A tractor test data acquisition and fault diagnosis method and system

Abstract

The present application relates to the technical field of agricultural fault diagnosis, and particularly relates to a tractor test data acquisition and fault diagnosis method and system, comprising deploying a multi-source sensor group at key nodes of a tractor hydraulic system to acquire operating state data; performing time-frequency joint denoising and adaptive missing value filling on the operating state data, then extracting time domain features, frequency domain features and working condition associated features to construct a standardized feature vector; establishing a hybrid training set that fuses digital twin simulation data and field test fault data, and training a hydraulic fault diagnosis model based on an attention mechanism based on the hybrid training set; inputting the standardized feature vector into the hydraulic fault diagnosis model to output a fault level, a fault type and a fault location; triggering a hierarchical early warning according to the fault level, and generating corresponding troubleshooting and maintenance strategies, which can diagnose the hydraulic system using multi-source data to improve diagnosis efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of agricultural fault diagnosis technology, and in particular to a method and system for tractor test data acquisition and fault diagnosis. Background Technology

[0002] As the core power equipment for agricultural production, the hydraulic system of a tractor undertakes key functions such as suspension operation, steering control, and braking, and is a core component ensuring the stability and reliability of tractor operation. Tractors operate in complex and variable environments, often facing harsh conditions such as field bumps, dust interference, and temperature fluctuations. The hydraulic system is prone to various failure modes, including increased internal leakage, valve core sticking, oil contamination and blockage, and seal failure. If these failures are not detected and addressed in a timely manner, they can lead to decreased tractor operating efficiency, increased downtime losses, and in severe cases, even safety accidents, affecting the progress of agricultural production.

[0003] Currently, the acquisition of test data and fault diagnosis of tractor hydraulic systems mainly rely on traditional manual testing and conventional monitoring methods, which have many technical defects and are difficult to meet the needs of modern agriculture for intelligent and precise tractor operation and maintenance.

[0004] In terms of data acquisition, existing technologies mostly use a single type of sensor or deploy sensors at non-critical nodes, collecting only single-dimensional signals such as hydraulic oil pressure. Furthermore, they do not achieve synchronous triggering of data acquisition for the four typical operating conditions of tractor suspension lifting, lowering, neutral, and pressure drop and holding. This results in problems such as missing information and low fit to operating conditions in the collected data, which cannot comprehensively and accurately reflect the actual operating status of the hydraulic system. At the same time, noise and mechanical vibration interference from the complex field environment, as well as missing and abnormal data transmission, further reduce data quality and cause significant interference to subsequent fault diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for tractor test data acquisition and fault diagnosis, which can use multi-source data to diagnose hydraulic systems, thereby improving diagnostic efficiency and accuracy.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for tractor test data acquisition and fault diagnosis, comprising deploying a multi-source sensor group at key nodes of the tractor hydraulic system, and triggering synchronous acquisition of operating status data according to four operating conditions of the tractor suspension: lifting, lowering, neutral, and pressure drop and pressure holding; the operating status data includes hydraulic oil pressure, temperature, flow rate, hydraulic valve core displacement signal, and vibration signal of hydraulic pump housing; The operating status data is subjected to joint time-frequency denoising and adaptive imputation of missing values. Then, time-domain features, frequency-domain features and operating condition-related features are extracted to construct a standardized feature vector. Establish a hybrid training set that integrates digital twin simulation data and field measured fault data, and train an attention-based hydraulic fault diagnosis model based on the hybrid training set; The standardized feature vector is input into the hydraulic fault diagnosis model, which outputs the fault level, fault type and fault location. Based on the fault level, a graded early warning is triggered, and corresponding troubleshooting and maintenance strategies are generated.

[0007] The specific steps for deploying a multi-source sensor group at key nodes of the tractor hydraulic system and triggering synchronous acquisition of operating status data according to four operating conditions of the tractor suspension: lifting, lowering, neutral, and pressure drop and pressure holding include: Pressure sensors, temperature sensors, flow sensors, hydraulic valve core displacement sensors, and hydraulic pump housing vibration sensors are respectively arranged at key nodes of the tractor hydraulic system, such as the hydraulic pump outlet, multi-way valve inlet and outlet, suspension cylinder chamber, hydraulic pump housing, and hydraulic oil tank, to form a multi-source sensor group. Based on four typical working conditions of tractor suspension (lifting, lowering, neutral, and pressure drop and holding), working condition identification conditions are set respectively. The current working condition is identified by real-time detection of suspension control commands and hydraulic system operating parameters. When any typical working condition is identified, data is collected to obtain the operating status data.

[0008] The specific steps of performing time-frequency joint denoising and adaptive imputation of missing values ​​on the operating status data, and then extracting time-domain features, frequency-domain features, and operating condition-related features to construct a standardized feature vector include: The operating status data is subjected to time-frequency joint denoising processing to generate a clean time-series signal; Identify missing data segments in the pure time-series signal, and perform adaptive imputation of missing values ​​according to the type of operation condition in which the missing data segments are located, to generate a complete time-series sequence; Based on the complete time series, time-domain features, frequency-domain features, and operating condition-related features are extracted respectively, and then fused to construct an initial feature set. The initial feature set is standardized and dimensionality reduced to construct a standardized feature vector for fault diagnosis.

[0009] The specific steps for performing time-frequency joint denoising processing on the operating status data to generate a clean time-series signal include: The original data is decomposed into several intrinsic mode components using a variational mode decomposition algorithm; Calculate the permutation entropy value of each intrinsic mode component and its correlation coefficient with the original data; Based on the permutation entropy value and correlation coefficient, the intrinsic mode components are classified into high-frequency noise mode, mixed mode, and trend term mode; The high-frequency noise modes are directly removed; the mixed modes are denoised using adaptive wavelet thresholding; and the trend terms are filtered using moving average filtering. The processed components are reconstructed to obtain the pure timing signal.

[0010] The specific steps for establishing a hybrid training set that integrates digital twin simulation data and field-measured fault data, and training an attention-based hydraulic fault diagnosis model based on the hybrid training set, include: A high-fidelity digital twin of a tractor hydraulic system is constructed. By injecting preset fault mode parameters, a digital twin simulation dataset covering all working conditions and all fault types is generated. Multi-source sensor data were collected under actual field conditions and then cleaned and labeled to form a field-measured fault dataset. The digital twin simulation dataset and the field measured fault dataset are subjected to distribution alignment and domain adaptation processing, and then fused to construct a hybrid training set; Based on the attention mechanism, the fault diagnosis model is trained and fine-tuned in stages using the hybrid training set to optimize the model parameters until convergence, thus obtaining the final fault diagnosis model.

[0011] The specific steps involved in constructing a high-fidelity digital twin of a tractor hydraulic system, and generating a digital twin simulation dataset covering all operating conditions and all fault types by injecting preset fault mode parameters, include: A mechanistic model incorporating the nonlinear characteristics of hydraulic pumps, multi-way valves, cylinders, and pipelines was established, and baseline parameters for normal conditions were set. Define a set of failure modes, which include increased internal leakage coefficient, valve core jamming, oil contamination and blockage, and seal failure. The Monte Carlo simulation method is used to randomly perturb key physical parameters under preset fault modes and run simulations under four typical operating conditions. The output is synchronous time-series data consistent with the actual sensor type, and includes fault labels and fault injection timestamps.

[0012] The specific steps for performing distribution alignment and domain adaptation processing on the digital twin simulation dataset and the field measured fault dataset, and fusing them to construct a hybrid training set, include: Extract the data source domain of the digital twin simulation; Calculate the maximum difference in mean between two domains; A transfer learning strategy is introduced to map the feature distribution of the source domain data through nonlinear transformation, thereby minimizing the difference in marginal probability distribution between the source and target domains. The hybrid training set is constructed by splicing the distribution-corrected simulation data and the measured data according to a preset ratio.

[0013] The specific steps for training the attention-based hydraulic fault diagnosis model using a hybrid training set include: The input layer receives the standardized multi-channel temporal feature vector; The first level is a 1D-CNN feature extraction layer, which contains multiple one-dimensional convolutional kernels and pooling layers, used to extract local spatial features and high-frequency transient impact features from multi-source signals in parallel. The second level is the GRU temporal evolution layer, which is used to receive the feature sequence output by 1D-CNN and use the gating mechanism to capture the long-term and short-term dependencies and dynamic trends of hydraulic faults over time. The third level is the attention mechanism layer, which is used to calculate the attention weight matrix of the hidden state sequence output by the GRU, adaptively focusing on the key time steps with significant fault characteristics and suppressing the interference of background noise under normal operating conditions. The weighted feature vector is mapped to the fault category probability space, and the fault diagnosis result is output, which includes fault type, fault probability and confidence level.

[0014] The specific steps for triggering graded early warnings based on fault levels and generating corresponding troubleshooting and maintenance strategies include: The fault diagnosis model outputs fault type, fault probability, and confidence level in real time. Combined with the current operating conditions and historical degradation trends, a comprehensive fault risk index is calculated. Based on the comprehensive fault risk index, the fault level is divided into four levels: minor warning level, moderate alarm level, severe shutdown level, and emergency danger level, and corresponding graded warning signals are triggered. Based on the hierarchical early warning signals, matching investigation paths and maintenance strategies are retrieved from a pre-constructed knowledge graph; The troubleshooting path and maintenance strategy will generate a digital maintenance work order containing visual guidance, operation steps and security constraints, and push it to the remote operation and maintenance platform.

[0015] Secondly, the present invention also provides a tractor test data acquisition and fault diagnosis system, which is applied to the aforementioned tractor test data acquisition and fault diagnosis method.

[0016] The present invention discloses a method and system for tractor test data acquisition and fault diagnosis. By deploying a multi-source sensor group at key nodes of the tractor hydraulic system and combining synchronous triggering acquisition of four typical working conditions of tractor suspension lifting, lowering, neutral, and pressure drop and pressure holding, the method achieves comprehensive and accurate acquisition of hydraulic system operating status data, solves the problem of information loss caused by single signal acquisition, and improves the integrity and accuracy of test data.

[0017] By combining time-frequency denoising and adaptive imputation of missing values, environmental and mechanical interference is effectively filtered out, data interruptions and abnormal defects are compensated for, and the quality of operational status data is significantly improved, providing a reliable data foundation for subsequent feature extraction and fault diagnosis.

[0018] By combining time domain, frequency domain, and operating condition-related features to construct standardized feature vectors, the ability to characterize fault features is enhanced, enabling accurate capture of early weak faults and complex fault features, thereby improving the sensitivity and accuracy of fault identification.

[0019] By integrating digital twin simulation data with field-measured fault data to construct a hybrid training set, and combining it with an attention mechanism to train a hydraulic fault diagnosis model, the problems of difficulty and high cost in obtaining fault samples are solved, while ensuring the model's fit with actual working conditions are ensured. This significantly improves the model's generalization ability and enables precise location and quantitative diagnosis of fault level, fault type, and fault location, reducing the difficulty of manual analysis.

[0020] By triggering graded early warnings based on fault levels and generating corresponding troubleshooting and maintenance strategies, the hydraulic system can be transformed from "post-event maintenance" to "pre-event early warning and on-demand maintenance," effectively extending the service life of the hydraulic system, reducing downtime losses and maintenance costs, and improving the safety and uptime of tractor operations. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0022] Figure 1 This is a flowchart of a tractor test data acquisition and fault diagnosis method according to the present invention.

[0023] Figure 2 This invention is a flowchart of deploying a multi-source sensor group at key nodes of the tractor hydraulic system and triggering synchronous acquisition of operating status data according to four working conditions: tractor suspension lifting, lowering, neutral, and pressure drop and pressure holding.

[0024] Figure 3 This invention describes a flowchart of the process of performing time-frequency joint denoising and adaptive imputation of missing values ​​on the operating status data, and then extracting time-domain features, frequency-domain features and operating condition-related features to construct a standardized feature vector.

[0025] Figure 4 This is a flowchart of the present invention for performing time-frequency joint denoising processing on the operating status data to generate a clean time-series signal.

[0026] Figure 5This is a flowchart of the present invention for establishing a hybrid training set that integrates digital twin simulation data and field measured fault data, and for training an attention-based hydraulic fault diagnosis model based on the hybrid training set.

[0027] Figure 6 The present invention constructs a high-fidelity digital twin of a tractor hydraulic system, and generates a flowchart of a digital twin simulation dataset covering all working conditions and all fault types by injecting preset fault mode parameters.

[0028] Figure 7 This is a flowchart of the present invention for performing distribution alignment and domain adaptation processing on the digital twin simulation dataset and the field measured fault dataset, and fusing them to construct a hybrid training set.

[0029] Figure 8 This is a flowchart of the hydraulic fault diagnosis model based on an attention mechanism trained using a hybrid training set, as described in this invention.

[0030] Figure 9 This invention is a flowchart that triggers graded early warnings based on fault levels and generates corresponding troubleshooting and maintenance strategies. Detailed Implementation

[0031] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0032] First embodiment: Please see Figures 1-9 This invention provides a method for tractor test data acquisition and fault diagnosis, including... S101 deploys a multi-source sensor group at key nodes of the tractor's hydraulic system, triggering synchronous acquisition of operating status data according to four working conditions: tractor suspension lifting, lowering, neutral, and pressure drop and holding. The operating status data includes hydraulic oil pressure, temperature, flow rate, hydraulic valve core displacement signal, and hydraulic pump housing vibration signal. The specific steps for deploying a multi-source sensor group at key nodes of the tractor hydraulic system and triggering synchronous acquisition of operational status data according to four operating conditions—tractor suspension lifting, lowering, neutral, and pressure drop and pressure holding—include the following: S201 places pressure sensors, temperature sensors, flow sensors, hydraulic valve core displacement sensors, and hydraulic pump housing vibration sensors at key nodes of the tractor hydraulic system, including the hydraulic pump outlet, multi-way valve inlet and outlet, suspension cylinder chamber, hydraulic pump housing, and hydraulic oil tank, to form a multi-source sensor group. Pressure and flow sensors are precisely positioned at the outlet of the hydraulic pump to capture the pump's output pulsation and volumetric efficiency changes in real time. At the same time, a high-frequency vibration acceleration sensor is installed on the surface of the hydraulic pump housing to monitor abnormal vibration signals caused by bearing wear, cavitation, and internal mechanical imbalance.

[0033] Pressure sensors are deployed at the inlet and outlet of the multi-way valve to calculate the pressure drop at the valve port; hydraulic valve spool displacement sensors (such as LVDT or magnetostrictive sensors) are integrated into the actuator inside or outside the multi-way valve to directly provide feedback on the valve spool opening degree, response hysteresis and jamming.

[0034] Pressure sensors are installed in the rodless chamber and rod chamber (i.e., upper and lower chambers) of the suspension cylinder to accurately measure the load pressure and back pressure; and temperature sensors are arranged near the cylinder or in the return oil line to monitor the oil temperature rise caused by throttling heating or friction.

[0035] Temperature sensors and level / pressure sensors are installed at the bottom of the hydraulic oil tank or in the main return oil circuit as a reference for the system's thermal balance and oil suction status.

[0036] The aforementioned dispersed sensors converge via shielded twisted-pair cables or industrial buses (such as CAN Bus) to form a multi-source sensor group consisting of five types of signals: pressure, temperature, flow, displacement, and vibration. All sensors are calibrated to ensure that their measurement range covers extreme operating conditions of the tractor and that they are resistant to electromagnetic interference and oil contamination.

[0037] S202 sets working condition identification conditions according to four typical working conditions of tractor suspension: lifting, lowering, neutral, and pressure drop and pressure holding. It identifies the current working condition by real-time detection of suspension control commands and hydraulic system operating parameters. When any typical working condition is identified, data is collected to obtain the operating status data.

[0038] For four typical operating conditions of tractor suspension systems, predefine their logical recognition conditions: The identification conditions for the suspension lifting condition are: the suspension control handle / electrical signal issues an "up" command, and the pressure in the rodless chamber of the hydraulic cylinder rapidly rises above the load threshold, with the valve core displacement signal showing an opening direction consistent with the lifting condition; the identification conditions for the suspension lowering condition are: an "down" command is issued, the flow rate in the rod chamber or return oil circuit increases significantly, and the system pressure exhibits specific unloading or controlled descent characteristics; the identification conditions for the neutral condition are: no operation command input, the valve core displacement returns to zero (neutral position), the pressure in each chamber remains relatively stable, and the flow rate approaches zero (only internal leakage exists); the identification conditions for the pressure drop and pressure holding condition are: the system is in a heavy load holding state, the control command remains neutral, but the pressure in the hydraulic cylinder chamber remains at a high pressure level, and a slight pressure decay occurs over time (used to evaluate sealing performance).

[0039] The controller polls the suspension control command signal and basic operating parameters of the hydraulic system at high frequency. Real-time data is input into a preset state machine model and compared with the feature vectors of the four operating conditions. Once a parameter combination is detected that meets the entry conditions for any operating condition (e.g., a transition from "neutral" to "lifting"), the current operating condition type is immediately determined. At the instant the operating condition is confirmed, the system sends a global synchronization trigger signal. This signal simultaneously activates the A / D converters of all channels, ensuring that pressure, temperature, flow, displacement, and vibration data are strictly aligned on the time axis at the same moment. The acquired data packets are automatically tagged with an "operating condition label" (e.g., Mode: Lifting) and a timestamp, stored in a circular buffer, or transmitted to the host computer, ensuring that subsequent analysis can clearly distinguish the system's dynamic response characteristics under different operating modes.

[0040] S102 performs time-frequency joint denoising and adaptive missing value filling on the operating status data, and then extracts time-domain features, frequency-domain features and operating condition-related features to construct a standardized feature vector; The specific steps include: S301 performs time-frequency joint denoising processing on the operating status data to generate a clean time-series signal; The specific steps include: S401 uses a variational mode decomposition algorithm to decompose the original data into several intrinsic mode components; Variational Mode Decomposition (VMD) is introduced to adaptively decompose the original noisy operating state data x(t) into K intrinsic mode components with specific sparse characteristics. By constructing a variational constraint model, the center frequency and bandwidth of each component can be found in the frequency domain, minimizing the sum of the estimated bandwidths of all components. Compared with Empirical Mode Decomposition (EMD), VMD effectively avoids mode aliasing and ensures that each component represents a physical oscillation mode at different scales in the original signal (such as the pump engagement frequency, valve switching impact, random noise, etc.).

[0041] S402 calculates the permutation entropy value of each intrinsic mode component and its correlation coefficient with the original data; To quantify the noise content and information contribution of each IMF component, a dual evaluation index system is established: calculating the permutation entropy value of each IMF component. The PE value reflects the complexity and randomness of the time series. High-frequency noise usually exhibits high disorder and corresponds to a higher PE value; while deterministic signals containing the main fault characteristics have relatively low PE values.

[0042] Calculate the Pearson correlation coefficient between each IMF component and the original signal x(t). This index measures the contribution of a component to the energy and morphology of the original signal; a high correlation coefficient means that the component contains the main trend or key abrupt change information of the signal.

[0043] S403 classifies the intrinsic mode components into high-frequency noise modes, mixed modes, and trend term modes based on the permutation entropy value and correlation coefficient; Based on the calculated PE values ​​and correlation coefficients, an adaptive threshold is set to accurately classify the K IMF components into three categories: High-frequency noise mode: characterized by "high permutation entropy and low correlation coefficient". This type of component mainly consists of sensor electronic noise and high-frequency electromagnetic interference, and contains almost no effective fault information.

[0044] Hybrid modes: characterized by "moderate permutation entropy and moderate to high correlation coefficients". These components contain both useful fault impact components and some background noise, making them a key target for denoising.

[0045] Trend Term Mode: Characterized by "low permutation entropy and high correlation coefficient". This type of component represents macroscopic trends such as low-frequency pressure fluctuations and temperature drift in hydraulic systems, with extremely low noise content.

[0046] S404 directly removes the high-frequency noise mode; applies adaptive wavelet threshold denoising to the mixed mode; and applies moving average filtering to the trend term mode. For different types of modal components, implement targeted filtering strategies: High-frequency noise mode rejection: This involves directly discarding such components, cutting off the propagation path of high-frequency random noise at its source.

[0047] Hybrid-mode adaptive wavelet denoising: An adaptive wavelet thresholding denoising algorithm is used for the retained hybrid modes. Based on the noise variance estimation of each component, the optimal wavelet basis function (such as the dbN series) and the number of decomposition levels are dynamically selected. Soft thresholding or semi-soft thresholding functions are used to suppress residual noise in the components, while preserving the amplitude and phase characteristics of the fault impact signal to the greatest extent.

[0048] Trend Term Mode Smoothing Filter: The trend term mode is processed by moving average filtering or low-pass filtering to further smooth low-frequency glitches and ensure the stability of the baseline trend.

[0049] S405 reconstructs the processed components to obtain the pure timing signal.

[0050] All effective IMF components (denoised mixed mode + smoothed trend term mode) after the above differential processing are linearly superimposed and reconstructed to synthesize the final clean time series signal. This signal retains the integrity of fault characteristics in the time domain and significantly improves the signal-to-noise ratio (SNR) in the frequency domain.

[0051] S302 identifies missing data segments in the clean time-series signal and performs adaptive filling of missing values ​​according to the type of operation condition in which the missing data segments are located, generating a complete time-series sequence; This step addresses the data gap issue caused by packet loss during wireless transmission, momentary power failure of sensors, or asynchronous sampling under extreme conditions.

[0052] Using a timestamp continuity detection algorithm, a clean time series signal is scanned, and data segments with time intervals exceeding a preset threshold (such as 2 × sampling period) are located and marked as missing intervals.

[0053] In static / steady-state conditions (such as neutral and pressure-holding conditions), the rate of change of the signal before and after the missing data segment is small. Cubic spline interpolation or local linear regression is used to fill the gaps to ensure a smooth transition of the curve.

[0054] In dynamic / transient operating conditions (such as the instantaneous switching between lifting and lowering), the signal exhibits strong nonlinearity and rapid abrupt changes, and simple interpolation can smooth out critical fault impacts. Therefore, a deep learning prediction model based on a Long Short-Term Memory (LSTM) network is employed. Utilizing historical data from windows before and after the missing point, along with the current operating condition label, the model is trained to predict the waveform of the missing segment, thereby faithfully reconstructing the dynamic characteristics of the transient process and generating a complete time-series sequence without breaks and with logical continuity.

[0055] S303 Based on the complete time series, extract time-domain features, frequency-domain features, and operating condition-related features respectively, and fuse them to construct an initial feature set; Based on the complete time series, we comprehensively mine the state information contained in the data from three dimensions: Time-domain feature extraction: Calculate indicators that reflect the amplitude distribution and statistical characteristics of the signal, including the root mean square value (RMS, reflecting energy) and the peak factor (reflecting impact).

[0056] Frequency domain feature extraction: Perform Fast Fourier Transform (FFT) or Power Spectral Density (PSD) analysis on the signal to extract the amplitude of the dominant frequency component, the centroid frequency, and the frequency variance.

[0057] Operating condition related feature extraction: Combine operating condition labels to extract dynamic response features across operating conditions. For example: pressure rise time, overshoot, valve core displacement hysteresis time, and average energy efficiency under specific operating conditions during the switch from "neutral" to "upgrade".

[0058] S304 performs standardization and dimensionality reduction filtering on the initial feature set to construct a standardized feature vector for fault diagnosis.

[0059] To eliminate the influence of dimensions and remove redundant information, an optimal feature vector is constructed: Standardization: Z-Score standardization is used to map all features to the same order of magnitude (e.g., [0,1] or mean 0 and variance 1), eliminating weight bias caused by different physical dimensions such as pressure, temperature, and vibration.

[0060] Calculate the correlation coefficient matrix between features, remove redundant features with high collinearity (correlation coefficient > 0.95), and then use the maximum correlation minimum redundancy algorithm to select the subset of key features with the highest discrimination and robustness against fault categories.

[0061] The selected key features are arranged in a fixed order to construct the final standardized feature vector. This vector has moderate dimensionality and high information density, and can be directly used as the input interface for subsequent machine learning or deep learning fault diagnosis models.

[0062] S103 establishes a hybrid training set that integrates digital twin simulation data and field measured fault data, and trains an attention-based hydraulic fault diagnosis model based on the hybrid training set. The specific steps include: S501 constructs a high-fidelity digital twin of a tractor hydraulic system. By injecting preset fault mode parameters, it simulates and generates a digital twin simulation dataset covering all working conditions and all fault types. The core of this step lies in establishing a virtual mapping space capable of accurately reproducing the nonlinear dynamic characteristics of a tractor's hydraulic system, and generating standardized simulation data covering "all operating conditions, all fault types, and all severity levels" in batches through parametric fault injection technology. Specific steps include: S601 establishes a mechanism model that includes the nonlinear characteristics of hydraulic pumps, multi-way valves, cylinders, and pipelines, and sets the baseline parameters for normal conditions. Based on the principles of fluid mechanics, thermodynamics, and mechanical dynamics, high-precision mathematical models of the core components are established. Considering the nonlinear characteristics of volumetric efficiency as a function of pressure / temperature, leakage coefficient variables and cavitation models are introduced to simulate flow pulsation and pressure fluctuations.

[0063] Multi-way valve model: A nonlinear model based on the valve orifice flow equation is established to accurately describe the functional relationship between valve core displacement and flow area, including dead zone, saturation and hydrodynamic feedback effects.

[0064] Suspension cylinder model: Considering friction (Stribeck effect), elastic load and chamber compressibility, the hysteresis characteristics of piston movement are simulated.

[0065] Pipeline network model: Using distributed parameter or lumped parameter methods, the pressure wave propagation delay, friction loss and fluid capacitive reactance effect of long pipelines are simulated.

[0066] The above sub-models are integrated into a simulation platform (such as AMESim, Simulink, or a custom solver). The standard physical parameters of the tractor under rated operating conditions (such as rated speed, standard oil viscosity, nominal leakage, etc.) are input, the simulation is run, and the simulation is compared and iterated with the measured data under normal conditions until the error between the simulation output and the measured reference signal is controlled within a preset threshold (such as <3%), and the reference parameters under normal conditions are established.

[0067] S602 defines a set of failure modes, which include increased internal leakage coefficient, valve core sticking, oil contamination and blockage, and seal failure. Construct a structured set of failure modes, mapping physical failure phenomena to adjustable parameters in the model: Increased internal leakage coefficient: This corresponds to wear in the internal clearances of pumps, valves, or cylinders, manifested as a leakage coefficient C. leak The value is gradually increased from the nominal value to the failure threshold.

[0068] Valve core jamming: Simulates the obstruction or mechanical deformation of pollutant particles, manifested as a sudden change in the friction term in the valve core motion equation or a limitation on the maximum displacement.

[0069] Oil contamination and clogging: Simulating filter element clogging or reduced valve orifice throttling area, manifested as a significant increase in local resistance coefficient or effective flow area A. eff attenuation.

[0070] Seal failure: Simulated aging and cracking of the seal ring, manifested as a sharp increase in external leakage flow and a decrease in the system's pressure holding capacity.

[0071] Severity grading: For each fault mode, three levels are set: minor, moderate, and severe, forming a fine-grained fault labeling system.

[0072] The S603 uses the Monte Carlo simulation method to randomly disturb key physical parameters under preset fault modes and runs simulations under four typical operating conditions. It outputs synchronous time-series data consistent with the actual sensor type, along with fault tags and fault injection timestamps.

[0073] Random perturbation strategy: The Monte Carlo simulation method is adopted to conduct large-scale random sampling within a preset range of fault parameters. Not only are the fault parameters perturbed (such as the leakage rate being randomly distributed between 0% and 50%), but environmental uncertainties are also introduced (such as oil viscosity fluctuating with temperature, random changes in load mass, and random spectra of ground excitation) to simulate the complexity of the real world.

[0074] The digital twin is controlled to sequentially execute four typical operating conditions: "suspension lifting, descent, neutral, and pressure drop holding," and faults are injected during the transient processes of condition switching. The simulation solver is configured to output data at the same sampling frequency (e.g., 1kHz) as the physical sensors. The generated dataset includes: Multi-source timing signals: pressure, temperature, flow rate, valve core displacement, vibration acceleration, etc., with formats completely consistent with measured data.

[0075] Metadata tags: Each data record is accompanied by detailed fault tags (fault type, severity), specific timestamp of fault injection, current operating condition type, and environmental parameter vector.

[0076] This allows for the generation of large-scale, well-balanced digital twin simulation datasets, covering all preset fault combinations and extreme boundary conditions.

[0077] The S502 collects multi-source sensor data under actual field operation conditions, and after cleaning and labeling, it forms a field measured fault dataset. During actual tractor operations in the field (including different scenarios such as tilling, transportation, and lifting), the multi-source sensor group deployed by S101 continuously collects operational data.

[0078] We focus on capturing failure cases that occur naturally or are artificially induced (such as pre-set failures in a controlled test field) to ensure that we collect data on the actual failure evolution process.

[0079] The time-frequency joint denoising and missing value imputation algorithm described in S102 is applied to standardize the original measured data, eliminating outliers caused by electromagnetic interference and communication packet loss to ensure data quality. Then, time synchronization calibration is performed to eliminate phase differences between different sensor channels.

[0080] Hydraulic system experts, combining maintenance records, fault descriptions, and oscilloscope waveform analysis, meticulously reviewed each segment of the cleaned data. They then confirmed the start and end times of the faults, their types, and severity, assigning accurate ground-truth labels. Normal operating condition data was labeled "healthy state." The processed measured data was then categorized and organized according to "operating condition - fault type," forming a field-measured fault dataset. Although this dataset has a relatively small sample size, it possesses extremely high authenticity and authority, primarily used for constructing a validation set for model training and for domain adaptation fine-tuning in transfer learning.

[0081] S503 performs distribution alignment and domain adaptation processing on the digital twin simulation dataset and the field measured fault dataset, and merges them to construct a hybrid training set; The specific steps include: S701 extracts the data source domain of the digital twin simulation; Data preparation: Source domain: Extract the digital twin simulation dataset generated by S501. , where x s For multi-source time-series signal segment y sFor accurate fault labeling. The source domain has a large amount of data and balanced categories, but the distribution P... S(x) It deviates from the real environment.

[0082] Target domain: Extract field-measured fault dataset processed by S502. (Note: During the domain adaptation phase, some target domain data may only utilize its unlabeled portion for distribution alignment, or utilize a small amount of labeled data for semi-supervised alignment.) The target domain data volume is small, the classes are imbalanced, but the distribution P... T(x) It represents the true laws of physics.

[0083] Feature embedding: Using a pre-trained shallow feature extraction network (such as 1D-CNN or autoencoder), the original high-dimensional temporal data x is mapped to a low-dimensional latent feature space H to obtain the source domain feature representation. and target domain feature representation This step aims to remove high-frequency random noise from the original signal while preserving the core manifold structure that reflects the dynamic characteristics of the system.

[0084] S702 calculates the maximum mean difference between two domains; Measurement principle: In order to quantify the source domain distribution P S Distribution P of the target domain T To measure the difference between the two means, the maximum mean difference is introduced as a distance metric. The Euclidean distance between the two mean vectors distributed in this space is calculated by mapping the data to the reproducing kernel Hilbert space (RKHS).

[0085] The larger the MMD value, the more significant the difference in statistical characteristics between the simulation data and the measured data (e.g., the pressure fluctuation frequency of the simulation data is too high, or the amplitude distribution of the vibration signal is too narrow); when the MMD value is close to 0, it means that the two have achieved distribution alignment in the feature space.

[0086] S703 introduces a transfer learning strategy, which maps the feature distribution of the source domain data through nonlinear transformation to minimize the difference in marginal probability distribution between the source and target domains; By introducing a deep domain adaptation mechanism, a learnable feature transformer is constructed. The goal is to find the optimal parameters. This minimizes the MMD distance between the transformed source domain features and the target domain features.

[0087] Constructing a joint loss function Among them, L task The loss for the classification task of the source domain data (such as cross-entropy loss) ensures that the model retains its fault detection capability.

[0088] MMD(G(ZS ),Z T The domain alignment loss forces the feature distribution of the source domain to converge toward the target domain.

[0089] As a balancing factor, it dynamically adjusts the weights of task performance and domain alignment.

[0090] Parameters are updated iteratively using the backpropagation algorithm. During this process, the model automatically learns a nonlinear linear transformation strategy to correct systematic biases in the simulation data caused by model simplification, making the corrected simulation data statistically approach the measured data infinitely.

[0091] S704 constructs the hybrid training set by splicing the distribution-corrected simulation data and the measured data according to a preset ratio.

[0092] The specific steps for training the attention-based hydraulic fault diagnosis model based on a hybrid training set include: The S801 input layer receives the standardized multi-channel timing feature vector; Data input: The model's input layer receives standardized multi-channel time-series data from the S704 mixed training set. Each sample is represented as a matrix X∈R. T ×C, where T is the time step (sequence length) and C is the number of sensor channels (including pressure, flow, temperature, displacement, vibration, etc.).

[0093] Optionally, the original signal of each channel is linearly projected through a fully connected layer and mapped to a unified high-dimensional embedding space to eliminate subtle differences remaining between different physical dimensions and to initially fuse the correlation information between channels.

[0094] The first stage of S802 is a 1D-CNN feature extraction layer, which contains multiple one-dimensional convolutional kernels and pooling layers, used to extract local spatial features and high-frequency transient impact features from multi-source signals in parallel. Architecture design: Construct a one-dimensional convolutional neural network (1D-CNN) module containing 2-3 layers.

[0095] Multi-scale convolution kernels: Convolution kernels of different sizes are used in parallel (e.g., k=3, 5, 7) to capture local features at different time scales. Small convolution kernels are good at capturing high-frequency abrupt changes (such as pressure spikes during valve switching or high-frequency vibrations caused by cavitation collapse), while large convolution kernels can cover a wider range of waveform patterns.

[0096] Each convolutional layer is followed by a ReLU activation function to introduce non-linearity, and then downsampling is performed through a max pooling layer to reduce the data dimensionality while retaining the most significant local response features, thereby enhancing the model's invariance to small time shifts.

[0097] This layer primarily acts as a "feature detector," extracting local spatial features (coupling relationships between multiple sensors) and high-frequency transient impact features in parallel from the original multi-source signals, outputting a feature sequence H rich in detail. cnn ∈R T′×Dcnn .

[0098] The second stage of S803 is the GRU temporal evolution layer, which is used to receive the feature sequence output by 1D-CNN and use the gating mechanism to capture the long-term and short-term dependencies and dynamic trends of hydraulic faults over time. The feature sequence H_{cnn}Hcnn output by the CNN is input into a bidirectional gated recurrent unit (Bi-GRU) network.

[0099] The update gate determines how much information from the current input is retained for the next time step, effectively capturing the slow drift of the hydraulic system state (such as viscosity changes caused by a gradual increase in oil temperature).

[0100] The reset gate determines how much historical information to ignore, enabling the model to respond sensitively to sudden changes in operating conditions or failures.

[0101] Hydraulic failures are often a dynamic process that evolves over time (e.g., a leak progresses from minor to severe). The GRU layer, through its internal memory units, establishes long-term and short-term dependencies between consecutive time steps, captures the dynamic evolution trend of failure characteristics (e.g., changes in pressure drop rate, vibration energy accumulation process), and outputs a hidden state sequence containing temporal context information.

[0102] The third level of S804 is the attention mechanism layer, which is used to calculate the attention weight matrix of the hidden state sequence output by GRU, adaptively focusing on the key time steps with significant fault characteristics and suppressing the interference of background noise under normal operating conditions. S805 maps the weighted feature vector to the fault category probability space and outputs the fault diagnosis result, which includes the fault type, fault probability and confidence level.

[0103] Architecture design: Introduce a multi-head self-attention mechanism to process the sequence of GRU outputs.

[0104] Weight calculation logic: Calculate the query matrix (Q), key matrix (K), and value matrix (V).

[0105] pass The formula calculates the correlation score between any two time steps in the sequence.

[0106] The model automatically learns to assign high attention weights to time steps that contain significant fault characteristics (e.g., the moment of sudden pressure drop, the moment of abnormal vibration outbreak), while assigning low weights to background noise or steady segments under normal operating conditions.

[0107] By using weighted summation, a new feature representation is generated. This step effectively suppresses irrelevant background interference, greatly enhances the signal-to-noise ratio of fault features, and enables the model to "focus" on the key segments that determine the fault type.

[0108] The weighted feature vectors are compressed into fixed-length vectors using global average pooling.

[0109] A fully connected layer and a softmax activation function are used to map features to a fault category probability space. The final fault diagnosis result is then output, including: Fault type: The specific fault category predicted (such as "pump internal leakage - severe", "valve core jamming - moderate", etc.).

[0110] Failure probability: The probability distribution vector belonging to each category.

[0111] Confidence score: A diagnostic confidence score calculated based on the difference between the highest probability value and the second highest probability value, used to assess the reliability of the results.

[0112] S504 uses the hybrid training set based on the attention mechanism to perform phased training and fine-tuning of the fault diagnosis model, optimizing the model parameters until convergence, and obtaining the final fault diagnosis model.

[0113] This step aims to utilize the distributed consistency hybrid training set constructed by S503 to efficiently train and optimize the attention-based fault diagnosis model through a two-stage strategy of "pre-training-fine-tuning," ensuring that the model has both a deep understanding of general fault mechanisms and robustness to adapt to complex field environments.

[0114] Load simulation data, comprising approximately 80%-90% of the mixed training set after distribution correction. Leveraging the advantages of accurate data labels, balanced samples, and coverage of the entire fault spectrum, quickly initialize the weight parameters of the CNN-GRU-Attention network. This stage focuses on enabling the model to learn the fundamental feature mapping patterns of the hydraulic system under different fault modes (such as the nonlinear relationship between pressure pulsation and leakage, and the correspondence between vibration spectrum and cavitation intensity), establishing a general fault discrimination boundary. A larger learning rate and batch size can be used to accelerate model convergence to near the global optimum and avoid getting trapped in local minima.

[0115] Then, the model was switched to field-measured fault data, which comprised approximately 10%-20% of the mixed training set. The model was then fine-tuned to address unmodeled dynamics, sensor drift, and complex background noise present in the real environment. The parameters of the attention layer and classification head were updated as a key focus, enabling the model to automatically suppress interference and focus on key fault features in real-world noise environments, thus eliminating residual bias between simulation and field measurements.

[0116] S104 inputs the standardized feature vector into the hydraulic fault diagnosis model and outputs the fault level, fault type and fault location. This step deploys the trained model on an onboard edge computing unit or cloud server to achieve online real-time monitoring and diagnosis of the tractor's hydraulic system. It receives in real-time standardized feature vectors (including time-domain, frequency-domain, and operating condition-related features) generated in the preceding step S104.

[0117] The feature vectors are sequentially passed through the model's 1D-CNN layer (to extract local impulses), GRU layer (to capture temporal evolution), and Attention layer (to focus on key segments). The model's output layer calculates the posterior probability distribution of each fault category using the Softmax function.

[0118] The output includes the following: Fault type: Identify the specific fault mode (such as "internal leakage of main pump", "stuck valve core of multi-way valve", "cavitation of suction line", etc.).

[0119] Fault severity: Based on probability distribution and preset threshold, the severity of the fault is initially determined (early stage, development stage, failure stage).

[0120] Fault location: By combining the sensor layout and the propagation path of the fault characteristics, the physical component where the fault occurred (such as "left rear suspension cylinder" or "steering priority valve") can be accurately located.

[0121] Confidence score: Outputs the model's confidence level (0-1) for the diagnostic result, used to assist in human review and decision-making.

[0122] S105 triggers graded early warnings based on the fault level and generates corresponding troubleshooting and maintenance strategies.

[0123] The specific steps include: The S901 receives the fault type, fault probability and confidence level output by the fault diagnosis model in real time, and calculates the comprehensive fault risk index by combining the current working conditions and historical degradation trends. Data fusion: Real-time reception of fault type, fault probability, and confidence level output from S104. Simultaneously, external context variables are introduced. Current operating conditions: differentiate between high-load (e.g., heavy plowing) and low-load (e.g., empty transport) conditions. The risk of the same fault is significantly amplified under high load conditions.

[0124] Historical degradation trend: retrieve the health change curve of the component over the past NN hours, and calculate the degradation rate (first derivative) and acceleration (second derivative).

[0125] Index Calculation Model: Constructing a weighted risk assessment function to calculate the comprehensive failure risk index (RI): Among them, α, , These are the weighting coefficients.

[0126] Output: Generate quantized R I The value (normalized to the 0-100 range) serves as the sole basis for tiered early warning.

[0127] Based on the comprehensive fault risk index, S902 classifies the fault level into four levels: minor warning level, moderate alarm level, severe shutdown level, and emergency danger level, and triggers the corresponding graded warning signals. Based on the calculated R I The value categorizes the failure risk into four levels and triggers different response mechanisms for each level: Mild warning level (0) <RI≤30 ): Definition: Early minor faults or occasional anomalies that do not affect current operations but require attention.

[0128] Action: A yellow warning icon pops up on the cab display screen, the log is recorded, and the operation is not interrupted.

[0129] Moderate alarm level (30) <RI≤60 ): Definition: The fault is characterized by obvious features, a slight decrease in performance, and long-term operation may lead to component damage.

[0130] Action: Audible and visual alarm (yellow light flashing), suggests limiting engine speed or hydraulic flow, prompts "Arrange maintenance as soon as possible".

[0131] Severe shutdown level (60) <RI≤85 ): Definition: The malfunction has affected the core function, and continuing to operate will result in irreversible damage or safety hazards.

[0132] Action: Red alert, forcibly limits hydraulic system output power (Limp-home mode), announces "Please stop immediately and check", and automatically uploads alarm information to the remote platform.

[0133] Emergency Hazard Level (RI>85): Definition: The pipe is about to fail and there is a risk of bursting, going out of control, or fire.

[0134] Action: Highest level audible and visual alarm, immediately cuts off hydraulic pump power, locks operating handle, automatically dials emergency rescue number, and pushes precise coordinates to the operation and maintenance center.

[0135] S903 retrieves matching investigation paths and maintenance strategies from a pre-built knowledge graph based on hierarchical early warning signals; Using the fault type and warning level determined by S902 as the query entry point, and combined with the fault location information, multi-hop reasoning is performed in the graph. The most matching troubleshooting path (e.g., check oil level first -> measure pressure -> disassemble valve) and maintenance strategy (e.g., replace seal ring, clean filter element, adjust relief valve pressure) are retrieved.

[0136] Based on the current geographical location (field / workshop) of the agricultural machinery and available resources, the complexity of the recommended strategy is dynamically adjusted (for example, only temporary emergency measures are recommended in the field, and a thorough repair is carried out after returning to the workshop).

[0137] S904 will generate digital maintenance work orders containing visual guidance, operation steps and safety constraints based on the troubleshooting path and maintenance strategy, and push them to the remote operation and maintenance platform.

[0138] The retrieved strategies are converted into standardized digital maintenance work orders, which include: Visual guidance: Embed 3D exploded views, AR augmented reality markers, or short video tutorials to intuitively show the location of faulty parts and the order of disassembly and assembly.

[0139] Operating steps: List the standard operating procedures (SOP) step by step, including key parameters such as torque requirements and oil specifications.

[0140] Safety constraints: Highlight the safety precautions that must be followed (such as "Release residual pressure", "Wear safety goggles", "Do not disassemble at high temperatures").

[0141] Spare parts list: Automatically generates the model, quantity, and inventory query link of the required spare parts.

[0142] Afterwards, it can be displayed directly on the tractor's central control screen for the operator to view on-site, or it can be simultaneously pushed to the fleet manager's mobile APP and the manufacturer's after-sales service center, making it convenient to dispatch maintenance personnel or prepare spare parts.

[0143] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for tractor test data acquisition and fault diagnosis, characterized in that, Includes the following steps: Multi-source sensor groups are deployed at key nodes of the tractor hydraulic system to trigger synchronous acquisition of operating status data according to four working conditions: tractor suspension lifting, lowering, neutral, and pressure drop and holding. The operating status data includes hydraulic oil pressure, temperature, flow rate, hydraulic valve core displacement signal, and hydraulic pump housing vibration signal. The operating status data is subjected to joint time-frequency denoising and adaptive imputation of missing values. Then, time-domain features, frequency-domain features and operating condition-related features are extracted to construct a standardized feature vector. Establish a hybrid training set that integrates digital twin simulation data and field measured fault data, and train an attention-based hydraulic fault diagnosis model based on the hybrid training set; The standardized feature vector is input into the hydraulic fault diagnosis model, which outputs the fault level, fault type and fault location. Based on the fault level, a graded early warning is triggered, and corresponding troubleshooting and maintenance strategies are generated.

2. The method for tractor test data acquisition and fault diagnosis as described in claim 1, characterized in that, The specific steps for deploying a multi-source sensor group at key nodes of the tractor hydraulic system and triggering synchronous acquisition of operational status data according to four operating conditions—tractor suspension lifting, lowering, neutral, and pressure drop and pressure holding—include the following: Pressure sensors, temperature sensors, flow sensors, hydraulic valve core displacement sensors, and hydraulic pump housing vibration sensors are respectively arranged at key nodes of the tractor hydraulic system, such as the hydraulic pump outlet, multi-way valve inlet and outlet, suspension cylinder chamber, hydraulic pump housing, and hydraulic oil tank, to form a multi-source sensor group. Based on four typical working conditions of tractor suspension (lifting, lowering, neutral, and pressure drop and holding), working condition identification conditions are set respectively. The current working condition is identified by real-time detection of suspension control commands and hydraulic system operating parameters. When any typical working condition is identified, data is collected to obtain the operating status data.

3. The method for tractor test data acquisition and fault diagnosis as described in claim 2, characterized in that, The specific steps of performing time-frequency joint denoising and adaptive imputation of missing values ​​on the operating status data, and then extracting time-domain features, frequency-domain features, and operating condition-related features to construct a standardized feature vector include: The operating status data is subjected to time-frequency joint denoising processing to generate a clean time-series signal; Identify missing data segments in the pure time-series signal, and perform adaptive imputation of missing values ​​according to the type of operation condition in which the missing data segments are located, to generate a complete time-series sequence; Based on the complete time series, time-domain features, frequency-domain features, and operating condition-related features are extracted respectively, and then fused to construct an initial feature set. The initial feature set is standardized and dimensionality reduced to construct a standardized feature vector for fault diagnosis.

4. The tractor test data acquisition and fault diagnosis method as described in claim 3, characterized in that, The specific steps for performing time-frequency joint denoising processing on the operating status data to generate a clean time-series signal include: The original data is decomposed into several intrinsic mode components using a variational mode decomposition algorithm; Calculate the permutation entropy value of each intrinsic mode component and its correlation coefficient with the original data; Based on the permutation entropy value and correlation coefficient, the intrinsic mode components are classified into high-frequency noise modes, mixed modes, and trend term modes; The high-frequency noise modes are directly removed; the mixed modes are denoised using adaptive wavelet thresholding; and the trend terms are filtered using moving average filtering. The processed components are reconstructed to obtain the pure timing signal.

5. The tractor test data acquisition and fault diagnosis method as described in claim 4, characterized in that, The specific steps for establishing a hybrid training set that integrates digital twin simulation data and field-measured fault data, and for training an attention-based hydraulic fault diagnosis model based on the hybrid training set, include: A high-fidelity digital twin of a tractor hydraulic system is constructed. By injecting preset fault mode parameters, a digital twin simulation dataset covering all working conditions and all fault types is generated. Multi-source sensor data were collected under actual field conditions and then cleaned and labeled to form a field-measured fault dataset. The digital twin simulation dataset and the field measured fault dataset are subjected to distribution alignment and domain adaptation processing, and then fused to construct a hybrid training set; Based on the attention mechanism, the fault diagnosis model is trained and fine-tuned in stages using the hybrid training set to optimize the model parameters until convergence, thus obtaining the final fault diagnosis model.

6. The tractor test data acquisition and fault diagnosis method as described in claim 5, characterized in that, The specific steps for constructing a high-fidelity digital twin of a tractor hydraulic system, and generating a digital twin simulation dataset covering all operating conditions and all fault types by injecting preset fault mode parameters, include: A mechanistic model incorporating the nonlinear characteristics of hydraulic pumps, multi-way valves, cylinders, and pipelines was established, and baseline parameters for normal conditions were set. Define a set of failure modes, which include increased internal leakage coefficient, valve core jamming, oil contamination and blockage, and seal failure. The Monte Carlo simulation method is used to randomly perturb key physical parameters under preset fault modes and run simulations under four typical operating conditions. The output is synchronous time-series data consistent with the actual sensor type, and includes fault labels and fault injection timestamps.

7. The method for tractor test data acquisition and fault diagnosis as described in claim 6, characterized in that, The specific steps for performing distribution alignment and domain adaptation processing on the digital twin simulation dataset and the field measured fault dataset, and fusing them to construct a hybrid training set, include: Extract the data source domain of the digital twin simulation; Calculate the maximum difference in mean between two domains; A transfer learning strategy is introduced to map the feature distribution of the source domain data through nonlinear transformation, thereby minimizing the difference in marginal probability distribution between the source and target domains. The hybrid training set is constructed by splicing the distribution-corrected simulation data and the measured data according to a preset ratio.

8. The method for tractor test data acquisition and fault diagnosis as described in claim 7, characterized in that, The specific steps for training the attention-based hydraulic fault diagnosis model based on a hybrid training set include: The input layer receives the standardized multi-channel temporal feature vector; The first level is a 1D-CNN feature extraction layer, which contains multiple one-dimensional convolutional kernels and pooling layers, used to extract local spatial features and high-frequency transient impact features from multi-source signals in parallel. The second level is the GRU temporal evolution layer, which is used to receive the feature sequence output by 1D-CNN and use the gating mechanism to capture the long-term and short-term dependencies and dynamic trends of hydraulic faults over time. The third level is the attention mechanism layer, which is used to calculate the attention weight matrix for the hidden state sequence output by the GRU, adaptively focusing on the key time steps with significant fault characteristics and suppressing the interference of background noise under normal operating conditions. The weighted feature vector is mapped to the fault category probability space, and the fault diagnosis result is output, which includes fault type, fault probability and confidence level.

9. The method for tractor test data acquisition and fault diagnosis as described in claim 8, characterized in that, The specific steps for triggering graded early warnings based on fault levels and generating corresponding troubleshooting and maintenance strategies include: The fault diagnosis model outputs fault type, fault probability, and confidence level in real time. Combined with the current operating conditions and historical degradation trends, a comprehensive fault risk index is calculated. Based on the comprehensive fault risk index, the fault level is divided into four levels: minor warning level, moderate alarm level, severe shutdown level, and emergency danger level, and corresponding graded warning signals are triggered. Based on the hierarchical early warning signals, matching investigation paths and maintenance strategies are retrieved from a pre-constructed knowledge graph; The troubleshooting path and maintenance strategy will generate a digital maintenance work order containing visual guidance, operation steps and security constraints, and push it to the remote operation and maintenance platform.

10. A tractor test data acquisition and fault diagnosis system, characterized in that, The method for acquiring test data and diagnosing faults in a tractor, as described in any one of claims 1 to 9.

Images