A gear fault degree evaluation method based on multi-modal features and domain adaptation
By employing a multimodal feature and domain adaptation approach, the problem of being unable to quantify the severity and changes in operating conditions in gear fault diagnosis is solved, enabling accurate assessment and efficient predictive maintenance under complex operating conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2025-12-01
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies cannot accurately quantify the severity of gear faults and lack the ability to generalize models when operating conditions change, leading to performance degradation under complex conditions.
A multimodal feature and domain adaptation approach is adopted. By preprocessing vibration, acoustic, current and torque sensor data, physical features that are not sensitive to speed changes are extracted. A random forest regressor is used for adaptive feature selection, and an ensemble regression model is constructed to assess the severity of the fault.
It enables accurate quantitative assessment of gear failures under complex operating conditions, enhances the practical application value of predictive maintenance, reduces the risk of model overfitting, and improves the transparency and interpretability of diagnosis.
Smart Images
Figure CN121502274B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment fault diagnosis and predictive maintenance technology, and in particular to a method for assessing the degree of gear faults based on multimodal features and domain adaptation. Background Technology
[0002] As a core component of mechanical transmission systems, the health of gears directly affects the operational safety and reliability of the entire equipment. Therefore, gear fault diagnosis has always been a key technology in the field of industrial equipment condition monitoring and predictive maintenance. In recent years, with the advancement of artificial intelligence technology, data-driven methods based on deep learning have achieved significant results in fault diagnosis. These methods can typically automatically learn features from sensor data such as vibration and sound, enabling intelligent identification and classification of fault types.
[0003] However, existing technologies still face two major challenges in practical industrial applications. First, most methods treat fault diagnosis as a classification problem, only able to determine the existence or type of a fault, but unable to accurately quantify the severity of the fault. In predictive maintenance practice, quantitatively assessing the development stage of a fault is crucial for accurately predicting remaining life and formulating precise maintenance strategies, far exceeding the value of simple qualitative judgments. Second, existing models generally lack generalization ability. They typically assume that training data and future test data are under the same operating conditions, but in real-world scenarios, fluctuations in operating conditions such as equipment speed and load can cause significant "domain shift" problems, leading to a sharp decline in model performance under new operating conditions. Summary of the Invention
[0004] To address the problems mentioned in the background art, this invention provides a gear fault severity assessment method based on multimodal features and domain adaptation, aiming to achieve accurate quantitative estimation of gear fault severity under complex working conditions.
[0005] This invention adopts the following technical solution: This invention provides a gear fault severity assessment method based on multimodal features and domain adaptation, including: S1: Preprocessing multi-sensor time-series data of the gear system to obtain segmented and aligned multimodal data. S2: Extracting physical features insensitive to speed changes from the preprocessed multi-sensor data. S3: Performing adaptive feature selection on the physical features extracted in step S2 to filter out a subset of features that are fault-sensitive and domain-invariant. S4: Inputting the feature subset into an integrated regression model to output the gear fault severity assessment result.
[0006] Step S1 preprocessing includes: performing data cleaning, outlier handling, and standardization on the time-series data collected by the vibration sensor, acoustic sensor, current sensor, and torque sensor, respectively. The processed sensor data are then segmented according to fixed lengths and overlap rates, and aligned at the segment level.
[0007] In step S1, the signals of each sensor channel are... According to a fixed length and overlap rate Segmentation, step size The first segment after segmentation The segment is:
[0008]
[0009] Where M is the total number of segments, .
[0010] Physical characteristics that are insensitive to changes in rotational speed include vibration characteristics, acoustic characteristics, current characteristics, and torque characteristics.
[0011] Vibration characteristics include traditional time-domain and frequency-domain statistical characteristics, gear system-specific frequency-domain characteristics, and order ratio analysis characteristics. Traditional time-domain and frequency-domain statistical characteristics include mean, standard deviation, RMS, kurtosis, skewness, peak factor, impulse factor, and margin factor. The formula for calculating gear system-specific frequency-domain characteristics is as follows:
[0012]
[0013]
[0014] in, It represents the meshing harmonic ratio. The sideband strength ratio.
[0015] In step S2, the vibration signal is resampled at an angle and subjected to a fast Fourier transform to obtain the order ratio spectrum. Based on the order ratio spectrum, the energy proportions near multiple characteristic orders are calculated as order ratio analysis features. The calculation formula for the order ratio analysis features is as follows:
[0016] .
[0017] Acoustic characteristics include the statistics of the Mel-frequency coefficients and the spectral centroid. Current characteristics include the total harmonic distortion rate and the amplitude ratio of characteristic harmonics to the fundamental frequency. Torque characteristics include the ripple intensity and the dominant ripple frequency.
[0018] In step S3, the adaptive feature selection adopts a feature importance evaluation method based on random forest regressors to select features with importance higher than a preset threshold, forming a feature subset.
[0019] The preset threshold is the product of the average of all feature importance scores and the threshold coefficient.
[0020] The ensemble regression model in step S4 is a histogram-based gradient boosting regressor that learns a nonlinear mapping from features to fault severity by minimizing the squared loss function.
[0021] Compared to existing technologies, the beneficial effects are as follows:
[0022] 1. This invention breaks through the limitations of traditional fault classification by constructing an ensemble regression model to directly establish a nonlinear mapping relationship between multimodal characteristics and fault severity, enabling the output of continuous quantitative assessment results. This overcomes the shortcomings of classification methods in reflecting the fault development process, providing key data support for more accurate remaining service prediction and maintenance strategy formulation, and greatly enhancing the practical application value of predictive maintenance.
[0023] 2. This invention innovatively extracts physical features that are insensitive to changes in rotational speed, such as order ratio analysis features and harmonic ratios, from multi-source sensor data including vibration, acoustics, current, and torque. This feature design based on physical mechanisms weakens the impact of changes in operating conditions such as rotational speed and load on feature distribution at the source, effectively overcoming the domain offset problem. This allows the model to maintain high-precision evaluation performance under complex operating conditions such as varying rotational speed and varying load without relying on target domain labeled data.
[0024] 3. This invention, through adaptive feature selection based on random forests, can automatically filter out a subset of features that are strongly correlated with faults and remain stable under different operating conditions. This process not only eliminates redundant and irrelevant features, reducing the risk of model overfitting and improving computational efficiency, but more importantly, it clarifies key diagnostic features, enhances the transparency and interpretability of the model's decision-making process, helps engineers understand and trust the model's output, and facilitates practical industrial deployment and application. Attached Figure Description
[0025] Figure 1 This is a flowchart of a gear fault severity assessment method based on multimodal features and domain adaptation provided in an embodiment of this application;
[0026] Figure 2 This is a comparison chart of experimental results provided in an embodiment of this application;
[0027] Figure 3 This is another comparative chart of experimental results provided in the embodiments of this application. Detailed Implementation
[0028] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0029] Reference Figure 1 This application provides a method for evaluating the degree of gear failure based on multimodal features and domain adaptation. The method includes:
[0030] S1: Preprocess the multi-sensor time-series data of the gear system to obtain segmented and aligned multimodal data.
[0031] For example, the preprocessing in step S1 includes: performing data cleaning, outlier handling, and standardization on the time-series data collected by the vibration sensor, acoustic sensor, current sensor, and torque sensor, respectively. The processed sensor data are then segmented according to a fixed length and overlap rate, and aligned at the segment level.
[0032] Data cleaning removes invalid data generated during sensor acquisition due to environmental interference, such as electromagnetic noise, ambient noise, and temporary equipment fluctuations. Outlier handling eliminates extreme anomalies caused by sensor drift and transient malfunctions, laying a data foundation for subsequent accurate extraction of fault-sensitive features. Standardization maps data from various sensors to the same scale, ensuring that the physical features extracted in step S2 have fair weights in the feature space, supporting the effective fusion of multimodal features.
[0033] The original time-series data is a continuous long sequence, which cannot be directly used to extract local physical features, such as gear meshing harmonic ratio and torque fluctuation intensity. By segmenting the continuous data into discrete data segment analysis units according to a fixed length, step S2 can accurately extract time-domain, frequency-domain, or order-ratio features from each segment of data. At the same time, the standardized segment length also provides a unified input dimension for subsequent model training, avoiding model training errors caused by inconsistent data lengths.
[0034] As one possible implementation, in step S1, the signals of each sensor channel are... According to a fixed length and overlap rate Segmentation, step size The first segment after segmentation The segment is:
[0035]
[0036] Where M is the total number of segments, .
[0037] S2: Extract physical features that are insensitive to changes in rotational speed from the preprocessed multi-sensor data.
[0038] Examples of physical characteristics that are insensitive to changes in rotational speed include vibration characteristics, acoustic characteristics, current characteristics, and torque characteristics.
[0039] As one possible approach, vibration characteristics include traditional time-domain and frequency-domain statistical characteristics, gear system-specific frequency-domain characteristics, and order ratio analysis characteristics. Traditional time-domain and frequency-domain statistical characteristics include mean, standard deviation, RMS, kurtosis, skewness, peak factor, impulse factor, and margin factor.
[0040] The formula for calculating the specific frequency domain characteristics of a gear system is as follows:
[0041]
[0042]
[0043] in, It represents the meshing harmonic ratio. The sideband strength ratio.
[0044] As one possible implementation, in step S2, the vibration signal is resampled at an angle and subjected to a fast Fourier transform to obtain the order spectrum. The energy proportions near multiple characteristic orders are calculated based on the order ratio spectrum, serving as the characteristics for order ratio analysis. The formula for calculating the order ratio analysis characteristics is as follows:
[0045] .
[0046] In some embodiments, acoustic features include statistics (mean, standard deviation) and spectral centroid of Mel spectral coefficients (MFCCs). Current features include total harmonic distortion (THD) and the amplitude ratio of characteristic harmonics to the fundamental frequency. Torque features include ripple intensity and dominant ripple frequency.
[0047] Gear malfunctions, such as tooth wear or broken teeth, produce abnormal sounds at specific frequencies, such as high-frequency friction noise and impact pulse sounds. However, these acoustic signals are easily masked by ambient noise. Mel-frequency coefficients, which simulate the human ear's sensitivity to different frequencies, can automatically focus on the mid-to-high frequency bands relevant to gear malfunctions while filtering out irrelevant low-frequency ambient noise. Furthermore, statistical measures such as mean and standard deviation can further smooth out instantaneous noise fluctuations, extract stable characteristics of the malfunction sound over a period of time, avoid misjudgments caused by instantaneous noise, and improve the malfunction identification accuracy of acoustic features.
[0048] The spectral centroid reflects the main frequency distribution of acoustic signal energy: during normal gear operation, sound energy is concentrated near a stable meshing frequency; when gears experience wear, broken teeth, or other faults, abnormal friction or impact increases high-frequency energy, causing the spectral centroid to shift towards higher frequencies, and the magnitude of this shift is positively correlated with the severity of the fault. This change in energy distribution can be directly quantified through the spectral centroid, providing an intuitive and quantifiable acoustic basis for the subsequent step S4, continuous regression of fault severity.
[0049] Gears are the core component of the transmission chain, and their failures are transmitted to the drive motor through the transmission system, causing distortion of the motor's current waveform. During normal operation, the current waveform is close to a sine wave, but during a fault, load fluctuations introduce harmonic components. The total harmonic distortion (THD) rate can comprehensively measure the degree of distortion in the current waveform, directly reflecting the impact of gear failure on the motor load. Characteristic harmonics such as the 3rd and 5th harmonics are typical indicators of abnormal motor load; their amplitude ratio with the fundamental frequency can accurately pinpoint the type of load anomaly caused by the fault, avoiding misdiagnosis of current distortion caused by faults in other devices, such as the motor itself, as gear failure, thus improving the specificity of the characteristics.
[0050] Vibration and acoustic sensors need to be deployed close to the gearbox, while current sensors can be directly installed in the motor power supply circuit, making deployment more convenient and cost-effective. By capturing the indirect transmission signal of gear faults through current characteristics, the load-related information of gear faults can be supplemented without increasing the difficulty of sensor deployment. This complements the direct fault signals from vibration and acoustic sensors, enriches the multimodal feature dimensions, and reduces the risk of diagnostic interruptions caused by single sensor failures.
[0051] Torque is the core physical quantity for power transmission in gears. Gear faults, such as tooth surface spalling or broken teeth, directly lead to discontinuous torque transmission, manifesting as periodic fluctuations in the torque signal: slight wear only causes small torque fluctuations, while broken teeth trigger drastic and regular torque abrupt changes. The intensity of the fluctuation directly measures the amplitude of the torque fluctuation, while the dominant frequency of the fluctuation corresponds to the periodicity of the fault. Combining the two can both quantify the severity of the fault and help pinpoint the type of fault.
[0052] S3: Perform adaptive feature selection on the physical features extracted in step S2 to filter out a subset of features that are fault-sensitive and domain-invariant.
[0053] For example, the adaptive feature selection in step S3 adopts a feature importance evaluation method based on random forest regressors to select features with importance higher than a preset threshold to form a feature subset.
[0054] As one possible implementation, the preset threshold is the product of the average of all feature importance scores and the threshold coefficient, as shown below:
[0055]
[0056] in For feature importance scores, This is the threshold coefficient.
[0057] S4: Input the feature subset into the integrated regression model and output the severity assessment results of gear failure.
[0058] The ensemble regression model in step S4 is a histogram-based gradient boosting regressor, which learns a non-linear mapping from features to fault severity by minimizing the squared loss function.
[0059] To evaluate the gear fault severity assessment method based on multimodal features and domain adaptation proposed in this invention, the embodiments of this application employ the coefficient of determination (R²). 2 The coefficient of determination (COD) and root mean square error (RMSE) are used as core quantitative indicators. The COD measures the model's ability to explain changes in fault severity; a value closer to 1 indicates a higher goodness of fit. RMSE quantifies the absolute deviation between predicted and actual values; a smaller value indicates higher prediction accuracy. These two indicators together ensure a comprehensive assessment of model performance.
[0060] To comprehensively verify the basic performance and anti-interference capability of this invention in a simulation environment, this application conducted system tests on a self-constructed high-fidelity multi-sensor gear fault simulation dataset. This dataset includes successive fault severity levels and introduces random fluctuations in speed and load to simulate uncertainties in real-world operating conditions. The tests focused on evaluating the robustness of the method under varying speeds and various noise interference scenarios.
[0061] For example, such as Figure 2 As shown, in the cross-speed task (600 RPM→1500 RPM±15%), the method of this invention achieved R... 2 With an excellent performance of 0.9412, it significantly outperforms the comparative method (CNN-CLS-Reg) of 0.8106, achieving a performance improvement of 17.4%, which fully demonstrates the effectiveness of the constant rotational speed feature of the order ratio analysis in overcoming the distribution shift caused by large-scale rotational speed changes.
[0062] In the anti-interference capability test, the method provided in this application exhibits stronger robustness under interference of varying intensities. Particularly in the extreme case of 10% data loss, the R0 of the method provided in this application is significantly improved. 2The performance remains at 0.8584, while the comparative method drops to 0.7627. Under interference conditions such as Gaussian noise (10%), sensor drift (5%), and temperature characteristic drift (20%), the performance degradation of the method provided in this application is much smaller than that of the comparative method.
[0063] For example, to verify the effectiveness of the method provided in this application, it is compared with several representative methods on the public dataset UM-GearEccDataset. These representative methods include: CNN-CLS-Reg (a classification-regression method based on a one-dimensional convolutional neural network), LGB-CLS-Reg (a classification-regression method based on LightGBM), and TL-CNN-Reg (a transfer learning regression method based on a convolutional neural network). See [reference]. Figure 3 The method provided in this application achieves state-of-the-art performance in both cross-domain tasks: In task 1, R... 2 The RSI reached 0.9167, an improvement of 8.3% compared to the comparison method; in Task 2, the most challenging scenario, RSI... 2 The result still reaches 0.7925, a significant improvement of 16.8% compared to the comparison method. This indicates that the method provided in this application effectively overcomes the domain offset problem by extracting rotational speed invariant features through order ratio analysis and combining multi-sensor physical feature fusion and adaptive feature selection, demonstrating excellent cross-condition generalization ability.
[0064] The above description, in conjunction with preferred technical solutions, provides a further detailed explanation of the present invention and should not be construed as limiting the specific implementation of the invention to these descriptions. For those skilled in the art, simple deductions and substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A method for assessing gear fault severity based on multimodal features and domain adaptation, characterized in that, The process includes: S1: Preprocessing the multi-sensor time-series data of the gear system to obtain segmented and aligned multimodal data; S2: Extracting physical features insensitive to speed changes from the preprocessed multi-sensor data; wherein, the physical features insensitive to speed changes include vibration features, acoustic features, current features, and torque features; the acoustic features include the statistics of Mel-frequency coefficients and the spectral centroid; the current features include the total harmonic distortion rate and the amplitude ratio of characteristic harmonics to the fundamental frequency; the torque features include the fluctuation intensity and the dominant frequency of the fluctuation; the vibration features include traditional time-domain and frequency-domain statistical features, gear system-specific frequency-domain features, and order analysis features; wherein, the traditional time-domain and frequency-domain statistical features include mean, standard deviation, RMS, kurtosis, skewness, peak factor, impulse factor, and margin factor; the calculation formula for the gear system-specific frequency-domain features is as follows: in, The meshing harmonic ratio; The sideband intensity ratio is used. In step S2, the vibration signal is resampled at an angle and subjected to a fast Fourier transform to obtain the order ratio spectrum. Based on the order ratio spectrum, the energy proportion near multiple characteristic orders is calculated as the order ratio analysis feature. The calculation formula for the order ratio analysis feature is as follows: S3: Perform adaptive feature selection on the physical features extracted in step S2 to filter out a subset of features that are fault-sensitive and domain-invariant; S4: Input the subset of features into the integrated regression model and output the severity assessment result of the gear fault.
2. The gear fault severity assessment method based on multimodal features and domain adaptation according to claim 1, characterized in that, The preprocessing in step S1 includes: cleaning, outlier handling, and standardizing the time-series data collected by the vibration sensor, acoustic sensor, current sensor, and torque sensor respectively; segmenting the processed sensor data into segments with fixed lengths and overlap rates, and aligning them at the segment level.
3. The gear fault severity assessment method based on multimodal features and domain adaptation according to claim 2, characterized in that, In step S1, the signals of each sensor channel are... According to a fixed length and overlap rate Segmentation, step size The first segment after segmentation The segment is: Where M is the total number of segments, .
4. The gear fault severity assessment method based on multimodal features and domain adaptation according to claim 1, characterized in that, The adaptive feature selection in step S3 adopts a feature importance evaluation method based on random forest regressors to select features with importance higher than a preset threshold, which constitute the feature subset.
5. The gear fault severity assessment method based on multimodal features and domain adaptation according to claim 4, characterized in that, The preset threshold is the product of the average of all feature importance scores and the threshold coefficient.
6. The gear fault severity assessment method based on multimodal features and domain adaptation according to claim 1, characterized in that, The ensemble regression model in step S4 is a histogram-based gradient boosting regressor that learns a nonlinear mapping from features to fault severity by minimizing the squared loss function.