Electric drive transmission system comprehensive life prediction method based on AI

Abstract

The invention relates to the technical field of electric drive transmission system service life prediction, and discloses an AI-based electric drive transmission system comprehensive service life prediction method, which comprises the following steps: acquiring electric drive system operation parameters and mechanical part degradation data; arranging the data into a system operation full-period input tensor and a component degradation full-period input tensor, and respectively extracting multi-order feature tensors through a bidirectional recurrent neural network and a gated time convolutional network; performing cross-modal fusion by using a cross attention fusion network in combination with physical constraints to obtain a system-component degradation fusion feature vector; and after dimension reduction, inputting a multi-layer perceptron network output life prediction value, and generating a residual service life interval with confidence in combination with historical cases. The method integrates multi-source data, realizes cross-domain feature dynamic association and physical constraint fusion, improves the precision and reliability of life prediction, is suitable for intelligent operation and maintenance of the electric drive system, and provides a scientific basis for equipment maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of life prediction of electric drive transmission systems, and specifically to a comprehensive life prediction method for electric drive transmission systems based on AI. Background Art

[0002] With the rapid development of new energy technologies, electric drive transmission systems are increasingly used in electric vehicles, industrial automation equipment and other fields. As a core power component, its life directly affects the reliability, safety and maintenance cost of the equipment. However, the operating environment of the electric drive transmission system is complex and changeable, involving the coupling of multiple fields such as electrical and mechanical, and traditional life prediction methods face many challenges.

[0003] From the perspective of the limitations of single parameter monitoring, traditional methods usually rely only on single-dimensional analysis of electrical signals such as current and speed or mechanical parameters such as wear, and cannot fully capture the dynamic relationship between system operating status and component degradation. For example, when judging the health status of a motor only by current intensity, the impact of torque fluctuations caused by gear wear on the overall life of the system may be ignored, resulting in a large deviation between the predicted results and the actual life.

[0004] In terms of cross-domain data fusion, the electric drive system operating parameters and mechanical component degradation data belong to different physical domains, and traditional methods lack an effective cross-domain feature association mechanism. Abnormal fluctuations in system operating parameters (such as operating temperature and torque output) may indicate early degradation of mechanical components (such as bearings and gears), but traditional algorithms find it difficult to mine the potential mapping relationship between data in different domains and build a comprehensive degradation assessment model.

[0005] The lack of multi-scale time series modeling capabilities is also an important defect of traditional methods. The degradation process of the electric drive system has multi-time scale characteristics. The speed fluctuation in a short period of time may reflect the real-time load change, while the long-term wear accumulation reflects the gradual degradation trend of the components. Traditional statistical analysis based on fixed windows or simple neural network models cannot effectively capture the evolution of features at different time scales, resulting in insufficient sensitivity to early degradation features.

[0006] The problem of combining physical constraints with data-driven is also prominent. Existing data-driven models often ignore the physical operating laws of electric drive systems, such as the thermodynamic coupling relationship between motor heating and bearing lubrication status, which may generate prediction results that violate physical principles. Methods based solely on physical models are difficult to adapt to parameter drift and uncertainty under complex working conditions, and have poor prediction robustness.

[0007] In addition, the requirements for the accuracy and reliability of life prediction are continuously increasing. In industrial applications, misjudgment may lead to over-maintenance or failure of critical components, causing huge economic losses. The single life prediction value generated by traditional methods lacks confidence evaluation, cannot provide comprehensive risk quantification information for maintenance decisions, and is difficult to meet the requirements of modern intelligent operation and maintenance for the interpretability and reliability of prediction results. Summary of the Invention

[0008] The purpose of the present invention is to provide an integrated life prediction method for an electric drive transmission system based on AI to solve the problems raised in the above background technology.

[0009] To achieve the above purpose, the present invention provides the following technical solution: An integrated life prediction method for an electric drive transmission system based on AI, the method comprising: Obtain the operation parameters of the electric drive system at multiple monitoring time points within a predetermined period collected by operation sensors, and the degradation data of mechanical components at the multiple monitoring time points collected by a state detection device; Extract features and perform cross-domain association on the operation parameters at the multiple monitoring time points and the degradation data of mechanical components at the multiple monitoring time points to obtain a system-component degradation fusion feature vector; Generate a life prediction evaluation result based on the system-component degradation fusion feature vector; Among them, extracting features and performing cross-domain association on the operation parameters at the multiple monitoring time points and the degradation data of mechanical components at the multiple monitoring time points to obtain a system-component degradation fusion feature vector includes: Arrange the operation parameters at the multiple monitoring time points into a system operation full-cycle input tensor according to the time dimension and parameter type dimension, and arrange the degradation data of mechanical components at the multiple monitoring time points into a component degradation full-cycle input tensor according to the time dimension and degradation index dimension; Perform multi-scale time series modeling on the system operation full-cycle input tensor and the component degradation full-cycle input tensor respectively to obtain a system operation multi-order feature tensor and a component degradation multi-order feature tensor; Perform cross-modal fusion on the system operation multi-order feature tensor and the component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector.

[0010] Preferably, the operation parameters include current intensity, working temperature, torque output, and rotational speed fluctuation.

[0011] Preferably, the degradation data includes wear amount, material fatigue index, and lubrication state parameter.

[0012] Preferably, multi-scale temporal modeling is respectively performed on the input tensor of the entire system operation cycle and the input tensor of the entire component degradation cycle to obtain the multi-order characteristic tensor of system operation and the multi-order characteristic tensor of component degradation, including: Input the input tensor of the entire system operation cycle into a multi-level feature extraction network based on a bidirectional recurrent neural network to obtain the multi-order characteristic tensor of system operation; Input the input tensor of the entire component degradation cycle into a degradation feature extraction network based on a gated temporal convolutional network to obtain the multi-order characteristic tensor of component degradation.

[0013] Preferably, cross-modal fusion is performed on the multi-order characteristic tensor of system operation and the multi-order characteristic tensor of component degradation to obtain the system-component degradation fusion feature vector, including: Input the multi-order characteristic tensor of system operation and the multi-order characteristic tensor of component degradation into a cross-attention fusion network to obtain a system-component degradation dynamic correlation matrix; Input the system-component degradation dynamic correlation matrix into an optimization module based on physical constraints to obtain a constraint-corrected correlation matrix; Input the multi-order characteristic tensor of system operation and the multi-order characteristic tensor of component degradation into a residual feature correction network respectively to obtain an optimized multi-order characteristic tensor of system operation and an optimized multi-order characteristic tensor of component degradation; Perform feature concatenation coding on the optimized multi-order characteristic tensor of system operation and the optimized multi-order characteristic tensor of component degradation to obtain the system-component degradation fusion feature vector.

[0014] Preferably, input the multi-order characteristic tensor of system operation and the multi-order characteristic tensor of component degradation into a cross-attention fusion network to obtain a system-component degradation dynamic correlation matrix, including: Calculate the self-attention weight matrix of the multi-order characteristic tensor of system operation and the mutual-attention weight matrix of the multi-order characteristic tensor of component degradation, and generate the system-component degradation dynamic correlation matrix through matrix multiplication operation.

[0015] Preferably, input the multi-order characteristic tensor of system operation and the multi-order characteristic tensor of component degradation into a residual feature correction network respectively to obtain an optimized multi-order characteristic tensor of system operation and an optimized multi-order characteristic tensor of component degradation, including: Perform channel dimension compression on the multi-order characteristic tensor of system operation to obtain a system operation feature compression vector; Perform weighted fusion on the constraint-corrected correlation matrix and the system operation feature compression vector, and then perform a feature expansion operation to obtain an optimized multi-order characteristic tensor of system operation; The spatially attention-weighted multi-order feature tensor of the component degradation is subjected to feature modulation with the constraint correction correlation matrix to obtain an optimized multi-order feature tensor of the component degradation.

[0016] Preferably, based on the system-component degradation fusion feature vector, a life prediction evaluation result is generated, including: Inputting the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value; Generating the life prediction evaluation result including the remaining service life interval according to the life prediction value.

[0017] Preferably, inputting the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value, including: Performing multi-dimensional dimensionality reduction processing on the system-component degradation fusion feature vector to obtain a low-dimensional feature coding vector; Inputting the low-dimensional feature coding vector into a multi-layer perceptron network for non-linear relationship mapping, and outputting the life prediction value associated with the time series.

[0018] Preferably, generating the life prediction evaluation result including the remaining service life interval according to the life prediction value, including: Performing similarity matching between the life prediction value and a historical degradation case base to obtain a degradation mode matching weight distribution; Performing probability calibration on the life prediction value based on the degradation mode matching weight distribution to generate the remaining service life interval with confidence.

[0019] Compared with the prior art, the beneficial effects of the present invention are: By operating sensors and condition detection devices to synchronously collect the operating parameters of the electric drive system (such as current intensity, working temperature, torque output, rotational speed fluctuation) and mechanical component degradation data (such as wear amount, material fatigue index, lubrication state parameters), the present invention constructs a multi-source heterogeneous data set including the electrical domain and the mechanical domain. By respectively arranging the operating parameters and degradation data as the system operation full-cycle input tensor and the component degradation full-cycle input tensor, the structured expression of multi-dimensional data in the time and parameter dimensions is realized, laying a foundation for subsequent cross-domain feature fusion. This multi-source data integration mechanism overcomes the limitations of traditional single-parameter monitoring and can comprehensively reflect the coupling relationship between the system operation state and component degradation.

[0020] In the feature extraction stage, a bidirectional recurrent neural network (BRNN) is used to perform multi-level feature extraction on the input tensor of the entire system operation cycle. By leveraging the bidirectional modeling ability of BRNN for time series data, the forward and backward dependencies of parameters such as current and rotational speed in the time series are effectively captured, and multi-order dynamic feature tensors are extracted, which can sensitively identify short-term operation features such as sudden load changes and abnormal fluctuations. At the same time, the gated time convolutional network (GatedTCN) is used to process the input tensor of the entire component degradation cycle. By utilizing the causal convolution and gating mechanism of TCN, the trend features of degradation parameters such as wear amount and fatigue index during the long-term accumulation process are captured, realizing multi-scale time series modeling. This domain-specific modeling strategy can specifically explore the time series characteristics of different types of data and enhance the richness of feature expression.

[0021] In the cross-modal fusion stage, the self-attention weight matrix and cross-attention weight matrix of the system operation features and component degradation features are calculated through a cross-attention fusion network to generate a dynamic correlation matrix, realizing cross-modal interaction between the electrical domain and mechanical domain features and clarifying the influence weights between different parameters. For example, the strong correlation between the gear wear rate and torque output under high-temperature environments can be identified. An optimization module based on physical constraints is introduced to correct the correlation matrix, embedding physical prior knowledge such as the motor heat conduction law and the tribology principle of mechanical components into the model to avoid generating feature correlations that violate physical laws and improve the physical interpretability and robustness of the model. The feature tensor is optimized through a residual feature correction network, and through operations such as channel compression and spatial attention weighting, the expression ability of key features is further enhanced, and noise interference is suppressed.

[0022] In the remaining useful life prediction stage, a multi-layer perceptron (MLP) network performs non-linear mapping on the fused low-dimensional feature encoding vector to output a remaining useful life prediction value related to the time series, which can capture complex degradation patterns. The predicted value is matched with the historical degradation case database, and a remaining useful life interval with confidence is generated through probability calibration, providing a quantitative risk assessment basis for maintenance decisions. For example, the confidence level of the prediction result can be clarified, avoiding the uncertainty of single-value prediction, and supporting users to formulate differentiated maintenance strategies according to their risk preferences.

[0023] The comprehensive prediction framework constructed by the present invention realizes the full-process intelligence from data acquisition, feature engineering to prediction evaluation, and can be widely applied to scenarios such as electric vehicle drive systems and industrial motor drive devices. Through multi-domain data fusion and physical constraint embedding, the deficiencies of traditional methods in cross-domain correlation analysis and physical interpretability are solved, the generalization ability and engineering reliability of the prediction model are improved, the equipment maintenance cost can be effectively reduced, the risk of sudden failures can be reduced, and advanced technical support is provided for the intelligent operation and maintenance of electric drive systems. Description of the Drawings

[0024] Figure 1 This is the working schematic diagram of the AI-based comprehensive life prediction method for the electric drive transmission system of the present invention; Figure 2 This is the design diagram of cross-modal feature fusion; Figure 3 This is the design diagram of feature optimization and cascading; Figure 4 This is the design diagram of generating life prediction values; Figure 5 This is the design diagram of generating life assessment results. Specific implementation manners

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0026] Please refer to Figures 1 - 5 , the AI-based comprehensive life prediction method for the electric drive transmission system involved in the present invention is specifically implemented as follows: Collect the operating parameters of the electric drive system through multiple monitoring time points of the operating sensor within a predetermined period; at the same time, collect the degradation data of the mechanical components using the condition detection device at the same multiple monitoring time points. Among them, the collection time points of the operating parameters and the degradation data are strictly aligned to ensure the time relevance of subsequent data processing.

[0027] Extract features and perform cross-domain association on the above-collected operating parameters and degradation data to obtain a system-component degradation fusion feature vector. The specific implementation manners are as follows: Arrange the operating parameters at multiple monitoring time points according to the time dimension and parameter type dimension to form a system operation full-cycle input tensor; arrange the mechanical component degradation data at multiple monitoring time points according to the time dimension and degradation index dimension to form a component degradation full-cycle input tensor. The time series features and multi-dimensional attributes of the data are retained through the tensor structure.

[0028] Perform multi-scale time series modeling on the system operation full-cycle input tensor and the component degradation full-cycle input tensor respectively to obtain a system operation multi-order feature tensor and a component degradation multi-order feature tensor. The multi-scale modeling process is realized through a deep learning network, aiming to capture the data change rules at different time scales.

[0029] Perform cross-modal fusion on the multi-order feature tensors of system operation and the multi-order feature tensors of component degradation, and realize the association between system-level features and component-level features through a specific fusion mechanism, and finally generate a system-component degradation fusion feature vector.

[0030] Based on the system-component degradation fusion feature vector, generate a life prediction evaluation result through an AI model. This result includes a quantitative prediction and evaluation interval of the remaining service life of the electric drive transmission system, providing a scientific basis for system maintenance and replacement.

[0031] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1: On the basis of the above overall solution, the operating parameters specifically include current intensity, working temperature, torque output, and rotational speed fluctuation. Among them, the current intensity is collected by a Hall current sensor, which is designed based on the Hall effect principle and consists of a magnetic core, a Hall element, and a signal conditioning circuit. The magnetic core is sleeved outside the main circuit cable of the electric drive system. When current passes through the cable, the magnetic core generates a magnetic field proportional to the current. The Hall element senses the magnetic field intensity and outputs a weak voltage signal, which is amplified and filtered by the signal conditioning circuit and then converted into a standard analog or digital signal. Its measurement range can cover the 0-150% interval of the rated current of the electric drive system, and the resolution reaches 0.1 A, capable of capturing the current transient changes under conditions such as motor startup, acceleration, and load mutation in real time. For example, it can quickly respond to the sudden increase in current at the moment of motor stall.

[0032] The working temperature is monitored by a distributed temperature sensor. The sensor type is selected as Pt100 platinum thermal resistance or K-type thermocouple, and the installation method is adapted according to the environmental characteristics of the monitoring location: the motor stator winding is installed in an embedded manner. When winding the wire, a micro Pt100 element is embedded in the slot and fixed with insulating paint; the bearing housing and the gearbox housing are installed in a surface-mounted manner, and the gap between the sensor and the measured surface is filled with thermal conductive silicone grease to ensure the heat conduction efficiency; for high-temperature areas (such as the heat sink of power electronic devices), a high-temperature-resistant thermocouple is selected and measured in cooperation with an air-cooled sleeve. The temperature acquisition module supports multi-channel synchronous sampling, the sampling period can be set from 1 s to 10 s, and the measurement accuracy is ±0.5 °C (Pt100) or ±1 °C (thermocouple), capable of continuously recording the temperature curves of each key part at different operation times, such as the temperature rise process during long-term full-load operation of the motor and the temperature decay trend during the shutdown cooling stage.

[0033] The torque output is directly measured by a torque sensor. The sensor adopts the rotational strain gauge torque measurement principle. Strain gauges are pasted on the output shaft of the transmission system to form a Wheatstone bridge. When the shaft bears torque, the strain gauges deform, causing the bridge to become unbalanced and output a voltage signal proportional to the torque. The two ends of the sensor are connected to the drive shaft and the load shaft through elastic couplings, which can compensate for the coaxiality error during the installation of the shafting. The measurement range is designed according to the rated torque of the electric drive system as , with an accuracy class of 0.2, and can monitor the dynamic torque fluctuations in real time under conditions such as vehicle starting, climbing, and shifting. For example, during the vehicle acceleration process, the torque sensor can capture the change law that the torque gradually increases with the increase of the throttle opening, as well as the sudden drop and recovery process of the torque caused by the power interruption during the shifting moment.

[0034] The rotational speed fluctuation is obtained by collecting the motor speed signal through an encoder and calculating the fluctuation coefficient. An incremental optoelectronic encoder is selected and installed at the shaft extension of the non-drive end of the motor, rotating synchronously with the rotor. The encoder disk is engraved with evenly distributed light-transmitting slits. When the disk rotates with the shaft, the light source irradiates the photosensitive element through the slits to generate pulse signals, and the pulse frequency is proportional to the rotational speed. The signal processing unit counts and discriminates the pulse signals to calculate the real-time rotational speed value. The sampling frequency can reach 10 kHz, and the rotational speed measurement error does not exceed ±1 rpm. The calculation of the rotational speed fluctuation coefficient is based on the time series analysis method. First, the rotational speed data per minute is collected to generate a rotational speed sequence at equal time intervals, and then the standard deviation or coefficient of variation of this sequence is calculated as an index to characterize the rotational speed stability. For example, when there is a fault of poor gear meshing in the transmission system, the rotational speed fluctuation coefficient will increase significantly, reflecting the phenomenon of intensified periodic rotational speed fluctuations.

[0035] In the data acquisition stage, the sampling frequencies of various parameters need to match the system operation cycle, and synchronous acquisition technology is adopted to ensure time alignment. The specific implementation method is as follows: Set a unified clock synchronization module. Through the IEEE1588 Precision Clock Protocol or a hardware synchronization trigger signal, each acquisition channel such as current, temperature, torque, and rotational speed starts sampling at the same time point. The sampling period is set to 1 s, that is, a complete multi-parameter data point is collected per second. To avoid data loss, a redundant storage mechanism is designed. The acquired data is stored in the local solid-state drive and the remote server simultaneously, and the storage format adopts the HDF5 format, which supports the efficient storage and fast retrieval of massive time-series data.

[0036] The collected data is first normalized to eliminate the influence of different parameter dimensions. For parameters with clear physical dimensions such as current intensity and torque output, the min-max normalization method is adopted, and the formula is:

[0037] where is the original data, and are respectively the minimum and maximum values of the parameter within a predetermined period. The normalized data range is mapped to the interval [0, 1]. For the temperature parameter, considering the temperature threshold differences in different parts, a piecewise normalization method is adopted. According to the motor insulation class (such as class F, class H), the temperature safety thresholds of each monitoring point are set, and the measured temperature is compared with the corresponding threshold to calculate the temperature warning coefficient, which takes values between 0 (normal) and 1 (exceeding the threshold). The rotational speed fluctuation coefficient itself is a dimensionless index and directly retains the original calculated value without additional normalization processing.

[0038] Data preprocessing also includes the link of outlier detection and repair. The Laida criterion based on three times the standard deviation (3σ) is used to identify outlier data points, and the points outside the range of the mean ± 3σ are determined as outliers. The outlier repair method is selected according to the data characteristics: for isolated outliers that occasionally appear in a single channel, the linear interpolation method of adjacent moment data is used for repair; for abnormal data segments at multiple consecutive moments, combined with the operation condition log of the electric drive system (such as start-stop state, load mode) for analysis. If it is the normal fluctuation in the non-steady operation stage of the system (such as start-up, braking), the original data is retained; if it is an abnormality caused by a sensor failure, the historical mean value of the parameter under the same working condition is used for filling.

[0039] After the preprocessing is completed, the operation parameters at multiple monitoring time points are arranged as the input tensor of the full cycle of system operation according to the time dimension and the parameter type dimension. The time dimension corresponds to the acquisition time series. Suppose there are monitoring time points within the predetermined period, then the length of the time dimension is ; the parameter type dimension contains 4 channels including current intensity, working temperature, torque output, and rotational speed fluctuation. Each channel has data points in the time dimension. Therefore, the dimension of the input tensor of the full cycle of system operation is , and this tensor structure retains the dynamic change characteristics of each parameter in the time series and the parallel correlation between parameters, providing a standardized input format for subsequent multi-scale time series modeling.

[0040] During the data transmission process, the industrial Ethernet communication protocol (such as PROFINET, EtherCAT) is adopted to achieve high-speed data transmission between the sensor and the data processing unit, and the communication rate is not less than 100 Mbps to ensure data real-time performance. At the same time, to prevent the influence of electromagnetic interference on the signal quality, the sensor signal line uses a twisted pair shielded cable, the two ends of the cable shield are reliably grounded, and it is physically isolated from the power cable with a distance of not less than 30 cm.

[0041] The entire operation parameter acquisition and preprocessing process strictly follows the principles of timing alignment, dimension normalization, and anomaly repair to ensure that the data input into the subsequent feature extraction module has high integrity, consistency, and reliability, laying a data foundation for system-level health status assessment. Through the collaborative monitoring and analysis of multi-dimensional operation parameters such as current, temperature, torque, and rotational speed, it can comprehensively reflect the electrical performance, mechanical load characteristics, and operation stability of the electric drive system, providing rich input information for constructing system-component degradation fusion features.

[0042] Embodiment 2: The degradation data of mechanical components specifically includes wear amount, material fatigue index, and lubrication state parameters. The wear amount is obtained through ultrasonic testing or oil spectroscopy analysis. Ultrasonic testing uses a contact ultrasonic thickness gauge, whose working principle is based on the propagation characteristics of ultrasonic waves in materials. When ultrasonic waves are emitted from the probe and enter the component to be measured, reflected echoes will be generated on the front and back surfaces of the component. The thickness gauge calculates the remaining thickness of the component by measuring the propagation time of ultrasonic waves in the material and combining the known sound speed. For key components such as gears and bearings, the reference thickness is determined in advance during their design stage. By regularly measuring the current thickness and comparing it with the reference value, the wear amount can be obtained. For example, for a certain type of reducer gear, the designed reference tooth thickness is 20 mm. After running for 500 hours, the measured tooth thickness is 19.8 mm, so the wear amount at that moment is 0.2 mm.

[0043] Oil spectroscopy analysis collects lubricating oil samples and uses a spectroscope to analyze the composition and concentration of metal particles in the oil. Different metal elements correspond to different component materials. For example, iron elements mainly come from gears and bearings made of steel, and copper elements may come from copper alloy cages or sliding bearings. By detecting the changing trend of the concentration of each element in the oil, the wear degree of the corresponding component can be indirectly inferred. For example, when the concentration of iron elements shows a continuous upward trend, it indicates that the wear of steel components is intensifying. At the same time, combined with a particle counter to analyze the size distribution of particles in the oil, particles larger than 10 μm are usually considered to be generated by abrasive wear, and an increase in their number means the deterioration of the wear state.

[0044] The material fatigue index is calculated by monitoring the stress-strain data of components under cyclic loads using strain sensors and combining with Miner's linear cumulative damage theory. The strain sensor uses a foil strain gauge, which is pasted on the surface stress concentration areas of components, such as the tooth roots of gears and the raceways of the outer rings of bearings, through special adhesives. When the component is deformed by force, the strain gauge deforms accordingly, and its resistance value also changes accordingly. The resistance change is converted into a voltage signal through a Wheatstone bridge, and then converted into a digital signal through an amplifier and a data acquisition card to obtain the strain value on the surface of the component.

[0045] During the actual monitoring process, the strain history of the component is continuously recorded within multiple working cycles. The complex strain-time history is decomposed into multiple single-level stress cycles through the rainflow counting method, and the amplitude and frequency of each stress cycle are statistically analyzed. According to the S-N curve (stress-life curve) of the material, the fatigue life of the material at each stress level is determined. The Miner linear cumulative damage theory holds that the fatigue damage of the material is the linear accumulation of the damage caused by each stress cycle. When the cumulative damage reaches 1, the material undergoes fatigue failure. Therefore, the material fatigue index is defined as the cumulative damage degree, and the calculation formula is:

[0046] where is the actual number of applications of the -level stress cycle, and is the fatigue life of the material at this stress level. By calculating the fatigue index in real time, the remaining fatigue life of the component can be predicted. For example, when the fatigue index reaches 0.5, it indicates that the component has experienced approximately 50% of its fatigue life.

[0047] Lubrication state parameters include lubricating oil viscosity, metal particle concentration, etc., which are monitored in real time through oil sensors. Lubricating oil viscosity is an important indicator reflecting the performance of lubricating oil, and the change in viscosity directly affects the lubrication effect and the wear degree of components. A vibrating viscosity sensor is used, and its working principle is based on the relationship between the vibration characteristics of the oscillator in the liquid and the liquid viscosity. The sensor has a tuning fork oscillator built-in. When the oscillator vibrates in the lubricating oil, its vibration frequency and amplitude are affected by the oil viscosity. By detecting the change in the vibration parameters of the oscillator, the viscosity value of the lubricating oil can be calculated.

[0048] The monitoring of metal particle concentration uses an inductive particle sensor, which detects ferromagnetic and non-ferromagnetic particles in the oil based on the principle of electromagnetic induction. There is an excitation coil and two detection coils inside the sensor. When the metal particles in the oil pass through the detection area, it will cause a change in the induced electromotive force in the detection coil. By analyzing the amplitude and frequency of the electromotive force change, the size and material of the particles can be distinguished. For example, ferromagnetic particles will cause a larger change in the electromotive force, while non-ferromagnetic particles cause a smaller change. By monitoring the particle concentration in real time, early signs of abnormal wear of components can be detected in a timely manner.

[0049] The acquisition of degraded data requires setting up dedicated detection devices for key mechanical components. For the reducer gear set, detection holes are drilled on the gearbox housing, and ultrasonic probes and oil sampling valves are installed to facilitate regular ultrasonic thickness measurement and oil sampling analysis. At the same time, strain gauges are pasted on the gear shafts to monitor the stress changes of the gears during meshing. For the motor bearings, temperature sensors and vibration sensors are installed on the bearing housings. Combining with the results of oil analysis, the health status of the bearings is comprehensively evaluated. The temperature sensor monitors the temperature change of the bearings in real time. When the bearings have poor lubrication or increased wear, the temperature will rise significantly; the vibration sensor identifies the early faults of the bearing components by detecting the spectral characteristics of the bearing vibration signals.

[0050] When arranging the input tensor for the full cycle of component degradation, the time dimension corresponds to the sequence of monitoring time points. Starting from when the equipment is put into operation, monitoring is carried out every 100 hours, and a total of monitorings are carried out, then the length of the time dimension is . The degradation index dimension includes parameters such as wear amount, material fatigue index, lubricating oil viscosity, and metal particle concentration. Each parameter has data points in the time dimension. In addition, composite indicators can also be derived according to actual needs, such as wear rate and fatigue damage rate. The wear rate is defined as the ratio of the change in wear amount between two adjacent monitorings to the time interval, reflecting the severity of wear; the fatigue damage rate is the change rate of the fatigue index over time, reflecting the speed of fatigue accumulation.

[0051] The formed multi-dimensional degradation feature matrix contains rich component degradation information, which can not only reflect the degradation state at the current moment but also reflect the historical evolution trend of the degradation process. For example, by analyzing the change curve of the wear amount over time, it can be judged whether the component is in the normal wear stage, the accelerated wear stage, or the near-failure stage; by comparing the change trends of different parameters, the correlation between parameters can be found. For example, a decrease in lubricating oil viscosity may be accompanied by an increase in metal particle concentration, indicating that the deterioration of lubrication performance leads to increased wear.

[0052] To ensure the accuracy and reliability of the degraded data, a series of quality control measures are taken during the data acquisition process. For ultrasonic detection, the probe is calibrated before each measurement, and a standard test block with a known thickness is used to verify the accuracy of the measurement system. During the oil sampling process, the sampling operation procedures are strictly followed to ensure that the collected oil samples are representative and to avoid the influence of external contamination on the analysis results. For strain sensors and oil sensors, zero calibration and sensitivity calibration are carried out regularly to ensure that the measurement accuracy of the sensors is within the specified range.

[0053] Meanwhile, a data verification mechanism is established to check the rationality of the collected degradation data. For example, when the wear amount at a certain moment suddenly increases and exceeds the normal fluctuation range, repeated measurements will be carried out for verification to exclude the possibility of measurement errors. If the results of multiple measurements are consistent, the operating status of the relevant components will be further checked, and comprehensive analysis will be carried out in combination with other parameters (such as temperature and vibration) to determine whether abnormal wear has actually occurred.

[0054] The entire degradation data collection and processing process strictly follows standardized operating specifications to ensure the authenticity, integrity, and traceability of the data. Through multi-dimensional degradation data monitoring of key mechanical components, the evolution process of the health status of the components can be comprehensively grasped, providing a solid data foundation for subsequent multi-scale time series modeling and life prediction. The combination of these degradation data and system operating parameters can more accurately reflect the comprehensive degradation status of the electric drive transmission system, thereby realizing more accurate life prediction and evaluation.

[0055] Example 3: The multi-scale time series modeling process is as follows: The full-cycle input tensor of the system operation is input into a multi-level feature extraction network based on a bidirectional recurrent neural network (Bi-RNN) to obtain a multi-order feature tensor of the system operation. The bidirectional recurrent neural network consists of two recurrent layers, a forward recurrent layer and a backward recurrent layer. The forward recurrent layer processes data from front to back in time series, and the backward recurrent layer processes from back to front, enabling the network to capture the dependencies of the time series context. Taking the bidirectional long short-term memory network (Bi-LSTM) as an example, its core unit is the LSTM memory unit. Each memory unit contains a forget gate, an input gate, and an output gate, and controls the retention and update of information through a gating mechanism, which can effectively alleviate the gradient disappearance problem of traditional recurrent neural networks and is suitable for capturing time-dependent features in long sequences.

[0056] The dimension of the full-cycle input tensor of the system operation is , where represents the length of the time dimension (i.e., the number of monitoring time points), represents the number of channels of the operating parameters (such as the number of parameter types such as current intensity and working temperature). The input tensor first passes through an embedding layer for dimension expansion and is converted into a three-dimensional tensor of , where is the embedding dimension, which is used to enhance the expression ability of features. Subsequently, the three-dimensional tensor is input into the forward and backward recurrent layers of the Bi-LSTM network. The hidden state dimension output by each recurrent layer is , where is the number of neurons in the hidden layer. By stacking multiple layers of Bi-LSTM networks (such as 2 layers), a multi-level feature extraction structure is formed: the first layer extracts low-order temporal features (such as the parameter fluctuation pattern within a single monitoring period), and the second layer extracts high-order temporal features (such as the trend change across multiple periods) based on the low-order features output by the first layer. Finally, the dimension of the multi-order feature tensor of the system operation output by the multi-level feature extraction network is , where represents the number of feature levels (i.e., the number of network layers), and this tensor contains the system operation features at different time scales (short-term, medium-term, long-term) and different abstraction levels.

[0057] For the input tensor of the full cycle of component degradation, its dimension is , where is the length of the time dimension (aligned with the time points of the system operation parameters), is the number of channels of the degradation index (such as the number of parameter types like wear amount, material fatigue index, etc.). It is input into the degradation feature extraction network based on the Gated Time Convolution Network (GatedTCN). The Gated Time Convolution Network combines the Time Convolution Network (TCN) and the gating mechanism. The time convolution layer expands the receptive field without increasing the number of parameters through the Dilated Convolution technology, and can capture long-range time-dependent relationships; the gating mechanism generates a gating signal through the sigmoid activation function to selectively filter the features output by the convolution layer and automatically learn the importance of features at different time steps.

[0058] In the specific implementation, the input tensor of the full cycle of component degradation first undergoes feature preliminary extraction through a one-dimensional convolution layer, and the convolution kernel size is , generating a feature map with a dimension of , where is the number of convolution kernels. Subsequently, it enters the gated time convolution module, and each module contains two parallel dilated convolution layers: one for extracting the linear part of the features, and the other for generating the gating signal. The output of the linear part is denoted as , and the gating signal is denoted as , where is the sigmoid function, is the weight of the gating convolution kernel, is the bias term, represents the convolution operation. The final output is , and in this way, the adaptive weighting of the features is achieved. By stacking multiple gated time convolution modules, and the dilation rate of each module increases exponentially (such as ), the network can capture the degradation temporal features at different scales from local to global. Assuming modules are stacked, and the feature dimension output by each module is ( ) Finally, the output features of each module are fused through a skip connection to obtain a multi-order feature tensor of component degradation, whose dimension is , where is the number of output feature channels, and represents the number of multi-scale features (features corresponding to different dilation rates).

[0059] During the network training process, an end-to-end supervised learning method is adopted. The mean square error (MSE) between the system-component degradation fusion feature vector and the life prediction label is used as the loss function, and the network parameters are optimized through the backpropagation algorithm. For the bidirectional recurrent neural network, the input data needs to be processed by sequence padding to ensure that time series of different lengths can be batch input into the network. The hidden state at the padding position is excluded from the calculation through the mask mechanism. The gated temporal convolutional network needs to handle the causal convolution problem, that is, the convolution operation at the current time step only depends on the data in the past and current moments, and does not use the data in the future moment to ensure the temporal rationality of the model in practical applications.

[0060] The multi-scale characteristics of the system operation multi-order feature tensor and the component degradation multi-order feature tensor are reflected in: the former captures the operation modes of different time spans (such as second-level current fluctuations and hour-level temperature trends) through the hierarchical stacking of recurrent neural networks, and the latter captures the local details (such as the wear amount change within a single monitoring period) and global trends (such as the long-term growth of cumulative fatigue damage) of the degradation features through the different receptive fields of dilated convolutions. The two feature tensors provide multi-level temporal feature representations from the two dimensions of the system operation state and the component degradation process respectively, laying a foundation for the subsequent cross-modal fusion. Through this multi-scale modeling method, the long-term and short-term dependencies hidden in the data can be effectively extracted, avoiding the limitations of single-scale feature representations, and thus more comprehensively depicting the degradation process of the electric drive transmission system.

[0061] Example 4: The cross-modal fusion process specifically includes the following operations: Input the multi-order feature tensors of system operation and component degradation into the cross-attention fusion network to obtain the system-component degradation dynamic correlation matrix. Taking the correlation analysis between the operating state of the motor and the gear wear of the reducer in the electric drive transmission system as an example, assume that the multi-order feature tensor of system operation contains multi-level abstract representations of features such as the current intensity and rotational speed fluctuation of the motor at different time points, and the multi-order feature tensor of component degradation contains multi-scale features of degradation indicators such as gear wear amount and lubricating oil metal particle concentration. The cross-attention fusion network first calculates the self-attention weight matrix of system operation features, which reflects the interdependence relationship between parameters such as current and rotational speed at different time points. For example, there is a strong correlation between current and rotational speed during the motor startup phase. At the same time, it calculates the cross-attention weight matrix between component degradation features and system operation features to capture the cross-modal correlation between the gear wear degree and the motor load parameters. For example, an increase in the gear wear rate may correspond to a high torque output period. Through the matrix multiplication operation of the self-attention matrix and the cross-attention matrix, a system-component degradation dynamic correlation matrix is generated. Each element in this matrix represents the correlation strength between a certain system operation feature and a certain component degradation feature at a specific time point. For example, the correlation value between the motor current feature at time t and the gear wear amount at time t-1.

[0062] The dynamic correlation matrix is input into the optimization module based on physical constraints to obtain the constraint-corrected correlation matrix. The physical constraint optimization module incorporates physical models of the electric drive system, such as the motor heat conduction model and the gear meshing mechanics model. Taking the motor heat conduction model as an example, this model describes the relationship between the motor winding temperature, current intensity, and operating time. When a negative correlation value between the current feature and the temperature feature at a certain moment appears in the dynamic correlation matrix (violating the physical law that an increase in current leads to an increase in temperature), the optimization module will automatically correct this correlation value to a reasonable range. Taking the gear meshing mechanics model as another example, the model defines a positive correlation relationship between torque output, gear contact stress, and wear amount. If the correlation matrix shows an abnormal correlation where a high torque period corresponds to a low wear amount, the module will adjust this correlation strength according to the physical mapping relationship between torque and wear. Through such physical constraint corrections, false correlations that violate basic principles such as energy conservation and material mechanics are eliminated, ensuring that the correlation matrix conforms to actual physical laws. For example, the corrected data should reflect the causal chain of "high load (high current, high torque) → high stress → accelerated component degradation (increase in wear amount, increase in fatigue index)".

[0063] The system operation multi - order feature tensor and the component degradation multi - order feature tensor are respectively input into the residual feature correction network to obtain the optimized feature tensor. For the system operation features, first, channel - dimension compression is performed. For example, a feature tensor containing 4 channels of current, temperature, torque, and rotational speed is compressed into a one - dimensional feature vector through a fully - connected layer. This vector aggregates the comprehensive representation of the system operation state. Then, the weighted fusion of the constrained and corrected correlation matrix and the system operation feature compression vector is carried out. Specifically, the row vectors related to the system features in the correlation matrix are extracted and element - wise multiplied with the compression vector, and then the non - linear correlation is enhanced through an activation function (such as ReLU). After that, a feature expansion operation is performed to restore the one - dimensional vector to the same channel dimension as the original feature tensor, obtaining the optimized system operation multi - order feature tensor. This process can be understood as: using the dynamic correlation information of system - component in the correlation matrix to assign weights to the system operation features, highlighting the parameters highly related to component degradation (such as the weight of the feature channel corresponding to high torque increases), and suppressing irrelevant or interfering features.

[0064] For the component degradation multi - order feature tensor, first, spatial attention weighting is performed. Taking 4 degradation indexes of gear wear amount, fatigue index, lubricating oil viscosity, and metal particle concentration as an example, the spatial attention mechanism generates the attention weights of each index in the time dimension through convolution operations. For example, when the metal particle concentration shows an upward trend recently, the weight corresponding to this index will increase significantly. Then, the weighted component degradation features are modulated with the constrained and corrected correlation matrix. Specifically, the column vectors related to the component degradation features are extracted from the correlation matrix as modulation parameters to scale and translate the features of each degradation index. For example, if the correlation matrix shows that the high - speed fluctuation of the motor in a certain period is highly correlated with the gear fatigue index, the feature of the fatigue index in this period will be multiplied by a scaling factor greater than 1 to enhance its importance in the fused features. Through this feature modulation, the component degradation features can adaptively fuse the influence of the system operation state, reflecting the physical logic of "system load changes driving component degradation".

[0065] Feature - level cascade encoding is performed on the optimized system operation multi - order feature tensor and the component degradation multi - order feature tensor. Taking a specific dimension as an example, assume that the dimension of the optimized system operation feature tensor is [T, C1, L1] (T is the number of time points, C1 is the number of system parameter channels, and L1 is the number of feature levels), and the dimension of the component degradation feature tensor is [T, C2, L2] (C2 is the number of degradation index channels, and L2 is the number of multi - scale features). Then, the feature - level cascade encoding concatenates the two along the channel dimension to form a new tensor of [T, C1 + C2, max(L1, L2)] (if the number of levels is inconsistent, the dimensions can be unified through zero padding or interpolation). For example, the long - term trend feature of the motor current (the high - level output of the system operation feature) and the global feature of the gear wear rate (the large - scale output of the component degradation feature) are concatenated at the same time point to form a fusion feature vector containing system - component cross - domain information. This cascade encoding method preserves the original structures of the two types of features. At the same time, through cross - modal association optimization, it enhances the physical correlation between features, enabling the fused features to reflect both the dynamic process of system operation and the historical accumulation of component degradation, providing more comprehensive input information for subsequent life prediction.

[0066] In practical applications, taking an electric drive system of a certain type of electric vehicle as an example, the multi - order feature tensor of system operation may capture the following patterns: Under the urban driving conditions where the vehicle starts and stops frequently, the motor current shows high - frequency pulsed fluctuations (short - term feature), and at the same time, long - term monitoring shows that the average torque output is 15% lower than that in high - speed driving conditions (long - term feature); The multi - order feature tensor of component degradation shows that the wear rate of the input shaft bearing of the reducer is 20% higher in the start - stop condition than in the constant - speed condition (local - scale feature), while the cumulative fatigue index shows a linear increase during the entire monitoring period (global - scale feature). The cross - attention fusion network calculates the self - attention and identifies the strong coupling between the current and the rotational speed fluctuations under the start - stop condition (the corresponding element values in the self - attention weight matrix are relatively high); Through the mutual attention, it discovers the positive correlation between the current pulse frequency and the bearing wear rate (the corresponding element values in the mutual - attention matrix are relatively high), thus strengthening the association path of "start - stop → current fluctuation → bearing wear acceleration" in the dynamic association matrix. The physical constraint optimization module, based on the physical model of load transfer between the motor and the bearing, confirms that the high current fluctuation does indeed correspond to the high - frequency change of the motor output torque, which in turn leads to an increase in the alternating load on the bearing. Therefore, it retains and enhances this association and eliminates false associations that violate mechanical principles, such as "low rotational speed corresponding to high wear".

[0067] In the process of processing the system operation characteristics by the residual feature correction network, the vector after channel compression will highlight the weight of the current fluctuation characteristics (because it is strongly associated with component degradation). After weighted fusion and expansion, the characteristic value of the current channel in the optimized system operation characteristics increases significantly during the start-stop period, reflecting its dominant role in component degradation; the processing of component degradation characteristics increases the weight of the bearing wear characteristics during the start-stop period by 30% through spatial attention, and through the modulation of the correlation matrix, the wear characteristic value of this period is linearly correlated and scaled with the current characteristic value at the same time, further strengthening the physical connection between the two. The final cascade coding corresponds the "start-stop-current fluctuation" characteristics of the system operation to the "bearing wear acceleration" characteristics of component degradation in a one-to-one time series, forming a fused feature vector containing cross-domain causal relationships, providing the life prediction model with input data that directly reflects the "system operating conditions-component degradation" coupling process.

[0068] The entire cross-modal fusion process uses the attention mechanism to mine the potential associations between data, combines physical constraints to ensure the rationality of the associations, and then achieves deep fusion of cross-domain information through feature correction and cascading. This method avoids the limitation of traditional data fusion that only performs simple splicing, and can dynamically capture the nonlinear and time-varying associations between systems and components. At the same time, it uses domain knowledge (physical models) to improve the interpretability and reliability of fused features, providing more physically meaningful feature representations for the life prediction of electric drive transmission systems.

[0069] Embodiment 5: The generation process of the life prediction evaluation results is as follows: First, the system-component degradation fusion feature vector is subjected to multi-dimensional dimensionality reduction processing. Taking a certain electric drive system as an example, it is assumed that the fused feature vector contains 100-dimensional features (such as multi-order features of system operation and multi-scale features of component degradation). These features may contain redundant information (such as similar trend features at different time points) or irrelevant noise (such as interference introduced by sensor measurement errors). Dimensionality reduction processing can use the principal component analysis (PCA) method to project high-dimensional features into low-dimensional space through linear transformation, retaining the characteristic components that best reflect the data variance. For example, reducing 100-dimensional features to 20 dimensions allows the low-dimensional feature encoding vector to retain the main information of the original data while reducing the computational complexity. The vector after dimensionality reduction contains a comprehensive representation of key information such as long-term current trends, gear wear rates, and bearing fatigue indexes.

[0070] Subsequently, the low-dimensional feature encoding vectors are input into a multi-layer perceptron (MLP) network for non-linear relationship mapping. The MLP network contains multiple fully connected layers. Assuming 3 hidden layers are set, the number of neurons in each layer is 128, 64, and 32 respectively, and the ReLU activation function is used. The first hidden layer receives a 20-dimensional input vector. Through the linear transformation of the weight matrix and bias term, plus the ReLU activation, it extracts the first-order non-linear combination between features (such as the synergistic effect of current and torque on component degradation); the second hidden layer further processes the first-order features to generate second-order non-linear features (such as the product effect of high-load duration and wear amount); the third hidden layer then integrates the second-order features into an abstract representation directly related to the remaining useful life. The final output layer outputs a scalar value as the life prediction value through linear transformation. This value corresponds to the predicted remaining time (unit: hours) from the current state of the electric drive system to failure. For example, for a certain system after running for 500 hours, the MLP outputs a life prediction value of 800 hours, indicating that it is expected to run for another 800 hours.

[0071] Furthermore, the life prediction value is matched with the historical degradation case library for similarity. The historical degradation case library stores the degradation data and corresponding actual life data of different electric drive systems under various operating conditions. Each case contains a fused feature vector, operating condition records (such as average load, start-stop frequency), degradation mode descriptions (such as gear abrasive wear dominant, bearing fatigue failure dominant), and actual remaining life. The similarity matching uses the cosine similarity calculation method. The current fused feature vector is compared with the feature vectors of all cases in the case library one by one, and the cosine value is calculated to measure the similarity. For example, the cosine similarity between the fused feature vector of the current system and the case numbered C-001 in the case library is 0.85, and the similarity with the case numbered C-005 is 0.78, indicating that the current degradation mode is more similar to the C-001 case (gear wear dominant under high load).

[0072] Based on the similarity matching results, the degradation mode matching weight distribution is obtained. The weight distribution reflects the contribution degree of each historical case to the current prediction. The higher the similarity, the greater the weight of the case. Suppose there are 10 cases in the case library. After matching, the weight of case C-001 is 30%, the weight of case C-005 is 25%, and the combined weight of other cases is 45%. Use this weight distribution to perform probability calibration on the life prediction value. The calibration process uses a Bayesian network framework, taking the remaining life of historical cases as the prior distribution and the current prediction value as the likelihood function to calculate the posterior probability distribution. For example, the actual remaining life distribution of case C-001 is 700 - 900 hours (mean 800 hours), and that of case C-005 is 650 - 850 hours (mean 750 hours). A confidence-level remaining useful life interval is generated through weighted average. Suppose the calibrated 95% confidence interval is 720 - 880 hours. This interval reflects the possible range of the remaining life of the current system, taking into account the diversity and uncertainty of historical cases.

[0073] Taking a specific application scenario as an example, the state monitoring of an electric drive system of an electric bus is carried out after 600 hours of operation. The collected operating parameters include: average current 80 A (rated current 100 A), motor temperature 75 °C (safety threshold 100 °C), torque output (peak torque ), rotational speed fluctuation coefficient 0.05 (normal range ≤ 0.1); the degradation data shows that the gear wear is 0.3 mm (initial tooth thickness 20 mm), the bearing fatigue index is 0.4, and the metal particle concentration of the lubricating oil is 50 ppm (warning value 80 ppm). After feature extraction and cross-modal fusion, a fusion vector containing 40-dimensional features is obtained. After dimensionality reduction to 15 dimensions, it is input into the MLP network, and the output life prediction value is 750 hours.

[0074] Three highly similar cases are matched in the historical case library: Case A (urban bus working condition, dominated by gear wear, remaining life 700 - 800 hours, similarity 0.88), Case B (frequent start-stop working condition, dominated by bearing fatigue, remaining life 720 - 850 hours, similarity 0.82), Case C (mixed working condition, compound degradation caused by poor lubrication, remaining life 680 - 820 hours, similarity 0.79). According to the weight distribution (Case A 40%, Case B 35%, Case C 25%), the calibrated remaining useful life interval is: lower limit = 700×0.4 + 720×0.35 + 680×0.25 = 704 hours, upper limit = 800×0.4 + 850×0.35 + 820×0.25 = 817.5 hours. After rounding, a 95% confidence interval of 700 - 820 hours is obtained. This result provides a clear maintenance time window for the operation and maintenance personnel, based on which an overhaul plan can be formulated (such as replacing the gear oil and checking the bearing status before the remaining life reaches 500 hours).

[0075] In the dimensionality reduction process, in addition to PCA, non-linear dimensionality reduction methods such as t-SNE can also be used, which are suitable for scenarios where there are complex non-linear relationships between features. For example, when there is a non-linear coupling between the system operating parameters and the component degradation index (such as the wear rate increasing exponentially after the temperature exceeds the threshold), t-SNE can better preserve the local structure of the data, making the low-dimensional features more accurately reflect the manifold distribution in the high-dimensional space. The structure of the MLP network can be adjusted according to the data characteristics. If the temporal correlation in the fused features is strong, a one-dimensional convolutional layer can be added before the MLP to first extract the local features in the time dimension and then perform non-linear mapping.

[0076] The construction of the historical case library needs to follow the principles of data diversity and representativeness, covering different types of electric drive systems (such as passenger car and commercial vehicle motors), various degradation modes (wear, fatigue, corrosion), and typical operating conditions (high speed, low speed, alternating load). The case data should include complete life cycle monitoring data, the full-cycle parameters from the equipment being put into operation to failure and scrapping, to ensure the reliability of similarity matching. For example, the case library needs to include cases of different failure modes such as early pitting of gears and cracks in the outer ring of bearings, as well as the corresponding evolution process of degradation characteristics.

[0077] During the probability calibration process, if there is a lack of sufficient historical cases, the Bootstrap resampling method can be used to generate virtual cases. By sampling the existing data with replacement, the number of cases is increased and the prediction uncertainty is estimated. In addition, expert knowledge can be introduced to correct the weight distribution. For example, when a new lubricating material is used in the current system, the expert can manually reduce the weight of the degradation cases related to lubrication to improve the rationality of case matching.

[0078] Through data dimensionality reduction, non-linear modeling, similarity matching, and probability calibration, the entire remaining life prediction and evaluation process transforms the abstract fused features into a remaining life prediction interval with practical engineering significance. This method not only utilizes the powerful fitting ability of the AI model but also combines the statistical laws of historical data. It can provide an uncertainty assessment based on group experience while considering the individual differences of the system, meeting the accuracy and reliability requirements of life prediction in industrial scenarios. Maintenance personnel can optimize the maintenance strategy according to the prediction results, avoid resource waste caused by over-maintenance, and at the same time prevent safety risks caused by component failures, realizing the intelligent operation and maintenance management of the electric drive system.

[0079] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0080] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based comprehensive life prediction method for an electric drive transmission system, characterized in that Including: Obtain the operating parameters of the electric drive system at multiple monitoring time points within a predetermined period collected by the running sensor, and the mechanical component degradation data at the multiple monitoring time points collected by the state detection device; Perform feature extraction and cross-domain association on the operating parameters at the multiple monitoring time points and the mechanical component degradation data at the multiple monitoring time points to obtain a system-component degradation fusion feature vector; Generate a life prediction evaluation result based on the system-component degradation fusion feature vector; Among them, performing feature extraction and cross-domain association on the operating parameters at the multiple monitoring time points and the mechanical component degradation data at the multiple monitoring time points to obtain a system-component degradation fusion feature vector includes: Arrange the operating parameters at the multiple monitoring time points into a system operating full-cycle input tensor according to the time dimension and parameter type dimension, and arrange the mechanical component degradation data at the multiple monitoring time points into a component degradation full-cycle input tensor according to the time dimension and degradation index dimension; Perform multi-scale time series modeling on the system operating full-cycle input tensor and the component degradation full-cycle input tensor respectively to obtain a system operating multi-order feature tensor and a component degradation multi-order feature tensor; Perform cross-modal fusion on the system operating multi-order feature tensor and the component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector.

2. The comprehensive life prediction method of the AI-based electric drive transmission system according to claim 1, wherein The operating parameters include current intensity, working temperature, torque output, and rotational speed fluctuation.

3. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 2, wherein The degradation data includes wear amount, material fatigue index, and lubrication state parameter.

4. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 3, wherein, Performing multi-scale time series modeling on the system operating full-cycle input tensor and the component degradation full-cycle input tensor respectively to obtain a system operating multi-order feature tensor and a component degradation multi-order feature tensor includes: Input the system operating full-cycle input tensor into a multi-level feature extraction network based on a bidirectional recurrent neural network to obtain the system operating multi-order feature tensor; Input the component degradation full-cycle input tensor into a degradation feature extraction network based on a gated time convolutional network to obtain the component degradation multi-order feature tensor.

5. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 4, wherein Performing cross-modal fusion on the system operating multi-order feature tensor and the component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector includes: Input the system operating multi-order feature tensor and the component degradation multi-order feature tensor into a cross-attention fusion network to obtain a system-component degradation dynamic association matrix; Input the system-component degradation dynamic association matrix into an optimization module based on physical constraints to obtain a constraint-corrected association matrix; Input the system operating multi-order feature tensor and the component degradation multi-order feature tensor into a residual feature correction network respectively to obtain an optimized system operating multi-order feature tensor and an optimized component degradation multi-order feature tensor; Perform feature concatenation coding on the optimized system operating multi-order feature tensor and the optimized component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector.

6. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 5, characterized in that, Inputting the system operating multi-order feature tensor and the component degradation multi-order feature tensor into a cross-attention fusion network to obtain a system-component degradation dynamic association matrix includes: Calculate the self-attention weight matrix of the multi-order feature tensor of the system operation and the cross-attention weight matrix of the multi-order feature tensor of the component degradation, and generate the system-component degradation dynamic correlation matrix through matrix multiplication operation.

7. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 6, wherein, Input the multi-order feature tensor of the system operation and the multi-order feature tensor of the component degradation into the residual feature correction network respectively to obtain the optimized multi-order feature tensor of the system operation and the optimized multi-order feature tensor of the component degradation, including: Perform channel dimension compression on the multi-order feature tensor of the system operation to obtain a system operation feature compression vector; Perform weighted fusion on the constraint correction correlation matrix and the system operation feature compression vector, and then perform a feature expansion operation to obtain the optimized multi-order feature tensor of the system operation; Perform spatial attention weighting on the multi-order feature tensor of the component degradation, and then perform feature modulation with the constraint correction correlation matrix to obtain the optimized multi-order feature tensor of the component degradation.

8. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 7, wherein Generate a life prediction evaluation result based on the system-component degradation fusion feature vector, including: Input the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value; Generate the life prediction evaluation result including the remaining service life interval according to the life prediction value.

9. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 8, characterized in that, Input the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value, including: Perform multi-dimensional dimensionality reduction processing on the system-component degradation fusion feature vector to obtain a low-dimensional feature encoding vector; Input the low-dimensional feature encoding vector into a multi-layer perceptron network to perform non-linear relationship mapping, and output the life prediction value associated with the time series.

10. The AI-based comprehensive life prediction method for an electric drive transmission system according to claim 9, wherein, Generate the life prediction evaluation result including the remaining service life interval according to the life prediction value, including: Perform similarity matching between the life prediction value and the historical degradation case library to obtain a degradation mode matching weight distribution; Perform probability calibration on the life prediction value based on the degradation mode matching weight distribution to generate the remaining service life interval with confidence.

Images