Online condition assessment method and system for high-speed motor

Abstract

The application provides an online state evaluation method and system of a high-speed motor, relates to the technical field of motor state monitoring and fault diagnosis, and the method comprises the following steps: performing feature fusion on time domain and frequency domain characteristic vectors to form a fusion characteristic vector; based on the fusion characteristic vector, identifying abnormal features in operating parameters and analyzing the time sequence evolution law thereof to obtain an abnormal feature mode; generating a motor state comprehensive evaluation index based on the abnormal feature mode, and matching a pre-stored fault feature library to obtain a motor fault type and a state grade; and obtaining a fault-tolerant control instruction according to the motor fault type and the state grade, wherein the fault-tolerant control instruction is used for controlling the motor to maintain safe operation. Through the complete technical chain from multi-modal data acquisition and fusion, adaptive feature extraction, abnormal mode identification to fault-tolerant control instruction generation, the application realizes closed-loop management from data processing to motor safety control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor state monitoring and fault diagnosis, in particular to an online state evaluation method and system for high-speed motors. BACKGROUND

[0002] As the core power components of high-end equipment such as new energy vehicles and aerospace, the operation reliability of high-speed motors is directly related to the safety and performance of the whole system. In the field of new energy vehicles, the drive motor needs to operate stably in the high-speed range of 10,000 to 15,000 rpm, while bearing wear, rotor imbalance, stator winding overheating and other faults are prone to occur under complex road conditions and load changes. If these faults cannot be detected and fault-tolerant control in time, it will cause safety accidents such as power interruption, and active safety control needs to be implemented for online accurate evaluation.

[0003] The existing state monitoring method has obvious defects in data processing when facing the real-time management and control needs of high-speed motors. Taking the drive motor of a new energy vehicle as an example, although its monitoring system collects multi-modal signals, it has three shortcomings: first, the data features are isolated, and the system processes each modal signal independently, which cannot build a multi-modal feature correlation domain, resulting in the cooperative signs of early bearing wear in multiple dimensions such as vibration, current, and temperature being ignored, making it difficult to accurately distinguish between winding inter-turn short circuit and bearing jamming and other heterogeneous faults; second, the feature extraction strategy is rigid, and a fixed parameter feature extraction network is used, which is difficult to adapt to the dynamic changes of data distribution during the process of motor starting to high-speed running, which may lead to insufficient feature perception sensitivity in transient conditions such as sudden acceleration; third, the system lacks a time sequence dynamic perspective, and the linkage between evaluation and control is poor, only based on instantaneous threshold for alarm, ignoring the duration and intensity change trend of abnormal features, which cannot distinguish between transient disturbances and real faults, leading to a lag in fault-tolerant control response, and it is difficult to meet the real-time safety protection needs in high-speed driving scenarios. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an online state evaluation method and system for high-speed motors, which realizes closed-loop management from data processing to motor safety control through a complete technical chain of multi-modal data acquisition and fusion, adaptive feature extraction, abnormal pattern recognition, and fault-tolerant control instruction generation.

[0005] To solve the above technical problems, the technical solution of the present application is as follows:

[0006] In a first aspect, an online state evaluation method for a high-speed motor is provided, the method comprising:

[0007] The three-phase stator current, rotor speed, stator winding temperature and shell vibration frequency of the high-speed motor are collected in real time, the phase, fluctuation, temperature gradient and main frequency characteristics are extracted and fused, and a multi-modal operating parameter set is obtained;

[0008] The multi-modal operating parameter set is preprocessed to obtain a preprocessed multi-modal operating parameter set.

[0009] Four characteristic key dimensions are selected from the preprocessed multi-modal operating parameter set, and a four-dimensional characteristic correlation domain is constructed based on the numerical distribution of the four characteristic key dimensions. The four-dimensional characteristic correlation domain is grid segmented, and the grid density weight is obtained according to the feature distribution density in the segmented grid. The grid density weight is decomposed into network structure parameters to configure the convolution kernel size of the time sequence convolution in the multi-branch feature extraction network and the feature weight proportion of the channel attention mechanism, and an optimized multi-branch feature extraction network is constructed.

[0010] The time domain and frequency domain feature vectors are extracted from the preprocessed multi-modal operating parameter set by the optimized multi-branch feature extraction network.

[0011] The time domain and frequency domain feature vectors are fused to form a fused feature vector. Based on the fused feature vector, the abnormal features in the operating parameters are identified and the time sequence evolution law is analyzed to obtain an abnormal feature mode.

[0012] Based on the abnormal feature mode, a motor state comprehensive evaluation index is generated, and a pre-stored fault feature library is matched to obtain a motor fault type and a state grade.

[0013] According to the motor fault type and the state grade, a fault-tolerant control instruction is obtained, which is used to control the motor to maintain safe operation.

[0014] In a second aspect, an online state evaluation system of a high-speed motor includes:

[0015] The acquisition module is configured to collect the three-phase stator current, rotor speed, stator winding temperature and shell vibration frequency of the high-speed motor in real time, extract the phase, fluctuation, temperature gradient and main frequency characteristics and fuse them to obtain a multi-modal operating parameter set.

[0016] The preprocessing module is configured to preprocess the multi-modal operating parameter set to obtain a preprocessed multi-modal operating parameter set.

[0017] The feature extraction module is configured to filter four characteristic key dimensions from the preprocessed multi-modal operation parameter set, construct a four-dimensional characteristic correlation domain based on the numerical distribution of the four characteristic key dimensions, perform grid segmentation on the four-dimensional characteristic correlation domain, and obtain grid density weights according to the feature distribution density in the segmented grid.

[0018] The identification and evaluation module is configured to perform feature fusion on the time-domain and frequency-domain feature vectors to form a fused feature vector, identify abnormal features in the operation parameters and analyze the time sequence evolution rule of the abnormal features based on the fused feature vector, and obtain an abnormal feature mode, generate a motor state comprehensive evaluation index based on the abnormal feature mode, and match a pre-stored fault feature library to obtain a motor fault type and a state grade.

[0019] The fault-tolerant control module is configured to obtain a fault-tolerant control instruction according to the motor fault type and the state grade, and the fault-tolerant control instruction is used to control the motor to maintain safe operation.

[0020] In a third aspect, a computing device includes:

[0021] One or more processors;

[0022] A storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0023] In a fourth aspect, a computer readable storage medium stores a program, and the program is executed by a processor to implement the method.

[0024] The above-mentioned scheme of the present application at least has the following beneficial effects:

[0025] This system acquires multimodal data such as three-phase stator current and rotor speed from a high-speed motor in real time. It extracts four key features (phase, fluctuation, etc.) and fuses them to form a multimodal operating parameter set. This integrates the core information of the multi-dimensional raw data, providing a comprehensive multimodal foundation for subsequent data processing. The multimodal operating parameter set is preprocessed to remove noise interference and outlier values, unify data format and numerical range, improve data quality, and ensure that subsequent analysis uses regular and reliable data, reducing the impact of invalid data on the processing results. Four key feature dimensions are selected from the preprocessed data to construct a four-dimensional feature association domain, breaking the limitations of isolated single-feature analysis. Through grid segmentation and density weight calculation, feature distribution characteristics are transformed into network structure parameters. The temporal convolution kernel size and channel attention weight ratio are dynamically configured, allowing the multi-branch feature extraction network to adapt to the distribution patterns of multimodal features, improving the network's adaptability. The optimized multi-branch feature extraction network extracts time-domain and frequency-domain feature vectors, leveraging adaptation... The feature distribution network structure accurately captures the changing patterns of data in the time-series dimension and the distribution characteristics in the frequency dimension, allowing the extracted features to fully reflect the core attributes of multimodal parameters. It fuses time-domain and frequency-domain feature vectors to form a fused feature vector, integrating complementary information from both dimensions. Based on the fused features, it identifies abnormal features and obtains abnormal feature patterns through time-series evolution analysis, incorporating the duration and intensity trends of the anomalies to make the anomaly patterns more complete. Based on the abnormal feature patterns, it generates a comprehensive motor status evaluation index, quantifying the overall impact of multiple abnormal features. By matching with a pre-stored fault feature library, it obtains the fault type and status level, using standard fault modes as a reference to provide a unified basis for fault judgment, clarifying the attributes and severity of the current motor fault. Based on the fault type and status level, it obtains fault-tolerant control commands, transforming the fault assessment results into targeted motor adjustment actions (such as adjusting torque and speed), ensuring that the commands fit the current fault situation and directly affect motor operation control, maintaining safe motor operation. Attached Figure Description

[0026] Figure 1 This is a schematic flowchart of an online condition assessment method for a high-speed motor provided by an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of an online condition assessment system for a high-speed motor provided by an embodiment of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] like Figure 1 As shown, an embodiment of the present invention proposes an online condition assessment method for a high-speed motor, the method comprising the following steps:

[0030] Step 100: Real-time acquisition of three-phase stator current, rotor speed, stator winding temperature and housing vibration frequency of high-speed motor; extraction and fusion of phase, fluctuation, temperature gradient and main frequency characteristics to obtain multi-mode operating parameter set;

[0031] Step 200: Preprocess the multimodal operating parameter set to obtain a preprocessed multimodal operating parameter set;

[0032] Step 300: Select four key feature dimensions from the preprocessed multimodal running parameters set; construct a four-dimensional feature association domain based on the numerical distribution of the four key feature dimensions; perform grid segmentation on the four-dimensional feature association domain; obtain grid density weights based on the feature distribution density within the segmented grids; decompose the grid density weights into network structure parameters to configure the kernel size of temporal convolution and the feature weight ratio of the channel attention mechanism in the multi-branch feature extraction network, and construct an optimized multi-branch feature extraction network.

[0033] Step 400: Extract time-domain and frequency-domain feature vectors from the preprocessed multimodal operating parameter set through the optimized multi-branch feature extraction network;

[0034] Step 500: Perform feature fusion on the time domain and frequency domain feature vectors to form a fused feature vector; based on the fused feature vector, identify abnormal features in the operating parameters and analyze their temporal evolution patterns to obtain abnormal feature patterns;

[0035] Step 600: Generate a comprehensive evaluation index of motor status based on abnormal feature patterns, and match it with a pre-stored fault feature library to obtain the motor fault type and status level;

[0036] Step 700: Based on the motor fault type and status level, a fault-tolerant control command is obtained. The fault-tolerant control command is used to control the motor to maintain safe operation.

[0037] In this embodiment of the invention, multiple parameters are acquired in real time, and core features such as phase, fluctuation, temperature gradient, and main frequency are extracted and fused to directly form a multimodal operating parameter set, improving the efficiency of initial data integration and reducing subsequent secondary feature extraction steps. The multimodal operating parameter set is preprocessed to optimize data quality and eliminate redundant interference, providing a high-quality data foundation for subsequent key dimension feature selection and network configuration. Key dimensions are selected based on the numerical distribution of the preprocessed data, and a four-dimensional feature association domain is constructed. Network parameters are configured using density weights obtained through grid segmentation, adapting the structure of the multi-branch feature extraction network to the data feature distribution and improving the network's adaptability to the target data. The optimized network, adapted to the data features, further enhances the network's performance. By extracting time-domain and frequency-domain feature vectors, the feature extraction process is enhanced to capture key information from multimodal data, thereby improving the relevance of feature vectors. The time-domain and frequency-domain feature vectors are fused to simultaneously identify anomalous features and analyze their temporal evolution patterns, integrating multi-dimensional feature information to enrich the descriptive dimensions of anomalous features and improve the completeness of anomalous feature patterns. Based on these anomalous feature patterns, comprehensive evaluation indicators are generated, and then matched with a fault feature library to obtain fault types and status levels, directly linking evaluation indicators with anomalous features and enhancing the data support for fault determination. Finally, fault-tolerant control commands are generated based on fault types and status levels, ensuring that control commands correspond to data-driven fault analysis results and improving the relevance and adaptability of fault-tolerant control.

[0038] In a preferred embodiment of the present invention, step 100 above involves real-time acquisition of the three-phase stator current, rotor speed, stator winding temperature, and housing vibration frequency of the high-speed motor, extracting and fusing their phase, fluctuation, temperature gradient, and main frequency characteristics to obtain a multi-mode operating parameter set, including:

[0039] Step 101 involves acquiring the instantaneous values ​​of the three-phase stator current using current sensors and calculating their phase characteristics based on these instantaneous values. Specifically, this includes: using AH400 series Hall current sensors, which are installed on the outside of the cables connecting the U, V, and W phase output terminals of the high-speed motor stator windings to the motor controller. The radial distance between the sensor probe and the cable is maintained at 5 to 8 mm, and the probe axis is ensured to be parallel to the cable axis. A data acquisition card of model NI9227 provides ±15V DC power to the sensors, and the instantaneous values ​​of the three-phase stator current are acquired in real time at a sampling frequency of 10kHz. Every 100μs, i.e., one sampling cycle, the acquired analog signals of the A-phase, B-phase, and C-phase currents are converted into 16-bit digital signals and stored in the local data cache module. This process is repeated for 500ms to form an initial dataset containing 5000 sets of instantaneous current data.

[0040] When calculating phase characteristics based on the initial dataset, the moment when the instantaneous current value crosses from negative to positive is first defined as the positive zero-crossing point, and the moment when it crosses from positive to negative is defined as the reverse zero-crossing point. The positive zero-crossing point is selected first for calculation. The positive zero-crossing points tA1, tA2, and tA3 of the A-phase current, tB1, tB2, and tB3 of the B-phase current, and tC1, tC2, and tC3 of the C-phase current are extracted from the initial dataset. The phase difference ΔφAB between phases A and B is calculated. Specifically, tA1 and tB1 are selected, and the time difference between them is calculated as ΔtAB = tB1 - tA1. Simultaneously, the time interval between adjacent zero-crossing points of phase A is calculated as ΔtA = tA2 - tA1. 1. The phase difference between phase A and phase B is obtained by ΔφAB = 360° × (ΔtAB / ΔtA). Similarly, the phase difference between phase B and phase C is calculated by ΔφBC = 360° × [(tC1-tB1) / (tB2-tB1)], and the phase difference between phase C and phase A is calculated by ΔφCA = 360° × [(tA1-tC1) / (tC2-tC1)]. When calculating the phase change rate, 10 consecutive sets of phase difference data are selected, corresponding to 10 sampling periods. The difference between two adjacent sets of phase differences is calculated and then divided by the time interval between the two sets of data, 10 × 100 μs = 1 ms, to obtain the phase change rate. Finally, the three phase differences and one phase change rate are integrated into the phase characteristics of the three-phase stator current.

[0041] Step 102 involves synchronously acquiring real-time rotor speed data using a speed sensor and extracting its fluctuation characteristics based on the real-time data. Specifically, this includes: using an SZMB-5 type magnetoelectric speed sensor, which is installed at the edge of the end face of the high-speed motor rotor shaft. The vertical distance between the sensor probe and the end face of the shaft is controlled between 0.5 and 1 mm, and the probe is directly facing 20 metal teeth uniformly machined along the circumferential direction on the end face of the shaft. The teeth are 2 mm high, 5 mm wide, and the tooth groove width is equal to the tooth width. A 10 MHz synchronous clock signal is output through a DS3231 type system clock module and sent to the NI9227 data acquisition card in step 101 and the signal processing circuit of the speed sensor in this step, respectively, to ensure that the deviation of the sampling start time of the two does not exceed 1 μs, thereby achieving synchronous sampling.

[0042] The raw data of rotor speed is collected at a sampling frequency of 5kHz, i.e., 200μs / sampling period. For every 20 sampling periods of raw data collected, corresponding to 4ms, the arithmetic mean of these 20 raw data is taken to obtain 1 speed data point. 5 speed data points are generated every 10ms and stored in the speed data storage unit. When extracting fluctuation features based on stored rotational speed data points, a 50ms sliding window is used, meaning each window contains 5 consecutive rotational speed data points. The average value μ of the 5 rotational speed data points within each window is calculated, and then the sum of squared deviations of each data point from the average value μ is calculated. The sum of squared deviations is divided by the number of data points, 5, and the square root is taken to obtain the standard deviation S of the rotational speed data within the window. At the same time, the maximum value Max and the minimum value Min among the 5 rotational speed data points within the window are identified, and the range R = Max - Min is calculated. The weighting coefficient of the standard deviation is set to 0.6, and the weighting coefficient of the range is set to 0.4. The fluctuation feature value is calculated as S × 0.6 + R × 0.4 to obtain the fluctuation feature of the rotor speed. Each sliding window outputs one fluctuation feature value, and the window sliding step size is consistent with the data point generation interval, both being 10ms.

[0043] Step 103: Collect temperature distribution data of the stator winding using temperature sensors, and calculate its temperature gradient characteristics based on the distribution data. Specifically, this includes: using WRNK-191 type K thermocouple temperature sensors, a total of 8, arranged in a distributed temperature measurement array on the outer wall of the high-speed motor stator winding as follows: along the axial direction of the motor housing, i.e., the rotor axis direction, one axial measuring point group is set at the front 1 / 4 length, the middle 1 / 2 length, and the rear 3 / 4 length of the housing, respectively. The axial measuring point groups at the front and rear ends each... It contains two thermocouples, and the central axial measuring point group contains four thermocouples. The thermocouples in each axial measuring point group are evenly distributed along the circumference of the housing. The central angle deviation between adjacent thermocouples is 90° or 180°. The central measuring point group has a central angle deviation of 90°, and the front and rear measuring point groups have a central angle deviation of 180°. The probe of each thermocouple is attached to the outside of the insulation layer of the stator winding with thermally conductive silicone. The distance between the probe and the winding conductor is controlled at 3 to 5 mm, and the outside of the probe is wrapped with high-temperature resistant insulating tape to avoid contact with the housing.

[0044] The ADAM-4018 data acquisition module provides signal acquisition channels for eight thermocouples. The millivolt-level signals output by the thermocouples are converted into temperature values ​​at a sampling frequency of 1 kHz (1 ms / sampling period), with a resolution of 0.1℃. Temperature data from eight measuring points are recorded every 1 ms, continuously for 1 second to form a dataset containing 1000 sets of temperature distribution data. When calculating the temperature gradient characteristics based on this dataset, the spatial distance between adjacent measuring points is first determined: the spatial distance between axially adjacent measuring points is 1 / 4 of the motor housing length. For example, the distance between axially adjacent measuring points such as those at the front 1 / 4 and the middle 1 / 2 is 50 mm (assuming a housing length of 200 mm). The radial distance between adjacent measuring points (such as those within the middle measuring point group) is also determined. The spatial distance between adjacent measuring points is 1 / 4 of the circumference of the casing. Assuming the casing diameter is 80mm and the circumference is approximately 251.2mm, the distance is approximately 83.7mm. When calculating the axial temperature gradient, the temperature values ​​T1 (front end) and T2 (middle end) of adjacent axial measuring points in the same circumferential direction are selected. The temperature gradient in this direction is calculated as axial temperature gradient = (T2-T1) / 50mm, with units of ℃ / mm. When calculating the radial temperature gradient, the temperature values ​​T3 and T4 of adjacent radial measuring points in the same axial position are selected. The temperature gradient in this direction is calculated as radial temperature gradient = (T4-T3) / 83.7mm. Finally, all axial and radial temperature gradient values ​​are arranged in order of measuring point number to form a stator winding temperature gradient feature containing 8 gradient components.

[0045] Step 104: Collect the spectral data of the vibration frequency of the housing using vibration sensors, and extract its dominant frequency characteristics based on the spectral data. Specifically, this includes: using two PCB352C33 piezoelectric vibration sensors, which are respectively installed on the outer side walls of the high-speed motor housing near the front and rear bearing seats; during installation, first apply thread-locking adhesive to the pre-set M5 mounting screw holes on the housing, and then fix the sensor in the screw holes with M5 bolts. The bolt tightening torque is controlled at 5 N·m to ensure that the sensor is rigidly connected to the housing, and that the sensitive axis direction of the sensor is consistent with the radial direction of the motor rotor.

[0046] The sensor is powered and signal acquisition is performed using an LMSSCADAS Mobile vibration data acquisition unit. Vibration acceleration signals of the housing are acquired at a sampling frequency of 20 kHz (50 μs / sampling period), with units of m / s². 2A vibration signal sequence containing 20,000 data points was obtained by continuously acquiring data for 1 second. To extract the dominant frequency features, the signal sequence was first divided into 19 segments of 1024 data points each. If the last segment had fewer than 1024 data points, it was padded with zeros to reach 1024 data points. Each signal segment was windowed using a Hanning window function to reduce spectral leakage. The windowed time-domain signal was then converted to a frequency-domain signal using a Discrete Fourier Transform, yielding spectral data with a frequency range of 0 to 10 kHz. Each frequency in the spectral data... Each point corresponds to one vibration amplitude. Vibration amplitudes with frequencies ranging from 100Hz to 5kHz are selected from the spectrum data, excluding low-frequency interference and high-frequency noise. The amplitudes are sorted from largest to smallest, and the frequencies corresponding to the first three amplitudes are selected. If the difference between the largest and second largest amplitudes among the first three frequencies exceeds 10% of the largest amplitude, the frequency corresponding to the largest amplitude is determined as the dominant frequency feature. If the difference between the largest and second largest amplitudes does not exceed 10%, the arithmetic mean of the first two frequencies is taken as the dominant frequency feature. Finally, one dominant frequency value is output as the dominant frequency feature of the shell vibration frequency.

[0047] Step 105: The phase characteristics, fluctuation characteristics, temperature gradient characteristics, and dominant frequency characteristics are fused to form a multimodal operating parameter set with a unified timestamp. Specifically, this includes: reading phase characteristic data from the local data cache module, which includes 3 phase differences and 1 phase change rate; reading fluctuation characteristic data from the rotational speed data storage unit, which includes 1 fluctuation characteristic value; reading temperature gradient characteristic data from the ADAM-4018 data acquisition module, which includes 8 gradient components; and reading dominant frequency characteristic data from the LMSSCADASMobile vibration acquisition instrument, which includes 1 dominant frequency value. All characteristic data are accompanied by a system timestamp of the acquisition time with an accuracy of 1 μs.

[0048] Using 10ms as a time interval, the start time of each time interval is set to t0 and the end time to t0+10ms. All feature data whose timestamps fall within the interval are filtered out. If the timestamp of a feature data does not fall within the interval, such as the timestamp of the phase feature being t0+12ms and the timestamp of the fluctuation feature being t0+3ms, then the feature value in the interval is supplemented by linear interpolation. Specifically, the rate of change of the feature between two adjacent valid timestamps is calculated, and then the interpolation value corresponding to the midpoint time of the target time interval is calculated based on the rate of change and the midpoint time of the target time interval, i.e., t0+5ms. If there are multiple data points for a feature in the interval, the arithmetic mean of these data points is taken as the feature value in the interval.

[0049] Within each 10ms time interval, the 13 characteristic parameters are integrated into a 13-dimensional feature vector in the order of phase difference 1, phase difference 2, phase difference 3, phase change rate, fluctuation characteristic value, temperature gradient component 1, temperature gradient component 2, ..., temperature gradient component 8, and dominant frequency value. The midpoint timestamp of the time interval, i.e., t0+5ms, is added to the front of each feature vector. The consecutive feature vectors are integrated into a multi-dimensional array according to the chronological order of the time intervals. This multi-dimensional array is the set of multimodal operation parameters with a unified timestamp.

[0050] In a preferred embodiment of the present invention, step 200 above, which preprocesses the multimodal operating parameter set to obtain a preprocessed multimodal operating parameter set, includes:

[0051] Step 201 involves calculating the convex hull boundary of the phase features, fluctuation features, temperature gradient features, and dominant frequency features in the multimodal operating parameter set, and constructing a convex hull boundary point set. Specifically, this includes extracting feature data corresponding to all timestamps from the multimodal operating parameter set. Each timestamp corresponds to a set of 13-dimensional feature vectors containing phase features, fluctuation features, temperature gradient features, and dominant frequency features. The phase features consist of 3 phase differences and 1 phase change rate, the fluctuation features consist of 1 fluctuation feature value, the temperature gradient features consist of 8 gradient components, and the dominant frequency features consist of 1 dominant frequency value. Each set of feature vectors is used as a 13-dimensional data point to form a multimodal data point set containing N data points, where N is the total number of timestamps in the multimodal operating parameter set.

[0052] First, select the data point with the smallest phase change rate from the data point set as the reference point. If there are multiple points with the smallest phase change rate, further select the point with the smallest fluctuation characteristic value as the reference point. Sort the remaining data points clockwise according to the polar angle of the line connecting them to the reference point. The polar angle is calculated with the reference point as the origin and the phase change rate characteristic dimension as the positive X-axis direction. Starting from the first sorted data point, add the data points to the convex hull candidate point sequence one by one. For each new data point added, determine whether the broken line formed by the point and the last two points in the candidate point sequence is counterclockwise. The sign of the cross product result is used to determine whether the cross product result is positive (counterclockwise) or negative (clockwise). If it is clockwise, delete the last point in the candidate point sequence until any three consecutive points in the candidate point sequence are counterclockwise. Repeat the above process until all data points have been traversed. The data points in the final candidate point sequence are the convex hull boundary points. Store these boundary points in order of polar angle sorting to form the convex hull boundary point set, and record the timestamp and complete feature vector information corresponding to the point set.

[0053] Step 202: Based on the set of convex hull boundary points, calculate the relative distance from each feature point to the convex hull boundary; perform outlier detection and removal based on the relative distance and a preset distance threshold to obtain a cleaned multimodal operating parameter set, specifically including: Based on the set of convex hull boundary points, first calculate the maximum diameter of the convex hull: traverse all pairwise combinations of boundary points in the convex hull boundary point set, calculate the Euclidean distance between each pair of boundary points, that is, the square root of the sum of the squares of the differences in the corresponding dimensions in the 13-dimensional feature vector, and select the largest Euclidean distance as the maximum diameter D of the convex hull.

[0054] For each data point in the multimodal data point set, the shortest perpendicular distance from the data point to each triangular facet is calculated by traversing all three consecutive boundary points in the convex hull boundary point set. The minimum value among all perpendicular distances is taken as the absolute distance d from the data point to the convex hull boundary. The relative distance of each data point is calculated as relative distance = absolute distance d / maximum diameter D of the convex hull.

[0055] Based on historical multimodal data during normal operation of the high-speed motor, the relative distance threshold is determined through statistical analysis. Specifically, 100 sets of multimodal data under normal operating conditions are selected, and the relative distance of all data points in each set of data is calculated using the method described above. The maximum value of these relative distances is taken as the preset distance threshold. For example, if the maximum relative distance under normal operating conditions is found to be 0.125, the preset distance threshold is set to 0.15.

[0056] Iterate through each data point in the multimodal data point set. If the relative distance of a data point is greater than a preset distance threshold, the data point is determined to be an outlier and removed from the data point set. After removal, the remaining data points are reordered according to timestamp order to form a cleaned multimodal data point set. Then, combined with the timestamp and complete feature vector of each data point, a cleaned multimodal running parameter set is generated.

[0057] Step 203: Based on the cleaned multimodal operating parameter set, extract the extreme points from the convex hull boundary point set to determine the actual numerical boundaries of each feature dimension; based on the actual numerical boundaries, calculate the normalized scaling parameters for each feature dimension; using the normalized scaling parameters, map the cleaned multimodal operating parameter set to a unified numerical range to obtain a standardized multimodal operating parameter set. Specifically, from the convex hull boundary point set, for each feature dimension included in the cleaned multimodal operating parameter set (a total of 13 feature dimensions, namely 3 phase differences, 1 phase change rate, 1 fluctuation feature value, 8 temperature gradient components, and 1 dominant frequency value), extract the feature values ​​of all boundary points under each dimension, select the maximum value among the feature values ​​of each dimension as the upper boundary value of the dimension, and the minimum value as the lower boundary value of the dimension, thereby determining the actual numerical boundaries of each of the 13 feature dimensions.

[0058] Set the target numerical range for standardization to [0, 1], and calculate the normalization scaling parameter for each feature dimension: For the i-th feature dimension (i=1 to 13), set the upper boundary value of its actual numerical boundary to Max. i The lower boundary value is Min i According to the scaling factor k i =(1-0) / (Max i -Min i Calculate the scaling factor for this dimension, based on the offset b. i =0-Min i ×k i Calculate the offset for this dimension, and the scaling factor k. i With offset b i Together they constitute the normalization scaling parameter of the i-th feature dimension.

[0059] Extract the values ​​of each feature dimension corresponding to each timestamp from the cleaned multimodal operating parameter set. Then, for each feature dimension value, standardize the value as follows: Value = Original Value × k i +b i Perform mapping calculations; after mapping all feature dimension values, check whether each standardized value is within the [0, 1] interval. If there are values ​​outside the interval, readjust the upper and lower boundary values ​​of the corresponding dimension. The upper boundary value is 1.05 times the maximum value of the original value of the dimension, and the lower boundary value is 0.95 times the minimum value of the original value of the dimension. Recalculate the scaling parameters and map again until the standardized values ​​of all feature dimensions are within the [0, 1] interval. Integrate the standardized feature vectors corresponding to all timestamps in timestamp order to generate a standardized multimodal operating parameter set.

[0060] Step 204: Based on the standardized multimodal operating parameter set, establish an initial time series window; based on the spatial distribution range of the convex hull boundary point set, calculate the window scaling factor of the initial time series window; update the window size of the initial time series window using the window scaling factor to obtain the target time series window; based on the target time series window, perform time series segmentation on the standardized multimodal operating parameter set to obtain the preprocessed multimodal operating parameter set, specifically including: the timestamp interval based on the standardized multimodal operating parameter set is set to 10ms, consistent with the time interval interval in step 105; the basic size of the initial time series window is set to feature data containing 50 consecutive timestamps, that is, the initial window duration is 50 × 10ms = 500ms, and the initial window contains 50 sets of standardized feature vectors.

[0061] For the 13 feature dimensions of the convex hull boundary point set, namely 3 phase differences, 1 phase change rate, 1 fluctuation feature value, 8 temperature gradient components, and 1 dominant frequency value, the upper boundary value of each feature dimension is the maximum value among all boundary point feature values ​​of that dimension, and the lower boundary value is the minimum value among all boundary point feature values ​​of that dimension. The range of each dimension is calculated by subtracting the lower boundary value from the upper boundary value. The ranges of the 13 dimensions are added together to obtain the total range. The total range is then divided by 13 to obtain the average range S of the convex hull space distribution.

[0062] The standard value S0 of the average range of the convex hull spatial distribution during normal operation of the high-speed motor is selected. Specifically, 100 sets of normal operating condition data of the high-speed motor running continuously for 1 hour at rated speed and rated load are selected. The rated speed is, for example, 12000 rpm. The convex hull boundary point set of each set of data is constructed according to the method in step 201. The average range of the convex hull spatial distribution of each set of data is calculated. The arithmetic mean of these 100 average ranges is taken as the standard value S0. The window scaling factor f of the initial timing window is calculated according to the ratio of the average range S of the convex hull spatial distribution to the standard value S0, that is, f equals S divided by S0. If the calculated f is less than 0.5, f is adjusted to 0.5. If f is greater than 2, f is adjusted to 2 to avoid the window size deviating too much from the basic size.

[0063] The target window size is calculated by multiplying the initial window size by f, and the result is rounded to the nearest integer. For example, if the initial window size is 50 and f = 1.2, then the target window size is 60. The window sliding step is set to 1 / 2 of the target window size. For example, if the target window size is 60 and the sliding step is 30, then the window moves forward 30 timestamps each time it slides.

[0064] Starting from the first timestamp, standardized feature data containing consecutive timestamps of the target window size is extracted as the first time series segment. Moving backward by sliding step size, the second time series segment is extracted, and so on, until the last timestamp. If the number of timestamps in the last segment is less than 1 / 2 of the target window size, the segment is merged with the previous segment to ensure that the feature data in each time series segment can fully reflect the operating status of the motor in a certain time period. All time series segments are integrated in the extraction order to form a preprocessed multimodal operating parameter set. Each time series segment is accompanied by corresponding time interval information, namely the start timestamp and the end timestamp.

[0065] In this embodiment of the invention, convex hull boundaries are calculated for multiple features, and a set of convex hull boundary points is constructed. This visualizes the overall distribution range of multimodal feature data, providing a geometric reference based on the actual data distribution for subsequent outlier detection. The relative distances between feature points are calculated based on the convex hull boundary point set, and combined with a preset threshold, outlier detection and removal are achieved. This accurately identifies outliers that deviate from the overall data distribution while avoiding the accidental deletion of normal data within a reasonable distribution range, ensuring the accuracy of data cleaning. Extreme points are extracted from the convex hull boundary point set to determine the actual numerical boundaries of each feature. The normalized scaling parameters calculated accordingly better match the true data distribution, eliminating differences in the dimensions of different features while reducing distortion of the true feature distribution information during standardization. A window scaling factor is calculated based on the spatial distribution range of the convex hull boundary point set. The updated target time series window adapts to the actual data distribution density, avoiding segmentation bias caused by a fixed window size and improving the adaptability of time series segmentation to the time series features of multimodal data.

[0066] In a preferred embodiment of the present invention, step 300 involves selecting four key feature dimensions from the preprocessed multimodal operating parameter set, constructing a four-dimensional feature association domain based on the numerical distribution of the four key feature dimensions, performing grid segmentation on the four-dimensional feature association domain, obtaining grid density weights based on the feature distribution density within the segmented grid, and decomposing the grid density weights into network structure parameters to configure the kernel size of temporal convolutions and the feature weight ratio of channel attention mechanisms in the multi-branch feature extraction network, thereby constructing an optimized multi-branch feature extraction network, including:

[0067] Step 301 involves selecting phase characteristics, fluctuation characteristics, temperature gradient characteristics, and main frequency characteristics as four key feature dimensions from the preprocessed multimodal operating parameter set. Specifically, this includes extracting all feature data contained in each time segment from the preprocessed multimodal operating parameter set and selecting the four feature dimensions with the highest correlation to the core fault of the motor: where phase characteristics correspond to the phase difference and phase change rate of the three-phase stator current, fluctuation characteristics correspond to the fluctuation characteristic value of the rotor speed, temperature gradient characteristics correspond to the axial and radial temperature gradient components of the stator winding, and main frequency characteristics correspond to the main frequency value of the housing vibration frequency. The selection is based on the fact that the above features will show a synergistic change of vibration main frequency shift and current phase abnormality when the bearing wears, and will show a correlation characteristic of sudden increase in temperature gradient and increased speed fluctuation when the winding is overheated, thereby ensuring that the selected key feature dimensions can reflect the synergistic fault symptoms of multimodal parameters.

[0068] Step 302: Based on the numerical distribution of the four key feature dimensions, construct a four-dimensional feature association domain. Specifically, this includes: using the four key feature dimensions as four coordinate axes in a four-dimensional space, where the first dimension is the phase feature, with a value range of 0 to 360°, which is standardized and mapped to 0 to 1; the second dimension is the fluctuation feature, which is standardized and mapped to 0 to 1; the third dimension is the temperature gradient feature, which is standardized and mapped to 0 to 1; and the fourth dimension is the dominant frequency feature, which is standardized and mapped to 0 to 1. Iterate through all time-series segment feature data in the preprocessed multimodal operating parameter set, combining the four feature values ​​corresponding to each timestamp into a four-dimensional data point (x, y, z, w), where x is the phase feature value, y is the fluctuation feature value, z is the temperature gradient feature value, and w is the dominant frequency feature value. Arrange all four-dimensional data points in chronological order to form a set containing the coordinates of the data points and their corresponding time information. This set is the four-dimensional feature association domain, used to reflect the numerical correlation between the four feature dimensions.

[0069] Step 303: The four-dimensional feature correlation domain is uniformly divided into multiple grid cells. Specifically, this includes: dividing each dimension of the four-dimensional feature correlation domain at equal intervals: the first dimension is the phase feature, uniformly divided into 5 intervals from 0 to 1; the second dimension is the fluctuation feature, uniformly divided into 5 intervals from 0 to 1; the third dimension is the temperature gradient feature, uniformly divided into 5 intervals from 0 to 1; and the fourth dimension is the dominant frequency feature, uniformly divided into 5 intervals from 0 to 1. The interval length of each dimension is 0.2 (1 / 5), and the boundary values ​​of adjacent intervals are 0, 0.2, 0.4, 0.6, 0.8, and 1.0, respectively. Multiple closed four-dimensional grid cells are formed by the intersection of the interval boundaries of the four dimensions. The coordinate range of each grid cell is [x...]. i x {i+1} ]×[y j y {j+1} ]×[z k , z {k+1} ]×[w l w {l+1} ], where i, j, k, l are all integers from 0 to 4; finally, 5×5×5×5=625 uniformly distributed grid cells are formed, and each grid cell has a unique spatial coordinate identifier (i, j, k, l).

[0070] Step 304: Calculate the feature distribution density within each grid cell, and calculate the weight coefficient of each grid cell based on the feature distribution density. Specifically, this includes: traversing all four-dimensional data points in the four-dimensional feature association domain, and counting the number of data points contained in each grid cell: for grid cells with coordinate labels (i, j, k, l), determine whether the data point (x, y, z, w) satisfies x i ≤x <x {i+1} y j ≤y <y{j+1} z k ≤z <z {k+1} w l ≤w <w {l+1} If the condition is met, the data point is assigned to that grid cell; after the statistics are completed, each grid cell corresponds to a data point count value n. ijkl .

[0071] Calculate the characteristic distribution density for each grid cell. The volume of a grid cell is the product of the lengths of each dimensional interval, i.e., 0.2 × 0.2 × 0.2 × 0.2 = 0.0016. Based on the characteristic distribution density d... ijkl =n ijkl / 0.0016 is used to calculate the density value of each grid cell; weighting coefficients are calculated based on the characteristic distribution density, and the largest characteristic distribution density d among all grid cells is selected. max According to the weighting coefficient w ijkl =d ijkl / d max The weight coefficient of each grid cell is calculated, so that the weight coefficient ranges from 0 to 1. The denser the feature distribution of the grid cell, the larger the weight coefficient.

[0072] Step 305: Based on the spatial relationship between the weight coefficients and the grid cells, obtain the grid density weights, specifically including: calculating the spatial distance between each grid cell and all other grid cells based on the spatial coordinate identifiers (i, j, k, l) of the grid cells: the spatial distance is the square root of the sum of the squares of the differences in the four-dimensional coordinates, i.e., distance s = √[(i1-i2)]. 2 +(j1-j2) 2 +(k1-k2) 2 +(l1-l2) 2 ], where (i1, j1, k1, l1) and (i2, j2, k2, l2) are the coordinate identifiers of two grid cells; for each grid cell, select adjacent grid cells with a distance of less than 2, that is, grid cells whose sum of the absolute values ​​of the differences in each dimension of the coordinate identifier is less than or equal to 2, and calculate the average weight coefficient of these adjacent grid cells; the grid density weight of the grid cell is calculated as grid density weight = self-weight coefficient × 0.7 + average weight coefficient of adjacent grids × 0.3, thereby associating the self-density characteristics of the grid cell with the distribution characteristics of the surrounding grids to form a density weight value containing spatial correlation.

[0073] Step 306 decomposes the grid density weights into convolution kernel size parameters and channel attention weight allocation parameters. Specifically, this includes: decomposing the grid density weights of each grid unit according to the feature dimensions: for phase features (first dimension) and fluctuation features (second dimension), extract the weight components of the corresponding grid units in these two dimensions, and calculate the convolution kernel size parameters of the temporal convolutional layer according to the formula: convolution kernel size parameter = basic convolution kernel size (3×3) × (1 + weight component × 0.5), where the larger the weight component, the larger the convolution kernel size parameter; for temperature gradient features (third dimension) and main frequency features (fourth dimension), extract the weight components of the corresponding grid units in these two dimensions, and calculate the weight allocation parameters of the channel attention mechanism according to the formula: channel attention weight allocation parameter = weight component / sum of weight components of all dimensions, ensuring that the total proportion of attention weights in each feature dimension is 1.

[0074] Step 307: Based on the convolution kernel size parameter, configure the convolution kernel size of the temporal convolutional layer in the multi-branch feature extraction network to obtain the optimized temporal convolutional layer. Specifically, the multi-branch feature extraction network includes four dedicated branches: a first dedicated branch corresponding to phase features, a second dedicated branch corresponding to fluctuation features, a third dedicated branch corresponding to temperature gradient features, and a fourth dedicated branch corresponding to the main frequency features. Each dedicated branch has one temporal convolutional layer after its input layer. The initial number of input channels of each temporal convolutional layer is uniformly set to 16, the initial number of output channels is uniformly set to 32, the stride is set to 1, and zero padding is used to ensure that the size of the output feature map is consistent with the size of the input feature map.

[0075] In step 306, corresponding convolutional kernel size parameters have been generated for the four dedicated branches. The first dedicated branch corresponds to the convolutional kernel size parameters in the phase dimension, the second dedicated branch corresponds to the convolutional kernel size parameters in the fluctuation dimension, the third dedicated branch corresponds to the convolutional kernel size parameters in the temperature gradient dimension, and the fourth dedicated branch corresponds to the convolutional kernel size parameters in the main frequency dimension. For the temporal convolutional layer of each dedicated branch, the corresponding convolutional kernel size parameters are processed using a rounding method. If the decimal part of the parameter calculation result is greater than or equal to 0.5, it is rounded up to obtain the convolutional kernel size; if the decimal part is less than 0.5, it is rounded down to obtain the convolutional kernel size. At the same time, the value range of the convolutional kernel size is strictly limited to three integers: 3, 5, and 7. If the rounded result is less than 3, it is forcibly adjusted to 3; if the rounded result is greater than 7, it is forcibly adjusted to 7.

[0076] When the feature distribution density of the grid cells corresponding to a certain key feature dimension is greater than 1.2 times the average feature distribution density of all grid cells, the calculated result of the convolution kernel size parameter of the dedicated branch for that dimension is usually greater than the parameter result of other low-density dimension branches. After rounding, a larger convolution kernel size is obtained. For example, if the average grid cell density corresponding to the temperature gradient feature is 0.8, and the grid cell density of a certain region is 1.1, which is greater than 0.8 × 1.2 = 0.96, the calculated result of its corresponding convolution kernel size parameter is 5.3. After rounding, a convolution kernel size of 5 is obtained. Meanwhile, the parameter result corresponding to the fluctuation feature branch with a density of 0.7 is 3.2. After rounding, a convolution kernel size of 3 is obtained. In this way, richer temporal details on high-density feature dimensions can be captured by using a larger convolution kernel size.

[0077] After configuring the kernel size of all dedicated branch temporal convolutional layers, the key parameters such as kernel size, number of input channels, number of output channels, stride and padding method of each branch temporal convolutional layer have been determined, forming an optimized temporal convolutional layer that can be directly connected to a multi-branch feature extraction network. The optimized temporal convolutional layers of the first to fourth dedicated branches are adapted to the processing requirements of phase features, fluctuation features, temperature gradient features and main frequency features, respectively.

[0078] Step 308: Based on the channel attention weight allocation parameters, reconstruct the feature weight ratio of the channel attention mechanism in the multi-branch feature extraction network to obtain the optimized channel attention mechanism, specifically including:

[0079] The channel attention mechanism of the multi-branch feature extraction network adopts a three-level structure of feature compression, weight allocation, and feature weighting, and includes four independent feature channels, corresponding to phase features, fluctuation features, temperature gradient features, and dominant frequency features, respectively. Among them, phase features are channel 1, fluctuation features are channel 2, temperature gradient features are channel 3, and dominant frequency features are channel 4. Each channel receives the output features of the corresponding dedicated branch temporal convolutional layer. The input feature dimension is uniformly [batch size, seq len, 32], where batch size is the batch size, seq len is the temporal sequence length, and 32 is the channel output dimension.

[0080] First, output weight allocation parameters for each of the four feature channels, denoted as P1 (phase feature channel), P2 (fluctuation feature channel), P3 (temperature gradient feature channel), and P4 (dominant frequency feature channel). Calculate the sum of these four parameters: S = P1 + P2 + P3 + P4. If S is within the range of 0.98 to 1.02, then this set of parameters is directly used. If S is less than 0.98 or greater than 1.02, then the corrected parameters P4 are used. i’ =P i / S normalizes and corrects each parameter to ensure that the sum of the four after correction is 1, thus avoiding an imbalance in weight distribution.

[0081] Adjust the weight ratio of each channel based on the verified or corrected weight allocation parameters: Set the attention weight coefficient of channel 1 to P1 (or P 1’ Channel 2 is set to P2 (or P) 2’ Channel 3 is set to P3 (or P3’ Channel 4 is set to P4 (or P) 4’ For example, if step 306 outputs P1=0.2, P2=0.2, P3=0.35, P4=0.25, and the sum S=1.0, then the weight coefficients of channels 1 to 4 are directly configured to 0.2, 0.2, 0.35, and 0.25 respectively; if step 306 outputs P1=0.21, P2=0.21, P3=0.36, P4=0.26, and the sum S=1.04, then the corrected P... 1’ =0.21 / 1.04≈0.198、P 2’ =0.21 / 1.04≈0.198、P 3’ =0.36 / 1.04≈0.346、P 4’ =0.26 / 1.04≈0.25, configure the weight coefficients of each channel according to the corrected parameters.

[0082] After the weights are configured, each feature channel receives the output features from the corresponding dedicated branch. First, the features of dimension [batch size, seq len, 32] are compressed into channel feature vectors of dimension [batch size, 1, 32] through global average pooling. Then, this vector is multiplied by the weight coefficient of the corresponding channel to obtain the weighted channel feature vector. Finally, the weighted feature vectors of the four channels are concatenated according to the channel dimension to form a fused feature vector of dimension [batch size, seq len, 32], which serves as the output of the channel attention mechanism.

[0083] Step 309: Based on the optimized temporal convolutional layer and channel attention mechanism, an optimized multi-branch feature extraction network is constructed. Specifically, the overall architecture of the multi-branch feature extraction network includes an input layer, four dedicated branch layers, a channel attention layer, and an output layer, with each layer connected sequentially according to the data flow direction. The input layer receives a preprocessed multimodal operating parameter set, which is temporal segmented data. The dimension of a single segment is [1, T, 4], where 1 is the number of samples, T is the temporal length, and 4 are the four key feature dimensions. The input layer splits the data according to the feature dimensions, where the first dimension (phase feature) is input into the first dedicated branch, the second dimension (fluctuation feature) is input into the second dedicated branch, the third dimension (temperature gradient feature) is input into the third dedicated branch, and the fourth dimension (main frequency feature) is input into the fourth dedicated branch, achieving accurate matching between feature dimensions and branches.

[0084] Four dedicated branch layers are connected to optimized temporal convolutional layers. The temporal convolutional layer of the first dedicated branch receives phase feature data with an input dimension of [1, T, 1]. After being processed by temporal convolutional operations with kernel sizes of 3, 5, or 7, the output dimension is [1, T, 32]. The temporal convolutional layers of the second to fourth dedicated branches receive fluctuation feature, temperature gradient feature, and main frequency feature data, respectively, with an input dimension of [1, T, 1]. After being processed by their respective optimized convolutional kernels, the output dimension is [1, T, 32], ensuring that the output feature dimensions of each branch are consistent.

[0085] The channel attention layer incorporates an optimized channel attention mechanism. The four feature channels of this mechanism are connected to the outputs of four dedicated branches: Channel 1 receives the [1, T, 32]-dimensional phase features from the first dedicated branch; Channel 2 receives the [1, T, 32]-dimensional fluctuation features from the second dedicated branch; Channel 3 receives the [1, T, 32]-dimensional temperature gradient features from the third dedicated branch; and Channel 4 receives the [1, T, 32]-dimensional dominant frequency features from the fourth dedicated branch. After weight adjustment and feature weighting by the attention mechanism, a [1, T, 32]-dimensional fused feature is output. The output layer integrates the output features of the channel attention layer using a channel splicing operation, converting the [1, T, 32]-dimensional fused feature into a standard [1, T, 32]-dimensional output format, forming a unified feature output interface. The interface output data contains multimodal feature information enhanced by temporal convolution and weighted by temporal attention. Through the above hierarchical connections and feature transfer, an optimized multi-branch feature extraction network is constructed.

[0086] In this embodiment of the invention, four key feature dimensions—phase, fluctuation, temperature gradient, and main frequency—are selected from the preprocessed multimodal operating parameters. This focuses on the core characterizing features of the motor's operating state, reducing redundant interference from non-critical features and improving the targeting and efficiency of subsequent data processing. A four-dimensional feature association domain is constructed based on the numerical distribution of the four key feature dimensions, integrating the originally independent single-dimensional features into a multi-dimensional association structure. This preserves the collaborative association information between different features, providing a structured feature space for subsequent density analysis. The four-dimensional feature association domain is uniformly divided into multiple grid cells, discretizing the continuous multi-dimensional feature distribution into regular structured cells. This makes the statistical analysis of feature distribution density easier to operate, while ensuring the consistency and repeatability of the segmentation results. The feature distribution density within each grid cell is statistically analyzed, and a weight coefficient is calculated. This allows the weight coefficient to be directly linked to the actual distribution density of the features, enabling the weight to objectively reflect the differences in feature density within different grid cells and highlighting the importance of high feature density regions. The grid density weight is obtained based on the spatial positional relationship between the weight coefficient and the grid cells. The spatial distribution information of the features is incorporated into the density weight, making the weight representation more comprehensive.

[0087] The grid density weights are decomposed into convolutional kernel size parameters and channel attention weight allocation parameters, transforming the abstract density weights into concrete structural parameters that can be directly called by the multi-branch feature extraction network, thus building a bridge between data features and network configuration. Based on the convolutional kernel size parameters, the kernel size of the temporal convolutional layers is configured to adapt to the density characteristics of the feature distribution, improving the adaptability of temporal feature extraction. Based on the channel attention weight allocation parameters, the feature weight ratio of the channel attention mechanism is reconstructed, tilting attention resources towards channels with high feature density, strengthening the focus on key feature channels, and improving the effectiveness of the attention mechanism. Based on the optimized temporal convolutional layers and channel attention mechanism, an optimized multi-branch feature extraction network is constructed, ensuring that each key part of the network structure is adapted to the feature distribution characteristics, forming a data-driven network optimization closed loop, and improving the network's ability to process multimodal features.

[0088] In a preferred embodiment of the present invention, step 400 above, which extracts time-domain and frequency-domain feature vectors from the preprocessed multimodal operating parameter set through an optimized multi-branch feature extraction network, includes:

[0089] Step 401: Input the preprocessed multimodal operating parameter set into the optimized multi-branch feature extraction network to obtain the initial features of each operating parameter. Specifically, the preprocessed multimodal operating parameter set is time-series segmented data. Each time-series segment contains standardized data of four dimensions: phase features, fluctuation features, temperature gradient features, and dominant frequency features. The format of a single segment is [1, 60, 4], where 1 is the number of samples, 60 is the time series length, and 4 is the feature dimension. Input the parameter set into the optimized multi-branch feature extraction network in batches according to time order. Each batch contains 5 consecutive time-series segments, and the total input dimension is [5, 60, 4].

[0090] The network input layer performs feature splitting on the batch data, allocating the first-dimensional phase feature data to the first dedicated branch, the second-dimensional fluctuation feature data to the second dedicated branch, the third-dimensional temperature gradient feature data to the third dedicated branch, and the fourth-dimensional main frequency feature data to the fourth dedicated branch. The input layer of each dedicated branch performs dimensional expansion on the received single-dimensional data, converting the [5, 60, 1]-dimensional data into [5, 60, 16]-dimensional data, and achieving channel expansion through a 1×1 convolution kernel, which serves as the initial input feature for each branch.

[0091] After optimization by each dedicated branch, the temporal convolutional layer performs preliminary processing and outputs the initial features of each operating parameter: the first dedicated branch outputs the initial features of the phase feature with dimensions [5, 60, 32]; the second dedicated branch outputs the initial features of the fluctuation feature with dimensions [5, 60, 32]; the third dedicated branch outputs the initial features of the temperature gradient feature with dimensions [5, 60, 32]; and the fourth dedicated branch outputs the initial features of the dominant frequency feature with dimensions [5, 60, 32]. The initial features retain the original temporal variation trend of each parameter, while enriching the expression dimension of the features through channel expansion.

[0092] Step 402: Based on the initial features, temporal features are extracted through the optimized temporal convolutional layer in the optimized multi-branch feature extraction network to obtain a temporal feature vector. Specifically, this includes: based on the initial features of each parameter, temporal features are extracted through the optimized temporal convolutional layer: the temporal convolutional layer of the first dedicated branch uses a 3×1 convolutional kernel. To address the low fluctuation characteristics of the phase features, a sliding convolution is performed on the initial phase features, with a convolution stride of 1. Zero padding is used to maintain the temporal length, outputting phase temporal features of dimensions [5, 60, 32]. The temporal convolutional layer of the second dedicated branch uses... Using a 3×1 convolution kernel, the initial fluctuating features are convolved to target the moderate rate of change of the fluctuating features, outputting fluctuating temporal features of [5, 60, 32] dimensions. The temporal convolutional layer of the third dedicated branch uses a 5×1 convolution kernel (targeting the significant changing trend of the temperature gradient features) to convolve the initial temperature gradient features, outputting temperature gradient temporal features of [5, 60, 32] dimensions. The temporal convolutional layer of the fourth dedicated branch uses a 5×1 convolution kernel to target the high-frequency fluctuating characteristics of the dominant frequency features, convolving the initial dominant frequency features, outputting dominant frequency temporal features of [5, 60, 32] dimensions.

[0093] Each dedicated branch contains two optimized temporal convolutional layers. The output features of the first layer are processed by the ReLU activation function and then input into the second layer. The kernel size of the second layer is the same as that of the first layer, and the output feature dimension remains [5, 60, 32]. The output features of the second layer of the four branches are concatenated according to the channel dimension to obtain a fused temporal feature vector with dimensions [5, 60, 128]. This vector integrates the temporal variation details of the four parameters, including the temporal evolution of the phase difference, the continuous law of the rotation speed fluctuation, the spatial diffusion trend of the temperature gradient, and the frequency drift process of the main frequency.

[0094] Step 403: Based on the time-domain feature vector, frequency-domain feature extraction is performed through the optimized channel attention mechanism in the optimized multi-branch feature extraction network to obtain a frequency-domain feature vector. Specifically, this includes: firstly, performing a local Fourier transform on the fused time-domain feature vector to convert the feature value of each time-series position into a corresponding frequency component, resulting in a time-frequency feature matrix of dimensions [5, 60, 128, 32], where 32 represents the number of frequency bins; the four feature channels of the channel attention mechanism focus on different frequency ranges respectively, with channel 1 focusing on the low-frequency component from 0 to 50 Hz, corresponding to the fundamental frequency change of the phase feature; channel 2 focusing on the mid-frequency component from 50 to 200 Hz, corresponding to the harmonics of the rotational speed fluctuation; channel 3 focusing on the mid-high frequency component from 200 to 500 Hz, corresponding to the transmission characteristics of the temperature gradient; and channel 4 focusing on the high-frequency component from 500 to 2000 Hz, corresponding to the sideband of the vibration main frequency.

[0095] The weights of the mid-to-high frequency channels corresponding to temperature gradient features are 0.35, the weights of the high frequency channels corresponding to the main frequency features are 0.25, the weights of the low frequency channels corresponding to phase features are 0.2, and the weights of the mid-frequency channels corresponding to fluctuation features are 0.2. The time-frequency feature matrices of each channel are compressed into frequency feature vectors of [5, 1, 128, 32] dimensions by global average pooling. After multiplying with the corresponding channel weights, the vectors are concatenated according to the frequency dimension to obtain frequency domain feature vectors of [5, 60, 128]. This frequency domain feature vector highlights the high-frequency sideband characteristics of the vibration main frequency during bearing wear, the mid-to-high frequency transmission characteristics of the temperature gradient during winding overheating, and the frequency coordination change law when various parameters are abnormal, providing multi-band feature information for subsequent feature fusion.

[0096] In a preferred embodiment of the present invention, step 500 involves fusing the time-domain and frequency-domain feature vectors to form a fused feature vector; based on the fused feature vector, identifying abnormal features in the operating parameters and analyzing their temporal evolution patterns to obtain abnormal feature patterns, including:

[0097] Step 501 involves concatenating the time-domain feature vector and the frequency-domain feature vector to construct an initial fused feature matrix. Specifically, the time-domain feature vector is the [5, 60, 128]-dimensional data output in step 402, containing the temporal variation details of the four parameters: phase, fluctuation, temperature gradient, and dominant frequency. The frequency-domain feature vector is the [5, 60, 128]-dimensional data output in step 403, containing the frequency distribution characteristics of the above parameters in the 0 to 2000 Hz frequency band.

[0098] The time-domain feature vector and the frequency-domain feature vector are superimposed according to the last channel dimension, that is, the 128 channels of the time-domain feature and the 128 channels of the frequency-domain feature are arranged in sequence to form an initial fusion feature matrix of [5, 60, 256] dimensions; in this matrix, the first 128 channels retain the dynamic change information of each parameter over time, and the last 128 channels retain the energy distribution information of each parameter in different frequency bands, so as to realize the complementary integration of time-domain and frequency-domain features.

[0099] Step 502 involves performing principal component analysis on the initial fusion feature matrix to extract the main feature components and form a dimensionality-reduced fusion feature vector. Specifically, this includes: performing principal component analysis on the initial fusion feature matrix: first, converting the [5, 60, 256]-dimensional matrix into a two-dimensional data format, i.e., expanding it by sample × feature into a matrix of [5 × 60, 256] = [300, 256], where each row represents a 256-dimensional feature of a time series point; calculating the covariance matrix of this two-dimensional matrix, obtaining 256 eigenvalues ​​and corresponding eigenvectors through eigenvalue decomposition; sorting the eigenvalues ​​from largest to smallest, selecting the eigenvectors corresponding to the top K eigenvalues ​​with a cumulative variance contribution rate of 95%, for example, K = 64, to form the principal component transformation matrix.

[0100] The expanded two-dimensional matrix is ​​multiplied by the principal component transformation matrix to obtain dimensionality-reduced data of [300, 64] dimensions; this data is then restored to its temporal structure to obtain a dimensionality-reduced fused feature vector of [5, 60, 64] dimensions. This vector retains 95% of the original information while reducing the feature dimension from 256 to 64, reducing data redundancy and improving the computational efficiency of subsequent anomaly detection.

[0101] Step 503: Based on the dimensionality-reduced fused feature vector, calculate the confidence score of each anomalous feature using a multilayer perceptron network in the preset anomaly detection model. Specifically, the preset anomaly detection model is a 3-layer multilayer perceptron network, constructed as follows: The input layer contains 64 neurons, matching the dimension of the dimensionality-reduced fused feature vector. Neuron weights are initialized using the Xavier initialization method, and bias terms are initialized to 0. The first hidden layer contains 128 neurons, connected in a fully connected manner. Weights are also initialized using the Xavier initialization method, bias terms are initialized to 0, and the activation function... ReLU is chosen to enhance nonlinear feature extraction capability; the second hidden layer contains 64 neurons and is fully connected to the first layer. The weights are initialized in the same way as before, the bias term is 0, and the activation function is LeakyReLU with a negative slope of 0.01 to avoid gradient vanishing caused by the neuron output being 0; the output layer contains 3 neurons and is fully connected to the second layer. The weights are initialized with a normal distribution with a mean of 0 and a standard deviation of 0.01. The bias term is 0, and the activation function is Sigmoid to map the output to the range of 0 to 1, corresponding to three abnormality types: current fluctuation exceeding the threshold, sudden temperature rise, and abnormal vibration frequency.

[0102] The training data uses historical operating data of high-speed motors under various operating conditions, including 1000 sets of normal operating condition data and 800 sets of abnormal operating condition data. The 1000 sets of normal operating condition data include rated speed, variable load, and start / stop process. The 800 sets of abnormal operating condition data include 200 sets of current fluctuation exceeding the threshold, 300 sets of temperature surge, and 300 sets of vibration frequency abnormality. The historical data is processed according to steps 100 to 502 to generate a dimensionality-reduced fusion feature vector, which is used as the model input. Each set of input data is labeled with a corresponding abnormality label: the label for normal operating condition data is [0, 0, 0], the label for current fluctuation exceeding the threshold is [1, 0, 0], the label for temperature surge is [0, 1, 0], and the label for vibration frequency abnormality is [0, 0, 1].

[0103] The batch size was set to 32, the number of iterations to 200, the initial learning rate to 0.001, and the learning rate was decayed to 0.5 every 50 iterations. The loss function used was binary cross-entropy loss, which measures the difference between the confidence score of the model output and the label. The optimizer used was the Adam optimizer, with momentum parameters β1 set to 0.9, β2 set to 0.999, and weight decay coefficient set to 0.0001 to suppress overfitting.

[0104] During training, the model performance is evaluated every 10 iterations using a validation set (accounting for 20% of the training data). If the validation loss does not decrease after 20 consecutive iterations, training is terminated early. After training is completed, the model parameters are saved to obtain the preset anomaly detection model.

[0105] The fused feature vector of dimensions [5, 60, 64] obtained in step 502 is input into the trained multilayer perceptron network one by one according to the time sequence. Each time sequence is passed through the input layer to the first hidden layer, then through the ReLU activation layer to the second hidden layer, then through the LeakyReLU activation layer to the output layer, and finally through the Sigmoid activation layer to output three confidence scores, which correspond to the probability of three abnormal features respectively. For example, if the output score of a certain time sequence is 0.12, 0.89, 0.05, it means that the confidence of the temperature rise at that moment is 0.89, and the confidence of the current fluctuation and vibration frequency abnormality is low.

[0106] Step 504: Based on the comparison between the confidence score and the preset threshold, identify the characteristics of current fluctuation exceeding the threshold, temperature surge, and abnormal vibration frequency. Specifically, this includes: setting the confidence thresholds for the three abnormal characteristics: the threshold for current fluctuation exceeding the threshold is set to 0.7, which is determined based on the maximum confidence of current fluctuation under 100 sets of normal operating conditions; the threshold for temperature surge is set to 0.65, which is determined based on the confidence distribution of temperature changes under normal operating conditions; and the threshold for abnormal vibration frequency is set to 0.75, which is determined based on the confidence analysis of the normal vibration frequency range.

[0107] For each time series point output in step 503, the three confidence scores are compared with the corresponding thresholds: if the current fluctuation score is greater than or equal to 0.7, the time series point is marked as having a current fluctuation exceeding the threshold feature; if the temperature rise score is greater than or equal to 0.65, the temperature rise feature is marked as having a temperature rise feature; if the vibration frequency anomaly score is greater than or equal to 0.75, the vibration frequency anomaly feature is marked as having a vibration frequency anomaly feature. After traversing all time series points, a list of labels containing the times when each anomaly feature appears is obtained. For example, in the five time series segments, multiple time series points in the second and third segments are labeled as temperature rises, and the fourth segment is labeled as vibration frequency anomalies, thus achieving accurate identification of different types of anomaly features.

[0108] Step 505: Based on the aforementioned abnormal features, analyze the duration and intensity trends of the abnormal features using a sliding time window to obtain an abnormal feature pattern containing temporal characteristics. Specifically, this includes: using a sliding time window consistent with step 204, with a window size of 60 time points and a sliding step of 30 time points, performing temporal evolution analysis on the abnormal features marked in step 504: for each window, count the number of time points marked as the same abnormal feature within the window, calculate the abnormal duration (number × 10 ms, where 10 ms is the time interval of a single time point); simultaneously calculate the average confidence score of the abnormal feature within the window as an abnormal intensity index.

[0109] The abnormal duration and intensity trends of continuous windows are tracked. For example, the temperature surge feature has a duration of 100ms and an intensity of 0.68 in window 1, a duration of 300ms and an intensity of 0.79 in window 2, and a duration of 500ms and an intensity of 0.85 in window 3, showing a trend of increasing duration and intensity. The vibration frequency anomaly feature has a duration of 50ms and an intensity of 0.76 in window 4, and does not appear in window 5, showing a transient feature. The types of abnormal features, duration change curves, and intensity change curves are integrated to form an abnormal feature pattern that includes time-series characteristics, such as a temperature surge, duration increase, and intensity increase pattern, and a vibration frequency anomaly, instantaneous occurrence pattern, providing a time-series dynamic basis for distinguishing transient interference from real faults.

[0110] In a preferred embodiment of the present invention, step 600 above, which generates a comprehensive evaluation index of motor status based on abnormal feature patterns and matches it with a pre-stored fault feature library to obtain the motor fault type and status level, includes:

[0111] Step 601: Based on the duration and intensity change trends of the abnormal features in the aforementioned abnormal feature patterns, calculate the temporal evolution weight of each abnormal feature. Specifically, this includes: the abnormal feature patterns include three types: current fluctuation exceeding the threshold, sudden temperature rise, and abnormal vibration frequency. Each type corresponds to a set of duration change curves and intensity change curves, where the unit of the duration change curve is ms, and the value range of the intensity change curve is 0 to 1. When calculating the temporal evolution weight of each abnormal feature, the duration is first normalized: the longest duration of the abnormal feature in the sliding window analysis is taken as the benchmark value, and the duration of each window is divided by the benchmark value to obtain a value in the range of 0 to 1. The duration coefficient is calculated first; then the intensity change trend coefficient is calculated: by linearly fitting the intensity change curve, the slope k is obtained. If k is greater than 0, the intensity shows an increasing trend, and the trend coefficient is set to 1.2. If k=0, the trend coefficient is set to 1.0. If k is less than 0, the trend coefficient is set to 0.8. Finally, the weight value of each anomalous feature is calculated according to the time evolution weight = duration coefficient × 0.6 + intensity change trend coefficient × 0.4. The longer the duration and the greater the intensity of the anomalous feature, the greater its weight value. For example, the duration coefficient of the temperature rise feature is 0.8 and the trend coefficient is 1.2, and its weight is 0.8 × 0.6 + 1.2 × 0.4 = 0.96.

[0112] Step 602: Weighted fusion of each abnormal feature according to the time-series evolution weight to obtain a comprehensive evaluation index of motor status. Specifically, this includes: extracting the maximum confidence score of each abnormal feature from the output of step 503: the highest score of the current fluctuation exceeding the threshold feature is taken as S1, the highest score of the temperature rise feature is taken as S2, and the highest score of the vibration frequency abnormal feature is taken as S3; if an abnormal feature is not identified, that is, the scores of all time-series points are lower than the threshold, then its highest score is set to 0.

[0113] Based on the time-series evolution weights, the current fluctuation weight is denoted as W1, the temperature surge weight as W2, and the vibration frequency anomaly weight as W3. The motor condition comprehensive evaluation index is calculated using the weighted fusion formula: (S1×W1+S2×W2+S3×W3) / (W1+W2+W3), where the denominator is the sum of the weights to avoid index deviation caused by weight differences. The calculation result is mapped to a range of 0 to 1, where 0 indicates no anomaly and 1 indicates the highest degree of anomaly. For example, if S1=0.3, W1=0.6, S2=0.9, W2=0.96, S3=0.2, W3=0.5, then the comprehensive evaluation index is approximately 0.68: (0.3×0.6+0.9×0.96+0.2×0.5) / (0.6+0.96+0.5).

[0114] Step 603: The comprehensive evaluation index of motor condition is matched with the fault modes in the pre-stored fault feature library to obtain the similarity matching result. Specifically, this includes: selecting five typical fault types of high-speed motors, namely, early bearing wear, bearing jamming, rotor imbalance, stator winding inter-turn short circuit, and stator winding overheating; for each fault type, a fault simulation experiment is conducted in a laboratory environment. The experiment covers the motor's rated speed, 80% rated load, 50% rated load, and rapid acceleration conditions. The rated speed of the motor is 12,000 rpm, and the rapid acceleration condition is from 5,000 rpm to 15,000 rpm. Each condition is repeated 30 times.

[0115] For the multimodal operational data collected in each experiment, following the processing flow from steps 100 to 602, the following steps are performed: extracting abnormal feature combinations (e.g., in the early bearing wear experiment, abnormal vibration frequency accompanied by small current fluctuations), calculating comprehensive evaluation indicators (recording the indicator values ​​for each experiment), and analyzing time-series evolution characteristics (e.g., a sudden increase in the duration and intensity jump of abnormal vibration frequency during bearing jamming). Specifically, extracting abnormal feature combinations, such as the abnormal vibration frequency accompanied by small current fluctuations in the early bearing wear experiment, calculating comprehensive evaluation indicators involves recording the indicator values ​​for each experiment, and analyzing time-series evolution characteristics such as the sudden increase in the duration and intensity jump of abnormal vibration frequency during bearing jamming. The results of 30 experiments of the same fault type were statistically analyzed, and the item with the highest frequency of abnormal feature combination was taken as the standard abnormal feature combination of the fault. For example, in the 30 experiments of stator winding overheating, 28 of them showed a sudden temperature rise and slight current fluctuation, so it was determined as the standard combination. The 95% confidence interval of the comprehensive evaluation index was calculated as the standard index range. For example, the index value of stator winding overheating was concentrated between 0.6 and 0.85, so it was determined as the standard range. The common trend of the time evolution characteristics was taken as the standard time evolution characteristics. For example, in the 30 experiments of stator winding overheating, the sudden temperature rise showed an increasing trend in duration and intensity, so it was determined as the standard feature.

[0116] The above-mentioned standard abnormal feature combinations, standard comprehensive evaluation index ranges, and standard time-series evolution features are associated and stored to form a pre-stored fault feature library containing five typical fault modes, where mode 1 corresponds to early bearing wear, mode 2 corresponds to bearing jamming, mode 3 corresponds to rotor imbalance, mode 4 corresponds to stator winding inter-turn short circuit, and mode 5 corresponds to stator winding overheating.

[0117] The comprehensive evaluation index of motor condition (denoted as I) is compared with the standard index range of each mode in the fault feature library to calculate the similarity: if I falls within the standard range of a certain mode, the basic similarity is set to 0.8; if I is within the upper limit of 0.1 or the lower limit of -0.1, the basic similarity is calculated by subtracting the absolute value of the difference between I and the median of the range from 0.8 and then multiplying by 2. For example, if the median of the range for mode 5 is 0.725 and I = 0.88, then the similarity = 0.8 - |0.88 - 0.725| × 2 = 0.49; at the same time, the matching degree of the abnormal feature combination is combined, with 0.2 added for a complete match, 0.1 added for a partial match, and 0 added for a non-match, to obtain the final similarity of each mode. For example, if the current abnormality is a sudden temperature rise + slight current fluctuation, which is a complete match with the feature combination of mode 5, I = 0.68 falls within the range of mode 5, then the final similarity = 0.8 + 0.2 = 1.0.

[0118] Step 604: Based on the similarity matching results, determine the current state level and corresponding fault type of the motor, and obtain the motor fault type and state level. Specifically, the state level is divided into three levels: normal, slightly abnormal, and severely abnormal. Normal is when the comprehensive evaluation index is less than 0.3, slightly abnormal is when the index is less than 0.6 but greater than or equal to 0.3, and severely abnormal is when the index is greater than or equal to 0.6. Select the fault mode corresponding to the highest similarity from the similarity matching results as the current fault type. If the highest similarity is less than 0.5 and the comprehensive evaluation index is less than 0.3, it is determined to be fault-free.

[0119] For example, if the comprehensive evaluation index is 0.68, the highest similarity is 1.0, corresponding to mode 5, indicating stator winding overheating, and the index is greater than or equal to 0.6, then the current status level is determined to be severe abnormality, and the fault type is stator winding overheating. If the comprehensive evaluation index is 0.4, the highest similarity is 0.7, corresponding to mode 1, indicating early bearing wear, then the status level is slight abnormality, and the fault type is early bearing wear. Through this process, the comprehensive evaluation index is converted into specific fault types and status levels, clarifying the current abnormality degree and problem attributes of the motor.

[0120] In this embodiment of the invention, a time-series evolution weight is calculated based on the duration and intensity change trends of abnormal feature patterns. This transforms the time-series dynamic information of anomalies into quantifiable weight values, directly linking the importance of each abnormal feature to its real-time evolution degree (such as duration and intensity changes). A comprehensive evaluation index is obtained by weighted fusion of each abnormal feature according to the time-series evolution weight. This integrates information from multiple types of abnormal features and distinguishes the differences in the impact of different anomalies on the motor state through weights, forming a single evaluation dimension that comprehensively reflects the overall operating state of the motor, reducing the complexity of multi-feature dispersed analysis. The comprehensive evaluation index is then matched with a pre-stored fault feature library for similarity. Using pre-stored standard fault patterns as a reference, a standardized basis is provided for the correspondence between the evaluation index and fault types, ensuring that the matching process is based on a unified standard. Based on the similarity matching results, the state level and fault type are determined, transforming the abstract comprehensive evaluation index into specific and interpretable state information, such as fault type and severity level, clarifying the current problem and its impact on the motor.

[0121] In a preferred embodiment of the present invention, step 700 above, based on the motor fault type and state level, obtains a fault-tolerant control command, which is used to control the motor to maintain safe operation, including:

[0122] Step 701: Based on the motor fault type and state level, select the corresponding basic control strategy from the predefined fault-tolerant strategy library. Specifically, this includes: constructing the strategy library based on five typical fault types and three state levels. The five typical fault types are early bearing wear, bearing jamming, rotor imbalance, stator winding inter-turn short circuit, and stator winding overheating. The three state levels are normal, minor abnormality, and severe abnormality.

[0123] For each type of fault, different severity levels of fault states were simulated in a laboratory environment: minor abnormal states simulated the initial characteristics of the fault, such as a slight abnormality of stator winding overheating where the temperature is 5 to 10°C higher than the rated value, without a continuous upward trend; severe abnormal states simulated the characteristics of the fault development stage, such as a temperature more than 15°C higher than the rated value, and a continuous upward trend; for each simulated state, the effects of different control measures (such as adjusting speed, limiting torque, and distributing power) were tested, and control targets that could effectively alleviate the deterioration of the fault were recorded, such as suppressing temperature rise corresponding to stator winding overheating, and reducing vibration amplitude corresponding to bearing-related faults.

[0124] The baseline values ​​of the core adjustment parameters were determined through multiple sets of experiments: For the speed limit, the changes in fault characteristics at different speeds were tested, and the speed that could stabilize the fault characteristics without affecting basic operation was selected. For example, when the stator winding was severely overheated, the test found that 60% of the rated speed could effectively reduce the rate of temperature rise, so it was set as the target speed; For the torque reduction ratio, the reduction ratio was determined based on the balance between load demand and fault mitigation effect. For example, when the bearing was severely stuck, reducing the torque to 50% of the rated torque could significantly reduce vibration while meeting the minimum load requirement; For the monitoring interval, it was set according to the fault change rate. It was set to 50ms when the temperature changed rapidly and 100ms when the vibration changed slowly.

[0125] The fault type, status level, control objective, and core adjustment parameters are associated and stored to form a predefined fault-tolerant strategy library. Each strategy corresponds to a unique combination of fault type and status level. For example, the minor abnormality strategy of mode 5 (stator winding overheating) is associated with the control objective of limiting power output, and the core parameters are 80% of rated power and 50ms temperature monitoring interval. The severe abnormality strategy is associated with the control objective of forced speed reduction and cooling, and the core parameters are 60% of rated speed and 50% of rated torque.

[0126] Based on the motor fault type and state level, the corresponding strategy is called from the predefined fault-tolerant strategy library through the strategy matching algorithm: if the fault type is stator winding overheating and the state level is severe abnormality, the severe abnormality basic control strategy corresponding to mode 5 is matched; if the fault type is early bearing wear (mode 1) and the state level is slight abnormality, the slight abnormality basic control strategy corresponding to mode 1 is matched, ensuring that the basic strategy is directly related to the fault attributes (type, severity).

[0127] Step 702: Based on the severity of the state level, adaptively adjust the execution parameters of the basic control strategy to obtain an optimized control strategy. Specifically, this includes: quantifying the severity of the state level into a level coefficient: 0 for normal, 0.5 for minor anomaly, and 1.0 for severe anomaly; adaptively adjusting the core adjustment parameters of the basic control strategy according to the level coefficient: for speed limiting parameters, the adjustment formula is: adjusted parameter = basic parameter × (1 - level coefficient × 0.3). For example, in the basic strategy, if the target speed for stator winding overheating is 60% of the rated speed and the level coefficient is 1.0, and the state level is minor anomaly with a level coefficient of 0.5, then the adjusted target speed = 60. %×(1-0.5×0.3)=51% of rated speed; For torque limiting parameters, the adjustment formula is: adjusted parameter = basic parameter × (1-level coefficient × 0.2). For example, in the basic strategy, the torque limit value for bearing jamming is 70% of the rated torque. For severe abnormalities and minor abnormalities, it is adjusted to 70%×(1-0.5×0.2)=63% of the rated torque; After adjustment, the parameters are checked at the boundary to ensure that the speed is not lower than the minimum safe operating speed, such as 3000 rpm, and the torque is not lower than the minimum torque required by the current load, so as to avoid over-adjustment leading to operation interruption; After the check is passed, an optimized control strategy is formed, whose parameters change dynamically with the state level to adapt to faults of different severity.

[0128] Step 703: Based on the optimized control strategy and combined with the real-time operating parameters of the motor, a fault-tolerant control command is obtained, which specifically includes: the real-time operating parameters of the motor include the current speed, real-time torque, winding temperature, and current road load. The current speed is collected by a speed sensor with an accuracy of ±10 rpm, the real-time torque is calculated by a current sensor with an accuracy of ±2% of the rated torque, the winding temperature is collected by a thermocouple with an accuracy of ±1℃, and the current road load is obtained from the vehicle CAN bus, such as climbing load and flat road load.

[0129] If the optimization strategy is forced speed reduction and cooling, and the current speed is 12000 rpm, the target speed is 7200 rpm, which is 60% of the rated speed, then the speed adjustment rate is calculated. Based on the current temperature deviation, the rate increases by 500 rpm / s for every 5°C above the upper limit of the temperature, generating a speed control command to reduce the speed from 12000 rpm to 7200 rpm at a rate of 800 rpm / s. If the optimization strategy is to limit power output, and the current power is 80kW, the upper limit of the power is 64kW, which is 80% of the rated power, then the maximum allowable current is calculated based on the real-time load, generating a current limit command that the peak value of the three-phase current does not exceed 150A.

[0130] The control command format adopts the CAN bus message format that the motor controller can recognize, which includes the command type (speed / torque / power), target value, adjustment rate, and effective timestamp, ensuring that the controller can directly parse and execute it.

[0131] Step 704 involves executing the fault-tolerant control command through the motor controller to adjust the motor torque, speed, or power, thereby obtaining the adjusted real-time operating state of the motor. Specifically, this includes: after receiving the fault-tolerant control command, the motor controller converts the command into specific execution signals through an internal control algorithm. For speed control commands, the controller adjusts the PWM (Pulse Width Modulation) duty cycle of the inverter to change the voltage frequency of the input motor, thereby reducing the speed at a set rate. For example, when reducing from 12000 rpm to 7200 rpm, the PWM frequency gradually decreases from 200 Hz to 120 Hz. For torque limiting commands, the controller limits the dq axis current amplitude through current closed-loop control to ensure that the output torque does not exceed the limit value. For example, when the torque is limited to 63% of the rated torque, the dq axis current is limited to within 63% of the rated current.

[0132] During the adjustment process, the controller collects motor feedback signals in real time, such as actual speed and output torque, and compares them with the command target values. It eliminates the deviation through PID (proportional, integral, derivative) adjustment. For example, when the deviation between the actual speed and the target speed exceeds 500 rpm, the PWM duty cycle adjustment is increased. After the adjustment is completed, the controller outputs the real-time operating status of the motor after adjustment, such as the current speed of 7200 rpm, torque of 315 N·m, and winding temperature of 85℃.

[0133] Step 705: Based on the adjusted real-time operating state of the motor, the motor state changes are monitored in real time by sensors, and the parameters of the fault-tolerant control command are dynamically adjusted according to the feedback of the state changes. Specifically, the sensors monitor the adjusted motor state changes in real time, including collecting winding temperature, vibration amplitude, and three-phase current waveforms every 10ms, and calculating the temperature change rate, vibration amplitude change, and current fluctuation coefficient. The temperature change rate is in °C / s, the vibration amplitude change is in mm / s, and the current fluctuation coefficient is the ratio of the actual fluctuation value to the rated fluctuation value.

[0134] Based on the monitoring results, adjust the fault-tolerant control command parameters: if the stator winding overheats and the temperature change rate remains positive after adjustment, and the temperature continues to rise, further reduce the upper limit of power from 64kW to 56kW, i.e., 70% of the rated power, and reduce the target speed from 7200rpm to 6000rpm; if the vibration amplitude change is less than 0 after early bearing wear adjustment, i.e., the vibration is weakened, maintain the current torque limit value, and extend the monitoring interval to 100ms; if the monitored state returns to normal, the temperature drops to a safe range, and the vibration amplitude returns to the rated value, gradually release the restrictions, increase the power at a rate of 5% of the rated power / second, until the normal control strategy is restored.

[0135] The adjusted parameters are updated to the fault-tolerant control command and resent to the motor controller via the CAN bus, forming a closed-loop control of command execution, status monitoring, and parameter adjustment, ensuring that the fault-tolerant control continuously adapts to changes in motor status.

[0136] In this embodiment of the invention, a basic control strategy is selected from a predefined fault-tolerant strategy library based on the motor fault type and state level. Relying on the correspondence between fault attributes (type, level) and strategies in the predefined library, the basic strategy is ensured to accurately match the current fault characteristics, providing a targeted framework for subsequent control. The execution parameters of the basic control strategy are adaptively adjusted according to the severity of the state level, allowing the parameters to dynamically adapt to the severity of the anomaly, improving the strategy's adaptability to faults of different severity levels. Fault-tolerant control commands are obtained based on the optimized control strategy combined with the motor's real-time operating parameters. The optimized strategy is integrated with the motor's current actual operating conditions (such as real-time speed and load), ensuring that the commands fit the real-time operating state and improving the real-time effectiveness of the commands. The motor controller executes the fault-tolerant control commands to adjust the motor torque, speed, or power, transforming the optimized strategy into specific and operable motor adjustment actions that directly act on the motor's operation, quickly responding to fault situations and providing practical operational support for maintaining safe operation. Based on the feedback of the adjusted motor state monitoring results, the control command parameters are dynamically adjusted, forming a closed-loop mechanism of execution, monitoring, and adjustment. This mechanism can continuously optimize parameters according to changes in motor state, ensuring long-term adaptation to the motor's operating state and maintaining safe operation.

[0137] like Figure 2 As shown, embodiments of the present invention also provide an online condition assessment system for a high-speed motor, comprising:

[0138] The acquisition module is used to acquire the three-phase stator current, rotor speed, stator winding temperature and housing vibration frequency of the high-speed motor in real time, extract its phase, fluctuation, temperature gradient and main frequency characteristics and fuse them to obtain a multi-mode operating parameter set;

[0139] The preprocessing module is used to preprocess the multimodal operating parameter set to obtain a preprocessed multimodal operating parameter set.

[0140] The feature extraction module is used to select four key feature dimensions from the preprocessed multimodal operating parameter set, and construct a four-dimensional feature association domain based on the numerical distribution of the four key feature dimensions; the four-dimensional feature association domain is divided into grids, and grid density weights are obtained according to the feature distribution density within the grids; the grid density weights are decomposed into network structure parameters to configure the kernel size of temporal convolution and the feature weight ratio of the channel attention mechanism in the multi-branch feature extraction network, thereby constructing an optimized multi-branch feature extraction network; the optimized multi-branch feature extraction network extracts temporal and frequency domain feature vectors from the preprocessed multimodal operating parameter set.

[0141] The identification and evaluation module is used to fuse time-domain and frequency-domain feature vectors to form a fused feature vector; based on the fused feature vector, it identifies abnormal features in the operating parameters and analyzes their time-series evolution to obtain abnormal feature patterns; based on the abnormal feature patterns, it generates a comprehensive evaluation index of motor status and matches it with a pre-stored fault feature library to obtain the motor fault type and status level.

[0142] The fault-tolerant control module is used to obtain fault-tolerant control instructions based on the motor fault type and status level. These instructions are used to control the motor to maintain safe operation.

[0143] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0144] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0145] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for online condition assessment of a high-speed motor, characterized in that, The method includes: The three-phase stator current, rotor speed, stator winding temperature and shell vibration frequency of the high-speed motor are collected in real time. The phase, fluctuation, temperature gradient and main frequency characteristics are extracted and fused to obtain a multi-mode operating parameter set. The multimodal operating parameter set is preprocessed to obtain a preprocessed multimodal operating parameter set; Four key feature dimensions are selected from the preprocessed multimodal operating parameters. Based on the numerical distribution of the four key feature dimensions, a four-dimensional feature association domain is constructed. The four-dimensional feature association domain is divided into grids, and the grid density weights are obtained according to the feature distribution density within the grids. The grid density weights are decomposed into network structure parameters to configure the kernel size of temporal convolution and the feature weight ratio of channel attention mechanism in the multi-branch feature extraction network, thereby constructing an optimized multi-branch feature extraction network. The optimized multi-branch feature extraction network extracts time-domain and frequency-domain feature vectors from the preprocessed multimodal operating parameter set. The time-domain and frequency-domain feature vectors are fused to form a fused feature vector; based on the fused feature vector, abnormal features in the operating parameters are identified and their temporal evolution patterns are analyzed to obtain abnormal feature patterns; Based on abnormal feature patterns, a comprehensive evaluation index for motor status is generated and matched with a pre-stored fault feature library to obtain the motor fault type and status level. Based on the motor fault type and status level, a fault-tolerant control command is obtained, which is used to control the motor to maintain safe operation.

2. The online condition assessment method for high-speed motors according to claim 1, characterized in that, Real-time acquisition of three-phase stator current, rotor speed, stator winding temperature, and housing vibration frequency of the high-speed motor; extraction and fusion of phase, fluctuation, temperature gradient, and main frequency characteristics to obtain a multi-modal operating parameter set, including: The instantaneous values ​​of the three-phase stator current are acquired by current sensors, and their phase characteristics are calculated based on the instantaneous values. Real-time rotor speed data is collected synchronously using a speed sensor, and its fluctuation characteristics are extracted based on the real-time data. Temperature distribution data of the stator winding is collected by a temperature sensor, and its temperature gradient characteristics are calculated based on the distribution data. The vibration frequency spectrum data of the shell is collected by a vibration sensor, and its dominant frequency characteristics are extracted based on the spectrum data. The phase characteristics, fluctuation characteristics, temperature gradient characteristics, and dominant frequency characteristics are fused to form a multimodal operating parameter set with a unified timestamp.

3. The online condition assessment method for high-speed motors according to claim 2, characterized in that, The multimodal operating parameter set is preprocessed to obtain a preprocessed multimodal operating parameter set, including: Convex hull boundary calculations are performed on the phase characteristics, fluctuation characteristics, temperature gradient characteristics, and dominant frequency characteristics of the multimodal operating parameter set to construct a convex hull boundary point set; Based on the set of convex hull boundary points, the relative distance from each feature point to the convex hull boundary is calculated; outlier detection and removal are performed based on the relative distance and a preset distance threshold to obtain a cleaned multimodal operating parameter set; Based on the cleaned multimodal operating parameter set, the extreme points in the convex hull boundary point set are extracted to determine the actual numerical boundaries of each feature dimension; based on the actual numerical boundaries, the normalization scaling parameters of each feature dimension are calculated; using the normalization scaling parameters, the cleaned multimodal operating parameter set is mapped to a unified numerical range to obtain a standardized multimodal operating parameter set. An initial time-series window is established based on a standardized multimodal operating parameter set. The window scaling factor of the initial time-series window is calculated based on the spatial distribution range of the convex hull boundary point set. The window size of the initial time-series window is updated using the window scaling factor to obtain the target time-series window. Based on the target time-series window, the standardized multimodal operating parameter set is time-series segmented to obtain the preprocessed multimodal operating parameter set.

4. The online condition assessment method for high-speed motors according to claim 3, characterized in that, Four key feature dimensions are selected from the preprocessed multimodal operating parameters set. Based on the numerical distribution of these four key feature dimensions, a four-dimensional feature association domain is constructed. The four-dimensional feature association domain is then divided into grids, and grid density weights are obtained based on the feature distribution density within the divided grids, including: Phase characteristics, fluctuation characteristics, temperature gradient characteristics, and dominant frequency characteristics were selected as four key feature dimensions from the preprocessed multimodal operating parameter set. Based on the numerical distribution of the four key feature dimensions, a four-dimensional feature association domain is constructed; The four-dimensional feature association domain is uniformly divided into multiple grid cells; The feature distribution density within each grid cell is statistically analyzed, and the weight coefficient of each grid cell is calculated based on the feature distribution density. The grid density weight is obtained based on the spatial relationship between the weight coefficient and the grid cell.

5. The online condition assessment method for high-speed motors according to claim 4, characterized in that, The grid density weights are decomposed into network structure parameters to configure the kernel size of temporal convolutions and the proportion of feature weights in the channel attention mechanism in the multi-branch feature extraction network, thereby constructing an optimized multi-branch feature extraction network, including: The grid density weights are decomposed into convolution kernel size parameters and channel attention weight allocation parameters; Based on the kernel size parameter, the kernel size of the temporal convolutional layer in the multi-branch feature extraction network is configured to obtain the optimized temporal convolutional layer. Based on the channel attention weight allocation parameters, the feature weight ratio of the channel attention mechanism in the multi-branch feature extraction network is reconstructed to obtain the optimized channel attention mechanism. An optimized multi-branch feature extraction network is constructed based on the optimized temporal convolutional layers and channel attention mechanism.

6. The online condition assessment method for a high-speed motor according to claim 5, characterized in that, The optimized multi-branch feature extraction network extracts time-domain and frequency-domain feature vectors from the preprocessed multimodal operating parameter set, including: The preprocessed multimodal operating parameter set is input into the optimized multi-branch feature extraction network to obtain the initial features of each operating parameter; Based on the initial features, temporal features are extracted through the optimized temporal convolutional layer in the optimized multi-branch feature extraction network to obtain a temporal feature vector. Based on the time-domain feature vector, frequency-domain feature extraction is performed through the optimized channel attention mechanism in the optimized multi-branch feature extraction network to obtain the frequency-domain feature vector.

7. The online condition assessment method for a high-speed motor according to claim 6, characterized in that, Feature fusion is performed on the time-domain and frequency-domain feature vectors to form a fused feature vector; based on By fusing feature vectors, abnormal features in operating parameters are identified and their temporal evolution patterns are analyzed to obtain abnormal feature patterns, including: The time-domain feature vector and the frequency-domain feature vector are concatenated to construct an initial fused feature matrix; Principal component analysis is performed on the initial fusion feature matrix to extract the main feature components and form a dimensionality-reduced fusion feature vector; Based on the dimensionality-reduced fused feature vector, the confidence score of each abnormal feature is calculated through the multilayer perceptron network in the preset anomaly detection model. Based on the comparison between the confidence score and the preset threshold, characteristics such as current fluctuation exceeding the threshold, sudden temperature rise, and abnormal vibration frequency are identified. Based on the aforementioned abnormal features, the duration and intensity of the abnormal features are analyzed using a sliding time window to obtain an abnormal feature pattern that includes temporal characteristics.

8. The online condition assessment method for a high-speed motor according to claim 7, characterized in that, A comprehensive motor condition assessment index is generated based on abnormal feature patterns and matched with a pre-stored fault feature library to obtain the motor fault type and condition level, including: Based on the variation trends of the duration and intensity of the abnormal features in the aforementioned abnormal feature patterns, the temporal evolution weight of each abnormal feature is calculated. The abnormal features are weighted and fused according to the time-series evolution weights to obtain a comprehensive evaluation index of motor status; The motor condition comprehensive evaluation index is matched with the fault patterns in the pre-stored fault feature library to obtain the similarity matching result. Based on the similarity matching results, the current state level of the motor and the corresponding fault type are determined, thus obtaining the motor fault type and state level.

9. The online condition assessment method for a high-speed motor according to claim 8, characterized in that, Based on the motor fault type and status level, fault-tolerant control commands are obtained. These commands are used to control the motor to maintain safe operation, including: Based on the motor fault type and status level, select the corresponding basic control strategy from the predefined fault-tolerant strategy library; Based on the severity of the state level, the execution parameters of the basic control strategy are adaptively adjusted to obtain an optimized control strategy; Based on the optimized control strategy and combined with the real-time operating parameters of the motor, fault-tolerant control commands are obtained. The fault-tolerant control command is executed by the motor controller to adjust the motor torque, speed or power, and obtain the real-time operating status of the adjusted motor. Based on the adjusted real-time operating status of the motor, the changes in the motor status are monitored in real time by sensors, and the parameters of the fault-tolerant control command are dynamically adjusted according to the feedback of the status changes.

10. An online condition assessment system for a high-speed motor, the system implementing the method as described in any one of claims 1 to 8, characterized in that, include: The acquisition module is used to acquire the three-phase stator current, rotor speed, stator winding temperature and housing vibration frequency of the high-speed motor in real time, extract its phase, fluctuation, temperature gradient and main frequency characteristics and fuse them to obtain a multi-mode operating parameter set; The preprocessing module is used to preprocess the multimodal operating parameter set to obtain a preprocessed multimodal operating parameter set. The feature extraction module is used to select four key feature dimensions from the preprocessed multimodal operating parameters set, and construct a four-dimensional feature association domain based on the numerical distribution of the four key feature dimensions; the four-dimensional feature association domain is divided into grids, and grid density weights are obtained according to the feature distribution density within the grids; the grid density weights are decomposed into network structure parameters to configure the kernel size of temporal convolution and the feature weight ratio of channel attention mechanism in the multi-branch feature extraction network, thereby constructing an optimized multi-branch feature extraction network; The optimized multi-branch feature extraction network extracts time-domain and frequency-domain feature vectors from the preprocessed multimodal operating parameter set. The identification and evaluation module is used to fuse time-domain and frequency-domain feature vectors to form a fused feature vector; based on By fusing feature vectors, abnormal features in operating parameters are identified and their temporal evolution patterns are analyzed to obtain abnormal feature patterns. Based on the abnormal feature patterns, a comprehensive evaluation index of motor status is generated and matched with a pre-stored fault feature library to obtain the motor fault type and status level. The fault-tolerant control module is used to obtain fault-tolerant control instructions based on the motor fault type and status level. These instructions are used to control the motor to maintain safe operation.

Images