Nondestructive testing method for encapsulated bearing based on hyperspectral and multi-sensor information fusion

Abstract

This invention discloses a non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion, comprising the following steps: acquiring hyperspectral image data and various operating condition data of the rubber-coated bearing to be tested; preprocessing the hyperspectral image data and operating condition data, and inputting the preprocessed hyperspectral image data and operating condition data into a pre-trained deep learning model; wherein, the deep learning model is constructed to perform correlation and fusion analysis on the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data; based on the output of the deep learning model, obtaining the defect detection results of the rubber-coated bearing to be tested, wherein the defect detection results include at least one of defect type, defect location, and defect severity level, so as to achieve accurate detection of defects such as delamination, microcracks, and internal delamination.

Description

Technical Field

[0001] This invention relates to the field of rubber-coated bearing testing technology, specifically to a non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion. Background Technology

[0002] Rubber-coated bearings play a crucial role in many sectors of modern industry. In the semiconductor manufacturing industry, high-precision equipment such as lithography machines and etching machines place extremely high demands on the performance of rubber-coated bearings. These bearings must ensure extremely low vibration and noise while operating at high speeds to guarantee the accuracy of chip manufacturing. Rubber-coated bearings are equally indispensable in the motors, transmission systems, and lithium battery production equipment of new energy vehicles. Their excellent vibration damping, noise reduction, corrosion resistance, and high reliability play a key role in improving the performance of new energy vehicles and the production quality of lithium batteries. Furthermore, in automated production lines and medical equipment, rubber-coated bearings, with their unique advantages, have become critical components ensuring the stable operation of equipment.

[0003] During the injection molding process, the plastic in rubber-coated bearings can block the venting channels, trapping gas and preventing its escape, leading to insufficient load-bearing capacity. Furthermore, due to improper sealing, storage, or use, the workpiece surface is prone to defects such as scratches, dents, and corrosion, resulting in substandard products. Rubber-coated bearings are composed of a metal inner / outer ring and an elastic rubber layer (such as polyurethane or rubber), and are widely used in transmission systems. The integrity of the rubber layer (free from internal air bubbles and surface scratches) directly affects the bearing's load-bearing capacity and service life.

[0004] Rubber-coated bearings, as composite transmission components, rely heavily on the internal integrity (no bubbles, no delamination) and surface quality (no scratches, no missing adhesive) of their rubber layer (materials such as polyurethane and nitrile rubber) to directly affect their load-bearing capacity and service life, thus determining the reliability of equipment operation. They are widely used in automotive chassis, aerospace servo systems, industrial robot joints, and other fields.

[0005] Non-destructive testing techniques for rubber-coated bearings can be mainly divided into two categories: single testing techniques and early information fusion testing techniques. Traditional testing methods for rubber-coated bearings mainly include manual visual inspection, ultrasonic testing, magnetic particle testing, and radiographic testing.

[0006] Manual inspection relies heavily on the experience and skills of the inspectors, is highly subjective, and is not only inefficient but also has extremely limited ability to detect minute and internal defects, making it difficult to meet the inspection needs of large-scale production. While ultrasonic testing can detect internal defects, it has certain errors in judging the shape, size, and location of defects, and the results are greatly affected by the operator's skill level. Magnetic particle testing is only suitable for detecting surface and near-surface defects in ferromagnetic materials, limiting its application range; it is ineffective for rubber-coated bearings made of non-ferromagnetic materials. Although radiographic testing can detect internal defects, it poses radiation hazards, has strict requirements for the testing environment and equipment, and is costly. Furthermore, this method is also ineffective at detecting some minute defects and defects in complex structures.

[0007] It is evident that existing testing technologies typically cannot simultaneously meet the requirements of "material universality," "comprehensive coverage of internal and external defects," "high-precision positioning," and "automated testing." Furthermore, multi-sensor fusion technology only reaches the level of data overlay and fails to achieve in-depth correlation analysis between intrinsic material properties (such as composition and microstructure) and external operating conditions (such as vibration and temperature). Therefore, it cannot meet the requirements for comprehensive, high-precision, and automated non-destructive testing of rubber-coated bearings. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a non-destructive testing method for rubber-coated bearings based on the fusion of hyperspectral and multi-sensor information, thereby solving the problems in existing technologies.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion includes: The system simultaneously acquires hyperspectral image data and various operating condition data of the rubber-coated bearing under test. The hyperspectral image data is obtained by acquiring continuous spectral information of the bearing surface through a hyperspectral imaging device, which is used to characterize the bearing material composition and microstructure. The operating condition data is acquired by a multimodal sensor and includes vibration data and temperature data, which are used to characterize the mechanical operating state and thermal load of the bearing. The hyperspectral image data and the operating condition data are preprocessed, and the preprocessed hyperspectral image data and the operating condition data are respectively input into a pre-trained deep learning model; wherein, the deep learning model is constructed to be able to correlate and fuse the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data. Based on the output of the deep learning model, the defect detection results of the rubber-coated bearing under test are obtained; wherein, the defect detection results include at least one of defect type, defect location and defect severity level.

[0010] In one embodiment, hyperspectral image data and various operating condition data of the rubber-coated bearing under test are acquired simultaneously, including: Acquiring hyperspectral image data of a rubber-coated bearing to be tested includes: projecting continuous spectral light onto the surface of the rubber-coated bearing to be tested, collecting reflected or transmitted light and decomposing it into continuous monochromatic light according to wavelength, simultaneously collecting spectral data at each spatial location, and generating a hyperspectral image containing spatial and spectral information; wherein, each pixel of the hyperspectral image corresponds to a continuous spectral curve, which is used to characterize the reflectivity or absorptivity of the bearing at different wavelengths. The system acquires various operating condition data of the rubber-coated bearing under test, including: during static testing, collecting vibration baseline data and temperature stability values ​​of the bearing in a static state; and during dynamic operation, synchronously triggering a hyperspectral imaging device and a multimodal sensor through a photoelectric sensor.

[0011] In one embodiment, the hyperspectral imaging device has a spectral range of 400-1000nm and the acquired hyperspectral image has a pixel resolution of ≥300×300; The wavelength resolution of the continuous spectral curve is ≤0.1nm, in order to capture spectral anomalies caused by differences in bearing material composition or microstructural defects.

[0012] In one embodiment, preprocessing of the hyperspectral image data and the operating condition data includes: The hyperspectral image data is subjected to whiteboard correction and wavelet noise reduction. The vibration data were subjected to Fourier transform to extract frequency domain features; The temperature data is then filtered using a moving average and outlier removal.

[0013] In one embodiment, the deep learning model is a CNN-BiLSTM-Attention architecture, comprising: The hyperspectral branch of CNN includes 3 convolutional layers and 2 pooling layers, with convolutional kernel sizes of 3×3, 3×3 and 5×5, respectively, used to extract 256-dimensional spectral features; The working condition branch fully connected network includes two fully connected layers (128 neurons and 64 neurons respectively) to splice the vibration frequency domain features and temperature features to obtain 64-dimensional working condition features; The feature fusion layer includes at least one of a data-level fusion layer, a feature-level fusion layer, or a decision-level fusion layer; wherein the feature-level fusion layer is used to concatenate spectral features and operating condition features into a 320-dimensional comprehensive feature, and an attention mechanism is used to assign different importance weights to the spectral features and the operating condition features respectively. BiLSTM layers are used to extract temporal dependencies; The output layer uses the Softmax activation function to output the state prediction results of the rubber-coated bearing under test.

[0014] In one embodiment, the data-level fusion layer is used to associate and bind each pixel of the hyperspectral image with the corresponding operating condition data at that time. The decision-level fusion layer uses a voting method or a weighted average method to fuse the independent decision results from each data source.

[0015] In one embodiment, the defect type of the defect detection result includes at least one of degumming, cracks, internal delamination, bubbles, missing adhesive, and wear.

[0016] In one embodiment, it further includes: Based on the defect detection results and historical detection data, a time series prediction algorithm is used to generate an assessment of the remaining service life of the bearing, and a maintenance strategy is output according to the defect type and severity level.

[0017] A non-destructive testing system for rubber-coated bearings based on hyperspectral and multi-sensor information fusion includes: The acquisition module is used to simultaneously acquire hyperspectral image data and various operating condition data of the rubber-coated bearing under test. The hyperspectral image data is obtained by acquiring continuous spectral information of the bearing surface through a hyperspectral imaging device, which is used to characterize the bearing material composition and microstructure. The operating condition data is acquired by a multimodal sensor and includes vibration data and temperature data, which are used to characterize the mechanical operating state and thermal load of the bearing. The fusion analysis module is used to preprocess the hyperspectral image data and the operating condition data, and input the preprocessed hyperspectral image data and the operating condition data into a pre-trained deep learning model; wherein, the deep learning model is constructed to be able to correlate and fuse the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data. The output module is used to obtain the defect detection results of the rubber-coated bearing under test based on the output of the deep learning model. The defect detection results include at least one of the following: defect type, defect location, and defect severity level.

[0018] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention acquires the chemical composition and internal structure information of the bearing adhesive layer through hyperspectral imaging, and collects the spatial features of surface defects (scratches, bubbles) through machine vision. After preprocessing (spectral normalization, image denoising), the data is input into a CNN model. This model extracts spectral features and spatial features through a multi-branch network, outputs feature vectors through a fusion layer, and finally compares them with a preset grade threshold to achieve non-destructive grading. This solves the problem that traditional detection (ultrasound, eddy current) cannot take into account both internal and external defects and relies on manual grading. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the workflow of a non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion provided by this invention; Figure 2 This is a schematic diagram illustrating the workflow and signal transmission logic of the present invention; Figure 3 This is a schematic diagram illustrating the construction and prediction process of the CNN-BiLSTM-Attention model of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1 to 3 As shown, the present invention provides a non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion, comprising the following steps: S100: Acquire hyperspectral image data and various operating condition data of the rubber-coated bearing under test; specifically, including: S110. Continuous spectral information of the surface and shallow layer of the rubber-coated bearing under test is acquired using a hyperspectral imaging device to obtain hyperspectral image data; wherein, the hyperspectral image data is used to characterize the material composition and microstructure; specifically, it includes: S111, Project light onto the surface of the rubber-coated bearing to be tested; S112. Collect its reflected or transmitted light and decompose it into continuous monochromatic light according to wavelength to form a spectrum; S113. Synchronously acquire spectral data corresponding to each spatial location to generate a hyperspectral image containing spatial and spectral information; In this system, each pixel of the hyperspectral image data corresponds to a continuous spectral curve. The spectral curve is used to characterize the reflectance or absorptivity of the rubber-coated bearing under test at different wavelengths, so that the hyperspectral image data can characterize the chemical properties of the material composition of the rubber-coated bearing under test and the spectral anomalies caused by microstructural defects.

[0022] It should be noted that hyperspectral imaging technology is an advanced detection method that integrates spectroscopy and imaging technology, enabling the acquisition of detailed information about an object across a continuous spectrum. In the non-destructive testing of rubber-coated bearings, the hyperspectral imaging device mainly consists of a light source, an imaging spectrometer, and a detector. Light emitted from the light source illuminates the surface of the rubber-coated bearing; some light is absorbed by the bearing, and some is reflected back. The reflected light enters the imaging spectrometer, which uses dispersive elements such as gratings and prisms to decompose the reflected light into continuous monochromatic light according to wavelength, forming a spectrum. The detector simultaneously acquires spectral data corresponding to each spatial location (pixel), ultimately generating a hyperspectral image containing both spatial and spectral information. Each pixel corresponds to a complete spectral curve, which records information such as the reflectivity or absorptivity of the rubber-coated bearing at different wavelengths.

[0023] Because different materials have different light absorption and reflection characteristics, their spectral features also differ. Therefore, by analyzing the spectral curve of each pixel in a hyperspectral image, information such as the material composition, structure, and presence of defects in rubber-coated bearings can be obtained. For example, rubber in rubber-coated materials and metal bearings will exhibit different characteristic peaks in their spectra. The position and intensity of these characteristic peaks can reflect the microscopic properties of the materials.

[0024] In one embodiment, the hyperspectral imaging device has a spectral range of 400-1000nm, and the acquired hyperspectral image has a pixel resolution of ≥300×300; the wavelength resolution of the continuous spectral curve is ≤0.1nm, in order to capture spectral anomalies caused by differences in bearing material composition or microstructural defects (width ≤0.1mm).

[0025] By employing hyperspectral technology, we can achieve unique advantages in many aspects of rubber-coated bearing testing.

[0026] First, it can accurately detect differences in bearing material composition. Rubber-coated bearings typically consist of a metal bearing and a rubber coating. Different manufacturers may produce rubber coatings with slight variations in composition and formulation, leading to differences in their spectral characteristics. Hyperspectral technology, by analyzing spectral information, can accurately identify the type and composition of the rubber coating material, thereby determining whether it meets quality standards.

[0027] Secondly, hyperspectral technology has extremely high detection sensitivity for minute defects. Even extremely small defects such as cracks or delamination can alter the light reflection characteristics of the rubber-coated bearing surface, thus exhibiting anomalies in the spectrum. Hyperspectral imaging can capture these subtle changes, enabling early detection and precise location of minute defects.

[0028] Compared to traditional testing methods, hyperspectral technology offers significant advantages. Traditional visual inspection struggles to detect internal and minute defects, and suffers from low efficiency and high subjectivity. While methods such as ultrasonic testing, magnetic particle testing, and radiographic testing can detect some defects, they have limitations in accuracy and applicability. Hyperspectral technology, with its high resolution, provides rich spectral detail, enabling comprehensive and meticulous inspection of rubber-coated bearings. It can detect not only surface defects but also gain in-depth understanding of the internal structure through spectral analysis, achieving a holistic assessment of the bearing's quality.

[0029] S120. Vibration and temperature data of rubber-coated bearings during static testing or dynamic operation are collected by multi-modal sensors to obtain various operating condition data. Among them, vibration data is used to characterize the mechanical operating state and abnormal frequency of rubber-coated bearings, and temperature data is used to characterize the friction state and thermal load of rubber-coated bearings.

[0030] It should be noted that multiple sensors are deployed simultaneously to collect multimodal data of the rubber-coated bearing during static testing or dynamic operation. For example, the vibration sensor uses a piezoelectric sensor with a sampling frequency of 10kHz to acquire the bearing's time-domain vibration signal; the temperature sensor uses an infrared temperature measurement module with an accuracy of ±0.1℃ to collect the temperature of the bearing surface and core area in real time; and the pressure sensor monitors the bearing's operating load pressure.

[0031] It should be noted that the hyperspectral image data acquisition includes: projecting continuous spectral light onto the surface of the rubber-coated bearing under test, collecting the reflected or transmitted light and decomposing it into continuous monochromatic light according to wavelength, simultaneously collecting spectral data at each spatial location, and generating a hyperspectral image containing spatial and spectral information; each pixel of the hyperspectral image corresponds to a continuous spectral curve, which is used to characterize the reflectivity or absorptivity of the bearing at different wavelengths. The operational data acquisition includes: during static testing, acquiring vibration baseline data and temperature stability values ​​of the bearing under static conditions; during dynamic operation, synchronously triggering the hyperspectral imaging device and multimodal sensor through photoelectric sensors, and aligning the data acquisition time based on hardware timestamps to eliminate time synchronization errors.

[0032] Specifically, for dynamic detection scenarios, this invention employs a synchronous triggering mechanism based on photoelectric sensors: When the photoelectric sensor detects that the bearing is in place and sends a trigger signal, the hyperspectral imaging device, vibration sensor, and temperature sensor simultaneously start data acquisition. During the acquisition process, the acquisition time of each data point is recorded by hardware timestamp (accuracy ≤1ms) to ensure that the acquisition time of each pixel in the hyperspectral image corresponds one-to-one with the acquisition time of the working condition data, with a time alignment error of ≤10ms, thereby achieving accurate synchronization of multi-source data in dynamic scenarios.

[0033] By addressing the differences in data collection under static / dynamic scenarios and establishing a dynamic data synchronization mechanism, the problem of asynchronous dynamic detection data in existing technologies can be solved.

[0034] S200. Preprocess the hyperspectral image data and operating condition data, and input the preprocessed hyperspectral image data and operating condition data into a pre-trained deep learning model; wherein, the deep learning model is constructed to be able to correlate and fuse the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data. The preprocessing of hyperspectral image data and operating condition data includes: S210. Perform whiteboard correction and wavelet denoising on the hyperspectral image data; specifically, the acquired hyperspectral data is preprocessed using a "spectral correction-noise removal-normalization" workflow, such as: Eliminate errors caused by uneven lighting by using a standard whiteboard for calibration; High-frequency noise in spectral data is removed using wavelet transform algorithm; Normalizing the spectral reflectance to the [0,1] interval ensures data scale consistency and provides high-quality data for subsequent model input.

[0035] S220. Perform Fourier transform on the vibration data to extract frequency domain features; extract key indicators such as peak frequency and root mean square; the Fourier transform sampling rate is set to 10kHz. Since the abnormal vibration frequency during the operation of the rubber-coated bearing is mainly concentrated in the range of 0-5kHz, the 10kHz sampling rate can satisfy the Nyquist sampling theorem to ensure that the frequency domain features are extracted without distortion.

[0036] S230. Perform moving average filtering and outlier removal on the temperature data. The temperature data is smoothed using the moving average method (where the window size can be set to 5. The temperature data change period is usually about 0.5 seconds. The window size of 5 sampling points can effectively smooth short-term fluctuation noise while preserving the temperature change trend). Features such as temperature change rate and hotspot threshold are extracted. As needed, pressure data is combined with time series data to form a load feature vector.

[0037] To enable the constructed convolutional neural network to accurately detect defects in rubber-coated bearings, the model needs to be trained using a large amount of rubber-coated bearing sample data.

[0038] By collecting samples of rubber-coated bearings from different production batches and under various operating conditions, covering samples in normal states as well as those with various common defects, the diversity and representativeness of the training data were ensured. These sample data were acquired using a hyperspectral imaging system to obtain hyperspectral images, and were combined with other auxiliary detection methods to obtain relevant detection information, such as temperature and vibration data. This data underwent preprocessing, including normalization and standardization, to ensure a uniform scale and distribution, thereby improving the model's training effectiveness and convergence speed.

[0039] For example, 1,000 rubber-coated bearing samples were collected from different production batches and under different working conditions. Among them, 500 were normal samples, including 500 samples with common defects such as delamination (100), cracks (100), internal delamination (80), bubbles (70), missing glue (80), and wear (70). The samples covered different rubber coating materials such as polyurethane and silicone rubber, as well as different sizes from φ20 to φ50mm, simulating the diverse scenarios in actual production.

[0040] The accuracy of defect location marking is ≤0.1mm (pixel-level marking based on high-resolution images); the severity of defects is divided into three levels: mild (defect size <0.2mm, does not affect basic use), moderate (defect size 0.2-0.5mm, requires monitoring), and severe (defect size >0.5mm, prohibits use). The marking process is cross-verified by 3 senior inspection engineers to ensure the accuracy of the marking.

[0041] The preprocessed sample data is divided into training, validation, and test sets. The training set is used for model parameter updates and training, while the validation set is used to monitor the model's training process and adjust hyperparameters such as learning rate and number of iterations to prevent overfitting. For example, the data can be divided into three sets: 70% training set (700 samples), 15% validation set (150 samples), and 15% test set (150 samples). Stratified sampling is used to ensure consistent distribution of defect types and severity levels across the datasets.

[0042] During training, the cross-entropy loss function is chosen as the metric to measure the difference between the model's predictions and the true labels. For binary classification problems, the cross-entropy loss function effectively measures the deviation between the model's predicted probabilities of being positive and negative and the true labels; for multi-class classification problems, it comprehensively considers the differences between the model's predicted probabilities for each class and the true class. By minimizing the cross-entropy loss function, the model's parameters are continuously adjusted to make the model's predictions as close as possible to the true labels.

[0043] Specifically: the cross-entropy loss function is chosen as the metric to measure the difference between the model's predicted results and the true labels; the Adam optimizer is used, with an initial learning rate of 0.001. When the validation set accuracy does not improve for 5 consecutive iterations, the learning rate automatically decays to 1 / 10 of its original value (learning rate adjustment strategy); the convergence condition for model training is: the number of iterations reaches 500 or the validation set loss function value is lower than 0.01; L1 and L2 regularization are used (λ=0.001), and a Dropout mechanism (dropout=0.5) is introduced in the fully connected layer to prevent the model from overfitting.

[0044] During training, regularization techniques can be employed to prevent overfitting. For example, L1 and L2 regularization, by adding regularization terms to the loss function, constrain the model's parameters, preventing them from becoming excessively large and thus avoiding the model learning overly complex patterns, thereby improving its generalization ability. Simultaneously, introducing Dropout mechanisms into fully connected layers, which randomly discard some neurons and their connections with a certain probability, reduces co-adaptation among neurons, preventing the model from over-relying on certain specific neurons and further enhancing its generalization ability.

[0045] During training, model performance metrics such as accuracy, recall, and F1 score are periodically evaluated on the validation set. When the model's performance on the validation set stops improving or shows a downward trend, hyperparameters are adjusted or training is stopped promptly to prevent overfitting on the training set and resulting in poor performance on the validation and test sets. The test set is used for the final performance evaluation. After training, the model is tested on the test set to obtain metrics such as accuracy and recall on unknown data, measuring the model's actual detection and generalization abilities. By continuously adjusting the model's structure, parameters, and training methods, the model achieves high detection accuracy and reliability on the test set, meeting the practical application requirements of non-destructive testing of rubber-coated bearings.

[0046] In the non-destructive testing of rubber-coated bearings, intelligent driving is achieved through machine learning algorithms and data analysis techniques. Machine learning algorithms are the core of this intelligent driving system, enabling the testing system to learn features and patterns from large amounts of data, thus possessing the ability for automatic decision-making and adaptive adjustment. By analyzing a large amount of historical testing data, potential relationships and patterns between data points are uncovered. For example, using association rule mining algorithms in data mining, the correlation between the production process parameters and defect types of rubber-coated bearings is analyzed. It is discovered that a specific rubber coating material formulation is more prone to delamination defects under specific injection molding temperature and pressure conditions. Based on these analysis results, process parameters can be adjusted in a timely manner during production to prevent defects. Simultaneously, time series analysis techniques are used to analyze the testing data of rubber-coated bearings at different operating stages to predict their fatigue strength trends. Based on the trends in vibration, temperature, and other data of the rubber-coated bearings over a period of time, it is predicted whether they are likely to fail in the future, thus allowing for proactive maintenance measures and avoiding equipment downtime.

[0047] In the non-destructive testing of rubber-coated bearings, to further improve the reliability and accuracy of the testing, a multi-source information fusion strategy is adopted, which integrates hyperspectral data with data from other sensors (such as vibration sensors, temperature sensors, pressure sensors, etc.). Specifically, the multi-source information fusion strategy of this invention can employ a single fusion strategy, namely a feature-level fusion layer.

[0048] It should be noted that the multi-source information fusion strategy of the present invention can also be determined according to a preset fusion strategy, which includes: In static detection scenarios, feature-level fusion (weighted fusion of spectral features and operating condition features) should be prioritized. In dynamic operation monitoring scenarios, a combination of "data-level fusion and decision-level fusion" is adopted (first, data-level fusion is used to bind pixels with operating condition data, and then decision-level fusion is used to verify the detection results). When the confidence level of a single data source detection result is lower than 0.8, the three-level fusion mode is automatically activated to ensure detection reliability.

[0049] Data-level fusion is the most direct fusion method, operating at the raw data level. It merges hyperspectral image data acquired by a hyperspectral imaging system with vibration signal data from vibration sensors and temperature data from temperature sensors during the preprocessing stage (based on a dynamic scene timestamp synchronization mechanism). For each pixel in the hyperspectral image, corresponding vibration and temperature data from the same moment are fused. When detecting internal defects in rubber-coated bearings, the hyperspectral image may show spectral anomalies in local areas, but relying solely on spectral information is insufficient to accurately determine the nature and severity of the defect. In this case, combining the vibration data of that area reveals abnormal fluctuations in the vibration signal (e.g., peak frequency deviating from the normal range by ±5%), indicating a possible structural defect in that area leading to altered vibration characteristics. Furthermore, combining the temperature data reveals an abnormal temperature increase (e.g., exceeding the normal operating temperature range by ±3℃), further suggesting increased friction, stress concentration, or other issues in that area, leading to a comprehensive assessment of a significant internal defect. Through data-level fusion, the raw, detailed information from each sensor can be fully utilized, providing a richer data foundation for subsequent analysis and processing.

[0050] Feature-level fusion is a process of fusing data extracted from various sensors. First, features are extracted from hyperspectral data, vibration data, and temperature data separately. For hyperspectral data, spectral and spatial features are extracted using methods such as convolutional neural networks. For vibration data, frequency and temporal features are extracted using signal processing methods such as Fourier transform and wavelet transform. For temperature data, features such as temperature change trends and hotspot regions are extracted. Then, these extracted features are fused to form a comprehensive feature vector. The spectral feature vector from the hyperspectral data, the frequency feature vector from the vibration data, and the trend feature vector from the temperature data are concatenated sequentially to form a new feature vector. This comprehensive feature vector is then input into a classifier (such as a support vector machine or neural network) for analysis and judgment. When determining whether a rubber-coated bearing has a fault, comprehensively considering the material property changes reflected in the hyperspectral data, the mechanical motion characteristics reflected in the vibration data, and the thermal characteristics reflected in the temperature data allows for a more comprehensive and accurate identification of the fault type and location, improving the reliability of detection.

[0051] Decision-level fusion involves each sensor independently analyzing and making decisions, then fusing these results. For example, a hyperspectral imaging system analyzes hyperspectral image data to determine if a rubber-coated bearing has a defect and the type of defect, outputting a decision result (confidence level 0-1). A vibration sensor analyzes vibration signals to determine if the bearing has a mechanical fault, also outputting a decision result (confidence level 0-1). The same applies to the temperature sensor. These decision results from different sensors are then fused using specific rules, such as voting or weighted averaging. For instance, using voting, if two or more of the three sensors determine that the rubber-coated bearing has a problem, then the bearing is ultimately determined to be faulty. Using a weighted averaging method (hyperspectral decision weight 0.5, vibration decision weight 0.3, temperature decision weight 0.2, determined based on the detection accuracy percentage of each data source), a fused confidence level ≥ 0.8 is considered a valid detection result; 0.5 ≤ confidence level < 0.8 triggers a second detection; and a confidence level < 0.5 is considered a detection failure, requiring data re-acquisition. The advantages of decision-level fusion are high flexibility, each sensor can work independently without interfering with each other, and the requirements for communication bandwidth and data processing capabilities are relatively low, making it highly practical in real-world applications.

[0052] To verify the effectiveness of the present invention, a series of non-destructive testing and model prediction experiments were conducted on actual rubber-coated bearing samples.

[0053] The experiment selected 1,000 rubber-coated bearing samples from different production batches and operating conditions, including 500 normal samples and 500 samples with various common defects (such as delamination, cracks, and wear). These samples covered different rubber coating materials (such as polyurethane and silicone rubber), different size specifications (φ20-φ50mm), and different operating environments (normal temperature, high temperature, and humidity) to fully simulate the diversity in actual production and application.

[0054] Data was acquired for each rubber-coated bearing sample using a hyperspectral imaging system to obtain its hyperspectral image. Simultaneously, vibration and temperature data of the samples during operation were collected using vibration and temperature sensors. After preprocessing, the acquired data was divided into three sets: 70% for training, 15% for validation, and 15% for testing.

[0055] The convolutional neural network model is trained using training set data. During training, the model's parameters and hyperparameters, such as learning rate, number of iterations, and kernel size, are continuously adjusted to improve model performance. The training process is monitored using validation set data to prevent overfitting. Training is stopped and the optimal model parameters are saved when the model's performance on the validation set no longer improves or shows a downward trend.

[0056] The trained model was tested using test set data to obtain its prediction results. Performance metrics such as accuracy, recall, and F1 score were calculated by comparing the predictions with the true labels. Experimental results show that the model exhibits high accuracy and reliability in the non-destructive testing of rubber-coated bearings. The model's accuracy reached over 95%, recall was over 90%, and the F1 score, reflecting both precision and recall, reached over 0.92. This means the model can accurately identify most normal and defective rubber-coated bearing samples, and effectively detect defective samples with few false negatives or missed detections.

[0057] Further detailed analysis of the model's prediction results revealed that the model's detection performance varies for different types of defects.

[0058] For debonding defects, the model can accurately identify debonding defects with an accuracy rate of over 98% because the debonded area exhibits obvious spectral feature changes in the hyperspectral image and has significant differences in features from other normal areas.

[0059] For crack defects, although the features of cracks in hyperspectral images are relatively subtle, the powerful feature extraction capability of convolutional neural networks can still effectively detect most crack defects with an accuracy of about 93%.

[0060] However, for some minute wear defects, the model's detection accuracy is relatively low, around 85%, because their impact on hyperspectral images and other detection data is relatively small. To address this, further optimization of the model's structure and parameters, or integration with other more sensitive detection technologies, can improve the detection capability for minute wear defects.

[0061] In one embodiment, the deep learning model is a CNN-BiLSTM-Attention architecture, including: Hyperspectral branch CNN is used to extract spectral features from hyperspectral images; the hyperspectral branch CNN includes multiple convolutional layers and pooling layers. The fully connected network of the working condition branch is used to perform feature splicing on the preprocessed working condition data to obtain working condition features; The feature fusion layer includes at least one of a data-level fusion layer, a feature-level fusion layer, or a decision-level fusion layer; wherein, the feature-level fusion layer is used to concatenate the spectral features output by the hyperspectral branch CNN and the working condition features output by the working condition branch fully connected network, and then use an attention mechanism to assign different importance weights to the spectral features and the working condition features respectively (e.g., hyperspectral feature weight 0.6, working condition feature weight 0.4, which can be determined based on the contribution analysis of the two to defect detection), forming a weighted fusion comprehensive feature; BiLSTM layers are used to extract long-term dependencies from time-related runtime state features; The output layer is used to output the state prediction results of the rubber-coated bearing under test.

[0062] It should be noted that the construction of the CNN-BiLSTM-Attention convolutional neural network model includes network structure design, training optimization, and prediction ranking. This network mainly consists of convolutional layers, pooling layers, and fully connected layers. These layers work together to achieve deep feature extraction and analysis of hyperspectral images and other detection information. Specifically, when predicting the state of rubber-coated bearings based on the trained convolutional neural network model, the following algorithm and process are used: After acquiring the hyperspectral image of the rubber-coated bearing to be tested and other relevant test data (such as vibration, temperature, etc.), the data are first subjected to the same preprocessing operation as the training data.

[0063] For hyperspectral images, normalization is performed to map the spectral value of each pixel in the image to the range of [0,1], in order to eliminate the brightness differences between different images caused by factors such as lighting and acquisition equipment, so that the model can focus more on the feature information of the image.

[0064] For data such as vibration and temperature, standardization is performed so that the mean is 0 and the standard deviation is 1. This ensures the consistency of scale for different types of data and avoids the influence of certain features on the model due to excessively large or small values.

[0065] The preprocessed data is input into the trained convolutional neural network model. The model processes the input data layer by layer according to the pre-trained weights and parameters.

[0066] The data first enters the convolutional layer, where the convolutional kernel slides across the image, extracting local features through convolution operations, such as the texture, edges, and spectral features related to defects on the surface of the rubber-coated bearing. These feature maps are then processed by an activation function (such as the ReLU function), preserving more representative feature information from the image and introducing nonlinear factors to enhance the model's expressive power.

[0067] The convolutional layer is the core component of the entire network, undertaking the crucial task of extracting local image features. In its design, multiple convolutional kernels of different sizes are selected, such as 3×3 and 5×5 kernels. The 3×3 kernel captures detailed image information, suitable for extracting edge features of minute defects; the 5×5 kernel helps obtain more macroscopic structural features, suitable for extracting features of the overall bearing shape and large-area defects. Smaller kernels capture detailed image information, while larger kernels help obtain more macroscopic structural features. For example, a 3×3 kernel can sensitively perceive the edge information of minute cracks on the surface of a rubber-coated bearing, while a 5×5 kernel can better grasp the overall shape and texture features of the bearing.

[0068] By stacking multiple convolutional layers, various features, from low-level to high-level, are extracted progressively from the initial input hyperspectral image. The first convolutional layer takes the hyperspectral image as input and performs a convolution operation using a 3×3 convolution kernel to obtain a series of feature maps. These feature maps contain preliminary features of the image, such as brightness variations at different wavelengths and preliminary texture contours. As the number of network layers increases, subsequent convolutional layers further extract more abstract and representative features based on the previous feature maps. Each convolutional operation adjusts the parameters of the convolution kernel based on the output of the previous layer to better adapt to the extraction requirements of different levels of features.

[0069] Next, the feature map is downsampled through pooling layers to reduce its spatial dimensionality, decrease computational cost, and retain important feature information. Max pooling selects the maximum value within each pooling window as the output, highlighting salient features in the feature map; average pooling averages all values ​​within the pooling window, smoothing the feature map and reducing noise. Through the pooling layer processing, the model becomes more robust to image transformations such as translation, rotation, and scaling.

[0070] Specifically, the pooling layer follows the convolutional layer, and its main function is to downsample the feature map, reducing its spatial dimensionality while preserving important feature information. A strategy combining max pooling and average pooling is employed. Max pooling highlights salient features in the feature map because it selects the maximum value within each pooling window as the output. For example, when detecting defects on the surface of a rubber-coated bearing, max pooling can more clearly highlight the features of the defect area, making it easier for the network to capture this key information. Average pooling, on the other hand, averages all values ​​within the pooling window, which can smooth the feature map to some extent, reduce the impact of noise, and preserve the overall features of the image. By appropriately setting the size and stride of the pooling window, such as the common 2×2 pooling window and stride of 2, the computational load is minimized while ensuring no loss of feature information. This adapts to the input dimensionality requirements of subsequent fully connected layers, effectively reducing computational load while maintaining the feature map's ability to express the key features of the original image. Pooling layers not only reduce the number of parameters in subsequent fully connected layers and lower computational complexity, but also enhance the model's robustness to transformations such as image translation, rotation, and scaling, enabling the model to maintain good performance in detecting rubber-coated bearings in different poses and positions.

[0071] After feature extraction through multiple convolutional and pooling layers, the resulting feature maps are flattened and transformed into one-dimensional vectors, which are then input into a fully connected layer. The fully connected layer performs a linear transformation on the input features using a weight matrix, integrating the previously extracted features to classify and predict the condition of the rubber-coated bearing.

[0072] Specifically, the fully connected layer is located at the end of the network. Its main function is to synthesize the features extracted by the previous convolutional and pooling layers to classify and predict defects in rubber-coated bearings. In the fully connected layer, each neuron is connected to all neurons in the previous layer. The input features are linearly transformed through a weight matrix, and then nonlinear factors are introduced through an activation function (such as the ReLU function) to enhance the network's expressive power. The feature maps output by the convolutional and pooling layers are flattened and transformed into one-dimensional vectors before being input into the fully connected layer.

[0073] Fully connected layers typically contain multiple hidden layers, each of which learns weight parameters to further combine and abstract the input features. For example, the first hidden layer might combine some related low-level features into a higher-level feature representation, and the second hidden layer would then perform deeper feature fusion and abstraction on this basis.

[0074] Finally, after layers of fully connected layers, the output layer generates a vector representing whether the rubber-coated bearing has a defect and the type of defect. For binary classification problems (i.e., determining whether a defect exists), the output layer can use the sigmoid activation function, outputting a probability value between 0 and 1, where 0 indicates no defect and 1 indicates a defect. For multi-class classification problems (i.e., distinguishing different types of defects), the softmax activation function is used, outputting the probability distribution for each category; the category with the highest probability is the predicted defect type.

[0075] The model's output is used to make decisions based on pre-set thresholds. For binary classification problems, if the model's output probability value is greater than the set threshold (e.g., 0.5), the rubber-coated bearing is determined to be defective; otherwise, it is determined to be defect-free. For multi-class classification problems, the category with the highest probability is directly selected as the prediction result. The prediction results are recorded and displayed. Furthermore, depending on actual needs, the prediction results can be fed back to the production system or relevant operators to facilitate appropriate measures, such as marking, isolating, or further inspecting and analyzing defective rubber-coated bearings.

[0076] It should be noted that this invention addresses the need for non-destructive testing of rubber-coated bearings by designing a convolutional neural network prediction model with a specific structure (CNN-BiLSTM-Attention). This network mainly consists of convolutional layers, pooling layers, and fully connected layers, with each layer working collaboratively. CNNs (Convolutional Neural Networks) are responsible for local feature extraction. By sliding one-dimensional convolutional kernels across the input sequence, they can effectively capture short-term dependencies, local patterns, and trends within the sequence. For example, they can capture short-term price fluctuations in stock prediction and phrase-level features in text classification.

[0077] Bi-LSTM (Bidirectional Long Short-Term Memory) is responsible for modeling long-term dependencies. LSTM itself excels at handling long-term dependencies in sequences, and its bidirectional structure allows it to consider both past and future information simultaneously, thus providing a more comprehensive understanding of the context at the current time point. Local feature sequences extracted by CNNs are used as input to the BiLSTM.

[0078] Attention mechanism: Responsible for weighting feature importance. Not all time steps contribute equally to the final prediction. The attention mechanism automatically learns the weights for each time step, allowing the model to "focus" on information more critical to the current task, thereby improving model performance and interpretability. It enables deep feature extraction and analysis of hyperspectral images and other detection information.

[0079] In one embodiment, the data-level fusion layer is used to associate and bind each pixel of the hyperspectral image with the corresponding operating condition data at that time. The decision-level fusion layer uses a voting method or a weighted average method to fuse the independent decision results from each data source.

[0080] In non-destructive testing of rubber-coated bearings, in order to further improve the reliability and accuracy of the test, a multi-source information fusion strategy (including data-level fusion layer, feature-level fusion layer or decision-level fusion layer) is adopted to fuse hyperspectral data with data from other sensors.

[0081] S300. Based on the output of the deep learning model, the defect detection results of the rubber-coated bearing under test are obtained. The defect detection results include at least one of the following: defect type, defect location, and defect severity level. The defect type includes at least one of the following: delamination, crack, internal delamination, bubbles, insufficient adhesive, and wear.

[0082] In one embodiment, it further includes: Based on defect detection results and historical detection data, a bearing remaining service life assessment (e.g., error ≤ ±10%) is generated using time series prediction algorithms (such as the ARIMA model). Maintenance strategies are then output according to the defect type and severity level (e.g., lubrication maintenance is recommended for minor wear, and bearing replacement is recommended for severe degumming).

[0083] It should be noted that, addressing the prominent shortcomings of existing non-destructive testing technologies for rubber-coated bearings, such as low detection accuracy, narrow applicability, insufficient intelligence, and weak multi-source information fusion capabilities, this invention overcomes the bottleneck in detecting minute and internal defects, improving detection accuracy and reliability. Specifically, in existing technologies, manual visual inspection cannot identify internal defects and micron-level minute defects; ultrasonic testing has large errors in judging defect morphology; and radiographic testing struggles to capture minute defects in complex structures, leading to high rates of missed and false detections. This invention uses hyperspectral technology to capture the differences in microscopic spectral characteristics of rubber-coated bearing materials at continuous wavelengths. Combined with the deep feature extraction capabilities of convolutional neural networks, it achieves precise location (position error ≤ 0.1mm) and qualitative detection of defects such as delamination, microcracks (width ≤ 0.1mm), and internal delamination, increasing the overall detection accuracy to over 95% and completely solving the technical pain points of traditional technologies that are "difficult to observe the surface and difficult to distinguish between large and small defects." This invention breaks through the limitations of materials and testing scenarios, broadening the scope of application of the technology. Traditional magnetic particle testing is only applicable to ferromagnetic materials and cannot be adapted to non-ferromagnetic coating materials such as polyurethane and silicone rubber; single-spectral technology, due to insufficient information dimensions, is difficult to cover the bearing testing needs under different operating conditions. On the one hand, this invention utilizes the universality of hyperspectral technology to achieve unified testing of ferromagnetic and non-ferromagnetic coated bearings; on the other hand, it integrates hyperspectral data with operating condition data such as vibration and temperature through a multi-source information fusion strategy, making the technology adaptable to multiple scenarios such as static quality testing, dynamic operation monitoring (speed ≤1000rpm), high-temperature environments (80℃), and oil contamination, solving the problem of "narrow material compatibility and limited scenario coverage" of traditional technologies. Furthermore, this invention enables intelligent detection processes and condition prediction, supporting preventative maintenance. Existing detection technologies rely on manual operation and experience-based judgment, resulting in low efficiency and an inability to predict the future condition of bearings, leading to high costs for "post-construction repairs" in industrial production. This invention utilizes an end-to-end detection process based on convolutional neural networks to achieve fully automated processing from data input to defect classification, increasing detection efficiency by more than 50 times compared to manual methods. It significantly enhances the level of technological intelligence and expands predictive capabilities. Based on historical detection data and time series analysis, a condition prediction model is constructed, providing early warnings of fault trends 1-3 operating cycles in advance. This provides data support for preventative maintenance of industrial equipment, reducing downtime losses and maintenance costs. Furthermore, this invention effectively reduces detection costs and safety risks, enhancing the feasibility of industrialization. Because X-ray detection poses radiation hazards and incurs high equipment costs, and ultrasonic detection requires specialized personnel, increasing labor costs and limiting large-scale adoption, this invention eliminates hazardous detection methods such as X-rays, employing a combination of hyperspectral imaging and conventional sensors to reduce equipment investment and maintenance costs. By automating the process, it minimizes manual intervention, ensuring the safety of testing personnel and reducing human error, thus improving the feasibility of industrial application in small and medium-sized enterprises. Ultimately, this achieves an optimized balance between economic efficiency and safety in detection.

[0084] This technology was compared with traditional non-destructive testing methods. Traditional ultrasonic testing methods, when detecting internal defects in rubber-coated bearings, suffer from reduced accuracy due to the attenuation and scattering of ultrasonic waves by the rubber material. This makes it difficult to accurately detect minute defects and defects in complex structures, with an accuracy rate of only about 80%. Magnetic particle testing is only suitable for detecting surface and near-surface defects in ferromagnetic materials and cannot detect non-ferromagnetic materials in rubber-coated bearings, thus limiting its application. Therefore: Existing technologies either focus solely on material composition (spectral analysis), surface morphology (visual perception), or operating parameters (vibration / temperature), failing to achieve multi-dimensional information coverage encompassing "material composition, structural state, and operating conditions." This incomplete information leads to biased defect detection. This application, however, fills this gap by fusing hyperspectral (material + structure) and vibration / temperature (operating conditions) sources.

[0085] Existing multi-sensor technologies only involve data overlay or threshold judgment, failing to achieve deep fusion at the feature and decision levels. This results in low-level data fusion, which cannot leverage the synergistic advantages of multi-source data. This application addresses the pain point of data "overlay rather than fusion" by constructing a three-level fusion architecture and extracting cross-modal features through a CNN model.

[0086] Existing technologies mostly employ traditional algorithms (SVM) or lightweight CNNs, which cannot extract deep features from high-dimensional data and lack predictive capabilities, resulting in insufficient intelligence. In contrast, this application achieves accurate defect identification (accuracy ≥ 95%) and trend prediction over 1-3 periods through deep CNNs and time series analysis.

[0087] Existing products are greatly affected by environmental factors such as materials, lighting, and oil contamination, making them unsuitable for the diverse scenarios of rubber-coated bearings and exhibiting poor industrial adaptability. This application, however, achieves full coverage of ferromagnetic / non-ferromagnetic materials and static / dynamic scenarios through the universality of hyperspectral materials and multi-source anti-interference design.

[0088] This technology, through an intelligent detection system based on hyperspectral imaging and a CNN-BiLSTM-Attention deep learning model, overcomes the aforementioned bottlenecks by employing a three-level fusion strategy at the data, feature, and decision levels. Its core lies in constructing a detection paradigm capable of simultaneously acquiring and deeply fusing multi-dimensional information such as the "chemical composition (hyperspectral), structural state (vibration), and operating conditions (temperature)" of rubber-coated bearings, thereby achieving accurate, rapid, and non-destructive defect identification and state prediction. It significantly outperforms traditional detection methods in terms of detection accuracy and applicability. Specifically: This invention utilizes hyperspectral technology to capture the microscopic spectral characteristics of rubber-coated bearings at continuous wavelengths. Combined with the multi-layer feature extraction capabilities of convolutional neural networks, it achieves accurate identification of defects such as delamination and micro-cracks. Experiments show that the overall detection accuracy exceeds 95%, with the accuracy rate for detecting delamination defects exceeding 98%, which is more than 15 percentage points higher than that of traditional ultrasonic testing (80% accuracy). This effectively reduces the risk of missed or false detections, and significantly improves detection precision and sensitivity compared to existing technologies.

[0089] This invention breaks through the limitations of traditional testing technologies in terms of materials and defect types. It can detect bearings made of ferromagnetic materials and is also compatible with non-ferromagnetic coating materials such as polyurethane and silicone rubber. It can simultaneously identify surface defects (such as wear), internal defects (such as delamination), and abnormal material composition. It solves the problems that magnetic particle testing is only applicable to ferromagnetic materials and radiographic testing is ineffective for detecting complex structures. It achieves full coverage testing of all types of coated bearings and has a significantly wider range of applications compared to existing technologies.

[0090] By employing an end-to-end detection process using convolutional neural networks and adaptive parameter adjustment of the Adam optimization algorithm, fully automated processing from data input to defect classification is achieved. No manual intervention is required for feature extraction and judgment, resulting in detection efficiency more than 50 times higher than manual visual inspection, meeting the real-time detection needs of large-scale industrial production lines. Simultaneously, the intelligent prediction function based on historical data can provide early warnings of bearing failure trends, supporting preventative maintenance. Compared to existing technologies, the level of intelligence and automation is significantly improved.

[0091] This invention employs a three-level fusion strategy—data-level, feature-level, and decision-level—to integrate hyperspectral data with multi-dimensional information such as vibration and temperature, avoiding the limitations of single-sensor data. For example, combining spectral anomalies identified by hyperspectral imaging with vibration signal fluctuations and temperature increases can accurately determine severe internal defects, reducing the false positive rate by more than 40% compared to single-spectral detection. The reliability of the detection results is significantly enhanced, and compared to existing technologies, multi-source information fusion can significantly improve detection reliability.

[0092] This invention eliminates the need for radiation-hazardous detection methods such as X-rays, as well as complex manual operation procedures, thereby reducing the investment and maintenance costs of detection equipment and ensuring the safety of inspection personnel. At the same time, by detecting defects in advance, it avoids equipment downtime losses. Calculations show that it can reduce the maintenance costs related to rubber-coated bearings in industrial production by more than 30%. Compared with existing technologies, it significantly reduces detection costs and safety risks.

[0093] This invention acquires information on the chemical composition and internal structure of the bearing adhesive layer through hyperspectral imaging, and uses machine vision to collect spatial features of surface defects (scratches, bubbles). After preprocessing (spectral normalization, image denoising), the data is input into a CNN model. This model extracts spectral and spatial features through a multi-branch network, outputs feature vectors through a fusion layer, and finally compares them with a preset grade threshold to achieve non-destructive grading. This invention solves the problems of traditional detection methods (ultrasound, eddy current) being unable to take into account both internal and external defects and relying on manual grading. The detection accuracy of this invention is ≥98.5%, and the detection efficiency is expected to be improved by more than 3 times. It can be used for the whole life cycle quality control of rubber-coated bearings in aerospace, automotive transmission systems and other fields.

[0094] like Figure 2As shown, the workflow and signal transmission logic of this invention are as follows: 1. Signal Trigger: Input: The photoelectric sensor detected that the bearing is in place.

[0095] Processing: Generate a 5V high-level signal with a duration of 100ms as the start command for the entire preprocessing process.

[0096] Output: Trigger signal.

[0097] 2. Multi-source data acquisition: This involves initiating data acquisition tasks from three sensors in parallel, and proceeding to the next stage after all tasks are completed. Data collected by hyperspectral camera: Input: Optical reflection from the bearing surface.

[0098] Processing: Acquire HDR format images of 300×300 pixels, spectral range 400-1000nm, acquisition time is 0.3 seconds.

[0099] Output: Raw hyperspectral image data.

[0100] Vibration sensor data acquisition: Input: Mechanical vibration generated by the operation of the bearing.

[0101] Processing: The time-domain signal was acquired at a sampling rate of 10kHz, with a total of 5000 sampling points for 0.5 seconds, and saved as a TXT file.

[0102] Output: The original time-domain signal of the vibration.

[0103] Temperature sensor data acquisition: Input: Temperature at 5 measuring points on the bearing surface.

[0104] Processing: Temperature was collected at 5 points at 0.1-second intervals, and the average value was calculated with an accuracy of ±0.1℃.

[0105] Output: Average temperature value.

[0106] 3. Hyperspectral preprocessing Whiteboard correction: Input: Raw hyperspectral image (containing sample spectrum, whiteboard spectrum, and dark current data).

[0107] Processing: Apply the formula reflectance = (sample spectrum - dark current) / (whiteboard spectrum - dark current) to normalize the data to the range [0,1].

[0108] Output: Reflectance image.

[0109] Wavelet noise reduction: Input: Reflectance image after whiteboard correction.

[0110] Processing: A 3-level decomposition was performed using the db4 wavelet basis, with a threshold of 0.02 set to remove high-frequency noise.

[0111] Output: Denoising-reduced hyperspectral feature data.

[0112] 4. Preprocessing of operating data (auxiliary data) Vibration signal processing: Input: The original time-domain signal of the vibration.

[0113] Processing: Perform Fourier Transform (FFT) to obtain the frequency domain signal from 0-5kHz. Extract the peak frequency (error ±1Hz) and root mean square (RMS) value.

[0114] Output: Vibration frequency domain characteristics.

[0115] Temperature data processing: Input: Temperature readings at 5 measuring points.

[0116] Processing: First, perform a moving average filter with a window size of 5, then remove outliers other than 3σ, and finally calculate the rate of temperature change.

[0117] Output: Stable temperature characteristic values.

[0118] 5. Data transmission and termination Input: All preprocessed data (hyperspectral features, vibrational features, temperature features).

[0119] Processing: Package the data and transmit it to the industrial motherboard via Ethernet protocol, ensuring network latency ≤100ms.

[0120] Output: Transmission complete signal, and triggers the next detection cycle.

[0121] like Figure 3 As shown, the construction and prediction process of the CNN-BiLSTM-Attention model of this invention is as follows: Process specifications: number of model layers, convolution kernel parameters, activation function, loss function, optimizer; Core process: Feature extraction → Feature fusion → Trend prediction → Result output; 1. Model training: Cross-entropy loss function, Adam optimizer (learning rate 0.001, β1=0.9, β2=0.999). 2. Regularization: Dropout=0.5 for fully connected layers, L2 regularization (λ=0.001) to prevent overfitting; 3. Dataset split: 70% training, 15% validation, 15% testing; test set accuracy ≥ 95%; 4. Prediction latency: Single sample processing time ≤ 500ms, meeting the requirements for real-time detection in pipelines.

[0122] Process steps breakdown: Step 1, Feature Extraction: Hyperspectral Branch (CNN): Convolutional Layer 1 (3×3×64, ReLU) → Pooling Layer 1 (2×2 max pooling) → Convolutional Layer 2 (3×3×128, ReLU) → Pooling Layer 2 (2×2 average pooling) → Convolutional Layer 3 (5×5×256, ReLU) → Output 256-dimensional spectral features; Working Condition Branch (Fully Connected): Vibration frequency domain features (32-dimensional) + Temperature features (32-dimensional) → Fully Connected Layer 1 (128 neurons) → Fully Connected Layer 2 (64 neurons) → Output 64-dimensional working condition features; Step 2, Feature Fusion: 256 + 64 = 320-dimensional comprehensive features → Z-score standardization (mean 0, standard deviation 1); Attention mechanism assigns weights to the comprehensive features (hyperspectral 0.6, working condition 0.4). Step 3, Trend Prediction: The BiLSTM layer inputs 320-dimensional features, has 128 hidden neurons, and a time step of 10 → captures temporal dependencies; the output layer: softmax activation → outputs a 3-level classification of "excellent / qualified / unqualified" (prediction in 1-3 cycles). Step 4, Output the result (rectangle): Local display: 10.1-inch touchscreen → Defect type (adhesion separation / crack / wear), location, size; External control: Send a 24V signal to the PLC (pass → release, fail → reject, lasting 100ms); Data storage: JSON format (including timestamp, bearing ID, defect parameters) → cloud + local SSD.

[0123] A non-destructive testing system for rubber-coated bearings based on hyperspectral and multi-sensor information fusion includes: The acquisition module is used to acquire hyperspectral image data and various operating condition data of the rubber-coated bearing under test; The fusion analysis module is used to preprocess hyperspectral image data and operating condition data, and then input the preprocessed hyperspectral image data and operating condition data into a pre-trained deep learning model. The deep learning model is constructed to perform correlation and fusion analysis on the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data. The output module is used to obtain the defect detection results of the rubber-coated bearing under test based on the output of the deep learning model.

[0124] The technical solution of this invention uses "data acquisition - information fusion - intelligent analysis - predictive output" as the core link, constructing an integrated detection system comprising a data acquisition module, a preprocessing module, an information fusion module, a convolutional neural network model module, and a predictive output module. Specifically, the data acquisition module is responsible for acquiring multi-source raw data; the preprocessing module performs data standardization and noise reduction; the information fusion module integrates features from multi-dimensional data; the convolutional neural network module achieves accurate defect identification; and the predictive output module provides bearing condition assessment and future trend judgment based on the detection results.

[0125] The foregoing description of specific exemplary embodiments of the present invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it is obvious that many changes and variations can be made based on the above teachings. Although embodiments of the invention have been shown and described, these specific embodiments are merely explanations of the invention and are not intended to limit it. The specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. The purpose of selecting and describing exemplary embodiments is to explain the specific principles of the invention and its practical application, so that those skilled in the art, after reading this specification, can make modifications, substitutions, variations, and various choices and changes to the embodiments as needed without departing from the principles and spirit of the invention, provided that such modifications, substitutions, variations, and choices and changes are within the scope of the claims of the invention and are protected by patent law.

Claims

1. A non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion, characterized in that, include: The system simultaneously acquires hyperspectral image data and various operating condition data of the rubber-coated bearing under test. The hyperspectral image data is obtained by acquiring continuous spectral information of the bearing surface through a hyperspectral imaging device, which is used to characterize the bearing material composition and microstructure. The operating condition data is acquired by a multimodal sensor and includes vibration data and temperature data, which are used to characterize the mechanical operating state and thermal load of the bearing. The hyperspectral image data and the operating condition data are preprocessed, and the preprocessed hyperspectral image data and the operating condition data are respectively input into a pre-trained deep learning model; wherein, the deep learning model is constructed to be able to correlate and fuse the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data. Based on the output of the deep learning model, the defect detection results of the rubber-coated bearing under test are obtained; wherein, the defect detection results include at least one of defect type, defect location and defect severity level.

2. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 1, characterized in that, Simultaneously acquire hyperspectral image data and various operating condition data of the rubber-coated bearing under test, including: Acquiring hyperspectral image data of a rubber-coated bearing to be tested includes: projecting continuous spectral light onto the surface of the rubber-coated bearing to be tested, collecting reflected or transmitted light and decomposing it into continuous monochromatic light according to wavelength, simultaneously collecting spectral data at each spatial location, and generating a hyperspectral image containing spatial and spectral information; wherein, each pixel of the hyperspectral image corresponds to a continuous spectral curve, which is used to characterize the reflectivity or absorptivity of the bearing at different wavelengths. The system acquires various operating condition data of the rubber-coated bearing under test, including: during static testing, collecting vibration baseline data and temperature stability values ​​of the bearing in a static state; and during dynamic operation, synchronously triggering a hyperspectral imaging device and a multimodal sensor through a photoelectric sensor.

3. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 2, characterized in that, The hyperspectral imaging device has a spectral range of 400-1000 nm and the pixel resolution of the acquired hyperspectral images is ≥300×300. The wavelength resolution of the continuous spectral curve is ≤0.1nm, in order to capture spectral anomalies caused by differences in bearing material composition or microstructural defects.

4. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 2, characterized in that, Preprocessing of the hyperspectral image data and the operating condition data includes: The hyperspectral image data is subjected to whiteboard correction and wavelet noise reduction. The vibration data were subjected to Fourier transform to extract frequency domain features; The temperature data is then filtered using a moving average and outlier removal.

5. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 1, characterized in that, The deep learning model is a CNN-BiLSTM-Attention architecture, including: The hyperspectral branch of CNN includes 3 convolutional layers and 2 pooling layers, with convolutional kernel sizes of 3×3, 3×3 and 5×5, respectively, used to extract 256-dimensional spectral features; The working condition branch fully connected network includes two fully connected layers (128 neurons and 64 neurons respectively) to splice the vibration frequency domain features and temperature features to obtain 64-dimensional working condition features; The feature fusion layer includes at least one of a data-level fusion layer, a feature-level fusion layer, or a decision-level fusion layer; wherein the feature-level fusion layer is used to concatenate spectral features and operating condition features into a 320-dimensional comprehensive feature, and an attention mechanism is used to assign different importance weights to the spectral features and the operating condition features respectively. BiLSTM layers are used to extract temporal dependencies; The output layer uses the Softmax activation function to output the state prediction results of the rubber-coated bearing under test.

6. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 5, characterized in that, The data-level fusion layer is used to associate and bind each pixel of the hyperspectral image with the corresponding operating condition data at that time. The decision-level fusion layer uses a voting method or a weighted average method to fuse the independent decision results from each data source.

7. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 1, characterized in that, The defect types in the defect detection results include at least one of the following: delamination, cracks, internal delamination, bubbles, missing adhesive, and wear.

8. The non-destructive testing method for rubber-coated bearings based on hyperspectral and multi-sensor information fusion according to claim 7, characterized in that, Also includes: Based on the defect detection results and historical detection data, a time series prediction algorithm is used to generate an assessment of the remaining service life of the bearing, and a maintenance strategy is output according to the defect type and severity level.

9. A non-destructive testing system for rubber-coated bearings based on hyperspectral and multi-sensor information fusion, characterized in that, include: The acquisition module is used to simultaneously acquire hyperspectral image data and various operating condition data of the rubber-coated bearing under test. The hyperspectral image data is obtained by acquiring continuous spectral information of the bearing surface through a hyperspectral imaging device, which is used to characterize the bearing material composition and microstructure. The operating condition data is acquired by a multimodal sensor and includes vibration data and temperature data, which are used to characterize the mechanical operating state and thermal load of the bearing. The fusion analysis module is used to preprocess the hyperspectral image data and the operating condition data, and input the preprocessed hyperspectral image data and the operating condition data into a pre-trained deep learning model; wherein, the deep learning model is constructed to be able to correlate and fuse the intrinsic material properties represented by the hyperspectral image data and the explicit operating state represented by the operating condition data. The output module is used to obtain the defect detection results of the rubber-coated bearing under test based on the output of the deep learning model. The defect detection results include at least one of the following: defect type, defect location, and defect severity level.

Images