Sample imbalance wind power gear box fault feature extraction method based on transfer learning

Abstract

The invention discloses a sample imbalance wind power gear box fault feature extraction method based on transfer learning, and relates to the technical field of transfer learning, and the method specifically comprises the following steps: collecting and aligning vibration, temperature and rotating speed signals, dividing stages according to a working condition change rate, and generating a multi-source signal set; the method comprises the following steps: performing frequency domain decomposition on a multi-source signal to extract frequency band features to generate a frequency band sparse feature set, judging a coupling frequency band based on sparsity and a working condition change rate to generate a frequency band associated information set, adjusting cross-domain mapping weight according to energy distribution to generate a feature extraction result, and performing weighted training in combination with the sparsity and a migration weight to generate an optimization result. According to the method, associated frequency bands are screened in combination with frequency band sparseness and energy concentration ratio through a multi-source signal alignment construction stage, cross-domain mapping weight is corrected according to energy distribution and kernel function distance, and feature stability is enhanced through weighted training of sparseness and migration weight under the condition that samples are rare.

Description

Technical Field

[0001] This invention relates to the field of transfer learning technology, and in particular to a method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning. Background Technology

[0002] Transfer learning technology involves using existing knowledge to improve the learning process for new tasks. The core issue in this field is how to solve new tasks, which may involve different data distributions or imbalanced samples, by transferring existing model parameters, structures, or knowledge. Transfer learning mainly includes two aspects: firstly, the algorithm design of transfer learning, which aims to improve the generalization ability of machine learning models through knowledge sharing between different tasks or domains; secondly, designing appropriate model structures and learning strategies for imbalanced data or tasks with few samples to enhance robustness in complex environments. Transfer learning is widely used in many fields such as speech recognition, image recognition, and natural language processing, especially when samples are insufficient, as it can improve learning performance by transferring pre-trained models.

[0003] Traditional transfer learning-based methods for extracting fault features from imbalanced wind turbine gearboxes involve using transfer learning techniques to address wind turbine gearbox fault diagnosis, extracting effective fault features even with imbalanced samples. This method primarily addresses the common imbalanced sample problem in wind turbine gearboxes during practical use, typically employing a transfer learning strategy to compensate for the lack of samples in the target domain by utilizing existing fault data and models. Specifically, the method selects appropriate source and target domain data for learning, transferring knowledge from the source domain to the target domain, thereby improving the accuracy of fault feature extraction. This process usually includes data preprocessing, feature selection, and model training, using transfer learning algorithms for model training and optimization to resolve the imbalanced sample problem.

[0004] Traditional transfer learning relies on limited vibration data and a single-domain mapping strategy when dealing with unbalanced wind turbine gearbox samples. It is difficult to divide the operating stages based on temperature and speed changes, the frequency domain features are not linked to the fluctuations in operating conditions, the coupling relationship between different frequency bands is difficult to show, the source domain experience is prone to shift in the target domain, the feature contribution lacks quantitative basis, the model update is lagging under multiple fault categories, and the feature extraction accuracy is prone to decrease when the proportion of fault samples is low. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning. The specific technical solution is as follows: A method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning includes the following steps: S1: Collect vibration, temperature and speed signal sequences of wind turbine gearbox and perform time alignment. Calculate the rate of change of operating conditions based on temperature and speed signals and divide the operating conditions into stages. Match the vibration signals of all stages with the corresponding operating conditions to generate a multi-source signal set. S2: Based on the multi-source signal set, perform frequency domain decomposition, analyze the frequency components of multiple signals and divide the frequency bands, calculate the energy concentration and sparsity index of each frequency band, extract key frequency domain features of multiple signals, and generate a frequency band sparse feature set. S3: Based on the sparsity index of the multi-band in the sparse feature set of the frequency band, compare the gradient with the operating condition change rate of the multi-source signal set to determine the coupling strength between frequency bands, filter the feature frequency bands coupled with operating condition fluctuations, and generate a frequency band association information set. S4: Based on the frequency band association information set, input the maximum mean difference migration model, calculate the kernel function distance according to the energy distribution of the same frequency band in the source domain and the target domain, adjust the feature mapping weight, analyze the feature contribution of vibration, temperature and speed signals and perform cross-domain correction, and generate the wind turbine gearbox fault feature extraction results. S5: Based on the wind turbine gearbox fault feature extraction results, the ratio of healthy samples to fault samples is calculated. When the proportion of fault samples does not reach the fault category proportion threshold, the sparsity index and migration weight of the corresponding frequency band are extracted, the loss function is weighted and adjusted and the gradient is updated. The model parameters are dynamically optimized and trained to generate the optimized wind turbine gearbox fault feature extraction results.

[0006] As a further embodiment of the present invention, the multi-source signal set includes vibration signal, temperature signal, rotational speed signal, operating condition change rate, and operating condition stage; the frequency band sparse feature set includes energy concentration, sparsity index, and key frequency domain features; the frequency band correlation information set includes the characteristic frequency bands of inter-band coupling strength and operating condition fluctuation coupling; the wind turbine gearbox fault feature extraction result includes frequency band mapping weight, vibration signal feature contribution, temperature signal feature contribution, and rotational speed signal feature contribution; and the optimized wind turbine gearbox fault feature extraction result includes a weighted adjusted loss function, an updated gradient direction, and dynamically optimized model parameters.

[0007] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect vibration, temperature and speed signal sequences of wind turbine gearbox in operation and align them in time. Synchronize the corresponding frames of multiple signal sequences according to the same time and map the time periods of vibration, temperature and speed signals to generate a time-aligned signal set. S102: Based on the time-aligned signal set, extract the temperature signal and the rotation speed signal and normalize them to calculate the signal difference value of adjacent sampling frames. Analyze the temperature change rate and the rotation speed change rate, and combine the operating condition change rate threshold to segment the rate sequence and generate the operating condition stage division result. S103: Based on the results of the operation phase division, the vibration signal frames in the time-aligned signal set are matched for phase affiliation. The temperature and speed identification information corresponding to multiple phases are called, and the matched vibration signal segments are paired with the operation phases to generate a multi-source signal set.

[0008] As a further aspect of the present invention, the operating condition change rate threshold is set by statistically analyzing the change rates of temperature and speed signals, calculating the change rate sequences of temperature and speed signals, and analyzing the sum of the mean of the change rate sequences and the standard deviation of a preset multiple.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the multi-source signal set, frequency domain transformation is performed on the vibration signal, temperature signal and rotation speed signal respectively, the amplitude sequence and phase sequence of the multiple signals on the frequency axis are extracted and the sequence is rearranged, the amplitude difference result between adjacent frequency points is continuously smoothed, and the spectral energy distribution result is generated. S202: Call the spectrum energy distribution results, divide the entire frequency range into bandwidth segments, aggregate the energy values ​​in multiple frequency bands, calculate the total energy and energy uniformity in each frequency band, identify the frequency bands whose energy concentration exceeds the preset concentration threshold, and generate a set of frequency band energy concentration parameters. S203: Based on the frequency band energy concentration parameter set, determine the sparsity of the energy distribution sequence in multiple frequency bands, calculate the joint index of the proportion of non-zero energy points and amplitude variance in the frequency band as the energy sparsity state, and statistically analyze the concentration and sparsity of all frequency bands as feature vectors to generate a frequency band sparsity feature set. The concentration threshold is set by statistically analyzing the concentration index of the spectral energy distribution and the sum of the standard deviations of the mean of the concentration index and a preset multiple.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the frequency band sparse feature set, extract the sparsity index corresponding to multiple frequency bands, and synchronously match the operating condition change rate of the signals in the multi-source signal set. Normalize the two sets of sequences and calculate the change gradient to generate the frequency band gradient corresponding data set. S302: Call the frequency band gradient corresponding data group, extract the gradient components of multiple frequency bands and the gradient components of the operating condition change rate, calculate the average gradient difference between the two, determine the coupling strength level between frequency bands based on the gradient difference, pair the frequency bands with the corresponding coupling strength, and generate a frequency band coupling strength parameter set. S303: Based on the frequency band coupling strength parameter set, retrieve frequency bands whose coupling strength level exceeds the set strength screening threshold, aggregate the frequency bands, and simultaneously record the matching range of the corresponding operating condition change rate to generate a frequency band association information set; The intensity screening threshold is set by analyzing the intensity distribution characteristics of the set of frequency band coupling intensity parameters and the sum of the standard deviations of the mean of all frequency band coupling intensities and a preset multiple.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the frequency band association information set, obtain the energy sequence of the same frequency band in the source domain and the target domain, calculate the time domain mean and variance of each frequency band, calculate the energy difference according to the statistical characteristics, normalize the difference value, and generate the frequency band energy distribution difference result. S402: Call the frequency band energy distribution difference results, extract the energy difference of the corresponding frequency band and perform Gaussian kernel function distance calculation, analyze the distance offset of each frequency band, match the offset with the frequency band, and generate a kernel function distance mapping set; S403: Based on the kernel function distance mapping set, adjust the feature mapping weights of corresponding frequency bands in the source and target domains, calculate the feature weight ratios of vibration, temperature and speed signals in multiple frequency bands, and perform cross-domain feature correction based on the weight ratios to generate wind turbine gearbox fault feature extraction results.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the wind turbine gearbox fault feature extraction results, count the number of healthy samples and fault samples and calculate the ratio. When the proportion of fault samples does not reach the fault category proportion threshold, extract the current frequency band and the corresponding sample proportion to generate the sample quantity ratio. S502: Call the sample quantity ratio, extract the sparsity index and migration weight under the corresponding frequency band data interval, perform a weighted average of the sparsity index and migration weight, and use the result as a feature weighting factor to generate a feature weighting factor set; S503: Based on the set of feature weighting factors, the weights in the loss function are adjusted by weighting, the gradient direction is calculated and updated, dynamic iterative optimization is performed on the training parameter set, and the parameter change vector is recorded after each update to generate the optimized wind turbine gearbox fault feature extraction result.

[0013] As a further aspect of the present invention, the fault category proportion threshold is set by statistically analyzing the proportion data of multiple fault categories in the sample set, and analyzing the sum of the standard deviations of the mean of the proportions of all fault categories and a preset multiple.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the operation phase is constructed by aligning multiple source signals and the correlation screening is carried out by combining frequency band sparsity and energy concentration. The coupling strength between frequency bands is quantitatively expressed. The cross-domain mapping weight is adaptively corrected based on the energy distribution and kernel function distance. The contribution of vibration, temperature and rotation speed features is kept in balance. In the case of scarce fault samples, the feature stability is enhanced by weighted training based on sparsity and migration weight. The output results are consistent under multiple fault classification conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning, comprising the following steps: S1: Collect vibration, temperature and speed signal sequences of wind turbine gearbox and perform time alignment. Calculate the rate of change of operating conditions based on temperature and speed signals and divide the operating conditions into stages. Match the vibration signals of all stages with the corresponding operating conditions to generate a multi-source signal set. S2: Based on the multi-source signal set, perform frequency domain decomposition, analyze the frequency components of multiple signals and divide the frequency bands, calculate the energy concentration and sparsity index of each frequency band, extract the key frequency domain features of multiple signals, and generate a frequency band sparse feature set. S3: Based on the sparsity index of multiple frequency bands in the sparse feature set, compare the gradient with the operating condition change rate of the multi-source signal set to determine the coupling strength between frequency bands, screen the feature frequency bands coupled with operating condition fluctuations, and generate a frequency band association information set. S4: Based on the maximum mean difference migration model of the frequency band correlation information set input, the kernel function distance is calculated according to the energy distribution of the same frequency band in the source domain and the target domain, and the feature mapping weight is adjusted. The feature contribution of vibration, temperature and speed signals is analyzed and cross-domain correction is performed to generate the wind turbine gearbox fault feature extraction results. S5: Based on the wind turbine gearbox fault feature extraction results, the ratio of healthy samples to fault samples is calculated. When the proportion of fault samples does not reach the fault category proportion threshold, the sparsity index and migration weight of the corresponding frequency band are extracted, the loss function is weighted and adjusted and the gradient is updated. The model parameters are dynamically optimized and trained to generate the optimized wind turbine gearbox fault feature extraction results.

[0023] The multi-source signal set includes vibration signals, temperature signals, speed signals, operating condition change rate, and operating condition stage. The frequency band sparse feature set includes energy concentration, sparsity index, and key frequency domain features. The frequency band correlation information set includes inter-band coupling strength and characteristic frequency bands of operating condition fluctuation coupling. The wind turbine gearbox fault feature extraction results include frequency band mapping weights, vibration signal feature contribution, temperature signal feature contribution, and speed signal feature contribution. The optimized wind turbine gearbox fault feature extraction results include a weighted adjusted loss function, updated gradient direction, and dynamically optimized model parameters.

[0024] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect vibration, temperature and speed signal sequences of wind turbine gearbox in operation and align them in time. Synchronize the corresponding frames of multiple signal sequences according to the same time and map the time periods of vibration, temperature and speed signals to generate a time-aligned signal set. First, vibration sensors, temperature sensors, and speed sensors are installed at predetermined locations on the wind turbine gearbox housing. Specifically, the vibration sensors are triaxial accelerometers installed near the high-speed bearing housing and planetary carrier bearing housing of the gearbox to collect vibration acceleration signals along the x, y, and z axes; the temperature sensors are thermocouple sensors installed in the gearbox lubrication oil sump and on the outer ring of the high-speed shaft bearing to collect lubrication oil temperature and bearing temperature signals; the speed sensors are photoelectric encoders coaxially connected to the high-speed shaft of the gearbox to collect real-time speed signals from the gearbox output shaft. Data acquisition is initiated, and different sampling frequencies are set for each sensor: the sampling frequency for vibration signals is set to 25600Hz, the sampling frequency for temperature signals is set to 1Hz, and the sampling frequency for speed signals is set to 1Hz. All sensors acquire data through a data acquisition unit (DAU) with a unified clock source. This clock source is synchronized with a standard time server via Network Time Protocol (NTP) to ensure that the timestamp accuracy of all data points reaches the millisecond level. The acquired raw signal sequence consists of its respective timestamp and signal value. For example, starting data acquisition at time T0 = 10:00:00.000 on November 20, 2024, 25,600 vibration signal points, 1 temperature signal point, and 1 rotation speed signal point will be acquired within 1 second. Next, a "synchronization time frame" is defined based on the sampling period (1 second) of the lowest frequency signal (in this example, the temperature and rotation speed signals, 1Hz). Time alignment processing is then performed on the original multi-signal sequence. Specifically, each synchronization time frame starting from T0 is iterated through; for example, the first time frame is (T0, T0+1 seconds). For this time frame, all signal data acquired within this time interval is extracted. For example, within the first time frame (10:00:00.000, 10:00:01.000), the temperature value of 55.2℃ at timestamp 10:00:00.500, the rotational speed value of 1502 rpm at timestamp 10:00:00.600, and all 25,600 vibration signal sample points collected within that second are extracted. The timestamp of the first sample point of each signal within this time frame (here, 10:00:00.000) is used as the alignment timestamp for that frame. Subsequently, this alignment timestamp, the temperature value, rotational speed value, and the vibration signal sample sequence within that frame are treated as a whole to form a synchronization signal frame. This process is repeated until all collected signal sequences have been processed. In this way, signals with different sampling rates are strictly aligned in time, and the dense time series of vibration signals is accurately mapped to the temperature and rotational speed states at corresponding moments, generating a time-aligned signal set.

[0025] Table 1 Example Table of Time-Aligned Signal Sets

[0026] As shown in Table 1, this table lists a portion of the generated time-aligned signal sets. Each row represents a synchronized signal frame, containing a unique frame number, a unified alignment timestamp, the specific temperature and rotational speed values ​​within that time frame, and a reference to the corresponding vibration signal data segment.

[0027] S102: Based on the time-aligned signal set, extract the temperature signal and speed signal, normalize and calculate the signal difference value of adjacent sampling frames, analyze the temperature change rate and speed change rate, combine the operating condition change rate threshold to segment the rate sequence, and generate the operating condition stage division result. Based on the generated time-aligned signal set, the temperature and rotational speed signal sequences are extracted. Taking the data in Table 1 as an example, the extracted temperature sequence is [55.2, 55.3, 55.3, 56.1, ...], and the rotational speed sequence is [1502, 1505, 1506, 1580, ...]. These two sequences are then normalized. The min-max normalization method is used to map the data to the [0, 1] interval. The min-max values ​​are determined based on the normal operating range recorded in the historical operating data of this type of wind turbine gearbox. Specifically, the normal operating range for temperature is set to [40°C, 90°C], and the normal operating range for rotational speed is set to [1000rpm, 1800rpm]. For the temperature value 55.2°C in frame number 1, its normalization calculation is: (55.2-40) / (90-40) = 15.2 / 50 = 0.304. For the rotational speed value of 1502 rpm in frame number 1, its normalized value is: (1502-1000) / (1800-1000) = 502 / 800 = 0.6275. This normalization calculation is performed on the temperature and rotational speed values ​​of all frames in sequence. Next, the normalized signal difference between adjacent sampled frames is calculated. For the i-th frame, its temperature change rate is the normalized temperature value of the i-th frame minus the normalized temperature value of the (i-1)-th frame; the rotational speed change rate is calculated similarly. For example, to calculate the change rate between the 2nd and 1st frames: the normalized temperature value of frame 2 is (55.3-40) / 50 = 0.306, and the temperature change rate is 0.306 - 0.304 = 0.002. The normalized rotational speed value for frame 2 is (1505-1000) / 800 = 0.63125, and the rotational speed change rate is 0.63125-0.6275 = 0.00375. Calculate the change rate between frame 3 and frame 4: The normalized temperature value for frame 3 is (55.3-40) / 50 = 0.306, and the normalized rotational speed value is (1506-1000) / 800 = 0.6325. The normalized temperature value for frame 4 is (56.1-40) / 50 = 0.322, and the normalized rotational speed value is (1580-1000) / 800 = 0.725. The temperature change rate is 0.322-0.306 = 0.016, and the rotational speed change rate is 0.725-0.6325 = 0.0925. Subsequently, a threshold for the operating condition change rate is set. This threshold is not a fixed value, but rather derived through experimental analysis of historical operating data. The experimental procedure is as follows: Temperature and speed data of this type of wind turbine were collected under specific transient conditions such as start-up and shutdown, emergency shutdown, and sudden changes in wind speed, as well as under long-term stable operating conditions, totaling 100 sets of transient condition samples and 100 sets of steady-state condition samples. The above-mentioned normalization and difference calculations were performed on these data to obtain the rate of change distribution under the two types of conditions. For the steady-state condition, 95% of the calculated normalized speed change rate values ​​were below 0.08, and 95% of the normalized temperature change rate values ​​were below 0.01.For transient conditions, the normalized rate of change of rotational speed is generally greater than 0.1, and the normalized rate of change of temperature is greater than 0.015 during periods of rapid power change. Based on this experimental data, the threshold values ​​for the rate of change of rotational speed and temperature are set to 0.08 and 0.015, respectively. The rate sequence is segmented according to these threshold values. The judgment logic is as follows: if the absolute value of the rate of change of temperature in the current frame is greater than 0.015, or the absolute value of the rate of change of rotational speed is greater than 0.08, then the transition phase between this frame and the previous frame is determined as a "transient phase"; otherwise, it is determined as a "steady-state phase". For frame 2, the rate of change of temperature is 0.002 < 0.015 and the rate of change of rotational speed is 0.00375 < 0.08, therefore frames 1 to 2 are steady-state phases. For frame 4, the rate of change of temperature is 0.016 > 0.015 and the rate of change of rotational speed is 0.0925 > 0.08, satisfying either condition, therefore frames 3 to 4 are transient phases. By judging the entire time series frame by frame, consecutive "steady-state phases" are merged into one operating phase, and "transient phases" are used as the dividing points between different steady-state phases to generate the operating phase division results, for example: [steady-state phase one: frames 1-3], [transient phase one: frames 3-4], [steady-state phase two: frames 5-80].

[0028] S103: Based on the results of the division of operating conditions, the vibration signal frames in the time-aligned signal set are matched for stage affiliation. The temperature and speed identification information corresponding to multiple stages are called. The matched vibration signal segments are combined and paired with the operating conditions to generate a multi-source signal set. Based on the generated operating condition stage division results, stage assignment matching is performed on the vibration signal frames in the generated time-aligned signal set. Specifically, each vibration signal frame (uniquely identified by its frame number) is traversed, and its corresponding stage is searched in the operating condition stage division results according to its frame number. For example, if the division result of S102 is "[Steady-state stage one: frames 1-3], [Transient stage one: frames 3-4], [Steady-state stage two: frames 5-80]", then vibration signal frames with frame numbers 1, 2, and 3 (i.e., data segments 1, 2, and 3) all belong to "Steady-state stage one"; vibration signal frame with frame number 4 (data segment 4) belongs to "Transient stage one"; and vibration signal frames with frame numbers 5 to 80 belong to "Steady-state stage two". After completing the assignment matching, identification information is generated for the vibration signals within each stage. This identification information consists of the average temperature and average rotational speed corresponding to that stage. The specific calculation process is as follows: For a defined operating phase, extract the temperature and rotational speed values ​​of all frames within that phase, and calculate their arithmetic mean. Taking "Steady-state Phase 1" (containing frames 1, 2, and 3) as an example: The temperature value set for this phase is [55.2, 55.3, 55.3]°C. The average temperature is (55.2 + 55.3 + 55.3) / 3 = 55.27°C. The rotational speed value set for this phase is [1502, 1505, 1506] rpm. The average rotational speed is (1502 + 1505 + 1506) / 3 = 1504.33 rpm. Therefore, the identification information for "Steady-state Phase 1" is {Phase Type: Steady-state, Average Temperature: 55.27°C, Average Rotational Speed: 1504.33 rpm}. For "Transient Phase 1" (which only includes the transition from frame 3 to frame 4), its identification information directly uses the operating condition parameters of the start and end frames, or is marked as "transient," without calculating the average value. Finally, the matched vibration signal segments are combined and paired with the identification information of the operating condition phase to which the segment belongs. Here, a vibration signal segment refers to the set of all continuous vibration signal frames belonging to the same operating condition phase. For example, for "Steady-State Phase 1," its corresponding vibration signal segments are the cascaded set of data segments 1, 2, and 3. This set will be paired with the identification information {Phase Type: Steady-State, Average Temperature: 55.27°C, Average Speed: 1504.33 rpm}. The generated multi-source signal set is a structured dataset, in which each data unit contains two parts: one part is continuous vibration signal data under a specific working condition stage, and the other part is identification information describing the key operating parameters (temperature and speed) of that working condition stage, such as "steady-state stage one" (vibration data [frames 1-3], {identifier: "steady-state stage one", temperature: 55.27, speed: 1504.33}).

[0029] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on a multi-source signal set, frequency domain transformation is performed on vibration signal, temperature signal and rotation speed signal respectively. The amplitude sequence and phase sequence of multiple signals on the frequency axis are extracted and the sequence is rearranged. The amplitude difference results between adjacent frequency points are continuously smoothed to generate the spectral energy distribution results. From the generated multi-source signal set, a data unit is selected for processing, for example, the vibration data [frames 1-3] corresponding to "steady-state stage one", {identifier: "steady-state stage one", temperature: 55.27, rotational speed: 1504.33}. The vibration signal in this data unit is continuous data for 3 seconds, with a sampling frequency of 25600Hz, containing a total of 76800 sampling points. First, frequency domain transformation is performed on this vibration signal segment, as well as the three temperature value sequences [55.2, 55.3, 55.3] and three rotational speed value sequences [1502, 1505, 1506] corresponding to this stage. For the vibration signal, a Fast Fourier Transform (FFT) is performed. Before the transform, a Hanning window function is applied to the 76800 data points, that is, the amplitude of each data point is multiplied by the window function coefficient at its corresponding position. The window function coefficients are distributed in a cosine shape with a center point of 1 and zeros at both ends. After the transform, a series of complex results are obtained, with a quantity of 76800. For each complex number result, its magnitude is calculated as the amplitude, and its argument is calculated as the phase, thus obtaining a single-sided amplitude and phase sequence of length 38400, with a frequency resolution of 25600Hz / 76800=0.333Hz. For low-frequency signals such as temperature and rotational speed, after transforming the sequences at the three points, the main energy is concentrated in the DC component at 0Hz, corresponding to their average values ​​of 55.27℃ and 1504.33rpm, respectively, with the amplitudes of the other frequency components close to zero. Next, the spectral sequences of different signals are rearranged. A new data structure is created, indexed by frequency. For each frequency point, such as 100Hz, the amplitude, phase, temperature, and rotational speed signals of the vibration signal at that frequency point are recorded. Since the amplitudes of the non-zero frequency components of the temperature and rotational speed signals are extremely small, subsequent processing mainly focuses on the amplitude sequence of the vibration signal. Taking the vibration signal amplitude sequence as an example, the amplitudes at 100Hz, 100.333Hz, 100.667Hz, 101Hz, and 101.333Hz are set to [0.051, 0.058, 0.055, 0.056, 0.053]g, respectively. The difference between the amplitudes of adjacent frequency points is calculated, resulting in a difference sequence: [0.007, -0.003, 0.001, -0.003]. Subsequently, continuous smoothing is performed on this difference result sequence. The smoothing process uses a moving average method, with a sliding window size of 5. For the i-th point in the difference sequence, its smoothed value is the arithmetic mean of itself and the two points before and after it (a total of 5 points). For example, for a difference value of -0.003, if its first two values ​​are 0.002 and 0.007, and its last two values ​​are 0.001 and -0.003, then its smoothed value is (0.002 + 0.007 - 0.003 + 0.001 - 0.003) / 5 = 0.0008. Applying this process to the entire amplitude difference sequence yields a smoother sequence with suppressed noise.The spectral energy distribution result is a multidimensional spectral data structure that has been rearranged and whose vibration amplitude has been smoothed.

[0030] S202: Call the spectrum energy distribution results, divide the entire frequency range into bandwidth segments, aggregate the energy values ​​in multiple frequency bands, calculate the total energy and energy uniformity in each frequency band, identify the frequency bands whose energy concentration exceeds the preset concentration threshold, and generate a set of frequency band energy concentration parameters. The generated spectral energy distribution results are used, specifically the vibration signal amplitude spectrum. Based on the gearbox's physical structure and design parameters, the entire frequency range (0Hz to 12800Hz) is divided into multiple frequency bands with clear physical meaning. The frequency band division is based on the meshing frequencies of each gear stage, the shaft rotation frequency and its harmonics, and the bearing fault characteristic frequency range. For example, a specific bandwidth segmentation scheme is as follows: Band 1 (0-50Hz) corresponds to the rotation frequency and its lower harmonics; Band 2 (50-800Hz) corresponds to the planetary gear meshing frequency range; Band 3 (800-2500Hz) corresponds to the intermediate stage gear meshing frequency range; Band 4 (2500-6000Hz) corresponds to the high-speed stage gear meshing frequency range; and Band 5 (6000-12800Hz) corresponds to the bearing fault characteristic frequency and high-frequency noise. Next, the energy values ​​within each frequency band are aggregated and calculated. First, the total energy within each frequency band is calculated. The energy value is obtained by squared amplitude and multiplying by a scaling factor. For simplicity, the squared amplitude is used directly as a relative measure of energy. Taking band 3 (800-2500Hz) as an example, the number of frequency points covered by this band is approximately (2500-800) / 0.333 ≈ 5105. Summing the squared amplitudes of these 5105 frequency points gives the total energy of band 3. For example, if the calculated total energy is 1.25g². Next, energy uniformity is calculated. Energy uniformity is measured by calculating the entropy of all energy points within the band. A more direct calculation is to use the reciprocal of the coefficient of variation, which is the ratio of the energy mean to the energy standard deviation. For example, if the energy mean in band 3 is 1.25 / 5105 = 0.000245g² and the energy standard deviation is 0.0011g², then the uniformity index is 0.000245 / 0.0011 = 0.223. Then, frequency bands with energy concentration levels exceeding a preset concentration threshold are identified. Energy concentration is defined as the ratio of the maximum energy value to the average energy value within a frequency band. The concentration threshold is set based on experimental data: data collection and analysis were performed on 50 healthy gearboxes of the same model and 50 gearboxes exhibiting early tooth surface pitting failure. Experimental data showed that 95% of the samples in the healthy gearboxes had an energy concentration index value below 8.0 across all frequency bands. However, in the faulty gearboxes, 95% of the samples had an index value above 15.0 in the frequency band containing its fault characteristic frequency (e.g., high-speed meshing frequency band 4). Therefore, the energy concentration threshold is set to 12.0. For frequency band 3, if its maximum energy value is 0.03 g² and its average energy value is 0.000245 g², then the energy concentration is 0.03 / 0.000245 ≈ 122.4. This value of 122.4 is much greater than the concentration threshold of 12.0, therefore frequency band 3 is identified as an energy concentration band. The calculation results for each frequency band are summarized to generate a frequency band energy condensed parameter set.

[0031] Table 2 Examples of Frequency Band Energy Concentration Parameters

[0032] As shown in Table 2, this table is an example of a set of parameters for concentrated energy in a frequency band, which records the energy statistical characteristics of each frequency band and the judgment results of whether it is concentrated.

[0033] S203: Based on the frequency band energy concentration parameter set, determine the sparsity of the energy distribution sequence in multiple frequency bands, calculate the joint index of the proportion of non-zero energy points and amplitude variance in the frequency band as the energy sparsity state, and statistically analyze the concentration and sparsity of all frequency bands as feature vectors to generate a frequency band sparsity feature set. Based on the generated set of parameters for concentrated energy in the frequency bands, the sparsity of the energy distribution sequence within each frequency band is further determined. Sparsity determination aims to quantify whether significant energy spectral lines within a frequency band are sparsely or densely distributed. This determination is accomplished by calculating a joint index that combines the proportion of non-zero energy points within the frequency band with the amplitude variance. First, a "non-zero energy point" is defined. Since background noise is prevalent in the spectrum, whether an energy point is "non-zero" is determined by comparing it with a noise baseline. The noise baseline is set as follows: in historical health data, a known frequency band without characteristic signals (e.g., 11000-12000Hz) is selected, the average energy within this frequency band is calculated, and this average value is multiplied by a coefficient of 1.5 to obtain the noise baseline value. For example, if the calculated average noise energy in a healthy state is 0.000001g², then the noise baseline is set to 0.0000015g². Spectral lines with energy values ​​higher than this baseline within the frequency band are counted as "non-zero energy points." Next, the joint index for each frequency band is calculated. Taking frequency band 3 (800-2500Hz) as an example in Table 2, it contains 5105 frequency points. The energy values ​​of these 5105 points are iterated, and the number of points higher than 0.0000015g² is counted. If 102 points are found to be higher than this baseline, then the proportion of non-zero energy points is approximately 102 / 5105 ≈ 0.02. Then, the original vibration amplitudes corresponding to these 102 "non-zero energy points" are extracted, and the variance of these amplitudes is calculated. The variance of these 102 amplitudes is calculated to be 0.009g². The joint index, i.e., the energy sparsity state, is calculated by weighting these two values. The calculation method is: Energy sparsity state = Weighting coefficient 1 × (1 - Proportion of non-zero energy points) + Weighting coefficient 2 × Amplitude variance. The setting of the weighting coefficients reflects the degree of importance attached to the two aspects of sparsity (number of spectral lines and difference in spectral line amplitudes). Here, weighting coefficient 1 is set to 0.6, and weighting coefficient 2 is set to 0.4. The weighting coefficients were analyzed and compared with 20 known sparse fault signals (such as cracks) and 20 known diffuse wear signals. It was found that the weighting combination could most effectively distinguish the sparse characteristics of the two types of signals. For frequency band 3, its energy sparsity state value is: 0.6×(1-0.02)+0.4×0.009=0.6×0.98+0.4×0.009=0.588+0.0036=0.5916. A high sparsity state value means that the energy in this frequency band is dominated by a few spectral lines with large amplitude differences. Finally, the concentration and sparsity of all frequency bands were statistically analyzed into feature vectors. The energy concentration sequence of all frequency bands was extracted from Table 2 of S202, namely [4.2, 7.8, 122.4, 9.5, 2.1]. At the same time, the same sparsity calculation as for band 3 is performed on all frequency bands (bands 1, 2, 4, 5) to obtain a sparsity sequence, such as [0.21, 0.45, 0.5916, 0.48, 0.15].Concatenating these two sequences creates a 10-dimensional feature vector: [4.2, 7.8, 122.4, 9.5, 2.1, 0.21, 0.45, 0.5916, 0.48, 0.15]. This vector comprehensively describes the energy distribution and structural characteristics of the "steady-state stage one" signal across the entire frequency domain. Repeating this process for each data unit in the multi-source signal set (e.g., different steady-state or transient stages) generates a sparse feature set, where each element is a feature vector associated with a specific operating condition segment.

[0034] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the sparse feature set of the frequency band, extract the sparsity index corresponding to multiple frequency bands, and synchronously match the operating condition change rate of the signal in the multi-source signal set. Normalize the two sets of sequences and calculate the change gradient to generate the data set corresponding to the frequency band gradient. Based on the generated frequency band sparse feature set, which contains feature vector sequences corresponding to each operating condition stage, feature vectors are extracted for four consecutive operating condition stages (stages A, B, C, and D), and sparsity indices for each frequency band are extracted to form a sparsity sequence. Stage A corresponds to the sparsity sequence calculated in S203: [0.21, 0.45, 0.5916, 0.48, 0.15]. The subsequent three stages B, C, and D undergo the same processing flow, resulting in the following sparsity sequences: Stage B [0.22, 0.46, 0.5930, 0.49, 0.14], Stage C [0.25, 0.49, 0.6100, 0.53, 0.18], and Stage D [0.26, 0.50, 0.6150, 0.54, 0.19]. These four stages' sparsity sequences are reorganized by frequency band into five time series, each with a length of 4. For example, the sparsity time series for frequency band 3 is [0.5916, 0.5930, 0.6100, 0.6150]. Simultaneously, the operating condition change rates corresponding to these four operating conditions are synchronously matched from the multi-source signals of S102. Here, the normalized speed change rate series is selected: [0.00375, 0.00125, 0.0925, 0.0025]. Next, these two sets of sequences (the sparsity time series and the operating condition change rate time series for each frequency band) are normalized. The normalization method uses min-max normalization, mapping the data to the [0, 1] interval. The normalization range is determined based on historical big data statistics; the statistical range for the sparsity index is [0.1, 0.8], and the statistical range for the operating condition change rate is [-0.1, 0.1]. Taking the sparsity sequence of frequency band 3 [0.5916, 0.5930, 0.6100, 0.6150] as an example, its normalization calculation process is as follows: First point: (0.5916-0.1) / (0.8-0.1)=0.4916 / 0.7=0.7023, Second point: (0.5930-0.1) / 0.7=0.7043, Third point: (0.6100-0.1) / 0.7=0.7286, Fourth point: (0.6150-0.1) / 0.7=0.7357. The normalized sparsity sequence of frequency band 3 is [0.7023, 0.7043, 0.7286, 0.7357]. The normalization calculation process for the operating condition change rate sequence [0.00375, 0.00125, 0.0925, 0.0025] is as follows: The first point: (0.00375-(-0.1)) / (0.1-(-0.1))=0.10375 / 0.2=0.51875, and the subsequent points are calculated as 0.50625, 0.9625, and 0.5125 respectively. The normalized operating condition change rate sequence is [0.51875, 0.50625, 0.9625, 0.5125].Finally, the gradient of each normalized sequence is calculated. The gradient is defined as the difference between values ​​at adjacent time points. The calculation of the sparsity gradient in band 3 is as follows: Gradient 1 (BA) = 0.7043 - 0.7023 = 0.0020, Gradient 2 (CB) = 0.7286 - 0.7043 = 0.0243, Gradient 3 (DC) = 0.7357 - 0.7286 = 0.0071. The gradient sequence of band 3 is [0.0020, 0.0243, 0.0071]. The calculation of the gradient of the rate of change of operating conditions is as follows: Gradient 1 (BA) = 0.50625 - 0.51875 = -0.0125, Gradient 2 (CB) = 0.9625 - 0.50625 = 0.45625, Gradient 3 (DC) = 0.5125 - 0.9625 = -0.4500. The gradient sequence of the operating condition change rate is [-0.0125, 0.45625, -0.4500]. The gradient sequences of all frequency bands are mapped to the gradient sequences of the operating condition change rate at specific time points to form a frequency band gradient correspondence data set.

[0035] S302: Call the frequency band gradient corresponding data group, extract the gradient components of multiple frequency bands and the gradient components of the operating condition change rate, calculate the average gradient difference between the two, determine the coupling strength level between frequency bands based on the gradient difference, pair the frequency bands with the corresponding coupling strength, and generate a frequency band coupling strength parameter set. The generated frequency band gradient corresponding data set is called. This data set contains gradient sequences for the sparsity of each frequency band and gradient sequences for the rate of change of operating conditions. Taking frequency band 3 as an example, its gradient sequence is [0.0020, 0.0243, 0.0071], and the gradient sequence for the rate of change of operating conditions is [-0.0125, 0.45625, -0.4500]. First, these two sequences are extracted. Then, the absolute value of the gradient difference between the two at each time point is calculated. The absolute value of the gradient difference at time point 1: |0.0020-(-0.0125)|=0.0145, the absolute value of the gradient difference at time point 2: |0.0243-0.45625|=0.43195, and the absolute value of the gradient difference at time point 3: |0.0071-(-0.4500)|=0.4571. Calculate the average of this set of gradient differences as the average gradient difference between band 3 and the operating condition changes: (0.0145 + 0.43195 + 0.4571) / 3 = 0.90355 / 3 = 0.3012. Repeat this calculation for all other bands (bands 1, 2, 4, and 5). For example, the calculated average gradient difference for band 1 is 0.4125, for band 2 it is 0.4053, for band 4 it is 0.3861, and for band 5 it is 0.4892. Next, determine the coupling strength level between the bands based on the calculated average gradient differences. The criteria for classifying the coupling strength level are obtained through verification on a dedicated test bench. The experimental scheme is as follows: using a gearbox test bench with a programmable load and injectable faults (such as tooth root cracks and bearing outer ring spalling), collect 100 sets of full life cycle data from healthy state to fault state. The calculations described in S301 and S302 are performed on these data to obtain the average gradient difference distribution between known fault frequency bands (e.g., the meshing frequency band corresponding to tooth root cracks) and other frequency bands. Experimental results show that for fault characteristic frequency bands directly related to operating condition changes (load changes), the average gradient difference is less than 0.35 for 95% of the samples; for frequency bands that are somewhat related to operating conditions but not directly related, the average gradient difference is distributed between 0.35 and 0.45; and for frequency bands that are basically unrelated, the average gradient difference is greater than 0.45. Based on this, coupling strength levels are defined as follows: "Strong Coupling": average gradient difference less than 0.35; "Medium Coupling": average gradient difference between 0.35 (inclusive) and 0.45 (inclusive); "Weak Coupling": average gradient difference greater than 0.45. Finally, each frequency band is paired with its calculated average gradient difference and the determined coupling strength level. Frequency band 1: average gradient difference 0.4125, determined as "medium coupling"; Frequency band 2: average gradient difference 0.4053, determined as "medium coupling"; Frequency band 3: average gradient difference 0.3012, determined as "strong coupling"; Frequency band 4: average gradient difference 0.3861, determined as "medium coupling"; Frequency band 5: average gradient difference 0.4892, determined as "weak coupling". These pairing results are then processed to generate a set of frequency band coupling strength parameters.

[0036] Table 3. Set of Frequency Band Coupling Strength Parameters

[0037] As shown in Table 3, this table displays the set of frequency band coupling strength parameters calculated in this embodiment. Each frequency band is assigned a quantitative coupling index and a qualitative strength level.

[0038] S303: Based on the frequency band coupling strength parameter set, retrieve frequency bands whose coupling strength level exceeds the set strength screening threshold, aggregate the frequency bands, and record the matching range of the corresponding operating condition change rate to generate a frequency band association information set; Based on the generated set of frequency band coupling strength parameters (as shown in Table 3), frequency bands with coupling strength levels exceeding the set strength screening threshold are retrieved. The strength screening threshold here is set to select frequency bands at the "strong coupling" level. This threshold is based on the fact that, in fault diagnosis applications, vibration characteristic frequency bands closely related to changes in operating conditions are crucial for locating and assessing fault states. The experimental verification process has been described in S302; selecting frequency bands at the "strong coupling" level (i.e., average gradient difference less than 0.35) can filter out more than 95% of known fault characteristic frequency bands. The search operation is performed, traversing Table 3 and filtering rows where the "coupling strength level" column is "strong coupling". In this example, only frequency band 3 meets this condition (average gradient difference 0.3012 < 0.35). The filtered frequency bands are aggregated. In this example, since only one frequency band is filtered out, the aggregation result is set {frequency band 3}. If in other analysis cases, frequency bands 2 and 4 are also determined to be "strongly coupled", the aggregation result will be set {frequency band 2, frequency band 3, frequency band 4}. Simultaneously, the matching range of operating condition change rates associated with these aggregated frequency bands is recorded. This process requires tracing back to the original data in S301, i.e., the unnormalized operating condition change rate sequence [0.00375, 0.00125, 0.0925, 0.0025]. The time points that cause significant changes in the gradient of the selected frequency band (band 3) are identified. In S301, the gradient sequence for band 3 is [0.0020, 0.0243, 0.0071], where the second gradient value of 0.0243 is the largest, corresponding to the change from operating condition stage B to C. The operating condition change rate corresponding to this change is the third value in the sequence, i.e., 0.0925. To determine a range rather than a single value, longer time series data needs to be analyzed. In a long-term monitoring dataset containing 1000 operating conditions, frequency band 3 was identified as "strongly coupled" 15 times. Each time the sparsity gradient peak occurred, the corresponding operating condition change rate (normalized rotational speed change rate) values ​​were [0.085, 0.092, 0.110, 0.098, ..., 0.089]. These 15 rate values ​​were statistically analyzed to find their minimum and maximum values; for example, the minimum was 0.081 and the maximum was 0.115. This interval [0.081, 0.115] is the recorded operating condition change rate matching range. Finally, the aggregated frequency bands were combined with the recorded matching range to generate a frequency band association information set. This information set clearly indicates which frequency bands' vibration characteristics are most sensitive to what degree of operating condition change. In this embodiment, the generated frequency band association information set is a single data entry: {aggregated frequency band: [frequency band 3 (800-2500Hz)], matching operating condition parameter: rotational speed change rate, matching range: [0.081, 0.115] (normalized value)}. This result indicates a strong coupling between the vibration sparsity characteristics of frequency band 3 and the dramatic variation of the normalized rotational speed change rate between 0.081 and 0.115.

[0039] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the frequency band association information set, obtain the energy sequence of the same frequency band in the source domain and the target domain, calculate the time domain mean and variance of each frequency band, calculate the energy difference according to the statistical characteristics, normalize the difference value, and generate the frequency band energy distribution difference result. Based on the generated frequency band correlation information set, this set indicates a strong coupling relationship between frequency band 3 (800-2500Hz) and the normalized rotational speed change rate within the interval [0.081, 0.115]. First, the source and target domains are defined. The source domain data comes from a wind turbine gearbox of the same model operating in a controlled laboratory environment, which has accumulated a large amount of health and fault data. The target domain data comes from a newly deployed wind turbine gearbox in a wind farm, awaiting condition assessment. From the source and target domain databases, 100 signal segments that meet the strong coupling condition are selected, meaning that the normalized rotational speed change rates corresponding to these signal segments all fall within the interval [0.081, 0.115]. For these 200 signal segments, the total energy of each frequency band calculated in S202 is called. Thus, for each frequency band, an energy sequence (length 100) of the source domain and an energy sequence (length 100) of the target domain are obtained. Next, the time-domain statistical characteristics, namely the mean and variance, are calculated for the energy sequence of each frequency band. Taking frequency band 3 as an example: the mean of 100 energy samples in the source domain frequency band 3 is calculated to be 1.25g², and the variance is 0.04(g²)². The mean of 100 energy samples in the target domain frequency band 3 is calculated to be 1.45g², and the variance is 0.09(g²)². The energy difference is calculated based on these two statistical characteristics. The energy difference is defined as the statistical distance between the source and target domain distributions, quantified here by the sum of the absolute values ​​of the mean difference and the variance difference: Energy difference in frequency band 3 = |Source domain mean - Target domain mean| + |Source domain variance - Target domain variance| = |1.25 - 1.45| + |0.04 - 0.09| = 0.20 + 0.05 = 0.25. This calculation is repeated for all 5 frequency bands to obtain an energy difference value vector. For example, the calculated energy difference values ​​for frequency bands 1 to 5 are [0.16, 0.12, 0.25, 0.14, 0.40]. Finally, this difference value vector is normalized. The maximum value of the normalization range is determined by the statistical distribution of differences calculated from a large number of different wind turbine (cross-domain) data pairs. The experiment selected data from 50 different wind turbines and calculated the energy difference values ​​for all frequency bands, finding that 99% of the difference values ​​were less than 2.0. Therefore, the maximum value of the normalization interval was set to 2.0, and the minimum value was 0. The normalized energy difference for frequency band 3 is 0.25 / 2.0 = 0.125. The difference values ​​for other frequency bands were also normalized in the same way: band 1: 0.16 / 2.0=0.080, band 2: 0.12 / 2.0=0.060, band 4: 0.14 / 2.0=0.070, band 5: 0.40 / 2.0=0.200. The resulting difference in frequency band energy distribution is a normalized vector: [0.080, 0.060, 0.125, 0.070, 0.200].

[0040] S402: Call the frequency band energy distribution difference results, extract the energy difference of the corresponding frequency band and perform Gaussian kernel function distance calculation, analyze the distance offset of each frequency band, match the offset with the frequency band, and generate a kernel function distance mapping set; The generated frequency band energy distribution difference results are retrieved, i.e., the vector [0.080, 0.060, 0.125, 0.070, 0.200]. Each value in this vector represents the energy distribution difference between the source and target domains for the corresponding frequency band. First, the energy difference value for each frequency band is extracted, and a Gaussian kernel function distance is calculated based on this. The distance calculation here does not directly apply the Gaussian kernel function formula; instead, the energy difference value calculated by S401 is used as the characteristic distance between the two domains in that frequency band. The bandwidth parameter of the Gaussian kernel function, denoted as sigma, affects the sensitivity of the distance metric. This parameter was determined experimentally: 10 sets of known "same-domain" data pairs with minimal inter-domain differences and 10 sets of known "different-domain" data pairs with significant differences were selected to test the value of sigma in the range of 0.01 to 1.0. When sigma is 0.1, the calculated kernel function value maximizes the distinction between "same-domain" and "different-domain" pairs. Next, the distance offset for each frequency band is analyzed. The distance offset here is directly taken from the normalized energy difference value calculated by S401. This value quantifies the degree of deviation of the target domain features from the source domain features. For band 1, the distance offset is 0.080. For band 2, the distance offset is 0.060. For band 3, the distance offset is 0.125. For band 4, the distance offset is 0.070. For band 5, the distance offset is 0.200. These offsets are matched with the corresponding bands. This matching process generates a data structure where each entry contains a band number and its corresponding distance offset.

[0041] Finally, the matching results of all frequency bands are combined to generate a kernel function distance mapping set.

[0042] Table 4 Kernel Function Distance Map Set

[0043] As shown in Table 4, this table is a specific kernel function distance mapping set. It clearly records the degree of deviation of the characteristic distribution of each vibration signal frequency band in the target domain from that in the source domain. The larger the distance offset, the more significant the difference between the two domains in that frequency band.

[0044] S403: Based on the kernel function distance mapping set, adjust the feature mapping weights of corresponding frequency bands in the source and target domains, calculate the feature weight ratios of vibration, temperature and speed signals in multiple frequency bands, and perform cross-domain feature correction based on the weight ratios to generate wind turbine gearbox fault feature extraction results. Based on the generated kernel function distance mapping set (as shown in Table 4), the feature mapping weights of each frequency band between the source and target domains are adjusted. First, an initial feature weight vector is set. Without any prior knowledge, it is assumed that all five frequency bands contribute equally to the final fault characteristics; therefore, the initial weight ratio is [0.2, 0.2, 0.2, 0.2, 0.2]. The adjustment logic is: the smaller the distance offset of a frequency band between the source and target domains, the more stable the characteristics of that frequency band and the greater its consistency across different domains; its weight should be higher. Conversely, the larger the distance offset, the less stable the characteristics, and the lower the weight should be. The weight adjustment is calculated by multiplying the initial weight by an attenuation factor determined by the distance offset. This attenuation factor is calculated as 1 divided by (1 plus the distance offset). Calculate the adjusted unnormalized weights for each frequency band: Band 1: 0.2 × (1 / (1+0.080)) = 0.2 × 0.9259 = 0.1852, Band 2: 0.2 × (1 / (1+0.060)) = 0.2 × 0.9434 = 0.1887, Band 3: 0.2 × (1 / (1+0.125)) = 0.2 × 0.8889 = 0.177 8. Band 4: 0.2×(1 / (1+0.070))=0.2×0.9346=0.1869, Band 5: 0.2×(1 / (1+0.200))=0.2×0.8333=0.1667. Calculate the sum of these adjusted weights: 0.1852+0.1887+0.1778+0.1869+0.1667=0.9053. Divide each adjusted weight by the sum and normalize to obtain the final feature weight ratios: Band 1: 0.1852 / 0.9053 ≈ 0.2046, Band 2: 0.1887 / 0.9053 ≈ 0.2084, Band 3: 0.1778 / 0.9053 ≈ 0.1964, Band 4: 0.1869 / 0.9053 ≈ 0.2065, Band 5: 0.1667 / 0.9053 ≈ 0.1841. The final feature weight ratios are [0.2046, 0.2084, 0.1964, 0.2065, 0.1841]. Based on these weight ratios, cross-domain feature correction is performed on the features of the target domain. Here, the total energy of each frequency band of a signal segment to be diagnosed in the target domain is used as the original feature vector, for example, [0.55, 0.62, 1.51, 0.93, 0.38]g². The correction process involves weighted summation of this feature vector and the adjusted weight ratio vector. The corrected feature value = (0.55×0.2046)+(0.62×0.2084)+(1.51×0.1964)+(0.93×0.2065)+(0.38×0.1841)=0.11253+0.129208+0.296564+0.192045+0.069958=0.8003.The single value of 0.8003 obtained after weighted summation is the final result of wind turbine gearbox fault feature extraction.

[0045] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the wind turbine gearbox fault feature extraction results, count the number of healthy samples and fault samples and calculate the ratio. When the proportion of fault samples does not reach the fault category proportion threshold, extract the current frequency band and the corresponding sample proportion to generate the sample quantity ratio. Based on the generated wind turbine gearbox fault feature extraction results, statistics were performed on a target domain training dataset containing 2000 labeled samples. The statistics showed 1800 healthy samples and 200 fault samples. The fault samples were further subdivided into two categories: "bearing wear" (120 samples) and "gear cracks" (80 samples). The proportion of each category to the total samples was calculated: "bearing wear" accounted for 120 / 2000 = 6%, and "gear cracks" accounted for 80 / 2000 = 4%. A fault category proportion threshold of 5% was set. This threshold was set based on training experiments on 50 gearbox fault datasets of different sizes and category distributions, which showed that when the sample category proportion was below 5%, the classifier's false negative rate increased by more than 15%. Therefore, 5% was used as the minimum sample proportion requirement to maintain model robustness. The calculated proportions of each fault sample were compared with the fault category proportion threshold of 5%. The proportion of "bearing wear" (6%) was higher than 5%, indicating a sufficient number of samples. The "gear crack" category, with a 4% share, falls below 5%, indicating insufficient sample size. For this category, we extract its associated frequency band information and sample proportion. Based on prior knowledge and the association analysis results of S303, the characteristic energy of the "gear crack" fault is known to be mainly concentrated in frequency band 3 (800-2500 Hz). Therefore, we extract the current imbalance frequency band as frequency band 3, with the corresponding fault category being "gear crack," and a sample proportion of 4%. The generated sample size proportion is represented by a data structure: {Associated frequency band: frequency band 3, associated fault: "gear crack", sample proportion: 0.04}.

[0046] S502: Call the sample quantity ratio, extract the sparsity index and migration weight under the corresponding frequency band data interval, perform a weighted average of the sparsity index and migration weight, and use the result as the feature weighting factor to generate a feature weighting factor set; The generated sample size ratio is used, i.e., {Associated frequency band: frequency band 3, associated fault: "gear crack", sample ratio: 0.04}. For frequency band 3 specified in this data structure, its sparsity index and transfer weight are extracted. The sparsity index is obtained from the calculation results of S203 and represents the intrinsic physical characteristics of the signal within this frequency band. For a typical "gear crack" fault sample, its energy sparsity value in frequency band 3 is 0.5916. The transfer weight is obtained from the calculation results of S403 and represents the stability of the frequency band features in the cross-domain process. The final feature weight ratio corresponding to frequency band 3 is 0.1964. A weighted average is calculated on the extracted sparsity index 0.5916 and the transfer weight 0.1964 to calculate the feature weighting factor. The weight coefficients of the weighted average are determined experimentally. The experiment involves using different weight combinations from 0.1 to 0.9 for 20 known imbalanced fault datasets, and then using these weights for subsequent training adjustments to evaluate the fault identification accuracy of the final model. Experimental results show that when the sparsity index is assigned a weight of 0.7 and the transfer weight a weight of 0.3, the model's recognition performance on minority class samples shows the most significant and stable improvement. Therefore, the weight coefficient of the sparsity index is set to 0.7, and the weight coefficient of the transfer weight is set to 0.3. The calculation process of the feature weighting factor is as follows: Feature weighting factor = (Sparseness index × 0.7) + (Transfer weight × 0.3) = (0.5916 × 0.7) + (0.1964 × 0.3) = 0.41412 + 0.05892 = 0.47304. The advantage of this calculation method is that by combining the sparsity index, which reflects the intrinsic physical characteristics of the signal, and the transfer weight, which reflects cross-domain stability, a comprehensive factor is generated that can both reflect the saliency of the fault itself and measure its cross-condition generalization ability. The calculated factors are associated with the corresponding frequency bands to generate a set of feature weighting factors. In this example, the set is: {Associated frequency band: frequency band 3, feature weighting factor: 0.47304}.

[0047] S503: Based on the feature weighting factor set, the weights in the loss function are adjusted by weighting, the gradient direction is calculated and updated, dynamic iterative optimization is performed on the training parameter set, and the parameter change vector is recorded after each update to generate the optimized wind turbine gearbox fault feature extraction result. Based on the generated feature weighting factor set, {Associated band: band 3, feature weighting factor: 0.47304}, the weights in the loss function are adjusted during the training of a classification model. The classification model includes a parameter set for feature extraction, with initial values ​​of [0.5000, -0.2000, 0.1000]. The loss function uses cross-entropy loss. When processing a minority class sample belonging to "gear crack" (i.e., associated band 3), its contribution to the total loss is multiplied by an adjustment factor. This adjustment factor is 1 plus the feature weighting factor, i.e., 1 + 0.47304 = 1.47304. For healthy samples or other majority class fault samples, this adjustment factor is 1. In one training iteration, when a "gear crack" sample is input, the calculated original loss gradient is [25.0, -10.0, 50.0]. Since this sample is a minority class, its updated gradient direction is calculated by multiplying the original gradient by an adjustment factor of 1.47304: Updated gradient = [25.0, -10.0, 50.0] × 1.47304 = [36.826, -14.7304, 73.652]. The learning rate is set to 0.001. Dynamic iterative optimization is performed on the training parameter set, that is, the parameters are updated according to the updated gradient and learning rate. First, the parameter change vector is calculated, which is the product of the negative learning rate and the updated gradient: Parameter change vector = -0.001 × [36.826, -14.7304, 73.652] = [-0.0368, 0.0147, -0.0737]. The original parameters are added to the parameter change vector to obtain the updated parameter set: New parameter set = [0.5000, -0.2000, 0.1000] + [-0.0368, 0.0147, -0.0737] = [0.4632, -0.1853, 0.0263]. This parameter change vector is recorded after each update. This iterative process is repeated for all samples in the training set until all training rounds are completed. The converged parameter set [0.4632, -0.1853, 0.0263] is the optimized wind turbine gearbox fault feature extraction result.

[0048] In this embodiment of the invention, the operation phase is constructed by aligning multiple source signals and the correlation screening is carried out by combining frequency band sparsity and energy concentration. The coupling strength between frequency bands is quantitatively expressed. The cross-domain mapping weight is adaptively corrected based on the energy distribution and kernel function distance. The contribution of vibration temperature and rotation speed features is kept in balance. In the case of scarce fault samples, the feature stability is enhanced by weighted training based on sparsity and migration weight. The output results are consistent under multiple fault classification conditions.

[0049] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning, characterized in that, Includes the following steps: S1: Collect vibration, temperature and speed signal sequences of wind turbine gearbox and perform time alignment. Calculate the rate of change of operating conditions based on temperature and speed signals and divide the operating conditions into stages. Match the vibration signals of all stages with the corresponding operating conditions to generate a multi-source signal set. S2: Based on the multi-source signal set, perform frequency domain decomposition, analyze the frequency components of multiple signals and divide the frequency bands, calculate the energy concentration and sparsity index of each frequency band, extract key frequency domain features of multiple signals, and generate a frequency band sparse feature set. S3: Based on the sparsity index of the multi-band in the sparse feature set of the frequency band, compare the gradient with the operating condition change rate of the multi-source signal set to determine the coupling strength between frequency bands, filter the feature frequency bands coupled with operating condition fluctuations, and generate a frequency band association information set. S4: Based on the frequency band association information set, input the maximum mean difference migration model, calculate the kernel function distance according to the energy distribution of the same frequency band in the source domain and the target domain, adjust the feature mapping weight, analyze the feature contribution of vibration, temperature and speed signals and perform cross-domain correction, and generate the wind turbine gearbox fault feature extraction results.

2. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 1, characterized in that, The multi-source signal set includes vibration signal, temperature signal, speed signal, operating condition change rate, and operating condition stage. The frequency band sparse feature set includes energy concentration, sparsity index, and key frequency domain features. The frequency band correlation information set includes the characteristic frequency bands of inter-band coupling strength and operating condition fluctuation coupling. The wind turbine gearbox fault feature extraction results include frequency band mapping weight, vibration signal feature contribution, temperature signal feature contribution, and speed signal feature contribution.

3. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect vibration, temperature and speed signal sequences of wind turbine gearbox in operation and align them in time. Synchronize the corresponding frames of multiple signal sequences according to the same time and map the time periods of vibration, temperature and speed signals to generate a time-aligned signal set. S102: Based on the time-aligned signal set, extract the temperature signal and the rotation speed signal and normalize them to calculate the signal difference value of adjacent sampling frames. Analyze the temperature change rate and the rotation speed change rate, and combine the operating condition change rate threshold to segment the rate sequence and generate the operating condition stage division result. S103: Based on the results of the operation phase division, the vibration signal frames in the time-aligned signal set are matched for phase affiliation. The temperature and speed identification information corresponding to multiple phases are called, and the matched vibration signal segments are paired with the operation phases to generate a multi-source signal set.

4. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 3, characterized in that, The operating condition change rate threshold is set by statistically analyzing the change rates of temperature and speed signals, calculating the change rate sequences of temperature and speed signals, and analyzing the sum of the mean of the change rate sequences and the standard deviation of a preset multiple, wherein the preset multiple is a value between 1 and 3.

5. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the multi-source signal set, frequency domain transformation is performed on the vibration signal, temperature signal and rotation speed signal respectively, the amplitude sequence and phase sequence of the multiple signals on the frequency axis are extracted and the sequence is rearranged, the amplitude difference result between adjacent frequency points is continuously smoothed, and the spectral energy distribution result is generated. S202: Call the spectrum energy distribution results, divide the entire frequency range into bandwidth segments, aggregate the energy values ​​in multiple frequency bands, calculate the total energy and energy uniformity in each frequency band, identify the frequency bands whose energy concentration exceeds the preset concentration threshold, and generate a set of frequency band energy concentration parameters. S203: Based on the frequency band energy concentration parameter set, determine the sparsity of the energy distribution sequence in multiple frequency bands, calculate the joint index of the proportion of non-zero energy points and amplitude variance in the frequency band as the energy sparsity state, and statistically analyze the concentration and sparsity of all frequency bands as feature vectors to generate a frequency band sparsity feature set. The concentration threshold is set by statistically analyzing the concentration index of the spectral energy distribution and the sum of the standard deviations of the mean of the concentration index and a preset multiple.

6. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Based on the frequency band sparse feature set, extract the sparsity index corresponding to multiple frequency bands, and synchronously match the operating condition change rate of the signals in the multi-source signal set. Normalize the two sets of sequences and calculate the change gradient to generate the frequency band gradient corresponding data set. S302: Call the frequency band gradient corresponding data group, extract the gradient components of multiple frequency bands and the gradient components of the operating condition change rate, calculate the average gradient difference between the two, determine the coupling strength level between frequency bands based on the gradient difference, pair the frequency bands with the corresponding coupling strength, and generate a frequency band coupling strength parameter set. S303: Based on the frequency band coupling strength parameter set, retrieve frequency bands whose coupling strength level exceeds the set strength screening threshold, aggregate the frequency bands, and simultaneously record the matching range of the corresponding operating condition change rate to generate a frequency band association information set; The intensity screening threshold is set by analyzing the intensity distribution characteristics of the set of frequency band coupling intensity parameters and the sum of the standard deviations of the mean of all frequency band coupling intensities and a preset multiple.

7. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the frequency band association information set, obtain the energy sequence of the same frequency band in the source domain and the target domain, calculate the time domain mean and variance of each frequency band, calculate the energy difference according to the statistical characteristics, normalize the difference value, and generate the frequency band energy distribution difference result. S402: Call the frequency band energy distribution difference results, extract the energy difference of the corresponding frequency band and perform Gaussian kernel function distance calculation, analyze the distance offset of each frequency band, match the offset with the frequency band, and generate a kernel function distance mapping set; S403: Based on the kernel function distance mapping set, adjust the feature mapping weights of corresponding frequency bands in the source and target domains, calculate the feature weight ratios of vibration, temperature and speed signals in multiple frequency bands, and perform cross-domain feature correction based on the weight ratios to generate wind turbine gearbox fault feature extraction results.

8. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 1, characterized in that, The method further includes: S5: Based on the wind turbine gearbox fault feature extraction results, the ratio of healthy samples to fault samples is calculated. When the proportion of fault samples does not reach the fault category proportion threshold, the sparsity index and migration weight of the corresponding frequency band are extracted, the loss function is weighted and adjusted and the gradient is updated. The model parameters are dynamically optimized and trained to generate the optimized wind turbine gearbox fault feature extraction results. The optimized wind turbine gearbox fault feature extraction results include a weighted adjusted loss function, an updated gradient direction, and dynamically optimized model parameters.

9. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 8, characterized in that, The specific steps of S5 are as follows: S501: Based on the wind turbine gearbox fault feature extraction results, count the number of healthy samples and fault samples and calculate the ratio. When the proportion of fault samples does not reach the fault category proportion threshold, extract the current frequency band and the corresponding sample proportion to generate the sample quantity ratio. S502: Call the sample quantity ratio, extract the sparsity index and migration weight under the corresponding frequency band data interval, perform a weighted average of the sparsity index and migration weight, and use the result as a feature weighting factor to generate a feature weighting factor set; S503: Based on the set of feature weighting factors, the weights in the loss function are adjusted by weighting, the gradient direction is calculated and updated, dynamic iterative optimization is performed on the training parameter set, and the parameter change vector is recorded after each update to generate the optimized wind turbine gearbox fault feature extraction result.

10. The method for extracting fault features of unbalanced wind turbine gearboxes based on transfer learning according to claim 9, characterized in that, The threshold for the proportion of fault categories is set by statistically analyzing the proportion data of multiple fault categories in the sample set, and by analyzing the sum of the standard deviations of the mean of the proportions of all fault categories and a preset multiple.

Images