A method and system for predicting operational failure of a hybrid tractor

Abstract

The application relates to the technical field of agricultural machinery, and discloses a hybrid power tractor operation fault prediction method and system, the method comprises the following steps: acquiring power switching data sets at the time of power switching and extracting multi-factor interaction features; performing wear state classification, determining a fault prediction signal in combination with a random forest algorithm, performing model training according to historical load fluctuation characteristics, and obtaining a wear identification model; extracting real-time load data for data processing and outputting a real-time fault prediction probability; if the real-time fault prediction probability is greater than a preset probability threshold, activating an alarm mechanism and obtaining an alarm activation state; generating a fault prediction report based on the alarm activation state and obtaining a final fault prediction report. The method can realize accurate identification of potential fault risks under the influence of frequent power system switching and complex environments.

Description

Technical Field

[0001] This invention relates to the field of agricultural machinery technology, and in particular to a method and system for predicting operational faults in a hybrid tractor. Background Technology

[0002] Currently, in the process of modern agricultural mechanization, hybrid tractors are important equipment for improving agricultural production efficiency and energy conservation and emission reduction. Fault prediction and health management of mechanical equipment to ensure stable operation are crucial for the sustainability of agricultural production.

[0003] In existing technologies, a single-parameter monitoring method based on a fixed threshold is typically used. This involves simple comparisons of isolated data such as engine speed and battery voltage collected by sensors, or reliance on periodic manual inspections and experience-based judgment. Existing technologies lack in-depth insights into the interaction of multiple factors during operation, especially in complex systems like hybrid tractors that integrate traditional fuel power and electric drive. Focusing solely on the operating data of a single component cannot comprehensively reflect the overall health of the equipment, leading to many potential problems being overlooked in their early stages. More importantly, the frequent switching between fuel and electric modes involves energy distribution balance issues and causes additional wear and stress accumulation on core components. This wear and stress are further affected by agricultural environmental factors such as mud, slopes, or extreme temperature differences, resulting in increased switching frequency and load fluctuations. Potential faults are difficult to accurately detect, potentially leading to downtime at critical agricultural periods and severely impacting agricultural production progress.

[0004] Existing technologies have the problem of failing to accurately identify potential failure risks under the influence of frequent power system switching and complex environments. Summary of the Invention

[0005] This invention provides a method and system for predicting operational faults in hybrid tractors, enabling accurate identification of potential fault risks under the influence of frequent power system switching and complex environments.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for predicting operational faults in a hybrid tractor, comprising: The power switching dataset is collected in real time by the sensor group during power switching. Based on the power switching dataset, interaction features based on coupling judgment are extracted to obtain multi-factor interaction features. Based on the aforementioned multi-factor interaction features, a support vector machine algorithm is used to classify wear status, and the classified wear index is obtained. Based on the wear index, high-risk areas are identified using a preset random forest algorithm model, and corresponding fault prediction signals are determined. Historical operation log data is extracted based on the fault prediction signals, and an interaction correlation matrix is ​​constructed. Based on the interaction correlation matrix and the fault prediction signals, data fusion is performed to obtain an interaction dataset. The interactive dataset is analyzed over time to obtain historical load fluctuation characteristics. The model is then trained based on these historical load fluctuation characteristics to obtain a wear identification model. Extract real-time load data from the power switching dataset, input the real-time load data into the wear identification model for data processing, and output the real-time fault prediction probability; If the real-time fault prediction probability is greater than the preset probability threshold, the alarm mechanism is activated, and the alarm is activated. Environmental data is collected based on the alarm activation status to obtain the current environmental variables. The fault prediction report is generated by combining the real-time load data and the current environmental variables to obtain the final fault prediction report.

[0007] Secondly, the present invention provides a fault prediction system for a hybrid tractor, comprising: The data feature extraction module is used to acquire the power switching dataset collected in real time by the sensor group during power switching, and to perform interaction feature extraction based on coupling judgment based on the power switching dataset to obtain multi-factor interaction features. The wear state classification module is used to classify the wear state based on the multi-factor interaction features using a support vector machine algorithm, and obtain the classified wear index. The signal data enhancement module is used to determine high-risk areas and corresponding fault prediction signals based on the wear index using a preset random forest algorithm model, extract historical operation log data based on the fault prediction signals, construct an interaction correlation matrix, and perform data fusion based on the interaction correlation matrix and the fault prediction signals to obtain an interaction dataset. The model optimization and training module is used to perform time-series analysis on the interactive dataset to obtain historical load fluctuation characteristics, and to train the model based on the historical load fluctuation characteristics to obtain a wear identification model. The real-time fault prediction module is used to extract real-time load data from the power switching dataset, input the real-time load data into the wear identification model for data processing, and output the real-time fault prediction probability. The alarm threshold judgment module is used to activate the alarm mechanism and obtain the alarm activation status if the real-time fault prediction probability is greater than the preset probability threshold. The fault report generation module is used to collect environmental data based on the alarm activation status, obtain the current environmental variables, and generate a fault prediction report by combining the real-time load data and the current environmental variables, thus obtaining the final fault prediction report.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts the multi-factor interaction features of the power switching dataset and uses the support vector machine algorithm to classify and obtain wear indicators. This breaks the limitation of the existing technology of monitoring data of a single component. It can capture the wear features of power components, energy distribution system and environmental factors during the fuel-electric power switching process, solve the problem that potential faults are easily overlooked in the incipient stage, and realize accurate identification of early wear status under multi-factor interaction.

[0009] (2) This invention generates high-risk area fault prediction signals through random forest algorithm, combines historical operation logs for data augmentation and retrains the wear identification model, which can supplement scarce working condition data in power switching and complex farmland environments. Model retraining can improve the adaptability to load fluctuations and environmentally induced wear, solve the problem that existing technologies are difficult to capture fault hazards caused by switching frequency and load fluctuations, and realize the enhanced model's ability to generalize to complex working conditions and accurately capture potential fault risks.

[0010] (3) The present invention obtains the probability of fault prediction by inputting real-time load data into the wear identification model. When the threshold is exceeded, a fault prediction report is generated by combining the current environmental variables. This can realize early warning of fault risk. Combining environmental variables can clarify the cause of fault, solve the problem that the existing technology is prone to causing downtime during critical agricultural seasons, realize real-time warning of potential faults and assist in accurate operation and maintenance decision-making. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of the operation fault prediction method for a hybrid tractor provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the operation fault prediction system for a hybrid tractor provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1The first embodiment of the present invention provides a method for predicting operational faults in a hybrid tractor, comprising the following steps: S11, acquire the power switching dataset collected in real time by the sensor group during power switching, and perform interaction feature extraction based on coupling judgment according to the power switching dataset to obtain multi-factor interaction features; S12, Based on the multi-factor interaction features, the support vector machine algorithm is used to classify the wear status and obtain the classified wear index. S13. Based on the wear index, high-risk areas are determined using a preset random forest algorithm model, and corresponding fault prediction signals are determined. Historical operation log data is extracted based on the fault prediction signals, and an interaction correlation matrix is ​​constructed. Based on the interaction correlation matrix and the fault prediction signals, data fusion is performed to obtain an interaction dataset. S14, perform time-series analysis on the interactive dataset to obtain historical load fluctuation characteristics, and train the model based on the historical load fluctuation characteristics to obtain a wear identification model; S15, extract the real-time load data from the power switching dataset, input the real-time load data into the wear identification model for data processing, and output the real-time fault prediction probability; S16, If the real-time fault prediction probability is greater than the preset probability threshold, then the alarm mechanism is activated and the alarm activation state is obtained; S17. Based on the alarm activation state, environmental data is collected to obtain the current environmental variables. The fault prediction report is generated by combining the real-time load data and the current environmental variables to obtain the final fault prediction report.

[0014] In step S11, it is necessary to obtain the power switching dataset collected in real time by the sensor group during power switching, and to perform interaction feature extraction based on coupling judgment according to the power switching dataset to obtain multi-factor interaction features.

[0015] It is worth noting that this step extracts multi-factor interaction features that reflect the operating status of the power system by real-time monitoring of the changes in multi-physics parameters during the power switching process, providing a data basis for subsequent wear status classification. This solves the problem of data redundancy and missing key information in traditional monitoring methods. Through targeted transient data extraction and cross-domain parameter fusion, it achieves a quantitative characterization of the coupling effect of mechanical stress and electrical shock during the power switching process.

[0016] In one implementation, the power switching dataset, which is collected in real time by the sensor array during power switching, includes: Acquire raw operation time-series data collected in real time by the sensor group, and extract the torque output ratio status of the raw operation time-series data; If the torque output ratio state is reversed, the original operation timing data is extracted to obtain the power switching dataset.

[0017] It is worth noting that the original operating time-series data is collected in real time by the sensor array during the operation of the hybrid tractor, including multi-dimensional time-series parameters such as speed, torque, oil temperature, current, voltage, and oil pressure. Among them, torque includes motor output torque, engine output torque, and total output torque; the torque output ratio status is the ratio of motor output torque and engine output torque to the total output torque in the hybrid system, used to determine the dominant state of the power source; the determination of torque output ratio reversal is that when the motor output torque ratio decreases for three consecutive sampling points and the engine output torque ratio increases for three consecutive sampling points, and the sum of the changes in the two ratios exceeds 15%, it is determined that the state is reversed; the sampling period is 10 milliseconds, and the sum of the motor torque ratio and the engine torque ratio is always 100%.

[0018] Specifically, the power switching dataset is a subset of data within a specific time window extracted from the original operating time-series data when the torque output ratio reverses. It includes key operating parameters at the moment of power mode transition, including but not limited to multi-dimensional parameters such as speed, torque, oil temperature, current, voltage, and oil pressure. For example, the system continuously collects original operating time-series data through a sensor array deployed in the powertrain. If it detects that the motor output torque ratio decreases from 70% to 30%, while the diesel engine output torque ratio increases from 30% to 70%, the system determines that the torque output ratio has reversed and then extracts data within two seconds before and after this point in time to form the power switching dataset.

[0019] In one implementation, interaction features based on coupling judgment are extracted from the power switching dataset to obtain multi-factor interaction features, including: Based on the power switching dataset, a component operating parameter matrix regarding voltage fluctuations and oil pressure changes is constructed; Acquire terrain data and fuse the terrain data with the component operating parameter matrix to obtain an initial dataset; The coupling coefficient is calculated based on the initial dataset to obtain the coupling coefficient. If the coupling coefficient is greater than the preset coupling threshold, then the multi-factor interaction features are extracted from the power switching dataset.

[0020] It is worth noting that the multi-factor interaction feature is a quantitative index obtained by fusing the component operating parameter matrix with terrain data and calculating the coupling coefficient. It is used to characterize the interaction strength between electrical system parameters and hydraulic system parameters under specific operating environment conditions, and serves as the input feature for subsequent wear state classification.

[0021] The component operating parameter matrix is ​​a two-dimensional numerical matrix constructed based on the power switching dataset. The row dimension corresponds to the time series sampling points, and the column dimension corresponds to the voltage fluctuation numerical sequence and the oil pressure change numerical sequence, respectively. The voltage fluctuation numerical sequence consists of battery terminal voltage data collected by the battery management system at a millisecond sampling frequency, and the oil pressure change numerical sequence consists of hydraulic circuit pressure data of the power shift clutch or transmission collected by the hydraulic system pressure sensor at the same sampling frequency. The matrix elements are the physical quantity values ​​at the corresponding sampling time.

[0022] Furthermore, the terrain data, including farmland slope, soil resistance coefficient, and surface flatness, is acquired in real time by onboard tilt sensors, soil resistance sensors, or geographic information systems. This data characterizes the current operating environment load conditions of the tractor and serves only as environmental background parameters for subsequent interpretation of coupling strength; it does not participate in the calculation of coupling coefficients. Data fusion refers to aligning the terrain data as an environmental feature vector with the component operating parameter matrix using timestamps to obtain the initial dataset.

[0023] Specifically, the timestamps of terrain data acquisition are matched with the time axis of the component operating parameter matrix, so that each row of operating parameters corresponds to the relevant terrain conditions, forming a multi-dimensional data set including electrical, hydraulic, and environmental parameters. The coupling coefficient is a normalized index used to quantify the correlation between voltage fluctuations and oil pressure changes. In one implementation, the coupling coefficient is obtained by calculating the absolute value of the Pearson correlation coefficient between the voltage fluctuation sequence and the oil pressure change sequence in the initial dataset, as follows.

[0024] in, Indicates the first Voltage values ​​at each sampling point This represents the average value of the voltage sequence. Indicates the first Oil pressure values ​​at each sampling point This represents the average value of the oil pressure sequence. This indicates the total number of sampling points.

[0025] It should be noted that the voltage fluctuation sequence and the oil pressure change sequence need to be aligned on the time axis before calculating the coupling coefficient. The sampling frequency of both voltage and oil pressure is set to 100 Hz by the sensor group, and the sampling times of the two are aligned by synchronizing the clock through the controller area network bus. The two sequences after alignment are respectively subjected to zero-mean normalization, that is, the value of each sampling point is subtracted from the mean of the sequence and then divided by the standard deviation, so as to eliminate the influence of the difference in dimensions on the calculation of the correlation coefficient. The coupling coefficient of the normalized sequence is then calculated according to the Pearson correlation coefficient formula mentioned above.

[0026] Furthermore, the preset coupling threshold is a discrimination boundary determined based on historical fault data statistics, typically set between 0.75 and 0.85. When the calculated coupling coefficient is greater than this threshold, it indicates a strong coupling relationship between the electrical and hydraulic systems under the current operating conditions, and the power coupling mechanism is at risk of abnormal wear. In one setting method, the system extracted 120 power switching events, of which 30 were subsequently detected as clutch overheating or slippage faults, and 90 were normal switching events. The calculated average coupling coefficient of the fault group was 0.89 with a standard deviation of 0.04, while the average coupling coefficient of the normal group was 0.62 with a standard deviation of 0.08. Through ROC curve analysis, when the threshold is set to 0.85, the true positive rate reaches 93%, and the false positive rate is only 5%. At this point, the Youden index is 0.88, which is greater than the Youden index under other values. Therefore, the preset coupling threshold is determined to be 0.85.

[0027] Specifically, from the historical power switching events with labeled fault categories, at least 50 fault events and 50 non-fault events are randomly selected as a sample set. After calculating the coupling coefficient for each sample, each candidate threshold is traversed in a range from 0.5 to 0.95 with a step size of 0.01. The true positive rate and false positive rate under each candidate threshold are calculated, and the Youden index, i.e., the true positive rate minus the false positive rate, is calculated. The candidate threshold that maximizes the Youden index is selected as the preset coupling threshold. If the threshold corresponding to the maximum Youden index is not unique, the minimum value is taken. In actual operation, this threshold can be updated every 30 days based on newly added labeled samples.

[0028] Multi-factor interaction features are derived from multi-dimensional data containing mechanical wear characteristics, thermal state characteristics, and electrical dynamic response characteristics extracted from the power switching dataset, provided the coupling coefficient exceeds a preset coupling threshold. These multi-dimensional data are then combined with historical data and subjected to min-max normalization to obtain the final multi-factor interaction features. These features are used as input to a support vector machine algorithm to classify wear states, including but not limited to clutch slip rate, transmission oil temperature change rate, and motor torque response time. Specifically, clutch slip rate refers to the relative slippage between the driving and driven discs during power switching engagement; transmission oil temperature change rate refers to the change in transmission lubricating oil temperature per unit time, collected by a temperature sensor installed in the transmission lubrication circuit; and motor torque response time refers to the time interval required from the motor controller issuing a torque command to the motor output shaft torque reaching 90% of the target torque value.

[0029] In step S12, the wear state needs to be classified using the support vector machine algorithm based on the multi-factor interaction features to obtain the classified wear index, including: A support vector machine classifier is constructed based on historical multi-factor interaction features to obtain the optimal classification hyperplane. The optimal classification hyperplane is used to perform classification operations on the multi-factor interaction features to determine the wear category; Based on the wear category, numerical quantification results are extracted from a preset database to obtain the wear index.

[0030] It is worth noting that this step uses a support vector machine algorithm to perform pattern recognition on the multi-factor interaction features extracted in the previous step, dividing the sample points in the high-dimensional feature space into different wear categories. Then, through a mapping mechanism, the discrete category labels are transformed into continuous wear indicators, providing a quantitative physical damage benchmark for subsequent stress trend prediction. The historical multi-factor interaction features are a set of feature vectors accumulated through long-term monitoring and clearly labeled as high-intensity wear categories or regular wear categories. Each feature vector contains dimensions such as clutch slip rate, transmission oil temperature change rate, motor torque response time, and coupling coefficient defined in the previous step.

[0031] In the construction of the support vector machine (SVM) classifier, the decision boundary is obtained by training the SVM on the historical multi-factor interaction features labeled with wear state. The system randomly divides the historical multi-factor interaction features into training and validation sets at an 8:2 ratio. During the training process, the radial basis function is used to map the low-dimensional features in the training set, including clutch slip rate, gearbox oil temperature change rate, and motor torque response time, to a high-dimensional feature space. This makes the different wear state samples, which were originally mixed in the low-dimensional space, form a linearly separable distribution in the high-dimensional space. The normal vector and bias parameters of the hyperplane are determined by maximizing the geometric margin from the class samples to the segmentation boundary, thereby constructing a decision function that can distinguish between high-intensity wear and regular wear. When the accuracy and recall of the decision function on the validation set reach more than 85% and more than 85%, the construction of the SVM classifier is considered complete, and the optimal classification hyperplane is determined.

[0032] Specifically, the wear category is determined by inputting real-time multi-factor interactive features, consisting of clutch slippage rate, transmission oil temperature change rate, and motor torque response time, into the trained support vector machine model. The sign of the function value and the geometric distance of this feature vector relative to the optimal classification hyperplane are calculated. If the function value is positive and exceeds a preset confidence threshold, it is classified as high-intensity wear; if the function value is negative, it is classified as normal wear. The preset confidence threshold is typically set between 0.5 and 1.0 to suppress misclassifications caused by vibration and noise during field operations. Specifically, at least 100 samples identified as high-intensity wear and at least 100 samples as regular wear are selected from historical data. Each sample is input into a trained support vector machine model, and its function value to the optimal classification hyperplane is calculated. The function values ​​of all samples are sorted, and the function value that maximizes the harmonic mean of precision and recall is taken as the preset reliability threshold. If the calculated threshold is lower than 0.5, it is forced to be set to 0.5; if it is higher than 1.0, it is forced to be set to 1.0. This threshold is recalculated every six months based on new samples.

[0033] Furthermore, the pre-built database is a mapping table established based on a large number of bench tests and actual working condition tests. It stores specific wear rate models corresponding to different wear categories, including physical quantities such as the single wear depth of friction plates, material fatigue accumulation rate, and wear amount of gear meshing surfaces. Specifically, in the bench test, newly manufactured clutches of the same model are pre-processed according to the standard break-in procedure, and different power switching conditions are simulated, including at least 500 engagement operations with parameter combinations such as engine speed, torque, oil temperature, and switching duration. After each engagement, a high-precision laser displacement sensor is used to measure the wear depth of the friction plate surface, and the wear category (high-intensity or conventional) corresponding to that switch is recorded. The measurement results are grouped according to the wear category, and the average wear depth and standard deviation of each group are calculated. The average value is stored in the database as the numerical quantification result of that category. For working conditions that cannot be covered by bench tests, finite element simulation is used to supplement them. The simulation model can only be used after being verified by at least 20 sets of measured data. Wear indexes refer to the numerical values ​​that use the category labels output by the support vector machine as indexes to retrieve pre-stored physical wear parameters, transforming abstract category information into specific wear values. This quantifies the degree of physical damage caused to transmission components during the current power switching process, such as the single-time wear depth or wear rate of friction plates. For example, when the system determines that the current operating condition belongs to the high-intensity wear category, the corresponding numerical quantization result extracted from the pre-set database is 0.003 mm, indicating that the single-time wear depth of the friction plate caused by this power switching is 0.003 mm; if it is determined to be the normal wear category, the numerical quantization result is extracted as 0.0005 mm. This mapping mechanism allows subsequent random forest trend prediction to perform regression calculations based on continuous physical quantities, rather than just making rough estimates based on discrete categories.

[0034] In step S13, high-risk areas need to be determined based on the wear index using a preset random forest algorithm model, and corresponding fault prediction signals need to be determined. Historical operation log data is extracted based on the fault prediction signals, and an interaction correlation matrix is ​​constructed. Based on the interaction correlation matrix and the fault prediction signals, data fusion is performed to obtain an interaction dataset.

[0035] In one implementation, based on the wear index, a high-risk area is determined using a preset random forest algorithm model, and the corresponding fault prediction signal is determined, including: Extract stress accumulation-related variables from the wear index; Based on the stress accumulation-related variables and combined with the preset random forest algorithm model, trend calculation is performed to obtain the potential stress accumulation trend. If the potential stress accumulation trend is greater than a preset trend threshold, then the area is marked to obtain a high-risk area; Historical monitoring data of the high-risk area is obtained, and the waveform characteristics of the historical monitoring data are extracted to determine the fault prediction signal.

[0036] It is worth noting that the stress accumulation-related variables are parameters characterizing changes in the internal stress state of materials, extracted from the wear index which contains comprehensive wear information through numerical analysis and physical model inversion. These parameters include the number of contact stress cycles, peak stress amplitude, or residual stress level, and are used as input features for the random forest regression model to predict future stress evolution trends. Specifically, the system separates the grinding wear component reflecting surface material loss from the stress accumulation component reflecting subsurface fatigue damage by inversely solving the clutch engagement dynamics equation, extracting stress accumulation-related variables specifically caused by frequent gear shifting impacts. For example, the system first obtains a wear index of 0.003 mm (characterizing the single wear depth under the high-intensity wear category). Through a preset wear mechanism decomposition model, it separates the component belonging to surface adhesive wear (approximately 0.001 mm) from the component belonging to fatigue wear (approximately 0.002 mm). Based on the fatigue wear component, it inversely calculates the stress accumulation-related variables with a current stress cycle count of 500 and a peak stress amplitude of 450 MPa.

[0037] It is worth noting that the pre-set wear mechanism decomposition model is based on the Acard wear theory and the fatigue damage accumulation theory. The total wear of the friction plate under different normal pressure, sliding speed and temperature conditions is measured by bench test. At the same time, scanning electron microscopy is used to analyze the wear surface morphology and distinguish the contribution ratio of adhesive wear and fatigue wear. Then, the empirical formulas for calculating the adhesive wear component and the fatigue wear component are fitted by using the multiple nonlinear regression method with normal pressure, sliding distance, temperature and stress cycle number as independent variables. The empirical formula is the wear mechanism decomposition model, and its prediction error needs to be controlled within ±15% of the measured value.

[0038] It should be noted that the clutch engagement dynamics equation adopts the following simplified single-degree-of-freedom torsional vibration model: the engine output torque minus the clutch transmitted torque equals the equivalent moment of inertia of the transmission system multiplied by the angular acceleration, where the angular acceleration is the derivative of the rotational speed with respect to time. In the reverse solution, the cumulative friction work in this switching process is derived from the measured wear depth of the friction plate in the wear index, combined with the Achad wear model, i.e., the wear volume is proportional to the normal pressure, sliding distance and wear coefficient. Then, the average sliding power in the clutch engagement process is derived from the cumulative friction work according to the relationship between the friction coefficient and the normal pressure, and then the number of contact stress cycles and the peak stress amplitude are calculated. The normal pressure, friction coefficient and wear coefficient required for the above calculations are all obtained from the preset database according to the current clutch model and oil temperature status.

[0039] Furthermore, the potential stress accumulation trend is a predicted value of the future stress state obtained by extrapolating the stress accumulation-related variables into a preset random forest algorithm model over time. The preset trend threshold is determined based on the product of the material's fatigue limit stress and the safety factor, typically 50%-70% of the material's yield strength. A high-risk area refers to a specific component location or operating state range marked as a concentration of potential faults in the physical or feature space when the potential stress accumulation trend exceeds the preset trend threshold. For example, by inputting the stress accumulation-related variables into a pre-trained random forest algorithm model and predicting the stress accumulation trend for the next 5 working cycles based on stress evolution data from the past 10 working cycles, a potential stress accumulation trend value of 0.82 is obtained; the preset trend threshold is 0.75. Since 0.82 is greater than 0.75, the system marks the sun gear portion of the planetary gear set as a high-risk area.

[0040] The random forest algorithm model training process includes data construction, decision tree base learner generation, and model ensemble optimization. Data construction refers to extracting sample pairs containing stress accumulation-related variables and corresponding future actual stress values ​​from the long-term operation records of hybrid tractors, forming the input-output mapping set required for supervised learning. Specifically, the system selects operation records of components that have failed or completed their full life cycle from the historical database, extracts the stress cycle number and peak stress amplitude at each time point as input features, and records the actual stress level verified by actual measurements after the next 5 to 10 working cycles as the target output, constructing a training dataset containing 500 to 1000 sets of samples, and dividing it into a training subset and a validation subset in an 8:2 ratio. The samples are from the historical records of hybrid tractors. In the historical operation database, complete life cycle records that have completed at least 2000 hours of operation or have experienced transmission component failures are selected. For each record, every 10 power switches are used as a sample window, and the stress accumulation-related variables of each switch within the window are extracted as input features. The maximum stress value measured in the next power switch cycle after the end of the window is used as the target output. The target output value is obtained by post-disassembly inspection or by fitting stress-life curves based on bench fatigue test data of the same type of component. To ensure sample diversity, the selected records must cover at least three terrain slope ranges, two soil moisture conditions, and an ambient temperature distribution from -10 degrees Celsius to 40 degrees Celsius. The labeled samples are divided in chronological order, with the first 80% used for training and the last 20% used for validation.

[0041] Furthermore, the generation of decision tree-based learners is a process of constructing multiple decision trees based on a training subset using a bootstrap sampling method. Specifically, a subsample with replacement, equal to the original sample size, is randomly drawn from the training subset. A single decision tree is constructed based on this subsample. When splitting at each node of the tree, a feature subset is randomly selected and the optimal split point is chosen. This process is repeated until 100 to 200 decision trees are generated. The depth of each tree is limited to 10 to 20 layers to prevent overfitting. Then, the parameter combination that minimizes the prediction error is determined by grid search or random search methods, and the model is ensembled and optimized to finally obtain a pre-trained random forest algorithm model.

[0042] Historical monitoring data consists of a continuous signal stream collected by high-frequency vibration or acoustic sensors installed on components corresponding to high-risk areas (such as the sun gear of a planetary gear set or the drive disc of a clutch). Waveform characteristics include the kurtosis coefficient, skewness index, impulse factor, or spectral energy concentration of the vibration signal. The fault prediction signal is an early fault identifier formed by pattern recognition of the waveform characteristics extracted from the historical monitoring data stream corresponding to the high-risk area. For example, for a high-risk area, vibration acceleration signals from the past 24 hours need to be retrieved. The kurtosis coefficient of the signal segment increases from 3.0 to 4.5, and the impulse factor increases from 1.2 to 1.8. The combination of these two waveform characteristics is determined as the fault prediction signal, indicating that the location has entered the initial stage of pitting damage.

[0043] In one implementation, historical operation log data is extracted based on the fault prediction signal, and an interaction correlation matrix is ​​constructed. Based on the interaction correlation matrix and the fault prediction signal, data fusion is performed to obtain an interaction dataset, including: Extract the time dimension marker of the fault prediction signal, and extract the historical operation log data based on the time dimension marker; By combining the historical operation log data and the fault prediction signal, a matrix is ​​constructed to obtain a synchronization time series matrix; Calculate the mutual information value of the synchronized time series matrix, and construct an interactive correlation matrix based on the mutual information value; Extract the working condition coupling factor from the interaction correlation matrix, and combine the vectors according to the working condition coupling factor to obtain the interaction feature vector; The interactive feature vector and the fault prediction signal are fused to obtain the interactive dataset.

[0044] It is worth noting that the time dimension marker refers to the timestamp information representing the moment the fault prediction signal occurs, including the date and time. Historical operation log data is a set of macroscopic operating parameters recorded by the vehicle control unit at a frequency of seconds or minutes, including motor output torque, engine speed, transmission oil temperature, hydraulic solenoid valve duty cycle, and operating gear information. The synchronous time series matrix is ​​a two-dimensional data structure constructed by uniformly processing the historical operation log data and the high-frequency monitoring data corresponding to the fault prediction signal in terms of time resolution.

[0045] For example, the fault prediction signal occurs at 14:32:15 on a certain day, and this time dimension marker is extracted; historical operation log data for 5 minutes before and after this moment is extracted from the vehicle control unit memory, including the motor output torque fluctuating between 200Nm and 350Nm, the engine speed maintained at 1800rpm, and the transmission oil temperature rising from 95℃ to 102℃ during this time period; at the same time, the corresponding high-frequency vibration data during this time period is extracted, with a sampling frequency of 10kHz and a total of 6000 sampling points; the second-level log data is expanded to millisecond-level resolution through linear interpolation, or the millisecond-level vibration data is compressed into second-level features, such as the kurtosis mean per second, through downsampling, and a synchronous time series matrix containing 6000 rows, corresponding to 6 seconds of valid data and 5 columns (motor torque, engine speed, oil temperature, vibration kurtosis, and vibration effective value) is constructed.

[0046] Furthermore, mutual information is a statistic used to quantify the degree of nonlinear dependence between two random variables. It measures the degree of information sharing between variables by calculating the joint probability distribution and marginal probability distribution among different variable columns in the synchronous time series matrix. The interaction correlation matrix is ​​a square matrix composed of mutual information values ​​as elements, with its rows and columns corresponding to different operating parameter variables. The matrix element values ​​represent the coupling strength between the corresponding two parameters.

[0047] Specifically, the operating condition coupling factor is a feature value or feature vector extracted from the interaction correlation matrix, representing the weight of the influence of a specific combination of operating conditions on the fault, reflecting the external operating condition pressure pattern that leads to the generation of the fault prediction signal. The enhanced interactive dataset is a high-dimensional dataset formed by concatenating the interactive feature vector composed of the operating condition coupling factor with the original fault prediction signal, containing both internal component damage information and external inducing condition information. For example, the feature vector corresponding to the largest eigenvalue is extracted from the interaction correlation matrix as the operating condition coupling factor; this operating condition coupling factor is concatenated with the original vibration waveform features to form an interactive feature vector containing 10 dimensions; this vector is then fused with the kurtosis value and impulse factor in the original fault prediction signal to finally obtain the enhanced interactive dataset.

[0048] It should be clarified that the operating condition coupling factor is obtained by performing principal component analysis on the interaction correlation matrix. By calculating the eigenvalues ​​and eigenvectors of the interaction correlation matrix, all eigenvalues ​​are sorted from largest to smallest, and the eigenvector corresponding to the largest eigenvalue is selected as the operating condition coupling factor. The eigenvector corresponding to the largest eigenvalue represents the direction with the largest variance in the interaction correlation matrix, that is, the comprehensive index that best explains the coupling change mode between various operating parameters. The value range of each element of this eigenvector is between negative one and positive one, and the larger the absolute value, the higher the contribution weight of the corresponding parameter to the current failure mode.

[0049] A simple example of data fusion using feature concatenation is shown below. The working condition coupling factor extracted from the interaction correlation matrix is ​​a 10-dimensional vector [0.85, 0.12, 0.03, 0.72, 0.18, 0.10, 0.91, 0.05, 0.04, 0.88]. Each element represents the contribution weight of motor torque fluctuation, engine speed stability, transmission oil temperature change rate, soil resistance coefficient, terrain slope, operating gear, hydraulic solenoid valve duty cycle, ambient temperature, humidity, and surface flatness to the current fault mode. The original fault prediction signal contains 2-dimensional waveform features, namely kurtosis coefficient 4.5 and impulse factor 1.8. These vectors are directly concatenated to obtain a 12-dimensional enhanced interaction dataset [0.85, 0.12, 0.03, 0.72, 0.18, 0.10, 0.91, 0.05, 0.04, 0.88, 4.5, 1.8]. The interactive datasets obtained in each work cycle will be stored in the historical operation log data so that they can be easily accessed when needed.

[0050] In step S14, the interactive dataset needs to be analyzed in a time series to obtain historical load fluctuation characteristics. Based on the historical load fluctuation characteristics, the model is trained to obtain a wear identification model.

[0051] It is worth noting that the historical load fluctuation features are a subset of multi-dimensional features extracted from the enhanced interactive dataset through time-series analysis. These features reflect segments of drastic load changes during tractor operation and are used as training samples for the retraining of the support vector machine model. For example, for the enhanced interactive dataset, direct analysis yields 500 sets of historical load fluctuation features, including torque standard deviation of 35 Nm, peak factor of 2.8, energy entropy of 2.5, transmission oil temperature and clutch surface temperature data for the corresponding time period, and operating condition coupling factor of 0.82, among other dimensions.

[0052] It should be noted that the time-series analysis extracts the time-series signals of torque, speed, and current from each set of data in the interactive dataset in chronological order. For each channel signal, a sliding window method is used to segment the signal, with a window width of 0.2 seconds and a sliding step size of 0.1 seconds. Within each window, the standard deviation, peak factor, and energy entropy are calculated. The standard deviation represents the fluctuation amplitude of the signal, the peak factor is the ratio of the maximum signal value to the root mean square value, and the energy entropy is the Shannon entropy of the signal power spectrum. The three channels and three statistics are combined to obtain a nine-dimensional historical load fluctuation feature vector. The dimension of this feature vector is consistent with the dimension of the real-time load fluctuation feature in the subsequent step S8, both being nine-dimensional, to ensure the compatibility between real-time input and model training input.

[0053] In one implementation, a wear identification model is obtained by training a model based on the historical load fluctuation characteristics, including: Calculate the thermal stress index of the historical load fluctuation characteristics; The historical load fluctuation characteristics are nonlinearly mapped to obtain a high-dimensional space mapping vector; The model is constructed based on the high-dimensional spatial mapping vector, and iterative optimization is performed using the thermal stress index as a sample weight correction factor to obtain the wear identification model.

[0054] It is worth noting that the thermal stress index is a physical quantity characterizing the cumulative effect of historical thermal load on components, calculated by accumulating temperature monitoring data corresponding to historical load fluctuation characteristics over a given time period. It is used as a sample weight correction factor during the support vector machine model training process, introducing thermodynamic damage mechanisms to improve the model's physical interpretability and prediction accuracy. In one example, the thermal stress index is calculated based on historical monitoring data of transmission oil temperature and clutch surface temperature using the following formula:

[0055] in, The thermal stress index, For the first The absolute temperature at each sampling time, in Kelvin. The thermal damage threshold temperature for the material is 523 Kelvin (250 degrees Celsius) for clutch friction plate materials and 393 Kelvin (120 degrees Celsius) for transmission lubricating oil. The sampling time interval is fixed at 1 second. The total number of sampling points; during summation, the excess amount multiplied by the time interval is only accumulated when the temperature exceeds the threshold; when it is below or equal to the threshold, this term is zero; the unit of thermal stress index is Kelvin seconds, which characterizes the degree of cumulative thermal damage suffered by the component per unit time. Furthermore, in a simplified embodiment, the thermal stress index can also be directly calculated by multiplying the cumulative time exceeding the threshold temperature by the over-temperature amplitude.

[0056] Furthermore, the high-dimensional space mapping vector is a vector representation located in the high-dimensional feature space obtained by nonlinearly mapping the historical load fluctuation characteristics. It is used to make the linearly inseparable wear state samples form a separable distribution in the high-dimensional space. The nonlinear mapping is implemented using a radial basis kernel function. The dimension of the mapped vector depends on the kernel function parameters and the original feature dimension. The specific calculation method is consistent with that described in step S12.

[0057] Specifically, the wear identification model is a classification model obtained by constructing a support vector machine model based on a high-dimensional spatial mapping vector and iteratively optimizing it using the thermal stress index as a sample weight correction factor. This model is used to identify the wear state of real-time load data. The model retraining refers to the process of incremental optimization using an enhanced interactive dataset based on the support vector machine model constructed in step S12: First, historical load fluctuation features are mapped to a high-dimensional spatial mapping vector through a radial basis kernel function; then, the thermal stress index for each sample in the corresponding time period is calculated, and the thermal stress index is normalized to the [0,1] interval and used as the sample weight correction coefficient. Weights are increased for wear samples experiencing high thermal stress and decreased for low-temperature, low-load samples. During normalization, the maximum value of the thermal stress index in the current training batch is used as the upper bound, and the minimum value as the lower bound. A minimum-maximum normalization method is used to map the thermal stress index to the [0,1] interval. If the thermal stress index of all samples in the batch is zero, all weight coefficients are set to 1. Based on the corrected sample weights, the quadratic programming problem of the optimal classification hyperplane is resolved. The objective function, after introducing sample weights, takes the following form:

[0058] in, For the first The weighting coefficients of the normalized thermal stress index for each sample; It is the normal vector of the optimal classification hyperplane; For the bias term, and Together, we determine the position of the hyperplane, that is, the offset of the hyperplane from the origin; For the first The slack variable for each sample is used in soft-margin support vector machines. It allows some sample points to fall within the margin band or even be misclassified. Its value is greater than or equal to 0. The larger the value, the greater the degree to which the sample violates the margin constraint. The total number of training samples is used; the model performance is evaluated through the validation set and the kernel function parameters and penalty parameters are iteratively optimized to finally obtain the optimal classification hyperplane, i.e., the wear recognition model.

[0059] In step S15, it is necessary to extract the real-time load data from the power switching dataset and input the real-time load data into the wear identification model for data processing, and output the real-time fault prediction probability.

[0060] It is worth noting that the real-time load data is raw time-series data collected in real time by the sensor array during the current power switching process. This data includes parameters such as drive shaft torque, speed, battery current, and voltage. It is used as online input for the wear identification model to calculate the current fault risk. The real-time fault prediction probability is a continuous value between 0 and 1 calculated by inputting the real-time load fluctuation characteristics into the wear identification model obtained in step S14. This value represents the likelihood that the current power switching process will cause high-intensity wear on the components.

[0061] In one implementation, the real-time load data is input into the wear identification model for data processing, and the real-time fault prediction probability is output, including: The real-time load data is decomposed in the time domain to obtain the real-time load fluctuation characteristics; The real-time load fluctuation characteristics are input into the wear identification model, and the hyperplane distance is output. The hyperplane distance is mapped to data to obtain the real-time fault prediction probability. It is worth noting that the real-time load fluctuation characteristics are feature vectors formed by extracting features from the components obtained by time-domain signal decomposition. Their dimensions are consistent with the historical load fluctuation characteristics in step S14, including the statistical characteristics (standard deviation, peak factor, energy entropy) and frequency domain energy distribution characteristics of each component, to ensure compatibility with the wear identification model input. Time-domain signal decomposition refers to a signal processing method that separates the various frequency components mixed in the real-time load data and extracts the intrinsic components reflecting the impact characteristics of power switching. It is used to remove background noise, separate the power switching impact components from the stable operation components, and extract feature vectors that match the dimensions of historical load fluctuation characteristics, ensuring consistency between real-time input and model training samples. In one embodiment, an empirical mode decomposition method is used for time-domain signal decomposition. Through a screening process, the signal is adaptively decomposed into a finite number of intrinsic mode function components. Each component represents an oscillation mode at different time scales in the signal, suitable for processing nonlinear and non-stationary mechanical vibration signals.

[0062] Furthermore, when the real-time load fluctuation characteristics are input into the wear identification model, the model maps these characteristics to a high-dimensional space using a radial basis function kernel. The function value of this mapping vector to the optimal classification hyperplane is then calculated, which is the hyperplane distance. Platt scaling is then used for data mapping, converting the hyperplane distance into a probability value, ultimately yielding the real-time fault prediction probability.

[0063] In step S16, if the real-time fault prediction probability is greater than a preset probability threshold, the alarm mechanism is activated, and the alarm is activated.

[0064] It is worth noting that the preset probability threshold is a discrimination boundary determined by statistical analysis of the real-time fault prediction probability distribution of historical fault cases. It is used to distinguish between high-risk operating conditions that require immediate intervention and routine operating conditions that can continue to be monitored. The determination method includes statistical analysis based on historical fault data and ROC curve optimization. For example, 200 power switching events that have occurred in the past two years are extracted, and the real-time fault prediction probability distribution at the moment before the fault occurs is calculated. The 25th percentile of this distribution is taken as the preset probability threshold to ensure that 75% of historical fault cases can be correctly triggered by alarms. Specifically, at least 100 power switching records that have been confirmed as fault events are selected from the historical operation database. For each fault record, the real-time fault prediction probability value of the last power switching process before the fault occurs is extracted to form a fault probability distribution. The first quartile, i.e., the 25th percentile, of this distribution is taken as the preset probability threshold. This threshold ensures that 75% of the historical fault events have a real-time fault prediction probability greater than or equal to this threshold. If there are not enough historical fault events, simulation data is used to supplement them. The simulation data is generated by the fault switching process simulated in the bench test, and the number is not less than 50. This threshold is recalculated and updated every 90 days based on newly confirmed fault events.

[0065] Specifically, the alarm mechanism refers to a series of operations performed by the system when the probability of a real-time fault exceeds a preset probability threshold, such as audible and visual warnings, power limiting, or data logging. These actions are used to alert operators or automatically intervene in equipment operation to reduce the risk of failure. The alarm activation state refers to the system's state after the alarm mechanism is triggered, represented by binary logic values ​​or level signals, and is used to drive subsequent environmental data acquisition and fault prediction report generation processes.

[0066] In one example, during the current power switching process, the real-time fault prediction probability output in step S15 is 0.78, which is greater than the preset probability threshold of 0.75. The system determines that the alarm triggering condition is met. Then, the alarm mechanism is executed, including: sending a command to the dashboard in the driver's cab via the controller area network bus to drive the red warning light to flash at a frequency of 2Hz; sending a torque limit request to the engine controller to limit the maximum output torque of the diesel engine to 80% of the rated value; sending a marking command to the data recording module to mark all sensor data before and after the current period as high-risk events and transfer them to non-volatile memory; at the same time, the system's internal status register is set, and the alarm activation state is represented by a logic high level "1". After the status signal is level-converted through an optocoupler, the environmental data acquisition and report generation process in step S17 is triggered.

[0067] In step S17, environmental data needs to be collected based on the alarm activation status to obtain the current environmental variables, and a fault prediction report is generated by combining the real-time load data and the current environmental variables to obtain the final fault prediction report.

[0068] It is worth noting that the current environmental variables are farmland operation condition parameters collected in real time by the vehicle-mounted environmental sensor group after the alarm activation state is triggered. These parameters include soil moisture, ground slope, ambient temperature, and atmospheric pressure, and are used to supplement real-time load data to comprehensively assess external inducing factors of failure risk. Environmental data acquisition refers to the process of activating the sampling circuits of the soil moisture sensor, tilt sensor, and temperature sensor in response to the logic level signal of the alarm activation state, and acquiring the current operating environment parameters at a preset frequency (usually 1Hz to 10Hz). For example, after the alarm activation state is triggered in step S16, the system immediately starts environmental data acquisition: the soil moisture sensor measures the current soil volumetric water content to be 25%, the tilt sensor measures the ground slope to be 5 degrees, and the temperature sensor measures the ambient temperature to be 32℃, which are combined to form a 3-dimensional current environmental variable.

[0069] In one implementation, a fault prediction report is generated by combining the real-time load data and the current environmental variables to obtain a final fault prediction report, including: The real-time load data and the current environmental variables are fused based on matrix construction to obtain a multi-factor interaction matrix; The multi-factor interaction matrix is ​​input into a preset random forest algorithm model for data calculation to obtain the final fault prediction report.

[0070] It is worth noting that the real-time load data is the raw signal data of the power switching process collected and decomposed in step S15, including the time-series changes of torque, speed, current, and voltage. The multi-factor interaction matrix is ​​a two-dimensional data structure, distinct from the aforementioned features, used to characterize the coupling relationship between real-time load and environmental conditions. This two-dimensional data structure is constructed by spatiotemporally aligning and cross-correlating the statistical characteristics of real-time load data with current environmental variables. Its row dimensions correspond to different categories of influencing factors (load-related, environment-related), and its column dimensions correspond to specific physical parameters. The matrix elements characterize the coupling weight or joint occurrence frequency between each parameter, used to comprehensively depict the joint context of the operating condition and environment when a fault occurs.

[0071] The multi-factor interaction matrix employs different data organization methods. The rows of the matrix correspond to different physical parameter types, specifically including six parameters: peak torque, standard deviation of torque, switching duration, soil moisture, ground slope, and ambient temperature. The columns of the matrix correspond to three equally divided intervals within the time window, with the first, middle, and last thirds of the power switching process each forming a column. The matrix elements are the mean values ​​of the corresponding parameters within the time interval, normalized to a minimum and maximum value, with a normalization range of 0 to 1. This matrix is ​​not subjected to eigenvalue decomposition but is directly flattened into a one-dimensional vector before being input into the random forest model.

[0072] For example, key statistical features of the real-time load data after time-domain decomposition in step S15 are extracted, including peak torque of 320 Nm, standard deviation of torque of 45 Nm, and switching duration of 180 ms. A 6×3 multi-factor interaction matrix is ​​constructed, with the row dimensions being load features peak torque, standard deviation of torque, and switching duration, and environmental features soil moisture, ground slope, and ambient temperature. The matrix elements are filled with normalized values, such as the element in the first row and second column representing the normalized mean of peak torque of 0.72 in the second time interval.

[0073] Furthermore, the preset random forest algorithm model is a decision-level classification model specifically trained to generate structured fault prediction reports, independent of the random forest model used for stress trend prediction in step S13. Its training data is based on the mapping relationship between the multi-factor interaction matrix and the corresponding manual diagnostic conclusions in historical fault cases. The output is natural language or structured data containing risk level assessment, fault location, and operational adjustment suggestions. The final fault prediction report is a comprehensive diagnostic document generated by inputting the multi-factor interaction matrix into the random forest model and making decisions through voting among multiple decision trees. It is used to guide operators or automatic control systems in making operational adjustments.

[0074] Specifically, the random forest algorithm model described in this step is independent of the random forest model in step S13. The model in step S13 is used to regress and predict the stress accumulation trend and output continuous values; the model in this step is used for multi-classification tasks and outputs risk level, fault location and recommended measures.

[0075] The training process of the random forest algorithm model includes data construction, decision tree generation, and model ensemble. Data construction involves extracting cases of malfunctions or completed maintenance verifications from the tractor's historical operating records. For each case, a multi-factor interaction matrix is ​​constructed as input features, and corresponding manual diagnostic conclusions, including risk level, fault location, and recommended measures, are recorded as the target output, forming 300 to 500 training samples. Decision tree generation involves randomly sampling subsamples with replacement from the training data using a bootstrap sampling method. A single decision tree is constructed based on these subsamples. When splitting nodes, a subset of features is randomly selected, and the optimal split point is chosen. This process is repeated to generate 200 decision trees. Model ensemble involves determining the final output through a voting mechanism among the multiple decision trees. Cross-validation is used to evaluate the model's accuracy and recall, and the number and depth of trees are adjusted to achieve a risk level classification accuracy of over 90%.

[0076] Specifically, the training process of this model involves extracting at least 300 samples from a historical failure case database. Each sample contains a multi-factor interaction matrix as input features and a corresponding manual diagnosis conclusion as output label. The manual diagnosis conclusion is confirmed based on the actual disassembly and detection results and includes three fields: risk level (low, medium, and high), fault location to the specific component name, and suggested measures as a Chinese description of no more than 50 characters. During training, the number of decision trees in the random forest is set to 200, the depth of each tree is limited to 15 layers, and the feature subset size is the square root of the input feature dimension. After the model is trained, five-fold cross-validation is used to evaluate the classification accuracy, requiring that the recall rate of high-risk level samples is not less than 90%. Otherwise, the number of trees or depth parameters are adjusted and the model is retrained.

[0077] Finally, the final fault prediction report is a comprehensive diagnostic document generated by inputting a multi-factor interaction matrix into the random forest model and making decisions through voting among multiple decision trees. It is used to guide operators or automatic control systems in making operational adjustments. For example, the output of the final fault prediction report includes, but is not limited to, a risk level assessment of high risk, a fault location of the planetary reducer sun gear, and recommended measures such as reducing the operating speed to below 3 km / h, increasing the motor power compensation ratio to 60%, and suggesting stopping the machine for inspection after 30 minutes.

[0078] In summary, this invention discloses a method for predicting operational faults in hybrid tractors, which solves the problem that existing technologies are prone to causing downtime during critical farming seasons, and enables real-time early warning of potential faults and assists in precise operation and maintenance decisions.

[0079] Reference Figure 2 The second embodiment of the present invention provides a fault prediction system for a hybrid tractor, comprising: The data feature extraction module is used to acquire the power switching dataset collected in real time by the sensor group during power switching, and to perform interaction feature extraction based on coupling judgment based on the power switching dataset to obtain multi-factor interaction features. The wear state classification module is used to classify the wear state based on the multi-factor interaction features using a support vector machine algorithm, and obtain the classified wear index. The signal data enhancement module is used to determine high-risk areas and corresponding fault prediction signals based on the wear index using a preset random forest algorithm model, extract historical operation log data based on the fault prediction signals, construct an interaction correlation matrix, and perform data fusion based on the interaction correlation matrix and the fault prediction signals to obtain an interaction dataset. The model optimization and training module is used to perform time-series analysis on the interactive dataset to obtain historical load fluctuation characteristics, and to train the model based on the historical load fluctuation characteristics to obtain a wear identification model. The real-time fault prediction module is used to extract real-time load data from the power switching dataset, input the real-time load data into the wear identification model for data processing, and output the real-time fault prediction probability. The alarm threshold judgment module is used to activate the alarm mechanism and obtain the alarm activation status if the real-time fault prediction probability is greater than the preset probability threshold. The fault report generation module is used to collect environmental data based on the alarm activation status, obtain the current environmental variables, and generate a fault prediction report by combining the real-time load data and the current environmental variables, thus obtaining the final fault prediction report.

[0080] It should be noted that the operation fault prediction system for a hybrid tractor provided in this embodiment of the invention is used to execute all the process steps of the operation fault prediction method for a hybrid tractor in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0081] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0082] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting operational faults in a hybrid tractor, characterized in that, include: The power switching dataset is collected in real time by the sensor group during power switching. Based on the power switching dataset, interaction features based on coupling judgment are extracted to obtain multi-factor interaction features. Based on the aforementioned multi-factor interaction features, a support vector machine algorithm is used to classify wear status, and the classified wear index is obtained. Based on the wear index, high-risk areas are identified using a preset random forest algorithm model, and corresponding fault prediction signals are determined. Historical operation log data is extracted based on the fault prediction signals, and an interaction correlation matrix is ​​constructed. Based on the interaction correlation matrix and the fault prediction signals, data fusion is performed to obtain an interaction dataset. The interactive dataset is analyzed over time to obtain historical load fluctuation characteristics. The model is then trained based on these historical load fluctuation characteristics to obtain a wear identification model. Real-time load data is obtained based on the power switching dataset, and the real-time load data is input into the wear identification model for data processing, and the real-time fault prediction probability is output. If the real-time fault prediction probability is greater than the preset probability threshold, the alarm mechanism is activated, and the alarm is activated. Environmental data is collected based on the alarm activation status to obtain the current environmental variables. The fault prediction report is generated by combining the real-time load data and the current environmental variables to obtain the final fault prediction report.

2. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The power switching dataset acquired in real time by the sensor group during power switching includes: Acquire raw operation time-series data collected in real time by the sensor group, and extract the torque output ratio status of the raw operation time-series data; If the torque output ratio state is reversed, the original operation timing data is extracted to obtain the power switching dataset.

3. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The step of extracting interaction features based on coupling judgment using the power switching dataset to obtain multi-factor interaction features includes: Based on the power switching dataset, a component operating parameter matrix regarding voltage fluctuations and oil pressure changes is constructed; Acquire terrain data and fuse the terrain data with the component operating parameter matrix to obtain an initial dataset; The coupling coefficient is calculated based on the initial dataset to obtain the coupling coefficient. If the coupling coefficient is greater than the preset coupling threshold, then the multi-factor interaction features are extracted from the power switching dataset.

4. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The wear state classification based on the multi-factor interaction features using the support vector machine algorithm yields the classified wear indices, including: A support vector machine classifier is constructed based on historical multi-factor interaction features to obtain the optimal classification hyperplane. The optimal classification hyperplane is used to perform classification operations on the multi-factor interaction features to determine the wear category; Based on the wear category, numerical quantification results are extracted from a preset database to obtain the wear index.

5. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The step of determining high-risk areas and corresponding fault prediction signals based on the wear index using a preset random forest algorithm model includes: Extract stress accumulation-related variables from the wear index; Based on the stress accumulation-related variables and combined with the preset random forest algorithm model, trend calculation is performed to obtain the potential stress accumulation trend. If the potential stress accumulation trend is greater than a preset trend threshold, then the area is marked to obtain a high-risk area; Historical monitoring data of the high-risk area is obtained, and the waveform characteristics of the historical monitoring data are extracted to determine the fault prediction signal.

6. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The step involves extracting historical operation log data based on the fault prediction signal, constructing an interaction correlation matrix, and then fusing the data based on the interaction correlation matrix and the fault prediction signal to obtain an interaction dataset, including: Extract the time dimension marker of the fault prediction signal, and extract the historical operation log data based on the time dimension marker; By combining the historical operation log data and the fault prediction signal, a matrix is ​​constructed to obtain a synchronization time series matrix; Calculate the mutual information value of the synchronized time series matrix, and construct an interactive correlation matrix based on the mutual information value; Extract the working condition coupling factor from the interaction correlation matrix, and combine the vectors according to the working condition coupling factor to obtain the interaction feature vector; The interactive feature vector and the fault prediction signal are fused to obtain the interactive dataset.

7. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The step of training the model based on the historical load fluctuation characteristics to obtain the wear identification model includes: Calculate the thermal stress index of the historical load fluctuation characteristics; The historical load fluctuation characteristics are nonlinearly mapped to obtain a high-dimensional space mapping vector; The model is constructed based on the high-dimensional spatial mapping vector, and iterative optimization is performed using the thermal stress index as a sample weight correction factor to obtain the wear identification model.

8. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The step of inputting the real-time load data into the wear identification model for data processing and outputting the real-time fault prediction probability includes: The real-time load data is decomposed in the time domain to obtain the real-time load fluctuation characteristics; The real-time load fluctuation characteristics are input into the wear identification model, and the hyperplane distance is output. The hyperplane distance is mapped to obtain the real-time fault prediction probability.

9. The method for predicting operational faults in a hybrid tractor according to claim 1, characterized in that, The process involves combining the real-time load data and the current environmental variables to generate a fault prediction report, resulting in a final fault prediction report, including: The real-time load data and the current environmental variables are fused based on matrix construction to obtain a multi-factor interaction matrix; The multi-factor interaction matrix is ​​input into a preset random forest algorithm model for data calculation to obtain the final fault prediction report.

10. A fault prediction system for a hybrid tractor, characterized in that, include: The data feature extraction module is used to acquire the power switching dataset collected in real time by the sensor group during power switching, and to perform interaction feature extraction based on coupling judgment based on the power switching dataset to obtain multi-factor interaction features. The wear state classification module is used to classify the wear state based on the multi-factor interaction features using a support vector machine algorithm, and obtain the classified wear index. The signal data enhancement module is used to determine high-risk areas and corresponding fault prediction signals based on the wear index using a preset random forest algorithm model, extract historical operation log data based on the fault prediction signals, construct an interaction correlation matrix, and perform data fusion based on the interaction correlation matrix and the fault prediction signals to obtain an interaction dataset. The model optimization and training module is used to perform time-series analysis on the interactive dataset to obtain historical load fluctuation characteristics, and to train the model based on the historical load fluctuation characteristics to obtain a wear identification model. The real-time fault prediction module is used to extract real-time load data from the power switching dataset, input the real-time load data into the wear identification model for data processing, and output the real-time fault prediction probability. The alarm threshold judgment module is used to activate the alarm mechanism and obtain the alarm activation status if the real-time fault prediction probability is greater than the preset probability threshold. The fault report generation module is used to collect environmental data based on the alarm activation status, obtain the current environmental variables, and generate a fault prediction report by combining the real-time load data and the current environmental variables, thus obtaining the final fault prediction report.

Images