim 1 is characterized in that: Also includes: Calibration and compensation steps: Deploy a laser Doppler vibrometer and establish a spatiotemporal conversion model through cross-correlation analysis; when sensor data is missing, use the implicit neural representation technology of the generative adversarial network to reconstruct the missing signal; develop a sensor health prediction module and use the long short-term memory network to analyze the sensor's historical data.

3. The bearing defect intelligent detection method based on deep learning according to claim 1 is characterized in that: Also includes: Model evolution steps: Build a real-time monitoring system for model performance and evaluate model uncertainty by calculating the Shannon entropy of the test results; When the uncertainty exceeds the threshold, the active learning mechanism is triggered, and the Bayesian active learning strategy is used to screen samples; a dynamic adjustment algorithm for the model structure is designed, and the number of network layers and connection methods are optimized based on the genetic algorithm.

4. The bearing defect intelligent detection method based on deep learning according to claim 1 is characterized in that: The diagnosis and prediction step uses a quantum convolutional neural network to replace the traditional convolutional layer and uses quantum bits to extract defect characteristics; Construct a defect semantic map, associate defect images, maintenance records and material properties, and jointly reason about multi-source information through a graph attention network; develop a sub-pixel positioning algorithm to locate defects based on super-resolution reconstruction technology.

5. The bearing defect intelligent detection method based on deep learning according to claim 1 is characterized in that: The decision-making and early warning steps use digital twin technology to construct a digital mirror of the bearing health, generate virtual detection data by simulating different working conditions, and expand the abnormality detection boundary; use the quantum Monte Carlo method to calculate the probability of defect occurrence, and quantify the uncertainty as the quantum state probability amplitude; design a graded early warning response mechanism, and when the remaining life is predicted to be lower than the threshold, automatically generate a maintenance work order and dispatch drones to deliver spare parts.

6. The bearing defect intelligent detection method based on deep learning according to claim 1 is characterized in that: The collaborative optimization step develops a model interpretability module and uses layer-by-layer correlation propagation technology to generate defect feature heat maps; builds a global inspection data blockchain network to share cross-enterprise model parameters through smart contracts; and uses quantum reinforcement learning to optimize model update strategies.

7. The bearing defect intelligent detection method based on deep learning according to claim 1 is characterized in that: Develop an intelligent defect repair module, generate maintenance robot operation paths based on reinforcement learning, and automatically replace bearing seals and add lubricating oil; build a predictive maintenance knowledge graph, use natural language processing technology to extract expert experience, and trace the cause of the fault; design an energy consumption optimization system by dynamically adjusting the sensor sampling frequency and algorithm calculation accuracy.

8. A bearing defect intelligent detection system based on deep learning according to any one of claims 1 to 7, characterized in that: Includes the following modules: Multimodal sensing module: Integrates a MEMS three-axis vibration sensor, a fiber optic acoustic sensor, an uncooled infrared focal plane array, and a hyperspectral imaging module, with a built-in field programmable gate array for real-time noise reduction and feature pre-extraction; Reconstruction processing module: adopts heterogeneous computing architecture, deploys dynamic network reconstruction engine, and automatically switches computing resource allocation strategy according to detection tasks; Design a hardware-level security unit to protect the security of detection data and model parameters through a trusted execution environment; Decision control module: builds a decision engine and integrates industry standards and expert experience; Develop a 3D visual diagnostic system, use AR to mark defect locations, and simulate repair plans; integrate industrial IoT interfaces and link PLC and SCADA systems; Evolutionary Operations and Maintenance Module: Build a federated learning server cluster to collaboratively train cross-regional models; deploy an automated testing platform to generate virtual test cases through digital twin technology; configure intelligent inspection robots to regularly calibrate system hardware and troubleshoot faults.

9. The deep learning-based intelligent bearing defect detection system according to claim 8, characterized in that: Also includes: Defect microscopic analysis module: equipped with a scanning electron microscope and an energy dispersive X-ray spectrometer, which automatically triggers microscopic inspection when suspected material defects are detected; Carbon footprint management module: Optimizes algorithm execution strategies by analyzing and detecting process energy consumption data; Supply chain collaboration module: When batch defects are detected, quality warnings are automatically sent to suppliers and the emergency dispatch process for spare parts is initiated.

10. The deep learning-based intelligent bearing defect detection system according to claim 8, characterized in that: The multimodal perception module adopts a self-powered design, integrating a vibration energy harvester and a thermoelectric power generation unit to convert energy; the reconstruction processing module supports dynamic loading and unloading of models, and enters a low-power sleep mode during idle periods; the evolutionary operation and maintenance module uses digital thread technology to record information throughout the entire life cycle and trace and reproduce problems.

Images