AI-based comprehensive life prediction method for electric drive transmission systems

Abstract

The present invention relates to the technical field of electric drive transmission system life prediction and discloses an AI-based comprehensive life prediction method for electric drive transmission systems. The method comprises: collecting electric drive system operating parameters and mechanical component degradation data; arranging the data into full-cycle system operation input tensors and full-cycle component degradation input tensors, respectively extracting multi-order feature tensors using a bidirectional recurrent neural network and a gated temporal convolutional network; utilizing a cross-attention fusion network combined with physical constraints for cross-modal fusion to obtain a system-component degradation fusion feature vector; and inputting the output of a life prediction value into a multi-layer perceptron network after dimensionality reduction. The remaining useful life interval is generated with confidence levels based on historical case studies. This method integrates multi-source data, realizes cross-domain dynamic feature association and physical constraint fusion, and improves the accuracy and reliability of life prediction. It is suitable for intelligent operation and maintenance of electric drive systems 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 an AI-based comprehensive life prediction method for electric drive transmission systems. Background Art

[0002] With the rapid development of new energy technologies, electric drive transmission systems are increasingly being used in applications such as electric vehicles and industrial automation equipment. As core power components, their lifespan directly impacts equipment reliability, safety, and maintenance costs. However, the complex and dynamic operating environment of electric drive transmission systems, involving the coupling of multiple electrical and mechanical domains, presents numerous challenges for traditional lifespan prediction methods.

[0003] Considering the limitations of single-parameter monitoring, traditional methods typically rely on single-dimensional analysis of electrical signals like current and speed, or mechanical parameters like wear, failing to fully capture the dynamic relationship between system operating status and component degradation. For example, when determining motor health solely based on current intensity, the impact of torque fluctuations caused by gear wear on the overall system lifespan may be overlooked, resulting in significant deviations between predicted and actual lifespan.

[0004] Regarding cross-domain data fusion, electric drive system operating parameters and mechanical component degradation data belong to different physical domains. Traditional methods lack effective cross-domain feature correlation mechanisms. 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). However, traditional algorithms struggle to exploit the potential mapping relationships between data from different domains, making it difficult to construct a comprehensive degradation assessment model.

[0005] A significant drawback of traditional approaches is the inadequate multi-scale time series modeling capabilities. The degradation process of electric drive systems exhibits multi-timescale characteristics. Short-term speed fluctuations may reflect real-time load changes, while long-term wear accumulation reflects the progressive degradation of components. Traditional fixed-window statistical analysis or simple neural network models cannot effectively capture the evolution of features across different time scales, resulting in insufficient sensitivity to early degradation signatures.

[0006] The integration of physical constraints and data-driven models is equally problematic. Existing data-driven models often ignore the physical operating laws of electric drive systems, such as the thermodynamic coupling between motor heating and bearing lubrication, potentially generating predictions that violate physical principles. Methods based solely on physical models struggle to adapt to parameter drift and uncertainty under complex operating conditions, resulting in poor predictive robustness.

[0007] Furthermore, the demand for higher accuracy and reliability in lifespan predictions continues to increase. In industrial applications, misjudgments can lead to excessive maintenance or failure of critical components, resulting in significant economic losses. Traditional methods generate single lifespan predictions that lack confidence assessments and cannot provide comprehensive risk quantification information for maintenance decisions. This makes it difficult to meet the interpretability and reliability requirements of modern intelligent operations and maintenance. Summary of the Invention

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

[0009] To achieve the above objectives, the present invention provides the following technical solution: an AI-based comprehensive life prediction method for electric drive transmission systems, the method comprising:

[0010] Acquiring operating parameters of the electric drive system at multiple monitoring time points within a predetermined period collected by an operating sensor, and mechanical component degradation data at the multiple monitoring time points collected by a state detection device;

[0011] performing feature extraction and cross-domain correlation on the operating parameters at the plurality of monitoring time points and the mechanical component degradation data at the plurality of monitoring time points to obtain a system-component degradation fusion feature vector;

[0012] generating a life prediction assessment result based on the system-component degradation fusion feature vector;

[0013] The process of extracting features and performing cross-domain correlation on the operating parameters at the plurality of monitoring time points and the mechanical component degradation data at the plurality of monitoring time points to obtain a system-component degradation fusion feature vector includes:

[0014] Arranging the operating parameters at the plurality of monitoring time points into a system operation full cycle input tensor according to a time dimension and a parameter type dimension, and arranging the mechanical component degradation data at the plurality of monitoring time points into a component degradation full cycle input tensor according to a time dimension and a degradation index dimension;

[0015] Performing multi-scale time series modeling on the full-cycle input tensor of the system operation and the full-cycle input tensor of the component degradation to obtain a multi-order feature tensor of the system operation and a multi-order feature tensor of the component degradation;

[0016] Cross-modal fusion is performed 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.

[0017] Preferably, the operating parameters include current intensity, operating temperature, torque output and speed fluctuation.

[0018] Preferably, the degradation data includes wear amount, material fatigue index and lubrication state parameters.

[0019] Preferably, multi-scale time series modeling is performed on the full-cycle input tensor of the system operation and the full-cycle input tensor of the component degradation to obtain a multi-order feature tensor of the system operation and a multi-order feature tensor of the component degradation, respectively, including:

[0020] Inputting the full-cycle input tensor of the system operation into a multi-level feature extraction network based on a bidirectional recurrent neural network to obtain a multi-order feature tensor of the system operation;

[0021] The full-cycle input tensor of component degradation is input into a degradation feature extraction network based on a gated temporal convolutional network to obtain a multi-order feature tensor of component degradation.

[0022] Preferably, cross-modal fusion of the system operation multi-order feature tensor and the component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector includes:

[0023] Inputting the system operation 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 correlation matrix;

[0024] Inputting the system-component degradation dynamic correlation matrix into a physical constraint-based optimization module to obtain a constraint-corrected correlation matrix;

[0025] Inputting the system operation multi-order feature tensor and the component degradation multi-order feature tensor into a residual feature correction network respectively to obtain an optimized system operation multi-order feature tensor and an optimized component degradation multi-order feature tensor;

[0026] Feature cascade encoding is performed on the optimized system operation multi-order feature tensor and the optimized component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector.

[0027] Preferably, the system operation multi-order feature tensor and the component degradation multi-order feature tensor are input into a cross attention fusion network to obtain a system-component degradation dynamic correlation matrix, including:

[0028] The self-attention weight matrix of the multi-order feature tensor of the system operation and the mutual attention weight matrix of the multi-order feature tensor of the component degradation are calculated, and the system-component degradation dynamic association matrix is ​​generated through matrix multiplication operation.

[0029] Preferably, 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 system operation multi-order feature tensor and the optimized component degradation multi-order feature tensor, including:

[0030] Performing channel dimension compression on the multi-order feature tensor of the system operation to obtain a system operation feature compression vector;

[0031] Performing a feature expansion operation after weighted fusion of the constraint correction correlation matrix and the system operation feature compression vector to obtain an optimized system operation multi-order feature tensor;

[0032] The component degradation multi-order feature tensor is subjected to spatial attention weighting and then feature modulated with the constraint correction association matrix to obtain an optimized component degradation multi-order feature tensor.

[0033] Preferably, generating a life prediction evaluation result based on the system-component degradation fusion feature vector includes:

[0034] Inputting the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value;

[0035] The life prediction evaluation result including the remaining service life interval is generated according to the life prediction value.

[0036] Preferably, inputting the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value includes:

[0037] Performing multi-dimensional dimensionality reduction processing on the system-component degradation fusion feature vector to obtain a low-dimensional feature encoding vector;

[0038] The low-dimensional feature encoding vector is input into a multi-layer perceptron network for nonlinear relationship mapping, and the life prediction value associated with the time series is output.

[0039] Preferably, generating the life prediction evaluation result including the remaining service life interval according to the life prediction value includes:

[0040] Performing similarity matching on the predicted lifespan value and a historical degradation case library to obtain a degradation pattern matching weight distribution;

[0041] The life prediction value is probability calibrated based on the degradation pattern matching weight distribution to generate the remaining service life interval with confidence.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This method uses operating sensors and status detection devices to synchronously collect operating parameters of the electric drive system (such as current intensity, operating temperature, torque output, and speed fluctuation) and mechanical component degradation data (such as wear, material fatigue index, and lubrication status parameters), constructing a multi-source heterogeneous dataset encompassing both the electrical and mechanical domains. By arranging the operating parameters and degradation data into full-cycle input tensors for the system and component degradation, respectively, this method achieves a structured representation of multidimensional data in both the temporal and parameter dimensions, laying the 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 coupled relationship between system operating status and component degradation.

[0044] In the feature extraction phase, a bidirectional recurrent neural network (BRNN) is used to perform multi-level feature extraction on the input tensor of the system's full operation cycle. Leveraging the BRNN's bidirectional modeling capabilities for time series data, it effectively captures the time series dependencies of parameters such as current and speed, extracting multi-order dynamic feature tensors that can sensitively identify short-term operating characteristics such as sudden load changes and abnormal fluctuations. Simultaneously, a gated temporal convolutional network (GatedTCN) processes the input tensor of the component degradation cycle. Leveraging the TCN's causal convolution and gating mechanisms, it captures the trend characteristics of degradation parameters such as wear and fatigue index over the long term, enabling multi-scale time series modeling. This domain-specific modeling strategy enables targeted exploration of the time series characteristics of different data types, enhancing the richness of feature expression.

[0045] During the cross-modal fusion phase, the self-attention weight matrix and mutual-attention weight matrix of the system's operational characteristics and component degradation characteristics are calculated through a cross-attention fusion network to generate a dynamic correlation matrix. This enables cross-modal interaction between electrical and mechanical domain features and clarifies the influence weights between different parameters. For example, the strong correlation between gear wear rate and torque output in high-temperature environments can be identified. A physical constraint-based optimization module is introduced to modify the correlation matrix, embedding physical prior knowledge such as the heat conduction laws of motors and the tribological principles of mechanical components into the model to avoid generating feature associations that violate physical laws and improve the model's physical interpretability and robustness. The feature tensor is optimized through a residual feature correction network. Combined with operations such as channel compression and spatial attention weighting, the expressive power of key features is further enhanced and noise interference is suppressed.

[0046] During the lifespan prediction phase, a multi-layer perceptron (MLP) network performs nonlinear mapping on the fused low-dimensional feature encoding vectors, outputting a lifespan prediction value correlated with the time series. This output is capable of capturing complex degradation patterns. The predicted value is then matched against a library of historical degradation cases for similarity, and a probability calibration is used to generate a confidence-level remaining useful life interval, providing a quantitative risk assessment basis for maintenance decisions. For example, the confidence level of the prediction result can be clearly defined, avoiding the uncertainty of a single numerical prediction and enabling users to develop differentiated maintenance strategies based on their risk preferences.

[0047] The comprehensive prediction framework constructed in this paper realizes intelligent full-process management from data acquisition and feature engineering to prediction and evaluation, and can be widely applied in scenarios such as electric vehicle drive systems and industrial motor transmissions. By integrating multi-domain data and embedding physical constraints, it addresses the shortcomings of traditional methods in cross-domain correlation analysis and physical interpretability, improves the generalization capability and engineering reliability of the prediction model, effectively reduces equipment maintenance costs and the risk of sudden failures, and provides advanced technical support for the intelligent operation and maintenance of electric drive systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a working principle diagram of the AI-based electric drive transmission system comprehensive life prediction method described in the present invention;

[0049] Figure 2 Design diagram for cross-modal feature fusion;

[0050] Figure 3 Design diagram for feature optimization and cascading;

[0051] Figure 4 Design diagrams generated for life prediction values;

[0052] Figure 5 Design plots generated for life assessment results. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] See also Figure 1-Figure 5 The present invention relates to an AI-based comprehensive life prediction method for electric drive transmission systems, and the specific implementation steps are as follows:

[0055] The operating sensors collect operating parameters of the electric drive system at multiple monitoring time points within a predetermined cycle. Simultaneously, the condition detection device collects degradation data of the mechanical components at the same multiple monitoring time points. The collection times of the operating parameters and degradation data are strictly aligned to ensure temporal correlation in subsequent data processing.

[0056] The above-collected operating parameters and degradation data are subjected to feature extraction and cross-domain correlation to obtain a system-component degradation fusion feature vector. The specific implementation is as follows:

[0057] The operating parameters at multiple monitoring time points are arranged according to the time dimension and parameter type dimension into a full-cycle input tensor for the system operation; the degradation data of mechanical components at multiple monitoring time points are arranged according to the time dimension and degradation index dimension into a full-cycle input tensor for the component degradation. The tensor structure preserves the time series characteristics and multidimensional properties of the data.

[0058] Multi-scale time series modeling is performed on the input tensors of the full system operation cycle and the full component degradation cycle, respectively, to obtain multi-order feature tensors of system operation and component degradation. The multi-scale modeling process is implemented through a deep learning network, aiming to capture the data variation patterns at different time scales.

[0059] The multi-order feature tensors of system operation and the multi-order feature tensors of component degradation are cross-modally fused, and the association between system-level features and component-level features is achieved through a specific fusion mechanism, and finally a system-component degradation fusion feature vector is generated.

[0060] Based on the system-component degradation fusion feature vector, an AI model generates a life prediction assessment result. This result includes a quantitative prediction and assessment range of the remaining service life of the electric drive transmission system, providing a scientific basis for system maintenance and replacement.

[0061] The present invention will be further described below in conjunction with Examples 1 to 5:

[0062] Example 1:

[0063] Based on the above overall plan, the operating parameters specifically include current intensity, operating temperature, torque output and 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 set outside the main circuit cable of the electric drive system. When the current passes through the cable, the magnetic core generates a magnetic field proportional to the current. The Hall element senses the magnetic field strength and outputs a weak voltage signal. After amplification and filtering by the signal conditioning circuit, it is converted into a standard analog or digital signal. Its measurement range can cover the range of 0-150% of the rated current of the electric drive system, with a resolution of 0.1A. It can capture the transient changes of current under conditions such as motor starting, 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.

[0064] Operating temperature is monitored using distributed temperature sensors, using either Pt100 platinum resistance thermometers or K-type thermocouples. The installation method is adapted to the environmental characteristics of the monitored area: The motor stator winding is embedded in the mounting system. Miniature Pt100 elements are inserted into slots and secured with insulating varnish when the winding is off the assembly line. The bearing seat and gearbox housing are surface-mounted, with thermal grease filling the gap between the sensor and the measured surface to ensure efficient heat transfer. For high-temperature areas (such as power electronics heat sinks), high-temperature-resistant thermocouples are used in conjunction with air-cooled sleeves for measurement. The temperature acquisition module supports multi-channel simultaneous sampling, with a sampling period adjustable from 1s to 10s. Measurement accuracy is ±0.5°C (Pt100) or ±1°C (thermocouple). It continuously records temperature profiles of key components over different operating periods, such as the temperature rise during long-term full-load operation and the temperature decay trend during shutdown and cooling.

[0065] The torque output is directly measured by a torque sensor. The sensor uses the principle of rotating strain gauge torque measurement. The strain gauge is attached to the output shaft of the transmission system to form a Wheatstone bridge. When the shaft is subjected to torque, the strain gauge deforms, causing the bridge to become unbalanced, and outputting 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 to compensate for the coaxiality error during shaft installation. The measurement range is designed according to the rated torque of the electric drive system. With an accuracy of 0.2, it can monitor dynamic torque fluctuations in real time during vehicle starting, climbing, and gear shifting. For example, during vehicle acceleration, the torque sensor can capture the gradual increase in torque as the throttle opening increases, as well as the sudden drop and recovery of torque caused by power interruption at the moment of gear shifting.

[0066] Speed ​​fluctuations are detected by an encoder that collects motor speed signals and calculates the fluctuation coefficient. The encoder uses an incremental photoelectric encoder, which is installed on the non-drive end shaft extension of the motor and rotates synchronously with the rotor. The encoder code disk is engraved with evenly distributed light-transmitting slits. When the code disk rotates with the shaft, light shines through the slits onto the photosensitive element to generate a pulse signal. The pulse frequency is proportional to the speed. The signal processing unit counts and identifies the direction of the pulse signal and calculates the real-time speed value. The sampling frequency can reach 10kHz, and the speed measurement error does not exceed ±1rpm. The calculation of the speed fluctuation coefficient is based on the time series analysis method. First, the speed data is collected every minute to generate a speed sequence with equal time intervals. The standard deviation or coefficient of variation of this sequence is then calculated as an indicator of speed stability. For example, when there is a gear meshing fault in the transmission system, the speed fluctuation coefficient will increase significantly, reflecting the phenomenon of intensified periodic fluctuations in speed.

[0067] During the data acquisition phase, the sampling frequency of each parameter must match the system's operating cycle, employing synchronous acquisition technology to ensure time alignment. This is achieved by setting up a unified clock synchronization module. Using the IEEE1588 precision clock protocol or hardware-synchronized trigger signals, each acquisition channel, including current, temperature, torque, and speed, initiates sampling at the same time. The sampling period is set to 1 second, meaning one complete multi-parameter data point is collected every second. To prevent data loss, a redundant storage mechanism is designed, with collected data stored simultaneously on a local solid-state drive and a remote server. The HDF5 format is used to support efficient storage and rapid retrieval of massive amounts of time-series data.

[0068] 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 minimum-maximum normalization method is used, and the formula is:

[0069]

[0070] in, is the original data, and The normalized data range is mapped to the interval [0, 1]. For temperature parameters, considering the differences in temperature thresholds at different locations, a segmented normalization method is used. Temperature safety thresholds are set for each monitoring point based on the motor insulation class (e.g., Class F or Class H). The measured temperature is compared with the corresponding threshold to calculate the temperature warning coefficient, which ranges from 0 (normal) to 1 (exceeding the threshold). The speed fluctuation coefficient is a dimensionless indicator, so the original calculated value is retained without additional normalization.

[0071] Data preprocessing also includes outlier detection and repair. The Laida criterion based on triple standard deviation (3σ) is used to identify anomalous data points, with points outside the range of ±3σ from the mean being considered outliers. The outlier repair method is selected based on the data characteristics: For isolated outliers that occur occasionally in a single channel, linear interpolation of data from adjacent moments is used to repair them. For abnormal data segments spanning multiple consecutive moments, analysis is conducted in conjunction with the electric drive system's operating condition logs (such as start-stop status and load mode). If the data represents normal fluctuations during the system's non-steady-state operation (such as starting and braking), the original data is retained. If the anomaly is caused by a sensor failure, the historical mean value of the parameter under the same operating conditions is used to fill in the gaps.

[0072] After preprocessing, the operating parameters of multiple monitoring time points are arranged into the input tensor of the system operation cycle according to the time dimension and parameter type dimension. The time dimension corresponds to the acquisition time series. Suppose there are a total of monitoring time points, the length of the time dimension is ; The parameter type dimension includes 4 channels: current intensity, operating temperature, torque output, and speed fluctuation. Each channel has a time dimension. data points. Therefore, the dimension of the input tensor for the full cycle of the system operation is ,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.

[0073] During data transmission, industrial Ethernet communication protocols (such as PROFINET and EtherCAT) are used to achieve high-speed data transmission between the sensor and the data processing unit, with a communication rate of no less than 100 Mbps to ensure real-time data. Furthermore, to prevent electromagnetic interference from affecting signal quality, the sensor signal line uses a twisted-pair shielded cable. The cable shield is reliably grounded at both ends and physically isolated from the power cable by at least 30 cm.

[0074] The entire operating parameter collection and preprocessing process strictly adheres to the principles of timing alignment, dimensional normalization, and anomaly repair, ensuring the integrity, consistency, and reliability of the data input to the subsequent feature extraction module, laying the data foundation for system-level health status assessment. Through the coordinated monitoring and analysis of multi-dimensional operating parameters such as current, temperature, torque, and speed, the electric drive system's electrical performance, mechanical load characteristics, and operational stability are fully reflected, providing rich input information for constructing system-component degradation fusion features.

[0075] Example 2:

[0076] The degradation data of mechanical components specifically include wear volume, material fatigue index, and lubrication status parameters. The wear volume is obtained through ultrasonic testing or oil spectral analysis. Ultrasonic testing uses a contact ultrasonic thickness gauge, whose working principle is based on the propagation characteristics of ultrasonic waves in materials. When the ultrasonic wave is emitted from the probe and enters the component being tested, a reflected echo is 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 the ultrasonic wave in the material and combining it with the known speed of sound. For key components such as gears and bearings, the baseline thickness is determined in advance during their design phase. By regularly measuring the current thickness and comparing it with the baseline value, the wear volume can be obtained. For example, for a certain model of reducer gear, the design baseline tooth thickness is 20mm. After 500 hours of operation, the tooth thickness is measured to be 19.8mm. The wear volume at that moment is 0.2mm.

[0077] Oil spectral analysis collects lubricating oil samples and uses a spectrometer to analyze the composition and concentration of metal particles in the oil. Different metal elements correspond to different component materials. For example, iron mainly comes from steel gears and bearings, while copper may come from copper alloy cages or sliding bearings. By detecting the concentration trend of each element in the oil, the degree of wear of the corresponding component can be indirectly inferred. For example, when the iron concentration shows a continuous upward trend, it indicates that the wear of steel components is increasing. At the same time, combined with a particle counter to analyze the size distribution of particles in the oil, particles larger than 10μm are generally considered to be produced by abrasive wear, and an increase in their number means that the wear condition is worsening.

[0078] The material fatigue index is calculated using strain sensors to monitor stress and strain data of components under cyclic loading, combined with Miner's linear cumulative damage theory. The strain sensors utilize foil strain gauges, which are affixed with a special adhesive to stress-concentrated areas on the component surface, such as gear tooth roots and bearing outer ring raceways. When the component deforms under stress, the strain gauge also deforms, and its resistance changes accordingly. This resistance change is converted into a voltage signal using a Wheatstone bridge circuit. This is then converted to a digital signal using an amplifier and a data acquisition card, resulting in the component's surface strain value.

[0079] During the actual monitoring process, the strain history of the component in multiple working cycles is continuously recorded. The complex strain-time history is decomposed into multiple single-stage stress cycles through the rain flow counting method, and the amplitude and frequency of each stress cycle are counted. According to the SN curve (stress-life curve) of the material, the fatigue life of the material at each stress level is determined. Miner's linear cumulative damage theory believes that the fatigue damage of the material is the linear accumulation of damage caused by each stress cycle. When the cumulative damage reaches 1, the material fails due to fatigue. Therefore, the material fatigue index is defined as the cumulative damage degree, and the calculation formula is:

[0080]

[0081] in, For the The actual number of stress cycles, The fatigue life of the material at this stress level. By calculating the fatigue index in real time, the remaining fatigue life of a 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.

[0082] Lubrication status parameters, including lubricant viscosity and metal particle concentration, are monitored in real time via oil sensors. Lubricant viscosity is a key indicator of lubricant performance, and changes in viscosity directly impact lubrication effectiveness and component wear. A vibration-type viscosity sensor is employed, whose operating principle is based on the relationship between the vibration characteristics of an oscillator in a liquid and the liquid's viscosity. The sensor contains a tuning fork-type oscillator. When the oscillator vibrates in the lubricant, its vibration frequency and amplitude are affected by the oil's viscosity. By detecting changes in the oscillator's vibration parameters, the lubricant's viscosity can be calculated.

[0083] Metal particle concentration is monitored using an inductive particle sensor, which detects ferromagnetic and non-ferromagnetic particles in the oil based on the principle of electromagnetic induction. The sensor contains an excitation coil and two detection coils. When metal particles in the oil pass through the detection area, they induce a change in the electromotive force (EMF) in the detection coils. By analyzing the amplitude and frequency of this EMF change, the particle size and material can be distinguished. For example, ferromagnetic particles cause a larger EMF change, while non-ferromagnetic particles cause a smaller change. Real-time monitoring of particle concentration can identify early signs of abnormal component wear.

[0084] The collection of degradation data requires the installation of dedicated detection devices for key mechanical components. For the reducer gear set, a detection hole is opened on the gearbox housing, and an ultrasonic probe and an oil sampling valve are installed to facilitate regular ultrasonic thickness measurement and oil sampling analysis. At the same time, strain gauges are attached to the gear shaft to monitor the stress changes of the gears during the meshing process. For motor bearings, temperature sensors and vibration sensors are installed on the bearing seat, and the health status of the bearings is comprehensively evaluated in combination with the oil analysis results. The temperature sensor monitors the bearing temperature changes in real time. When the bearing is poorly lubricated or wear is increased, the temperature will rise significantly; the vibration sensor identifies early failures of bearing components by detecting the spectral characteristics of the bearing vibration signal.

[0085] When arranging the input tensor for the full cycle of component degradation, the time dimension corresponds to the sequence of monitoring time points. From the time the equipment is put into operation, monitoring is performed every 100 hours, for a total of monitoring times, the length of the time dimension is The degradation index dimensions include wear volume, material fatigue index, lubricant viscosity, metal particle concentration and other parameters. Each parameter has a time dimension. data points. Furthermore, composite indicators such as wear rate and fatigue damage rate can be derived based on actual needs. The wear rate is defined as the ratio of the change in wear volume between two consecutive monitoring periods to the time interval, reflecting the severity of wear. The fatigue damage rate is the rate of change of the fatigue index over time, reflecting the rate of fatigue accumulation.

[0086] The resulting multidimensional degradation feature matrix contains a wealth of component degradation information, reflecting not only the current degradation state but also the historical evolution of the degradation process. For example, by analyzing the wear curve over time, it is possible to determine whether a component is in the normal wear stage, accelerated wear, or nearing failure. By comparing the changing trends of different parameters, correlations between them can be identified. For example, a decrease in lubricant viscosity may be accompanied by an increase in metal particle concentration, indicating that deteriorating lubrication performance has led to increased wear.

[0087] To ensure the accuracy and reliability of degradation data, a series of quality control measures are implemented during the data collection process. For ultrasonic testing, the probe is calibrated before each measurement, and the accuracy of the measurement system is verified using a standard test block of known thickness. During oil sampling, strict sampling procedures are followed to ensure representative samples and prevent external contamination from affecting the analysis results. For strain sensors and oil sensors, zero point and sensitivity calibration are regularly performed to ensure that the sensor's measurement accuracy is within the specified range.

[0088] At the same time, a data verification mechanism is established to verify the rationality of the collected degradation data. For example, if the wear volume at a certain moment suddenly increases and exceeds the normal fluctuation range, repeated measurements will be performed to eliminate the possibility of measurement error. If the multiple measurement results are consistent, the operating status of the relevant components will be further checked, and a comprehensive analysis combined with other parameters (such as temperature and vibration) will be conducted to determine whether abnormal wear has indeed occurred.

[0089] The entire degradation data collection and processing process strictly adheres to standardized operating procedures to ensure data authenticity, integrity, and traceability. By monitoring multi-dimensional degradation data of key mechanical components, we can fully understand the evolution of component health status, providing a solid data foundation for subsequent multi-scale time series modeling and life prediction. This degradation data, combined with system operating parameters, can more accurately reflect the comprehensive degradation state of the electric drive transmission system, thereby enabling more precise life prediction and assessment.

[0090] Example 3:

[0091] The multi-scale time series modeling process is as follows: the full-cycle input tensor of the system operation is fed 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: forward and reverse. The forward recurrent layer processes data from the front to the back of the time series, while the reverse recurrent layer processes data from the back to the front, enabling the network to capture dependencies between the context of the time series. Taking the Bidirectional Long Short-Term Memory Network (Bi-LSTM) as an example, its core unit is the LSTM memory cell. Each memory cell contains a forget gate, an input gate, and an output gate. This gated mechanism controls the retention and updating of information, effectively alleviating the vanishing gradient problem of traditional recurrent neural networks and is suitable for capturing time-dependent features in long sequences.

[0092] The dimension of the input tensor during the full cycle of system operation is ,in Indicates the length of the time dimension (i.e. the number of monitoring time points), The number of channels representing the operating parameters (e.g., the number of parameter types such as current intensity and operating temperature). The input tensor is first dimensionally expanded by the embedding layer and converted to A three-dimensional tensor of is the embedding dimension, which is used to improve the expressiveness of features. Subsequently, the three-dimensional tensor is input into the forward and reverse loop layers of the Bi-LSTM network, and the hidden state dimension of each loop layer output is ,in is the number of neurons in the hidden layer. By stacking multiple layers of Bi-LSTM networks (e.g., 2 layers), a multi-level feature extraction structure is formed: the first layer extracts low-order temporal features (e.g., parameter fluctuation patterns within a single monitoring cycle), and the second layer extracts high-order temporal features (e.g., trend changes across multiple cycles) based on the low-order features output by the first layer. Ultimately, the system operation multi-level feature tensor output by the multi-level feature extraction network has a dimension of ,in Represents the number of feature levels (i.e., the number of network layers). This tensor contains system operation characteristics at different time scales (short-term, medium-term, long-term) and different levels of abstraction.

[0093] For the full cycle input tensor of component degradation, its dimension is ,in is the length of the time dimension (aligned with the time point of the system operation parameters), is the number of degradation indicator channels (e.g., the number of parameter types such as wear volume and material fatigue index). This is input into a degradation feature extraction network based on a gated temporal convolutional network (GatedTCN). The GatedTCN combines a temporal convolutional network (TCN) with a gating mechanism. The temporal convolution layer uses dilated convolution technology to expand the receptive field without increasing the number of parameters, enabling it to capture long-range temporal dependencies. The gating mechanism generates a gating signal using a sigmoid activation function to selectively filter the features output by the convolution layer, automatically learning the importance of features at different time steps.

[0094] In the specific implementation, the full-cycle input tensor of the component degradation is firstly processed through a one-dimensional convolution layer for preliminary feature extraction. The convolution kernel size is , the generated dimension is The feature map of is the number of convolution kernels. Then it enters the gated temporal convolution module, each module contains two parallel dilated convolution layers: one for extracting the linear part of the feature and the other for generating the gating signal. The output of the linear part is recorded as , the gate signal is recorded as ,in is the sigmoid function, is the gated convolution kernel weight, is the bias term, Represents the convolution operation. The final output is By stacking multiple gated temporal convolution modules, and the expansion rate of each module increases exponentially (such as ), which enables the network to capture the degradation time series features at different scales from local to global. Assuming stacking modules, and the feature dimension of each module output is ( ), and finally the output features of each module are fused through the cross-layer connection (SkipConnection) to obtain the multi-order feature tensor of component degradation, whose dimension is ,in is the number of output feature channels, Indicates the number of multi-scale features (corresponding to features at different expansion rates).

[0095] During network training, an end-to-end supervised learning approach is employed, with the mean squared error (MSE) between the system-component degradation fusion feature vector and the life prediction label as the loss function. Network parameters are optimized via a backpropagation algorithm. For bidirectional recurrent neural networks, input data undergoes sequence padding to ensure that time series of varying lengths can be batched into the network. The hidden states at the padded positions are masked during calculations to eliminate their influence. Gated temporal convolutional networks address causal convolution, meaning that the convolution operation at the current time step relies solely on data from the past and current time steps, avoiding future data. This ensures the model's temporal rationality in practical applications.

[0096] The multi-scale characteristics of the multi-order feature tensors of system operation and component degradation are reflected in the following: the former captures operating modes of different time spans (such as current fluctuations in seconds and temperature trends in hours) through the hierarchical stacking of recurrent neural networks, while the latter captures local details of degradation features (such as changes in wear within a single monitoring cycle) and global trends (such as the long-term growth of cumulative fatigue damage) through the different receptive fields of dilated convolution. The two feature tensors provide multi-level temporal feature representations from the two dimensions of system operation status and component degradation process, laying the foundation for subsequent cross-modal fusion. Through this multi-scale modeling approach, the long-term and short-term dependencies implicit in the data can be effectively extracted, avoiding the limitations of single-scale feature representation, thereby more comprehensively characterizing the degradation process of the electric drive transmission system.

[0097] Example 4:

[0098] The cross-modal fusion process specifically involves the following operations: The multi-order system operation feature tensor and the multi-order component degradation feature tensor are input into a cross-attention fusion network to obtain a system-component degradation dynamic correlation matrix. Taking the correlation analysis between the motor operating state and reducer gear wear in an electric drive transmission system as an example, it is assumed that the multi-order system operation feature tensor contains a multi-level abstract representation of motor characteristics such as current intensity and speed fluctuation at different time points, while the multi-order component degradation feature tensor contains multi-scale features of degradation indicators such as gear wear and metal particle concentration in the lubricant. The cross-attention fusion network first calculates the self-attention weight matrix of the system operation features. This matrix reflects the interdependence of parameters such as current and speed at different time points, such as the strong correlation between current and speed during motor startup. Simultaneously, the cross-attention weight matrix between the component degradation features and the system operation features is calculated to capture the cross-modal correlation between gear wear and motor load parameters. For example, periods of high torque output may correspond to increased gear wear rate. The system-component degradation dynamic correlation matrix is ​​generated by the matrix multiplication operation of the self-attention matrix and the mutual attention matrix. Each element in this matrix represents the correlation strength between a system operation characteristic and a component degradation characteristic at a specific time point, such as the correlation value between the motor current characteristic at time t and the gear wear amount at time t-1.

[0099] The dynamic correlation matrix is ​​input into the physical constraint-based optimization module to generate a 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 mesh mechanics model. For example, the motor heat conduction model describes the relationship between motor winding temperature, current intensity, and operating time. If a negative correlation between current and temperature characteristics at a given moment appears in the dynamic correlation matrix (violating the physical law that increasing current leads to rising temperature), the optimization module automatically corrects this correlation to a reasonable range. For another example, the gear mesh mechanics model defines a positive correlation between torque output and gear contact stress and wear. If the correlation matrix shows an unusual correlation between high torque periods and low wear, the module adjusts the strength of this correlation based on the physical mapping between torque and wear. This type of physical constraint correction eliminates spurious correlations that violate fundamental principles such as energy conservation and material mechanics, 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 (increased wear and fatigue index)."

[0100] The multi-order feature tensors of system operation and component degradation are fed into the residual feature correction network to obtain optimized feature tensors. For system operation features, channel dimension compression is first performed. For example, the feature tensor containing four channels—current, temperature, torque, and speed—is compressed into a one-dimensional feature vector through a fully connected layer. This vector embodies a comprehensive representation of the system's operating state. Next, a weighted fusion is performed on the constraint-corrected correlation matrix and the compressed system operation feature vector. Specifically, the row vectors associated with the system features in the correlation matrix are extracted and element-wise multiplied with the compressed vector. Nonlinear correlations are then enhanced using an activation function (such as ReLU). Feature expansion is then performed to restore the one-dimensional vector to the same channel dimension as the original feature tensor, resulting in the optimized multi-order feature tensor of system operation. This process can be understood as leveraging the dynamic system-component correlation information in the correlation matrix to assign weights to system operation features, emphasizing parameters highly correlated with component degradation (e.g., increasing the weight of feature channels corresponding to high torque) while suppressing irrelevant or interfering features.

[0101] Spatial attention weighting is first performed on the multi-order component degradation feature tensor. Taking four degradation indicators—gear wear, fatigue index, lubricant viscosity, and metal particle concentration—as examples, the spatial attention mechanism generates attention weights for each indicator along the time dimension through convolution operations. For example, if metal particle concentration has recently shown an upward trend, the corresponding weight will be significantly increased. The weighted component degradation features are then modulated with the constraint-corrected correlation matrix. Specifically, column vectors associated with component degradation features are extracted from the correlation matrix and used as modulation parameters to scale and translate the features of each degradation indicator. For example, if the correlation matrix shows a high correlation between high motor speed fluctuations and gear fatigue index during a certain period, the fatigue index feature for that period is multiplied by a scaling factor greater than 1 to enhance its importance in the fused feature. This feature modulation enables component degradation features to adaptively incorporate the influence of system operating conditions, embodying the physical logic that "system load changes drive component degradation."

[0102] Feature cascade encoding is performed on the optimized multi-order feature tensors of system operation and component degradation. For example, assuming the dimensions of the optimized system operation feature tensor are [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 dimensions of the component degradation feature tensor are [T, C2, L2] (C2 is the number of degradation indicator channels, and L2 is the number of multi-scale features), feature cascade encoding involves concatenating the two along the channel dimension into 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 features of the motor current (a high-level output of the system operation features) and the global features of the gear wear rate (a large-scale output of the component degradation features) are concatenated at the same time point to form a fused feature vector containing cross-domain information between the system and the component. This cascade encoding method retains the original structure of the two types of features, while enhancing the physical correlation between features through cross-modal correlation optimization, so that the fused features can reflect both the dynamic process of system operation and the historical accumulation of component degradation, providing more comprehensive input information for subsequent life prediction.

[0103] In practical applications, taking the electric drive system of a certain electric vehicle as an example, the multi-order feature tensor of system operation may capture the following patterns: Under urban driving conditions with frequent vehicle starts and stops, the motor current exhibits high-frequency pulse-like fluctuations (short-term characteristics), while long-term monitoring shows that the average torque output is 15% lower than under high-speed conditions (long-term characteristics). The multi-order feature tensor of component degradation shows that the wear rate of the reducer input shaft bearing increases 20% higher under start-stop conditions than under uniform speed conditions (local-scale characteristics), while the cumulative fatigue index increases linearly throughout the monitoring period (global-scale characteristics). The cross-attention fusion network, by calculating self-attention, identifies the strong coupling between current and speed fluctuations under start-stop conditions (corresponding to higher values ​​in the self-attention weight matrix). Through mutual attention, it discovers a positive correlation between current pulse frequency and bearing wear rate (corresponding to higher values ​​in the mutual-attention matrix), thereby strengthening the association path of "start-stop → current fluctuation → accelerated bearing wear" in the dynamic association matrix. Based on the physical model of motor-bearing load transfer, the physical constraint optimization module confirms that high current fluctuations correspond to high-frequency changes in the motor's output torque, which in turn leads to increased alternating loads on the bearings. Therefore, this association is retained and enhanced, while false associations that violate mechanical principles, such as "low speed corresponds to high wear," are eliminated.

[0104] When the residual feature correction network processes system operation characteristics, the channel compression vector emphasizes the weight of the current fluctuation feature (due to its strong correlation with component degradation). After weighted fusion and expansion, the eigenvalue 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 component degradation characteristics are processed through spatial attention, which increases the weight of the bearing wear feature during the start-stop period by 30%. Through modulation of the correlation matrix, the wear eigenvalue during this period is linearly correlated and scaled with the current eigenvalue at the same time, further strengthening the physical connection between the two. The final cascade encoding method aligns the "start-stop-current fluctuation" characteristics of system operation with 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 input data for the life prediction model that directly reflects the "system operating condition-component degradation" coupling process.

[0105] The entire cross-modal fusion process uses an attention mechanism to explore potential connections between data, incorporates physical constraints to ensure the rationality of these connections, and then achieves deep fusion of cross-domain information through feature correction and cascading. This approach avoids the limitations of traditional data fusion, which relies solely on simple splicing. It can dynamically capture nonlinear and time-varying connections between systems and components, while leveraging domain knowledge (physical models) to enhance the interpretability and reliability of fused features, providing more physically meaningful feature representations for life prediction of electric drive transmission systems.

[0106] Example 5:

[0107] The process for generating life prediction assessment results is as follows: First, multi-dimensional dimensionality reduction is performed on the system-component degradation fusion feature vector. Taking an electric drive system as an example, assume 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 can be performed using principal component analysis (PCA). This linear transformation projects high-dimensional features into a low-dimensional space, 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 key information of the original data while reducing computational complexity. The reduced dimensionality vector contains a comprehensive representation of key information such as long-term current trends, gear wear rates, and bearing fatigue index.

[0108] The low-dimensional feature encoding vector is then input into a multi-layer perceptron (MLP) network for nonlinear relationship mapping. The MLP network consists of multiple fully connected layers, assuming three hidden layers, with 128, 64, and 32 neurons in each layer, respectively, and a Reluctant Unit (ReLU) activation function. The first hidden layer receives a 20-dimensional input vector and, through a linear transformation using the weight matrix and bias term, adds a Reluctant Unit (ReLU) activation to extract first-order nonlinear combinations of 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 nonlinear features (such as the multiplicative effect of high load duration and wear). The third hidden layer integrates the second-order features into an abstract representation directly related to the remaining useful life. The final output layer undergoes a linear transformation to produce a scalar value as the lifespan prediction, representing the estimated time (in hours) remaining from the electric drive system's current state until failure. For example, for a system that has operated for 500 hours, the MLP outputs a lifespan prediction of 800 hours, indicating an estimated remaining operating time of 800 hours.

[0109] Furthermore, the life prediction values ​​are matched with the historical degradation case library for similarity. The historical degradation case library stores the degradation data of different electric drive systems under various operating conditions and the corresponding actual life data. Each case contains a fused feature vector, operating condition records (such as average load, start-stop frequency), degradation mode description (such as gear abrasive wear-dominated, bearing fatigue failure-dominated), and actual remaining life. Similarity matching uses the cosine similarity calculation method to compare the current fused feature vector with the feature vectors of all cases in the case library one by one, and calculate the cosine value 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-dominated under high load).

[0110] Based on the similarity matching results, a degradation pattern matching weight distribution is derived. This weight distribution reflects the contribution of each historical case to the current prediction, with cases with higher similarity receiving greater weights. Assuming there are 10 cases in the case library, after matching, case C-001 is weighted 30%, case C-005 is weighted 25%, and the remaining cases are weighted 45% in total. This weight distribution is used to probabilistically calibrate the life prediction values. This calibration process employs a Bayesian network framework, using the remaining life of the 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 for case C-001 is 700-900 hours (mean 800 hours), and for case C-005 it is 650-850 hours (mean 750 hours). A weighted average is used to generate a confidence interval for the remaining useful life. Assuming a 95% confidence interval of 720-880 hours after calibration, this interval reflects the likely range of the current system's remaining useful life while accounting for the diversity and uncertainty of historical cases.

[0111] Taking a specific application scenario as an example, the electric drive system of an electric bus is monitored after 600 hours of operation. The collected operating parameters include: average current 80A (rated current 100A), motor temperature 75℃ (safety threshold 100℃), torque output (Peak torque ), speed fluctuation coefficient 0.05 (normal range ≤ 0.1); degradation data showed gear wear of 0.3mm (initial tooth thickness 20mm), bearing fatigue index 0.4, and lubricant metal particle concentration 50ppm (warning level 80ppm). After feature extraction and cross-modal fusion, a fused vector containing 40-dimensional features was obtained. This vector was then reduced to 15 dimensions and input into the MLP network, outputting a lifespan prediction of 750 hours.

[0112] Three highly similar cases were matched in the historical case database: Case A (urban bus operating conditions, dominated by gear wear, remaining life 700-800 hours, similarity 0.88); Case B (frequent start-stop operating conditions, dominated by bearing fatigue, remaining life 720-850 hours, similarity 0.82); and Case C (mixed operating conditions, poor lubrication leading to compound degradation, remaining life 680-820 hours, similarity 0.79). Based on the weight distribution (40% for Case A, 35% for Case B, and 25% for Case C), 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. Rounding this to obtain a 95% confidence interval of 700-820 hours. This result provides operation and maintenance personnel with a clear maintenance time window, which can be used to formulate maintenance plans (such as changing gear oil and checking bearing condition before the remaining life reaches 500 hours).

[0113] In addition to PCA, nonlinear dimensionality reduction methods such as t-SNE can also be used in dimensionality reduction. These methods are suitable for scenarios where complex nonlinear relationships exist between features. For example, when system operating parameters and component degradation indicators exhibit nonlinear coupling (such as exponential growth in wear rate after the temperature exceeds a threshold), t-SNE can better preserve the local structure of the data, allowing low-dimensional features to more accurately reflect the manifold distribution in the high-dimensional space. The structure of the MLP network can be adjusted based on data characteristics. If the fused features have strong temporal correlation, a one-dimensional convolutional layer can be added before the MLP to first extract local features in the time dimension before performing nonlinear mapping.

[0114] The construction of a historical case library must adhere to the principles of data diversity and representativeness, covering different types of electric drive systems (e.g., passenger car and commercial vehicle motors), multiple degradation modes (wear, fatigue, corrosion), and typical operating conditions (high speed, low speed, alternating load). Case data should include complete lifecycle monitoring data, encompassing parameters from the equipment's commissioning to failure and disposal, to ensure reliable similarity matching. For example, the case library must include cases of different failure modes, such as early pitting corrosion of gears and cracks in bearing outer rings, along with the corresponding degradation characteristic evolution processes.

[0115] During the probability calibration process, if there are insufficient historical cases, bootstrap resampling can be used to generate virtual cases. This method, through sampling with replacement from existing data, increases the number of cases and estimates the uncertainty of the prediction. Furthermore, expert knowledge can be incorporated to modify the weight distribution. For example, when a new lubricant is used in the current system, experts can manually reduce the weight of lubrication-related degradation cases to improve the rationality of case matching.

[0116] The entire life prediction and assessment process transforms abstract fusion features into remaining life prediction intervals with practical engineering significance through data dimensionality reduction, nonlinear modeling, similarity matching, and probability calibration. This method leverages the powerful fitting capabilities of AI models and incorporates the statistical laws of historical data. It can provide uncertainty assessments based on group experience while considering individual system differences, meeting the accuracy and reliability requirements of life prediction in industrial scenarios. Operations and maintenance personnel can optimize maintenance strategies based on the prediction results, avoiding resource waste caused by excessive maintenance while preventing safety risks caused by component failure, thereby achieving intelligent operation and maintenance management of electric drive systems.

[0117] It should be noted that, in this document, relational terms such as first and second, etc., are used only 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 terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0118] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based comprehensive life prediction method for electric drive transmission systems, characterized by: include: Acquiring operating parameters of the electric drive system at multiple monitoring time points within a predetermined period collected by an operating sensor, and mechanical component degradation data at the multiple monitoring time points collected by a state detection device; performing feature extraction and cross-domain correlation on the operating parameters at the plurality of monitoring time points and the mechanical component degradation data at the plurality of monitoring time points to obtain a system-component degradation fusion feature vector; generating a life prediction assessment result based on the system-component degradation fusion feature vector; The process of extracting features and performing cross-domain correlation on the operating parameters at the plurality of monitoring time points and the mechanical component degradation data at the plurality of monitoring time points to obtain a system-component degradation fusion feature vector includes: Arranging the operating parameters at the plurality of monitoring time points into a system operation full cycle input tensor according to a time dimension and a parameter type dimension, and arranging the mechanical component degradation data at the plurality of monitoring time points into a component degradation full cycle input tensor according to a time dimension and a degradation index dimension; Performing multi-scale time series modeling on the full-cycle input tensor of the system operation and the full-cycle input tensor of the component degradation to obtain a multi-order feature tensor of the system operation and a multi-order feature tensor of the component degradation; Performing 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; Cross-modal fusion is performed 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, including: Inputting the system operation 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 correlation matrix; Inputting the system-component degradation dynamic correlation matrix into a physical constraint-based optimization module to obtain a constraint-corrected correlation matrix; Inputting the system operation multi-order feature tensor and the component degradation multi-order feature tensor into a residual feature correction network respectively to obtain an optimized system operation multi-order feature tensor and an optimized component degradation multi-order feature tensor; Feature cascade encoding is performed on the optimized system operation multi-order feature tensor and the optimized component degradation multi-order feature tensor to obtain the system-component degradation fusion feature vector.

2. The AI-based electric drive transmission system comprehensive life prediction method according to claim 1 is characterized in that: The operating parameters include current intensity, operating temperature, torque output and speed fluctuation.

3. The AI-based electric drive transmission system comprehensive life prediction method according to claim 2 is characterized in that: The degradation data includes wear volume, material fatigue index and lubrication state parameters.

4. The AI-based electric drive transmission system comprehensive life prediction method according to claim 3 is characterized in that: Multi-scale time series modeling is performed on the full-cycle input tensor of the system operation and the full-cycle input tensor of the component degradation to obtain a multi-order feature tensor of the system operation and a multi-order feature tensor of the component degradation, including: Inputting the full-cycle input tensor of the system operation into a multi-level feature extraction network based on a bidirectional recurrent neural network to obtain a multi-order feature tensor of the system operation; The full-cycle input tensor of component degradation is input into a degradation feature extraction network based on a gated temporal convolutional network to obtain a multi-order feature tensor of component degradation.

5. The AI-based electric drive transmission system comprehensive life prediction method according to claim 1 is characterized in that: Inputting the system operation 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, including: The self-attention weight matrix of the multi-order feature tensor of the system operation and the mutual attention weight matrix of the multi-order feature tensor of the component degradation are calculated, and the system-component degradation dynamic association matrix is ​​generated through matrix multiplication operation.

6. The AI-based electric drive transmission system comprehensive life prediction method according to claim 5 is characterized in that: Inputting the system operation multi-order feature tensor and the component degradation multi-order feature tensor into the residual feature correction network respectively to obtain the optimized system operation multi-order feature tensor and the optimized component degradation multi-order feature tensor, including: Performing channel dimension compression on the multi-order feature tensor of the system operation to obtain a system operation feature compression vector; Performing a feature expansion operation after weighted fusion of the constraint correction correlation matrix and the system operation feature compression vector to obtain an optimized system operation multi-order feature tensor; The component degradation multi-order feature tensor is subjected to spatial attention weighting and then feature modulated with the constraint correction association matrix to obtain an optimized component degradation multi-order feature tensor.

7. The AI-based electric drive transmission system comprehensive life prediction method according to claim 6 is characterized in that: Based on the system-component degradation fusion feature vector, a life prediction assessment result is generated, including: Inputting the system-component degradation fusion feature vector into a regression prediction model to obtain a life prediction value; The life prediction evaluation result including the remaining service life interval is generated according to the life prediction value.

8. The AI-based electric drive transmission system comprehensive life prediction method according to claim 7 is characterized in that: 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 encoding vector; The low-dimensional feature encoding vector is input into a multi-layer perceptron network for nonlinear relationship mapping, and the life prediction value associated with the time series is output.

9. The AI-based electric drive transmission system comprehensive life prediction method according to claim 8, characterized in that: Generating the life prediction evaluation result including the remaining service life interval according to the life prediction value includes: Performing similarity matching on the life prediction value and the historical degradation case library to obtain a degradation pattern matching weight distribution; The life prediction value is probability calibrated based on the degradation pattern matching weight distribution to generate the remaining service life interval with confidence.

Images