Gearbox fault prediction and diagnosis method

Abstract

The invention provides a gearbox fault prediction and diagnosis method, which comprises the following steps of: acquiring sensor data of a preset point location of a gearbox, storing the sensor data according to a standardized format, and filtering interference noise such as strong electromagnetism and vibration of a wind power plant by customizing parameters of a Butterworth filter; inputting the standardized data set into a multi-branch parallel convolutional neural network model to extract feature sub-segments of each preset point location, and splicing global high-dimensional feature vectors; point feature weights are dynamically adjusted through point and segment training, fault high-incidence parts are focused, irrelevant data interference is reduced, fault types and severity scores are output in combination with an adaptive parameter random forest model, and the problems that a traditional model is poor in generalization ability and cannot adapt to gearboxes of different models / working conditions are solved. High-precision recognition and high-time-efficiency early warning across early faults of the gearbox are achieved, and the problem of gearbox fault monitoring and early warning lagging is solved.

Description

Gearbox Fault Prediction and Diagnosis Methods Technical Field

[0001] This invention relates to the field of wind turbine gearbox fault monitoring technology, and in particular to a method for predicting and diagnosing gearbox faults. Background Technology

[0002] The gearbox system of a wind turbine generator mainly consists of gears / gear shafts, bearings, locking plates, lubrication and cooling systems, heaters, air filters, and other components. With the continuous growth in the volume and installed capacity of wind power equipment, the stability and reliability of wind turbine generators, as high-load, continuously operating equipment, are receiving increasing attention from the industry.

[0003] The wind turbine gearbox, as the transmission component connecting the wind turbine and the generator, directly affects the wind turbine's energy efficiency and safety level. Because the gearbox operates under complex loads and high-speed meshing conditions for extended periods, it is highly susceptible to typical failure types such as tooth surface wear, bearing damage, abnormal lubrication, and oil contamination. A serious gearbox failure not only leads to unit shutdown and maintenance but can also cause significant economic losses and energy waste.

[0004] In existing technologies, wind turbine gearbox fault diagnosis methods mainly rely on vibration signal analysis, with some supplementary judgment based on oil analysis data. However, in vibration monitoring, early fault signals are weak and easily masked, signal processing is difficult, and the models have poor generalization ability and low interpretability. In oil analysis, fault location is inaccurate, the detection cycle is long and costly, and the results are greatly affected by sampling, resulting in low efficiency in gearbox fault monitoring and a delayed early warning system. Summary of the Invention

[0005] This invention provides a method for predicting and diagnosing gearbox faults, which solves the problem of delayed early warning in gearbox fault monitoring.

[0006] On one hand, this invention provides a method for predicting and diagnosing gearbox faults, comprising: collecting sensor data from multiple preset points in the gearbox and storing the sensor data in a preset standardized format to form an original dataset; matching the Butterworth filter parameters corresponding to the preset points and performing Butterworth filtering on the original dataset to obtain a standardized dataset; inputting the standardized dataset into a multi-branch parallel convolutional neural network model, outputting feature segments corresponding to the preset points, and concatenating the feature segments into a global high-dimensional feature vector; training the global high-dimensional feature vector by segmenting and dynamically adjusting the weights of the feature segments corresponding to the preset points to generate a weighted feature vector; inputting the weighted feature vector into a random forest model with adaptive parameters, and outputting a diagnostic result including fault type and severity score.

[0007] Optionally, sensor data from multiple preset points on the gearbox are collected, and the sensor data is stored in a preset standardized format to form an original dataset. This includes: collecting sensor data from multiple preset points located at different structural parts of the gearbox; the sensor data includes acceleration signals and oil state signals; binding the collected sensor data with corresponding point identification information and collection time information to obtain a structured data package; and splicing the structured data according to a time sequence to form the original dataset.

[0008] Optionally, matching the Butterworth filter parameters corresponding to the preset points and performing Butterworth filtering on the original dataset to obtain a standardized dataset includes: matching preset Butterworth filter parameters for the preset points according to the signal type and characteristic frequency range of the preset points; the Butterworth filter parameters include at least filter type, order, and cutoff frequency; applying the matched Butterworth filter parameters to the sensor data of the preset points to perform Butterworth filtering processing to obtain denoised data for each preset point; and normalizing the denoised data to form a standardized dataset.

[0009] Optionally, after matching the preset Butterworth filter parameters to the preset points based on the signal type and characteristic frequency range of the preset points, the method further includes: monitoring the signal noise amplitude of the original dataset; if the signal noise amplitude exceeds a preset threshold, adjusting the cutoff frequency of the Butterworth filter corresponding to the preset points by a preset step size.

[0010] Optionally, inputting the standardized dataset into a multi-branch parallel convolutional neural network model, outputting feature segments corresponding to the preset points, and concatenating the feature segments into a global high-dimensional feature vector includes: inputting the standardized data of the preset points into the corresponding convolutional neural network branches respectively; extracting local features of the standardized data through the convolutional layers of the convolutional neural network branches; reducing the dimensionality of the local features through the pooling layers of the convolutional neural network branches to obtain dimensionality-reduced features; converting the dimensionality-reduced features into a one-dimensional feature vector through the fully connected layers of the convolutional neural network branches to obtain the feature segments corresponding to the preset points; and concatenating the feature segments to obtain a global high-dimensional feature vector.

[0011] Optionally, the global high-dimensional feature vector is trained by segmenting it into points, and a weighted feature vector is generated by dynamically adjusting the weights of the feature segments corresponding to the preset points. This includes: splitting the global high-dimensional feature vector into multiple feature segments according to the number of preset points, with each feature segment corresponding to a preset point; assigning initial weights to the feature segments based on the fault contribution rate of the preset points in historical fault data; multiplying the feature segments by the initial weights and inputting the result into a random forest model to calculate the contribution of the feature segments to the fault diagnosis accuracy; adjusting the initial weights according to the contribution to obtain corrected weights; and combining the feature segments multiplied by the corrected weights to generate a weighted feature vector.

[0012] Optionally, adjusting the initial weights based on the contribution to obtain corrected weights includes: setting a step size range for weight adjustment; adjusting the weights of the corresponding feature segments within the step size range based on the contribution trend of the feature segments; adding a preset weight to the feature segment when the contribution of the feature segment ranks first in multiple consecutive training rounds; and stopping the weight adjustment process when the variance of the contribution of the feature segment is less than the preset variance.

[0013] Optionally, inputting the weighted feature vector into a random forest model with adaptive parameters and outputting a diagnostic result including fault type and severity score includes: configuring corresponding adaptive parameters for the random forest model based on the model attributes and operating condition parameters of the target gearbox, wherein the adaptive parameters include the number of decision trees, tree depth, and number of samples for node splits; inputting the weighted feature vector into the configured random forest model; voting on the weighted feature vector through multiple decision trees in the random forest model, and outputting the fault type classification result by majority vote; predicting the weighted feature vector through multiple regression trees in the random forest model, and outputting the severity score by the average of the regression results; and combining the fault type classification result with the severity score to form a diagnostic result.

[0014] Optionally, after outputting the diagnostic results including the fault type and severity score, the method further includes: generating early warning signals of different levels based on the severity score; and generating a maintenance work order containing fault information and handling suggestions based on the diagnostic results.

[0015] Optionally, it further includes: collecting monitoring data of the gearbox and the corresponding diagnostic results of the gearbox to form a verification dataset; inputting the verification dataset into the multi-branch parallel convolutional neural network model and the adaptive parameter random forest model to generate a diagnostic accuracy; if the diagnostic accuracy is lower than a preset threshold, then using a training set containing new data to adjust the parameters of the multi-branch parallel convolutional neural network model and the adaptive parameter random forest model.

[0016] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the gearbox fault prediction and diagnosis method as described above.

[0017] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the gearbox fault prediction and diagnosis method as described above.

[0018] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the gearbox fault prediction and diagnosis method as described above.

[0019] The gearbox fault prediction and diagnosis method provided by this invention collects sensor data from preset gearbox locations and stores it in a standardized format. After filtering the data using a customized Butterworth filter parameter to remove interference noise such as strong electromagnetic fields and vibrations from wind farms, the standardized dataset is input into a multi-branch parallel convolutional neural network model to extract feature segments from each preset location and concatenate them into a global high-dimensional feature vector. By dynamically adjusting the feature weights of each location through segmented training, the method focuses on high-incidence fault locations and reduces interference from irrelevant data. Combined with an adaptive parameter random forest model, the method outputs fault type and severity scores, solving the problems of weak generalization ability and inability to adapt to different gearbox models / operating conditions in traditional models. This enables high-precision identification and timely early warning of early faults across gearboxes, thus addressing the problem of delayed gearbox fault monitoring and early warning. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 is a schematic flowchart of the gearbox fault prediction and diagnosis method provided in an embodiment of the present invention; Figure 2 is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 is a schematic flowchart of the gearbox fault prediction and diagnosis method provided in an embodiment of the present invention.

[0024] As shown in Figure 1, the gearbox fault prediction and diagnosis method provided in this embodiment of the invention mainly includes the following steps: 101. Collect sensor data from multiple preset points of the gearbox and store the sensor data in a preset standardized format to form an original dataset.

[0025] The preset monitoring points are located at different structural parts of the gearbox. The sensors integrate both accelerometers and oil condition sensors; the accelerometers capture the impact characteristics in the vibration signals, while the oil condition sensors monitor oil temperature, viscosity, and particulate contamination levels in real time. During data acquisition, each sensor data packet is appended with a timestamp and a location identifier to ensure the spatiotemporal correlation of the data; the acquisition frequency is dynamically adjusted according to the gearbox rotational speed.

[0026] Specifically, sensor data from multiple preset points on the gearbox are collected and stored in a preset standardized format to form a raw dataset. This includes: collecting sensor data from multiple preset points located at different structural parts of the gearbox; binding the collected sensor data with corresponding point identification information and collection time information to obtain a structured data package; and splicing the structured data according to a time series to form the raw dataset.

[0027] This process involves collecting sensor data from multiple preset points located at different structural parts of the gearbox. The sensor data includes acceleration signals reflecting the vibration state of the components and oil state signals characterizing internal wear and lubrication. The collected sensor data is then bound to the corresponding point identification information and collection time information to form a traceable structured data package. Finally, all the structured data packages are sequentially spliced ​​together according to the time sequence to construct a well-organized, clearly sourced, and time-series complete original dataset. This ensures data traceability and provides a standardized data foundation for feature extraction.

[0028] For example, five pre-defined points are marked on the gearbox, with a sensor group deployed at each point to ensure independent and traceable data collection. These five points are: the high-speed shaft bearing housing, the intermediate-speed shaft bearing housing, the low-speed shaft bearing housing, the gear meshing area, and the oil tank return port. Each point is labeled with its corresponding sensor type. Specifically, a 3-axis accelerometer is deployed on the high-speed shaft bearing housing, with a measurement range of ±50g, accuracy of ±0.5%FS, and a sampling frequency of 20kHz, to capture fault signals such as pitting and spalling in the high-speed shaft bearing. A 3-axis accelerometer is deployed on the intermediate-speed shaft bearing housing, with a measurement range of ±50g, accuracy of ±0.5%FS, and a sampling frequency of 15kHz, focusing on the fault characteristics of the intermediate-speed shaft bearing. A 3-axis accelerometer is deployed on the low-speed shaft bearing housing, with a measurement range of ±50g, accuracy of ±0.5%FS, and a sampling frequency of 10kHz, to monitor the operating status of the low-speed shaft bearing. The system simultaneously deploys high-frequency vibration and temperature sensors in the gear meshing area. The vibration sensor has a measurement range of ±20g, an accuracy of ±0.3%FS, and a sampling frequency of 25kHz. The temperature sensor has a measurement range of -40℃ to 120℃ and an accuracy of ±0.5℃, respectively capturing gear meshing faults and temperature anomalies. An integrated oil sensor is deployed at the oil tank return port to monitor the concentration of 2μm-100μm metal particles and the viscosity of 2cSt-1000cSt (accuracy ±2%), collecting data once every 5 minutes to reflect the oil contamination and component wear status.

[0029] 102. Match the Butterworth filter parameters corresponding to the preset points, perform Butterworth filtering on the original dataset, and obtain a standardized dataset.

[0030] By using Butterworth filtering to eliminate noise interference and dimensional differences in the original dataset, effective fault characteristics are preserved to the maximum extent. Matching the Butterworth filter parameters corresponding to the preset points can be adjusted according to the signal characteristics and fault diagnosis requirements of different preset points. This ensures that the data preprocessing process at each preset point can effectively suppress environmental noise and electromagnetic interference while fully preserving fault characteristic information.

[0031] Specifically, the Butterworth filter parameters corresponding to the preset points are matched, and the original dataset is subjected to Butterworth filtering to obtain a standardized dataset. This includes: matching the preset Butterworth filter parameters for the preset points according to the signal type and characteristic frequency range of the preset points; the Butterworth filter parameters include at least the filter type, order, and cutoff frequency; applying the matched Butterworth filter parameters to the sensor data of the preset points to perform Butterworth filtering processing to obtain denoised data for each preset point; and normalizing the denoised data to form a standardized dataset.

[0032] The sensor signals at different preset points vary significantly due to differences in component characteristics and environmental interference. For example, the acceleration signal of the high-speed shaft bearing housing needs to focus on the bearing rolling characteristic frequency of 80Hz-6kHz, the high-frequency vibration signal of the gear meshing area needs to retain the tooth surface impact frequency of 400Hz-11kHz, while the oil status signal of the oil tank return port is mainly low-frequency below 50Hz. Based on this, the present invention first matches preset Butterworth filter parameters according to the signal type and characteristic frequency range of each preset point.

[0033] For example, high-speed shaft vibration is filtered using a 4th-order bandpass filter with a low cutoff frequency of 100Hz and a high cutoff frequency of 5kHz, preserving the characteristic frequencies of high-speed shaft bearing failure; medium-speed shaft vibration is filtered using a 4th-order bandpass filter with a low cutoff frequency of 50Hz and a high cutoff frequency of 3kHz, focusing on the characteristics of medium-speed shaft bearing failure; low-speed shaft vibration is filtered using a 4th-order bandpass filter with a low cutoff frequency of 20Hz and a high cutoff frequency of 1kHz, preserving the characteristics of low-speed shaft bearing failure; gear vibration is filtered using a 5th-order bandpass filter with a low cutoff frequency of 500Hz and a high cutoff frequency of 10kHz, capturing gear meshing frequencies and sideband characteristics; gear temperature is filtered using a 3rd-order low-pass filter with a cutoff frequency of 0.1Hz, filtering out instantaneous temperature fluctuations while preserving temperature change trends; oil data is filtered using a 3rd-order low-pass filter with a cutoff frequency of 0.05Hz, eliminating fluctuations caused by sampling errors while preserving concentration / viscosity change trends.

[0034] After matching the preset Butterworth filter parameters, the raw data of each preset point are filtered by applying the Butterworth filter parameters to effectively filter out irrelevant signals such as wind farm electromagnetic interference and box resonance noise, and obtain denoised data that retains the core fault characteristics; finally, through normalization processing, the difference in magnitude of different physical quantities is eliminated, forming a standardized dataset that can be directly input into the model.

[0035] In some embodiments, after matching the preset Butterworth filter parameters to the preset points according to the signal type and characteristic frequency range of the preset points, the method further includes: monitoring the signal noise amplitude of the original dataset; if the signal noise amplitude exceeds a preset threshold, adjusting the cutoff frequency of the corresponding preset point by a preset step size.

[0036] The preset step size can be set according to actual application scenarios and experience. For example, it can be set to a 10Hz adjustment step size for low cutoff frequencies and a 100Hz adjustment step size for high cutoff frequencies. When the signal and noise amplitude of the original dataset exceeds the preset threshold, the Butterworth filter cutoff frequency at the corresponding preset point is automatically adjusted according to the preset step size. If the signal and noise amplitude still exceeds the preset threshold after adjustment, the adjustment continues according to the preset step size until the signal and noise amplitude is within the preset threshold range or the maximum number of adjustments is reached. The maximum number of adjustments can also be preset according to actual conditions to avoid over-adjustment leading to poor filtering performance.

[0037] By adjusting the cutoff frequency of the Butterworth filter, the interference of signal noise on gearbox fault monitoring can be effectively reduced, thereby improving the accuracy and reliability of monitoring data.

[0038] 103. Input a standardized dataset to a multi-branch parallel convolutional neural network model, output the feature segments corresponding to preset points, and concatenate the feature segments into a global high-dimensional feature vector.

[0039] The multi-branch parallel convolutional neural network model includes several branches, each corresponding to a predefined power grid. The convolutional neural network structure within each branch is uniform, ensuring consistent feature dimensions. This multi-branch parallel approach allows for feature mining of standardized datasets from different perspectives, yielding richer feature information. Concatenating the feature segments output from each branch into a global high-dimensional feature vector integrates the features extracted from different branches, forming a comprehensive feature representation.

[0040] Specifically, the process involves inputting a standardized dataset into a multi-branch parallel convolutional neural network model, outputting feature segments corresponding to preset points, and concatenating these feature segments into a global high-dimensional feature vector. This includes: inputting the standardized data at preset points into the corresponding branches of the convolutional neural network, and extracting local features of the standardized data through the convolutional layers of the convolutional neural network branches.

[0041] The standardized data from each preset point in the gearbox are input into the corresponding convolutional branches of a multi-branch parallel convolutional neural network. The kernel parameters of different branches are customized according to the characteristics of the point signals. For example, for the vibration signal of the high-speed shaft bearing housing, 32 3×3 convolutional kernels are configured to focus on extracting the peak and valley changes of the vibration waveform in the time domain; for the high-frequency vibration signal of the gear meshing area, 64 3×3 convolutional kernels are configured to mine deeper coupling features, such as the coordinated changes of multi-axis vibration. Through point-adaptive convolution operations, the unique early weak features of faults in each part can be accurately separated and captured, avoiding feature confusion between signals from different points.

[0042] The dimensionality of local features is reduced by using the pooling layer of the convolutional neural network branch, resulting in dimensionality-reduced features.

[0043] In each branch of the convolutional neural network, the local features extracted by the convolutional layers contain a large amount of redundant information, such as repetitive normal vibration patterns and pseudo-features with residual noise. These need to be reduced in dimensionality using pooling layers. Specifically, max pooling with a 2×2 pooling window is used for the local features of the high-speed shaft bearing housing to retain peak information in the impact features; average pooling is used for the local features of the gear meshing area to smooth random noise in high-frequency vibrations and highlight stable meshing frequency features; and adaptive pooling is used for the local features of the oil state signal, dynamically adjusting the pooling window size according to the density of particle concentration fluctuations. Through targeted dimensionality reduction, while reducing data dimensionality and model computation, the core features crucial for fault diagnosis can be preserved to the maximum extent, avoiding the loss of effective information.

[0044] The dimensionality-reduced features are transformed into one-dimensional feature vectors through fully connected layers in a convolutional neural network branch, yielding feature segments corresponding to preset points. The features after dimensionality reduction via pooling layers are still multi-dimensional matrices, requiring further transformation into one-dimensional feature vectors by the fully connected layers of the convolutional neural network branch to form feature segments corresponding to each preset point. The fully connected layers perform non-linear mapping of the dimensionality-reduced features using weight matrices: for example, the dimensionality-reduced features of a high-speed shaft bearing housing are mapped to a 128-dimensional vector, with each dimension corresponding to a quantitative indicator of bearing failure; the dimensionality-reduced features of the gear meshing area are mapped to a 128-dimensional vector, covering feature parameters of faults such as tooth surface spalling and tooth root cracks; the dimensionality-reduced features of the oil state are mapped to a 128-dimensional vector, characterizing states such as lubrication failure and component wear rate. This condenses the complex features of each point into a fixed-length one-dimensional vector, achieving both quantitative integration of features and providing a unified structure for subsequent cross-point feature concatenation.

[0045] By concatenating the feature segments, a global high-dimensional feature vector is obtained.

[0046] This process involves concatenating the one-dimensional feature segments corresponding to each preset point in sequence to form a global high-dimensional feature vector. For example, a gearbox monitoring system with five points can generate a 640-dimensional global vector after concatenating its feature segments. This vector encompasses the vibration characteristics of high-speed shaft bearings, the high-frequency characteristics of gear meshing, and the wear characteristics of the oil. By concatenating feature segments to obtain the global high-dimensional feature vector, the limitations of single-point monitoring are overcome: when a complex fault occurs in the gearbox, the global vector can simultaneously integrate the impact characteristics of bearing vibration and the abnormal frequency characteristics of gear meshing, avoiding missed fault detection due to reliance on a single signal.

[0047] In addition, when training the multi-branch parallel convolutional neural network model, the historical data of three different gearbox models were used as the basis, and the data were divided into training set, validation set and test set in a ratio of 7:2:1. The data included four categories of labels: normal, tooth surface wear, bearing pitting, and poor lubrication.

[0048] The cross-entropy loss function and Adam optimizer are used, with an initial learning rate of 0.001, which decays to 0.8 times the previous learning rate every 5 epochs.

[0049] Training should be stopped when the accuracy on the validation set improves by less than 0.5% for three consecutive epochs to avoid overfitting.

[0050] By visualizing the output of the convolutional layers, we can verify that the multi-branch parallel convolutional neural network model can effectively capture key features such as bearing pitting vibration pulses and gear wear meshing frequency fluctuations.

[0051] 104. The global high-dimensional feature vector is trained by dividing it into points and segments. By dynamically adjusting the weights of the feature segments corresponding to the preset points, a weighted feature vector is generated.

[0052] The weighted feature vector enhances the ability of features to represent complex faults by fusing fault sensitivity information from different points. For example, for the coexistence of bearing pitting and gear wear, the weight coefficient of the vibration and shock feature segment can be automatically increased, while the weight of the normal meshing frequency segment can be reduced, making the distribution of the weighted vector in Euclidean space closer to the actual fault mode.

[0053] Specifically, the global high-dimensional feature vector is trained by dividing it into points and segments. By dynamically adjusting the weights of the feature segments corresponding to the preset points, a weighted feature vector is generated. This includes splitting the global high-dimensional feature vector into multiple feature segments according to the number of preset points.

[0054] In this system, each feature segment corresponds to a preset point. Based on the number of preset points in the gearbox, the global high-dimensional feature vector is split according to the dimensional boundaries of the feature segments at each point, resulting in feature segments that correspond one-to-one with the preset points. For example, if the global vector is 640-dimensional, the 640-dimensional feature is split into five feature segments: point A (1-128 dimensions), point B (129-256 dimensions), point C (257-384 dimensions), point D (385-512 dimensions), and point E (513-640 dimensions). The dimension of each feature segment is consistent with the feature segment of the corresponding preset point, ensuring that each feature segment independently carries the fault characteristics of a single preset point.

[0055] Initial weights are assigned to feature segments based on the fault contribution rate of preset points in historical fault data.

[0056] The process involves analyzing the historical fault database of the gearbox to determine the fault contribution rate of each preset point—the proportion of fault occurrences at a given point relative to the total number of faults. Initial weights for feature segments are then assigned based on this contribution rate. For example, historical data shows that faults in the gear meshing area due to tooth wear and scuffing account for 42% of total faults, and high-speed shaft bearing faults account for 28%. Therefore, the initial weights for the corresponding feature segments are set to 0.42 and 0.28, respectively. Meanwhile, the oil tank return port, which has a relatively low fault contribution rate (15%), has an initial weight of 0.15. By assigning initial weights, ineffective attention to low-contribution points is avoided, thus improving training efficiency.

[0057] The feature fragments are multiplied by the initial weights and then input into the random forest model to calculate the contribution of the feature fragments to the fault diagnosis accuracy.

[0058] The process involves training a random forest model with weighted feature input adaptive parameters by multiplying each feature segment by its corresponding initial weight. The contribution of each feature segment is then inferred from the diagnostic accuracy output by the random forest model. For example, after removing the feature segment from the gear meshing area, the random forest model's accuracy in identifying tooth surface spalling faults dropped from 95% to 78%, indicating that the feature segment from the gear meshing area contributed 17%, significantly higher than that of the high-speed shaft segment. By calculating the contribution of each feature segment to the fault diagnosis accuracy, the actual impact of each feature point on the diagnostic results is quantified, providing an objective basis for weight adjustment.

[0059] The initial weights are adjusted based on the contribution to obtain the corrected weights; the feature segments are multiplied by the corrected weights and then combined to generate a weighted feature vector.

[0060] In adjusting weights, if the contribution of a feature segment is more than 10% higher than the expected contribution corresponding to the initial weight, the weight is increased by a step of 0.05; if it is more than 10% lower, the weight is decreased by a step of 0.05, with each adjustment not exceeding ±0.1 to avoid drastic weight fluctuations. For example, if the actual contribution of the gear meshing area segment is 17%, lower than the expected 42%, the weight is decreased to 0.37; if the contribution of the high-speed shaft segment is 5%, close to the expected 28%, the difference is 23%, exceeding the threshold, and the weight is decreased to 0.23. If a segment ranks first in contribution for three consecutive times, an additional weight of 0.02 is added as a reward, ultimately forming a adjusted weight that takes into account both historical data and actual contribution.

[0061] Finally, the feature sub-vectors are concatenated in the original positional order to form the final weighted feature vector. For example, after multiplying the gear meshing area feature segment by a correction weight of 0.37, the proportion of gear meshing area features in the vector is enhanced; while the proportion of low-speed shaft segments with low contribution is weakened after multiplying by 0.12. The weighted combination highlights the features of fault-sensitive points in the feature segments, while suppressing irrelevant or low-contribution features, improving the sensitivity of early and subtle fault identification, and reducing diagnostic bias caused by feature equalization.

[0062] The process of adjusting the initial weights based on their contribution to obtain the corrected weights includes: setting a step size range for weight adjustment; adjusting the weights of corresponding feature segments within the step size range based on the trend of contribution changes of feature segments; adding a preset weight to a feature segment when its contribution ranks first in multiple consecutive training rounds; and stopping the weight adjustment process when the variance of the contribution of a feature segment is less than the preset variance.

[0063] When setting the step size range for weight adjustment, it is necessary to comprehensively consider the fluctuation of each feature segment in the historical fault data and the accuracy requirements of actual fault diagnosis. If the feature segments in the historical data fluctuate greatly and the diagnostic accuracy requirements are high, a larger step size range should be set; otherwise, a smaller step size range should be set.

[0064] When adjusting weights based on the trend of contribution changes of feature segments, if the contribution is on the rise, the weights should be increased appropriately within the step size range; if the contribution is on the fall, the weights should be decreased appropriately within the step size range.

[0065] When adding a preset weight to a feature segment that ranks first in contribution for multiple consecutive rounds, the magnitude of the preset weight can be determined based on the importance of the feature segment in the overall fault diagnosis. The higher the importance, the greater the added preset weight.

[0066] By adjusting the weights, the actual impact of feature segments on fault diagnosis can be more accurately reflected, avoiding diagnostic biases caused by unreasonable initial weight allocation. For example, in gearbox composite fault diagnosis, if the coupling effect of bearing pitting and gear wear is not fully considered during initial weight allocation, fault features will be weakened. Dynamically adjusting the weights automatically enhances the weights of feature segments that contribute significantly to fault diagnosis while reducing the weights of noisy or redundant features, making the weighted feature vector more discriminative in fault pattern recognition. Furthermore, stopping adjustment when the variance of the feature segment's contribution is less than a preset variance prevents overfitting to historical data from reducing generalization ability, ensuring that the weighted feature vector stably represents fault features under different operating conditions.

[0067] The final weighted feature vector not only retains the integrity of the global high-dimensional feature vector, but also highlights fault-sensitive features through weight optimization.

[0068] 105. Input the weighted feature vector into a random forest model with adaptive parameters, and output the diagnostic results including fault type and severity score.

[0069] The random forest model is constructed by integrating multiple decision trees. Each tree independently votes on the weighted feature vectors for classification, and the fault type is determined by combining the voting results of all trees. In addition, parameters such as the number of decision trees and tree depth are customized for different preset locations and operating conditions to improve the generalization ability of the random forest model across devices and scenarios.

[0070] Specifically, the weighted feature vector is input into the random forest model with adaptive parameters, and the output includes diagnostic results containing fault type and severity score. This includes configuring corresponding adaptive parameters for the random forest model based on the model attributes and operating condition parameters of the target gearbox. The adaptive parameters include the number of decision trees, tree depth, and number of samples for node splits.

[0071] The weighted feature vector is input into a random forest model with configured parameters; the weighted feature vector is voted on by multiple decision trees in the random forest model, and the fault type classification result is output by majority vote; the weighted feature vector is predicted by multiple regression trees in the random forest model, and the severity score is output by the average of the regression results; the fault type classification result and the severity score are combined to form the diagnostic result.

[0072] In practical applications, for gearboxes of different models and operating conditions, adaptive parameters are first configured for the random forest model. For example, for a 1.5MW gearbox under high load, since the failure modes are concentrated in the gear meshing area, 150 decision trees with a tree depth of 12 and 20 split samples per node are configured. For a 3MW gearbox under low load, since the bearing failure rate is higher, 200 decision trees with a tree depth of 15 and 25 split samples per node are configured to adapt to different failure feature distributions.

[0073] The weighted feature vector is then input into a random forest model. Multiple decision trees in the random forest model analyze the feature vector. If 210 out of 300 decision trees determine that the gear tooth is scuffed, the fault type is output according to the majority voting principle. At the same time, multiple regression trees predict the severity based on the wear degree feature in the feature vector. The average of 2.4, which is 2.3, 2.5, and 2.4, output by 50 regression trees, is taken as the severity score. Finally, the results are combined to form the diagnosis result of fault type gear tooth scuffed and severity score of 2.4, realizing accurate classification and quantitative assessment of gearbox faults in different scenarios.

[0074] In some embodiments, after outputting diagnostic results including fault type and severity score, the method further includes: generating warning signals of different levels based on the severity score; and generating a maintenance work order containing fault information and handling suggestions based on the diagnostic results.

[0075] When generating warning signals of different levels, multiple severity score ranges can be preset, with each range corresponding to a warning level. For example, the severity score can be divided into three ranges: low, medium, and high, corresponding to green, yellow, and red warning signals, respectively. When the output severity score falls into a certain range, a warning signal of the corresponding level is automatically generated.

[0076] When generating a maintenance work order containing fault information and handling suggestions, the fault type and severity score are extracted from the diagnostic results. The fault type and severity score are integrated into the maintenance work order. Then, based on the fault type, the corresponding handling suggestions are matched from the preset fault handling suggestion library and added to the maintenance work order to improve maintenance efficiency.

[0077] In some embodiments, the gearbox fault prediction and diagnosis method provided by the present invention further includes: collecting monitoring data of the gearbox and the corresponding diagnostic results of the gearbox to form a verification dataset; inputting the verification dataset into a multi-branch parallel convolutional neural network model and an adaptive parameter random forest model to generate a diagnostic accuracy; if the diagnostic accuracy is lower than a preset threshold, then using a training set containing new data to adjust the parameters of the multi-branch parallel convolutional neural network model and the adaptive parameter random forest model.

[0078] During the long-term operation of the gearbox, in order to cope with new fault modes or changes in operating conditions, monitoring data from various preset points of the gearbox and diagnostic results confirmed by manual inspection are continuously collected to form a verification dataset covering new scenarios. The verification dataset is then input into the deployed multi-branch parallel convolutional neural network model and adaptive parameter random forest model. By comparing the model output with the actual results, the diagnostic accuracy is calculated to determine whether the parameters of the multi-branch parallel convolutional neural network model and adaptive parameter random forest model need to be adjusted.

[0079] If the accuracy drops below the threshold, the training set including such new data will be re-input into the model, and the number of convolution kernels in the multi-branch parallel convolutional neural network model and the depth of the decision tree in the random forest will be adjusted so that the multi-branch parallel convolutional neural network model and the adaptive parameter random forest model can dynamically adapt to the changes in the gearbox operating state and maintain high diagnostic accuracy.

[0080] Figure 2 is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0081] As shown in Figure 2, the electronic device may include a processor 210, a communication interface 220, a memory 230, and a communication bus 240. The processor 210, communication interface 220, and memory 230 communicate with each other via the communication bus 240. The processor 210 can call logical instructions from the memory 230 to execute gearbox fault prediction and diagnosis methods.

[0082] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the gearbox fault prediction and diagnosis methods provided by the above methods.

[0084] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the gearbox fault prediction and diagnosis methods provided by the above methods.

[0085] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting and diagnosing gearbox faults, characterized in that, include: Sensor data from multiple preset points on the gearbox are collected and stored in a preset standardized format to form an original dataset. Match the Butterworth filter parameters corresponding to the preset points, perform Butterworth filtering on the original dataset to obtain a standardized dataset; input the standardized dataset into a multi-branch parallel convolutional neural network model, output the feature segments corresponding to the preset points, and concatenate the feature segments into a global high-dimensional feature vector; train the global high-dimensional feature vector by segmenting and dynamically adjusting the weights of the feature segments corresponding to the preset points to generate a weighted feature vector; input the weighted feature vector into a random forest model with adaptive parameters, and output a diagnostic result containing fault type and severity score.

2. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, The process involves collecting sensor data from multiple preset points on the gearbox and storing the sensor data in a preset standardized format to form a raw dataset. This includes: collecting sensor data from multiple preset points located at different structural parts of the gearbox; the sensor data includes acceleration signals and oil state signals; binding the collected sensor data with corresponding point identification information and collection time information to obtain a structured data package; and splicing the structured data according to a time sequence to form the raw dataset.

3. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, Matching the Butterworth filter parameters corresponding to the preset points and performing Butterworth filtering on the original dataset to obtain a standardized dataset includes: matching preset Butterworth filter parameters for the preset points according to the signal type and characteristic frequency range of the preset points; the Butterworth filter parameters include at least filter type, order, and cutoff frequency; applying the matched Butterworth filter parameters to the sensor data of the preset points to perform Butterworth filtering processing to obtain denoised data for each preset point; and normalizing the denoised data to form a standardized dataset.

4. The gearbox fault prediction and diagnosis method according to claim 3, characterized in that, Based on the signal type and characteristic frequency range of the preset point, after matching the preset Butterworth filter parameters for the preset point, the method further includes: monitoring the signal noise amplitude of the original dataset; if the signal noise amplitude exceeds a preset threshold, adjusting the cutoff frequency of the Butterworth filter corresponding to the preset point by a preset step size.

5. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, The process involves inputting the standardized dataset into a multi-branch parallel convolutional neural network model, outputting feature segments corresponding to the preset points, and concatenating the feature segments into a global high-dimensional feature vector. This includes: inputting the standardized data of the preset points into corresponding convolutional neural network branches; extracting local features of the standardized data through the convolutional layers of the convolutional neural network branches; reducing the dimensionality of the local features through the pooling layers of the convolutional neural network branches to obtain dimensionality-reduced features; converting the dimensionality-reduced features into a one-dimensional feature vector through the fully connected layers of the convolutional neural network branches to obtain the feature segments corresponding to the preset points; and concatenating the feature segments to obtain a global high-dimensional feature vector.

6. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, The global high-dimensional feature vector is trained by segmenting it into points. A weighted feature vector is generated by dynamically adjusting the weights of the feature segments corresponding to the preset points. This process includes: splitting the global high-dimensional feature vector into multiple feature segments based on the number of preset points, with each feature segment corresponding to a preset point; assigning initial weights to the feature segments based on the fault contribution rate of the preset points in historical fault data; multiplying the feature segments by the initial weights and inputting the result into a random forest model to calculate the contribution of the feature segments to the fault diagnosis accuracy; adjusting the initial weights based on the contribution to obtain corrected weights; and combining the feature segments multiplied by the corrected weights to generate a weighted feature vector.

7. The gearbox fault prediction and diagnosis method according to claim 6, characterized in that, Adjusting the initial weights based on the contribution to obtain corrected weights includes: setting a step size range for weight adjustment; adjusting the weights of the corresponding feature segments within the step size range based on the contribution trend of the feature segments; adding a preset weight to the feature segment when the contribution of the feature segment ranks first in multiple consecutive training rounds; and stopping the weight adjustment process when the variance of the contribution of the feature segment is less than the preset variance.

8. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, The process involves inputting the weighted feature vector into a random forest model with adaptive parameters, and outputting a diagnostic result containing fault type and severity score. This includes: configuring corresponding adaptive parameters for the random forest model based on the target gearbox's model attributes and operating condition parameters; the adaptive parameters including the number of decision trees, tree depth, and the number of samples for node splits; inputting the weighted feature vector into the configured random forest model; voting on the weighted feature vector using multiple decision trees in the random forest model, and outputting the fault type classification result using a majority vote; predicting the weighted feature vector using multiple regression trees in the random forest model, and outputting the severity score based on the average of the regression results; and combining the fault type classification result with the severity score to form a diagnostic result.

9. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, After outputting the diagnostic results including the fault type and severity score, the system further includes: generating early warning signals of different levels based on the severity score; and generating a maintenance work order containing fault information and handling suggestions based on the diagnostic results.

10. The gearbox fault prediction and diagnosis method according to claim 1, characterized in that, Also includes: The monitoring data of the gearbox and the corresponding diagnostic results of the gearbox are collected to form a verification dataset; the verification dataset is input into the multi-branch parallel convolutional neural network model and the adaptive parameter random forest model to generate a diagnostic accuracy; if the diagnostic accuracy is lower than a preset threshold, the parameters of the multi-branch parallel convolutional neural network model and the adaptive parameter random forest model are adjusted using a training set containing new data.

Images