Intelligent fault diagnosis method and system for agricultural tractor engine

Abstract

The application relates to the technical field of intelligent fault diagnosis, and discloses an intelligent fault diagnosis method and system for an agricultural tractor engine, the method comprising the following steps: collecting vibration, temperature and rotating speed original operation parameters of key positions of the tractor engine, obtaining a denoising parameter sequence through multi-point synchronous collection and denoising treatment; extracting time sequence characteristic analysis parameters to analyze parameter change trends, triggering abnormal detection when the parameters exceed a preset threshold value, and determining local abnormal components of the engine; combining a historical fault database to perform component fault potential correlation analysis and generate a cause distribution diagram, matching real-time correlation data to locate implicit damage components; adjusting engine load parameters, synchronously updating a fault library and optimizing fault matching rules, and generating fault diagnosis basis suitable for a current working environment. The method can realize accurate diagnosis, early warning and closed-loop self-optimization of diagnosis rules of engine faults under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of engine fault diagnosis technology, and in particular to an intelligent fault diagnosis method and system for agricultural tractor engines. Background Technology

[0002] Currently, in the field of engine fault diagnosis technology, the reliability of the engine of a tractor, as the core power equipment for farmland operations, directly determines the efficiency of agricultural production and the continuity of operations. Engine fault prediction and health management technology has become the core technical support for ensuring the stable operation of agricultural machinery throughout its entire life cycle and reducing the risk of downtime in field operations.

[0003] In one existing technology, tractor engine fault diagnosis mainly relies on single-parameter over-limit alarms with fixed thresholds. It only superficially records and simply compares operating data such as engine vibration and temperature, and finally outputs the fault diagnosis result. However, real-world farmland conditions are complex, and fault characteristics exhibit dynamic changes. Because tractors are affected by multiple factors such as soil resistance and weather conditions under different operating scenarios, the original operating parameters of the engine will show highly irregular changes. It is difficult to judge whether there is an abnormality using fixed standards. The single-point, single-dimensional detection mode cannot clarify the fault transmission relationship between multiple components, and it is difficult to accurately locate the root cause of the fault. As a result, existing engine fault diagnosis technology cannot analyze the complex correlation of fault causes.

[0004] In summary, existing technologies suffer from poor adaptability to complex operating conditions, making it impossible to accurately locate the root cause of faults. This results in a high risk of engine downtime and low maintenance efficiency, which seriously affects the continuity of agricultural production. Summary of the Invention

[0005] This invention provides an intelligent fault diagnosis method and system for agricultural tractor engines to solve the problems of delayed fault warning, low accuracy in locating hidden damage, and poor adaptability to complex working conditions in existing technologies, thereby realizing intelligent fault diagnosis and closed-loop optimization of engines.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides an intelligent fault diagnosis method for agricultural tractor engines, comprising: The original operating parameters of the engine are obtained, and noise removal processing is performed on the original operating parameters to obtain a noise removal parameter sequence; Temperature fluctuation sequence is extracted from the denoised parameter sequence, and the temperature fluctuation sequence is input into a preset time series prediction model to obtain the parameter change trend. By comparing the parameter change trend with a preset trend judgment threshold, anomaly detection is performed on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components; Based on the local abnormal components and the preset historical fault database, a potential correlation analysis of faults is performed to obtain the cause distribution map corresponding to the abnormal fluctuations. Damage comparison analysis is performed based on the cause distribution map and the denoising parameter sequence to obtain a damage distribution matrix. The damage distribution matrix is ​​then grouped into damage concentration areas to obtain the location of components with latent engine damage. Based on the location of the components, the engine load parameters are compensated and adjusted in real time according to timing.

[0007] Secondly, the present invention provides an intelligent fault diagnosis system for an agricultural tractor engine, comprising: The data acquisition module is used to acquire the engine's raw operating parameters, perform noise removal processing on the raw operating parameters, and obtain a denoised parameter sequence. The trend analysis module is used to extract the temperature fluctuation sequence from the denoised parameter sequence, input the temperature fluctuation sequence into a preset time series prediction model, and obtain the parameter change trend. An anomaly detection module is used to compare the parameter change trend with a preset trend judgment threshold, and to perform anomaly detection on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components. The fault analysis module is used to perform potential correlation analysis of the faults based on the local abnormal components and the preset historical fault database to obtain the cause distribution map corresponding to the abnormal fluctuations. The damage localization module is used to perform damage comparison analysis based on the cause distribution map and the denoising parameter sequence to obtain a damage distribution matrix, and to group the damage distribution matrix into damage concentration areas to obtain the component locations of hidden engine damage. The load control module is used to perform time-series compensation and real-time adjustment of engine load parameters based on the location of the components.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention collects multi-dimensional operating data of vibration, temperature and speed at multiple points in key parts of the engine simultaneously, uses Kalman filtering algorithm to remove field noise, and completes data denoising through frequency domain conversion and multi-dimensional verification. This solves the problems of poor anti-interference of existing data collection and the easy masking of abnormal signals by noise, and improves the accuracy of basic data for fault diagnosis.

[0009] (2) This invention extracts the time sequence characteristics of the original operating parameters to analyze the parameter change trend, triggers anomaly detection, completes multi-node data grouping comparison and synchronous analysis, and combines historical fault database to analyze the fault transmission link to generate a cause distribution map, which solves the problem that the existing technology cannot identify early hidden damage and has low fault source tracing accuracy, and realizes accurate location of fault root cause.

[0010] (3) This invention solves the problems of existing diagnostic rules having no self-iterative capability and poor adaptability to complex working conditions by performing time-series compensation and real-time adjustment of engine load parameters after confirming the damage location, synchronously updating the fault database and optimizing fault matching rules based on the working environment, thus forming a closed loop of the entire process of diagnosis-control-optimization. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of an intelligent fault diagnosis method for an agricultural tractor engine provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent fault diagnosis system for an agricultural tractor engine 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 1 The first embodiment of the present invention provides an intelligent fault diagnosis method for agricultural tractor engines, comprising the following steps: S11, Obtain the original operating parameters of the engine, and perform noise removal processing on the original operating parameters to obtain a noise removal parameter sequence; S12, extract the temperature fluctuation sequence from the denoised parameter sequence, input the temperature fluctuation sequence into a preset time series prediction model, and obtain the parameter change trend; S13, compare the parameter change trend with the preset trend judgment threshold, perform anomaly detection on the original operating parameters that exceed the trend judgment threshold, and obtain local abnormal components; S14, perform a potential correlation analysis of the fault based on the local abnormal component and the preset historical fault database to obtain the cause distribution map corresponding to the abnormal fluctuation; S15, perform damage comparison analysis based on the cause distribution map and the denoising parameter sequence to obtain a damage distribution matrix, and group the damage distribution matrix into damage concentration areas to obtain the component locations of hidden engine damage; S16, based on the location of the component, perform timing compensation and real-time adjustment of the engine load parameters.

[0014] In step S11, the original operating parameters of the engine are obtained, and noise removal processing is performed on the original operating parameters to obtain a noise-removing parameter sequence, including: By deploying a sensor network on the engine, vibration signals, temperature data and speed indicators are collected synchronously from multiple points to obtain raw operating parameters; The original operating parameters are subjected to noise removal using a Kalman filter algorithm to obtain a denoised parameter sequence; wherein the original operating parameters include vibration signals, temperature data, and rotational speed indicators.

[0015] It should be noted that the sensor network nodes are deployed at key locations on the tractor engine, including the cylinder head, crankshaft bearings, intake manifold, connecting rod bearings, and exhaust valve seats. Vibration signal acquisition uses piezoelectric accelerometers, with a sampling frequency set to 10kHz according to the industry standard for agricultural machinery engine fault diagnosis, covering a frequency range of 0Hz to 2000Hz. Temperature data acquisition uses patch-type platinum resistance temperature sensors, with a sampling interval set to 100ms, covering a measurement range of -40℃ to 150℃. Speed ​​index acquisition uses magnetoelectric speed sensors, with a sampling accuracy set to 1rpm, covering a measurement range of 0rpm to 5000rpm. The multi-point synchronous acquisition process is triggered synchronously by a hardware clock, and the time synchronization error of the data acquired by each node is controlled within ±1ms. The acquired raw parameter sequences are then stored in the local cache unit after being aligned by timestamps.

[0016] It is worth noting that the state equation and observation equation of the Kalman filter algorithm are constructed based on the variation law of the original engine operating parameters. The state vector includes three dimensions: vibration signal amplitude, real-time temperature value, and real-time speed value. The process noise covariance matrix and observation noise covariance matrix are statistically calibrated using multiple batches of field operation data, with the process noise covariance value set to 0.01 and the observation noise covariance value set to 0.1. The filtering process sequentially executes a prediction step and an update step. The prediction step calculates the prior state estimate and prior error covariance matrix at the current time based on the optimal state estimate value at the previous time step. The update step calculates the Kalman gain by combining the observation data at the current time step, and updates the current optimal state estimate and posterior error covariance matrix, completing the noise removal of the original parameter sequence and outputting the intermediate parameter sequence.

[0017] It is worth further explaining that the preset vibration judgment threshold is set based on the fault-free vibration reference value under the rated operating conditions of the tractor engine. After statistical analysis of more than 100 sets of fault-free operating condition data in field operations, the vibration judgment threshold is set to 8Hz. The vibration signal amplitude of the intermediate parameter sequence is compared with the vibration judgment threshold point by point. For intermediate parameter sequences whose amplitude exceeds the vibration judgment threshold, frequency domain transformation is performed. The frequency domain transformation adopts Fast Fourier Transform (FFT) to convert the time-domain vibration signal into a frequency-domain signal. The number of points of the FFT is set to 1024. During the transformation process, a Hanning window is added to suppress spectral leakage. After the transformation, a frequency domain feature sequence containing the amplitude and phase information of each frequency component is obtained.

[0018] Subsequently, data classification and validation employed a pre-trained support vector machine (SVM) classifier with a radial basis function (RBF) kernel. The training dataset contained 1500 labeled data sets under normal, warning, and fault conditions of the engine, with a training-to-validation ratio of 7:3. The classification accuracy was tested to be no less than 98%. Input features included peak frequencies and corresponding amplitudes of the frequency domain feature sequence, real-time values ​​and rates of change of temperature data in the intermediate parameter sequence, and real-time values ​​and steady-state deviations of the engine speed index. The classifier performed multi-dimensional feature fusion and state classification on the input data, removing invalid data segments with noise interference and retaining valid data segments with classification results indicating normal or abnormal operating conditions. The valid data segments were then concatenated along the time axis to obtain a denoised parameter sequence.

[0019] In step S12, a temperature fluctuation sequence is extracted from the denoised parameter sequence, and the temperature fluctuation sequence is input into a preset time series prediction model to obtain the parameter change trend, including: The denoised parameter sequence is subjected to wavelet packet decomposition to extract frequency domain energy features and construct a time-series feature set; A local data segment of the time-series feature set is extracted by a preset sliding time window, and the variance of the temperature data within the local data segment is calculated to obtain the temperature fluctuation sequence. The predicted weights of the temperature fluctuation sequence are corrected by combining a preset fluctuation increment threshold to obtain a corrected temperature sequence. The corrected temperature sequence is input into a preset time series prediction model to obtain the parameter change trend.

[0020] It should be noted that the frequency band division of the vibration signal adopts the wavelet packet decomposition method. The number of decomposition layers is set to 3 based on the sampling frequency and effective frequency range of the vibration signal. After decomposition, 8 equal-width frequency bands are obtained, each with a width of 1250Hz, covering the sampling frequency range from 0Hz to 10000Hz. The wavelet packet decomposition coefficients in each frequency band are reconstructed, and the sum of squares of the reconstructed signal is calculated to obtain the signal energy value of the corresponding frequency band. The energy values ​​of all frequency bands are arranged in ascending order of frequency to extract the energy distribution characteristics of the vibration signal.

[0021] It is worth noting that the time-series segmentation of the speed index adopts an equal-length segmentation method aligned with the time division of the vibration signal frequency band, with a single segment length set to 1 second. The number of speed data points within each segment is consistent with the number of sampling points in a single segment of the vibration signal. The steady-state deviation is calculated as the difference between the real-time speed value within a single segment and the engine's rated speed under current operating conditions. The mean and variance of all steady-state deviations within a single segment are calculated to extract the speed fluctuation characteristics within that segment. The energy distribution characteristics and speed fluctuation characteristics are concatenated along the same time segment dimension, and the concatenated feature vectors are arranged in chronological order to construct a time-series feature set.

[0022] The time window sliding method sets the window length to 10 minutes and the sliding step size to 1 minute, ensuring a 90% overlap rate between adjacent windows during the sliding process. For each sliding window, a local data segment of the time-series feature set is extracted, and the corresponding temperature data sequence is extracted. The variance of all temperature data within the sequence is calculated, and the variance values ​​calculated for each sliding window are arranged in chronological order of the window's end time to generate a temperature fluctuation sequence.

[0023] It should be further explained that the temperature fluctuation sequence is compared with a preset variance change threshold to determine whether the sequence shows an increasing trend. The variance change threshold was verified by 80 sets of engine thermal decay simulation experiments and heavy-load field operation data, and was set to an increment of 0.1 every 10 minutes. If the temperature fluctuation sequence shows an increasing trend, the energy distribution characteristics of the corresponding time period are extracted from the time series feature set, and the ratio of the difference in energy characteristics between adjacent time nodes to the corresponding time interval is calculated to obtain the vibration evolution rate. The vibration evolution rate is used as a dynamic adjustment factor to adaptively correct the weight coefficient of the temperature fluctuation sequence in the time series prediction model. The weight correction range is 0.3 to 0.8. After correction, the time series feature set and the corrected temperature sequence are input into the time series prediction model, and the predicted values ​​of the original operating parameters for the next 2 hours are output to determine the parameter change trend. If the temperature fluctuation sequence does not show an increasing trend, the time series feature set is directly input into the time series prediction model, and the predicted values ​​of the original operating parameters for the next 2 hours are output to determine the parameter change trend.

[0024] The time series prediction model employs a linear autoregressive model, adapted to the low-computing-power requirements of tractor-mounted controllers. The model order is determined using the Akaike Information Criterion, traversing orders from 1 to 12. The model training dataset covers normal engine operation data and latent damage / fault data under various horsepower models, all operating conditions, and multiple environmental conditions, with a total sample size of no less than 1200 sets. The model uses the least squares method to solve for the autoregressive coefficients, with minimizing the sum of squared residuals as the optimization objective. Cross-validation is performed using the leave-one-out method. After training, the average absolute error between the predicted and measured values ​​is controlled within 3%, meeting the accuracy requirements for fault warning.

[0025] In step S13, the parameter change trend is compared with a preset trend judgment threshold. Anomaly detection is performed on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components, including: Compare the parameter change trend with a preset trend judgment threshold. If the trend judgment threshold is exceeded, trigger the anomaly detection process and acquire the collected data of each sensor node that is aligned with the time period of the anomaly trend. The collected data is grouped according to node location and synchronously compared to obtain the transmission delay of each node. Synchronous calibration and anomaly identification are performed in combination with the transmission delay to obtain local overload hazard distribution data. By comparing the collected data corresponding to the overload hazard distribution data with the collected data of adjacent normal nodes, local abnormal components are obtained. If the trend determination threshold is not exceeded, return to the original operation parameter acquisition step.

[0026] It should be noted that the preset trend judgment thresholds are statistically calibrated based on historical operating data of the tractor engine under fault-free field operation conditions, with a calibration data volume of no less than 100 sets of valid operating condition data for a complete operation cycle. The trend judgment thresholds include trend limits for multiple dimensions of parameters: the short-term upward trend limit for coolant temperature is set at 5°C per 10 minutes, the short-term upward trend limit for vibration energy is set at 50% per 10 minutes, and the short-term upward trend limit for steady-state deviation of engine speed is set at 3% per 10 minutes. The parameter change trends are compared with the trend judgment thresholds dimension by dimension. When the trend values ​​of all dimensions do not exceed the corresponding limits, the operating state is judged to be normal, and the current engine operating state and data acquisition process are maintained. The sampling frequency, filtering parameters, and sliding window step size of the data acquisition process remain unchanged from their initial settings. Control parameters such as engine fuel injection quantity and intake throttle opening are maintained at the calibration values ​​of the current operating conditions, and the real-time acquisition, noise reduction, and trend analysis of the original operating parameters are continuously performed.

[0027] When the trend value of any dimension's parameter exceeds the corresponding trend judgment threshold, the anomaly detection process is immediately triggered. Real-time data on vibration acceleration, surface temperature, and rotational speed collected by each sensor node are acquired, aligned with the time period of the anomaly trend. The parameter data is then synchronously analyzed using a data transmission delay control mechanism. This mechanism sends a synchronization trigger signal to each sensor node through the main control unit, recording the time difference between signal transmission and data return to obtain the single-channel transmission delay of each node. The maximum delay compensation range is set to 0 to 100 ms. Based on the transmission delay values ​​of each node, the data timestamps are resampled and aligned. The aligned time window width is set to 500 ms. Within the same time window, non-stationary fluctuations occurring synchronously at different nodes with amplitudes exceeding the baseline are identified, generating overload hazard distribution data. This overload hazard distribution data includes the physical location of the sensor node corresponding to the abnormal fluctuation, the fluctuation amplitude, the fluctuation occurrence timestamp, and the deviation ratio of the fluctuation from the baseline value.

[0028] It is worth further explaining that, based on the abnormal time periods corresponding to the overload hazard distribution data, the actual parameter change trends within those time periods are extracted and compared with the theoretical change trends output by the engine standard operating condition model. The difference is then calculated to obtain the deviation value of the parameter change trends. The engine standard operating condition model is constructed based on the original factory design parameters of the equipment and its operating characteristics under different loads. The model inputs are the engine's real-time load, throttle opening, and ambient temperature, and the output is the theoretical change trend of the original operating parameters under the corresponding operating conditions. The preset deviation judgment threshold, verified through multiple batches of load surge tests and fault condition simulations, is set to 20ms. The deviation value is compared with the deviation judgment threshold, and corresponding operations are performed based on the comparison results.

[0029] Subsequently, when the deviation value exceeds the deviation judgment threshold, the vibration evolution rate and temperature fluctuation sequence within the corresponding time period are extracted, and the Pearson correlation coefficient between the two sets of sequences is calculated to obtain the correlation feature value between vibration and temperature. Multi-node grouping comparison is performed on the calculated correlation feature value. The grouping method is based on the engine's cylinder unit and crankshaft transmission unit for spatial division. Correlation data of cylinder head temperature, cylinder wall vibration, and injector speed of the same cylinder are divided into the same experimental group, while similar data from adjacent normal cylinders are divided into the control group. The differences in correlation feature values ​​between the experimental group and the control group are compared. When the absolute value of the correlation feature of the experimental group reaches twice or more the average value of the control group, it is determined that there is a local abnormal component of the engine in the corresponding area of ​​the experimental group. The result of the local abnormal component judgment, including the component location, overload degree, and correlation feature value, is output.

[0030] When the deviation value does not exceed the deviation judgment threshold, the current abnormal data is judged as data acquisition noise, the abnormal data segment in the current time window is discarded, the corresponding temporary buffer area is cleared, the original running parameter acquisition steps are returned, and the parameter acquisition, noise reduction and trend analysis process of the next time window is re-executed without triggering subsequent fault tracing and diagnosis operations.

[0031] In step S14, a potential correlation analysis of the fault is performed based on the local abnormal component and a preset historical fault database to obtain a cause distribution map corresponding to the abnormal fluctuation, including: Extract the fluctuation data corresponding to the local abnormal components and construct a fluctuation feature vector; The candidate fault dataset is obtained by retrieving the preset historical fault database based on the fluctuation feature vector. Extract potential fault correlations and transmission links between components from the candidate fault dataset; The fault contribution value of each component is obtained by weighted calculation based on the potential correlation of the fault and the transmission link. Combined with the engine topology and the fault contribution value, a cause distribution map corresponding to the abnormal fluctuation is generated.

[0032] It should be noted that, for the identified localized abnormal components, fluctuation data of the corresponding sensor points within 30 minutes before and after the occurrence of the abnormality are extracted. This fluctuation data includes the spectral distribution of the vibration signal, energy values ​​in each frequency band, real-time rate of change of cylinder wall temperature, instantaneous deviation of speed fluctuation, and peak and mean values ​​of vibration acceleration. The extraction process centers on the point when the abnormal fluctuation first exceeds the baseline value, extracting continuous data segments of 15 minutes before and after it. The sampling frequency of these data segments remains consistent with the original acquisition frequency to ensure the temporal continuity and integrity of the data. A multi-dimensional fluctuation feature vector is constructed based on the extracted fluctuation data. The vector dimensions include five dimensions: peak vibration frequency, rate of change of vibration energy, rate of temperature rise, steady-state deviation of speed, and vibration-temperature Pearson correlation coefficient. The values ​​of each dimension are normalized to 0-1 to eliminate the influence of different units on subsequent retrieval. The normalized feature vector elements range from 0 to 1.

[0033] A pre-built historical fault database stores fault labeling data for the entire lifecycle of tractor engines. Each fault record includes the faulty component identifier, fault type, feature vector under the corresponding operating condition, fault evolution time-series data, and fault root cause labeling information. The database contains no fewer than 500 sets of valid fault cases verified through actual maintenance. High-dimensional space retrieval employs a cosine similarity matching method, calculating the cosine similarity between the multidimensional fluctuation feature vector and the feature vectors of each fault record in the database. The retrieval process sorts fault records from highest to lowest cosine similarity value, selecting those with a similarity value of no less than 0.8 to form a candidate fault dataset. The maximum number of records in the candidate fault dataset is limited to 10 to avoid redundancy in subsequent calculations.

[0034] It is worth noting that, from the candidate fault dataset, matching fault records are selected, and the three records with the highest cosine similarity are chosen as the matching fault records. Pre-labeled fault occurrence time-series data, component state change nodes, and fault triggering conditions are extracted from these matching fault records to parse fault evolution information. The fault evolution information parsing process breaks down the causal chain of fault development in chronological order, clarifies the temporal correspondence between each preceding abnormal event and subsequent fault manifestation, distinguishes between root cause events and symptom events, and outlines the complete development process from micro-component performance degradation to macro-abnormal fluctuations. For example, the parsed fault evolution information shows that lubricating oil film rupture precedes local temperature rise, local temperature rise precedes a sudden increase in high-frequency vibration energy, and the sudden increase in high-frequency vibration energy ultimately triggers overall abnormal fluctuations in the engine's original operating parameters.

[0035] Based on the analyzed fault evolution information and combined with the engine's three-dimensional digital topology, potential fault correlations and fault transmission links between engine components are extracted. These potential correlations are pre-defined based on the engine's mechanical transmission relationships, fluid circuit connections, and thermodynamic conduction relationships, encompassing component association rules within the four major systems: lubrication, cooling, combustion, and crankshaft-connecting rod mechanisms. The fault transmission link extraction process starts from the abnormal fluctuation point and traces backward along the causal direction of the fault evolution information, tracking the physical transmission path of abnormal signals and clarifying the order in which the fault impact is transmitted from the root cause component to the manifest component, such as the complete transmission link from the oil pump to the main oil passage, then to the crankshaft bearing, and finally to the connecting rod journal.

[0036] Subsequently, a weighted calculation was performed based on the potential correlation of the fault and the fault transmission link. The weight coefficients were set according to the position of the component in the fault transmission link and its fault contribution. The weight coefficient of the root cause component was set to 1.0, and the weight coefficients of each related component along the transmission link decreased progressively by 0.8. This weight allocation method was verified by 100 sets of fault transmission path simulation experiments, and the matching degree was no less than 95%. In the weighted calculation process, the fault trigger frequency and abnormal contribution ratio of each component in the matched fault records were statistically analyzed, and the fault contribution value of each component was calculated by combining the weight coefficients. The fault contribution value ranged from 0 to 1. Based on the calculated fault contribution values ​​of each component, a cause distribution map of abnormal fluctuations was generated by combining the engine topology. The cause distribution map was marked in a visual form with the fault contribution of each component, the direction of the fault transmission link, and the position of the root cause component. Components with a fault contribution of no less than 0.7 were marked as major influencing components, and components with a fault contribution of 0.3 to 0.7 were marked as related influencing components.

[0037] In step S15, damage comparison analysis is performed based on the cause distribution map and the denoising parameter sequence to obtain a damage distribution matrix. The damage distribution matrix is ​​then grouped into damage concentration regions to obtain the component locations of latent engine damage, including: Based on the causal distribution diagram, obtain the abnormal association fields of potential components; The abnormal correlation field and the denoising parameter sequence are dynamically time-normalized and compared to obtain the damaged sequence set; Damage link analysis is performed on the damage sequence set to generate a damage distribution matrix; Based on the damage distribution matrix and the preset damage judgment threshold, the matrix elements that exceed the damage judgment threshold are clustered and grouped to obtain the core region of concentrated damage. The core region is mapped to the engine equipment topology to obtain the location of the components with hidden engine damage.

[0038] It should be noted that, based on the main influencing components and related influencing components marked on the causal distribution map, corresponding related data fields are obtained through synchronous real-time acquisition. These related data fields include the dynamic tension of the cooling system water pump drive belt, the real-time temperature difference between the radiator inlet and outlet, the exhaust manifold back pressure, the in-cylinder combustion pressure, and the lubricating oil supply pressure. The time range of the data acquisition is completely aligned with the time range of the denoising parameter sequence, and the sampling frequency is consistent with the main parameter acquisition frequency, ensuring the time synchronization and dimensional matching of the data. Dynamic time warping is performed on the related data fields and the denoising parameter sequence. The comparison process first performs Z-score standardization on both sets of sequences to eliminate matching errors caused by amplitude differences. Then, a dynamic programming algorithm is used to calculate the optimal matching path between the two sets of sequences, obtaining the warping distance between the sequences. Sequence segments with warping distances less than a preset matching threshold are selected to form a damage sequence set based on similarity matching. The preset matching threshold is set to 0.2 after verification through comparison of 80 sets of abnormal and normal data under the same operating conditions, ensuring that the initial sequence set retains effective data segments highly correlated with the abnormal operating conditions.

[0039] It is worth noting that multidimensional feature vectors containing the second derivative of the rate of change of temperature difference, the variance of high-frequency vibration of belt tension, and the amplitude of fuel supply pressure fluctuation are extracted from the damage sequence set. By analyzing the sequential response order of the feature vectors on the time axis, the potential propagation path of latent damage is deduced, clarifying the direction and temporal relationship of the transmission of abnormal features from the initial occurrence component to related components. Link analysis is performed on the potential propagation path based on preset component association rules. These rules are pre-constructed in conjunction with engine thermodynamic and mechanical transmission models, including spatial proximity between components, mechanical transmission relationships, and energy conduction path constraints. The analysis process calculates the cumulative thermal stress and mechanical fatigue of each related component node along the path, generating a damage distribution matrix with dimensions consistent with the number of components. Each element in the matrix corresponds to the quantified damage impact value of a single component, fully covering all related components in the causal distribution map.

[0040] It is worth further explaining that the eigenvectors within the damage distribution matrix undergo 0-1 normalization to eliminate the influence of different physical dimensions on the evaluation results. The Euclidean norm of the normalized eigenvectors is then calculated to obtain dimensionless damage assessment scalar values. These scalar values ​​are compared with a preset dimensionless damage judgment threshold, which is set to 0.7 after statistical calibration using over 100 sets of fault-free and fault-free operating data. Eigenvectors exceeding the damage judgment threshold are clustered using the K-means algorithm to identify the core regions of concentrated damage. The spatial coordinates of these core regions are mapped onto the engine's digital equipment topology to achieve precise physical location matching, ultimately yielding the location of components with latent engine damage. The positioning accuracy can be matched to the level of a single component in a single cylinder.

[0041] In step S16, based on the location of the component, the load parameters of the engine are subjected to time-series compensation and real-time adjustment, including: Based on the component location, extract the full-link transmission parameters of the corresponding sensor node, data transmission bus and engine actuator; Calculate the transmission delay time of the full-link transmission parameters; The load parameters of the engine are obtained, and the load parameters are compensated for timing advance based on the transmission delay time to obtain the load control command; The load control command is executed to adjust the engine's operating load.

[0042] It should be noted that, based on the component location, the full-link transmission parameters of the corresponding sensor node, data transmission bus, and engine actuator are extracted, and the total transmission delay time from data acquisition and triggering at the bottom layer to the main control unit completing the calculation and analysis is calculated. The total transmission delay time includes three parts: sensor signal acquisition and analog-to-digital conversion delay, CAN bus data transmission delay, and main control unit core calculation and processing delay. The calculation process sums up the delay times of each segment. The calibration values ​​of each segment delay were obtained through multiple batches of field operation tests, and the calibration range of the total transmission delay under normal operating conditions is 20ms to 50ms. Based on the calculated total transmission delay time, timing advance compensation is performed on the engine load parameters. The compensation duration is completely consistent with the total transmission delay time. For the cylinder unit and transmission mechanism corresponding to the hidden damaged components, four types of core load parameters, namely fuel injection pulse width, intake throttle opening, injection advance angle, and maximum speed limit, are precisely adjusted to generate load control commands with timing compensation values, thereby offsetting the control lag problem caused by the transmission delay.

[0043] It is worth noting that load control commands are sent to the engine electronic control unit in real time, driving the electronic fuel injection system and electronic throttle to perform corresponding operations, adjusting the engine's operating state, and forcing the engine into a torque-reducing protective operating mode. Latent damage is classified into three levels of risk based on its contribution to the failure: Level 1 risk, Level 2 risk, and Level 3 risk. Level 1 risk corresponds to a failure contribution of no less than 0.9, with a torque limitation ratio set at 30% of the rated torque. Level 2 risk corresponds to a failure contribution of 0.7 to 0.9, with a torque limitation ratio set at 50% of the rated torque. Level 3 risk corresponds to a failure contribution of 0.3 to 0.7, with a torque limitation ratio set at 70% of the rated torque, to prevent latent damage from further deteriorating under high-load conditions.

[0044] After executing the load control command, the method further includes: Real-time fluctuation data during load adjustment is collected, and the fluctuation data is structurally concatenated with the anomaly association field and written into the historical fault database to obtain updated fault records. Extract the updated fault records, generate a structured processing log, and parse the processing log to obtain an environmental feature vector; Based on the environmental feature vector, the preset fault matching rules are weighted and redistributed to obtain an updated set of rule weights. The fault matching logic is reconstructed based on the set of rule weights to obtain optimized matching rules; The processing logs are subjected to feature matching using the optimized matching rules to generate fault diagnosis criteria adapted to the current operating environment.

[0045] It should be noted that throughout the entire cycle of executing the load control command, high-frequency data on fluctuations during the load adjustment process are collected synchronously. The collected parameters include the in-cylinder combustion pressure of the corresponding cylinder, the vibration acceleration of damaged parts, surface temperature, lubricating oil film thickness, and real-time speed deviation. The collection frequency is consistent with the original operating parameter collection frequency of 10kHz. The timestamp of the collected data is strictly aligned with the issuance and execution time of the load control command. The collection duration covers the entire interval from 10 seconds before the command is executed to 30 seconds after the execution, ensuring that the collected data fully reflects the dynamic state changes of the engine during the load adjustment process.

[0046] Subsequently, the collected full-cycle fluctuation data and the associated data fields of corresponding components are structurally spliced ​​together according to five dimensions: timestamp, unique component identifier, operating condition parameters, abnormal characteristic values, and load control execution data, forming a standardized multi-dimensional event data packet. Based on the preset database update frequency, the optimal time window for data writing is calculated. The database update frequency, verified through over 100 sets of field operating conditions, is set to once every 200ms. The writing time window is set to the 15ms to 25ms range within the update cycle, avoiding peak periods of CAN bus data transmission and reducing interference from write operations on the engine's real-time control process. The spliced ​​event data packets are then prioritized according to the order of anomaly occurrence time, generating a data queue to be written.

[0047] It's worth noting that the system compares the current absolute system time with the write time window in real time. If the current system time falls within the write time window, all content in the data queue is appended to the historical fault database. This appended content includes the precise location of hidden damaged components, full-cycle data of abnormal fluctuations, related data fields, load control parameters, and complete fault handling process data. After writing, an updated fault record with a unique incremental iteration identifier is generated. If the current system time is not within the write time window, the entire data queue is cached in a preset non-volatile storage area. The caching process simultaneously performs cyclic redundancy checks to prevent data loss or corruption. Once the system clock reaches the write time window, the data write operation is automatically executed, updating the historical fault database and generating an updated fault record. This provides valuable annotation data for subsequent optimization of fault mode matching rules.

[0048] It should be noted that, based on the updated fault records, a data reorganization operation is performed according to a preset log format to generate a structured processing log with an iteration identifier. The updated fault records include a unique numerical identifier for the hidden damaged component, full-cycle time-series data of abnormal fluctuations, related data fields, load control execution parameters, and full-process time-series data of the abnormality handling. The preset log format predefines six fixed fields: unique event ID, iteration batch identifier, abnormality occurrence timeline, component characteristic data, control execution data, and operating environment data. The data reorganization process categorizes the discrete data in the updated fault records according to these fixed fields, performs integrity checks, removes redundant data, and completes missing timestamps and operating condition-related information. The iteration identifier uses an incrementing numerical number, corresponding one-to-one with the version iteration number of the historical fault database. The initial iteration identifier is 1, and it automatically increments by 1 after each database update, ensuring that each structured processing log has unique traceability. The log file is stored in a local non-volatile storage area, and cyclic redundancy checks are performed synchronously to prevent data corruption.

[0049] The generated structured logs are parsed to extract current operating environment parameters and construct an environmental feature vector. The extracted operating environment parameters include five core dimensions: ambient temperature, soil resistance coefficient, operating altitude, operating load level, and relative humidity. The value ranges of each parameter are calibrated according to general operating conditions for agricultural machinery in the field. Ambient temperature ranges from -40℃ to 50℃, soil resistance coefficient ranges from 0.2 to 1.0, operating altitude ranges from 0m to 3000m, operating load levels are divided into light, medium, and heavy loads, and relative humidity ranges from 10% to 100%. The extracted environmental parameters are normalized to 0-1 to eliminate the influence of different dimensions on weight calculations. The normalized parameters are then arranged in a fixed dimensional order to construct a 5×1 environmental feature vector, fully representing the external operating conditions of the current operation.

[0050] Next, based on the constructed environmental feature vector, the weight coefficients of the preset fault matching rules are calculated to obtain the updated rule weights. The preset fault matching rules are stored according to engine fault type, including five core categories: valve seat damage, connecting rod bearing wear, cylinder lubrication failure, cooling system abnormalities, and fuel supply failure. The initial weight coefficient for each rule is set to 0.3. The weight coefficient calculation process uses the environmental feature vector as input and combines it with the fault occurrence frequency statistics under corresponding environmental conditions in the historical fault database. A weighted summation method is used to calculate the fit of each fault matching rule. The calculation formula is as follows:

[0051] in, Let be the updated weight coefficient of the i-th fault matching rule. These are the initial weighting coefficients for this rule. The environmental impact factor, after multiple batches of operational condition verification, was set at 0.6. This corresponds to the historical occurrence count of this type of fault under the corresponding environmental conditions. This represents the total number of fault records under the corresponding environmental conditions. After calculation, the weight coefficients of all rules are normalized to a value range of 0.1 to 0.9, forming the set of updated rule weights.

[0052] Based on the updated rule weights, the original fault matching rules are logically restructured and optimized to obtain optimized matching rules. The logical restructuring employs a fault decision tree reconstruction approach, elevating the fault matching rule with the highest updated weight coefficient to the root node of the decision tree. The remaining rules are arranged sequentially at each level of the decision tree according to their weight coefficients, adjusting the priority order of fault matching. Rules with weight coefficients below 0.2 are moved to the last branch of the decision tree. Simultaneously, by incorporating the feature data from the updated fault records, the feature similarity threshold and anomaly judgment boundary parameters of the corresponding fault matching rules are corrected to improve the matching accuracy of fault features under the same operating conditions. The optimized fault matching rules are then synchronously updated to the local rule base, while the rule file of the previous version is retained for version rollback, ultimately resulting in optimized matching rules adapted to the current operating environment.

[0053] It should be noted that the structured logs were subjected to full-field feature scanning and matching analysis by optimizing the matching rules. The optimized matching rules were fault decision tree rules that had undergone weight optimization and logical reconstruction. The scanning process followed the priority order of the decision tree, sequentially extracting component anomaly feature data, time-series change data of original operating parameters, and load adjustment response data from the structured logs. These were then compared with the standard feature set of the corresponding fault type in the rule base using cosine similarity matching calculation. During the matching process, the temporal causal relationship and amplitude change range of the features were simultaneously verified, and interfering features that did not conform to the fault evolution information were eliminated. Valid features with a similarity of not less than 0.8 were selected, and anomaly features directly related to engine faults were identified. At the same time, the fault type, impact range, and risk level corresponding to the anomaly features were labeled.

[0054] It is worth noting that by combining environmental feature vectors with locked anomaly features, engine fault diagnosis criteria adapted to the current operating environment are generated. The environmental feature vector is a multi-dimensional vector representing the ambient temperature, soil drag coefficient, operating altitude, and load level of the current operation. During the generation process, the risk weights of anomaly features are adjusted based on the environmental feature vectors. For severe operating conditions such as high temperature, high drag, and heavy load, the weight coefficient of the fault deterioration risk is increased, while for light load and stable operating conditions, the influence weight of non-critical features is decreased. The final generated diagnostic criteria include the root cause component of the fault, the fault development trend, maintenance and handling suggestions adapted to the current operating conditions, and the fault risk level judgment result. It can be directly used to guide the field maintenance and fault repair of tractor engines, and at the same time, it provides standardized judgment for subsequent fault diagnosis under the same operating conditions.

[0055] In summary, this invention discloses an intelligent fault diagnosis method for agricultural tractor engines, comprising: acquiring the engine's original operating parameters; performing noise removal processing on the original operating parameters to obtain a denoised parameter sequence; extracting a temperature fluctuation sequence from the denoised parameter sequence; inputting the temperature fluctuation sequence into a preset time series prediction model to obtain parameter change trends; comparing the parameter change trends with a preset trend judgment threshold, and performing anomaly detection on the original operating parameters exceeding the trend judgment threshold to obtain locally abnormal components; performing potential correlation analysis of component faults based on the locally abnormal components and a preset historical fault database to obtain a causal distribution map corresponding to the abnormal fluctuations; performing damage comparison analysis based on the causal distribution map and the denoised parameter sequence to obtain a damage distribution matrix, and grouping the damage concentration areas according to the damage distribution matrix to obtain the component locations of hidden engine damage; and performing time-series compensation and real-time adjustment of engine load parameters based on the component locations. This invention achieves accurate diagnosis, early warning, and closed-loop self-optimization of diagnostic rules for engine faults under complex operating conditions through synchronous acquisition of multi-dimensional original operating parameters of tractor engines, time-series feature analysis, and potential fault correlation tracing.

[0056] Reference Figure 2 The second embodiment of the present invention provides an intelligent fault diagnosis system for agricultural tractor engines, comprising: The data acquisition module is used to acquire the engine's raw operating parameters, perform noise removal processing on the raw operating parameters, and obtain a denoised parameter sequence. The trend analysis module is used to extract the temperature fluctuation sequence from the denoised parameter sequence, input the temperature fluctuation sequence into a preset time series prediction model, and obtain the parameter change trend. An anomaly detection module is used to compare the parameter change trend with a preset trend judgment threshold, and to perform anomaly detection on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components. The fault analysis module is used to perform potential correlation analysis of the faults based on the local abnormal components and the preset historical fault database to obtain the cause distribution map corresponding to the abnormal fluctuations. The damage localization module is used to perform damage comparison analysis based on the cause distribution map and the denoising parameter sequence to obtain a damage distribution matrix, and to group the damage distribution matrix into damage concentration areas to obtain the component locations of hidden engine damage. The load control module is used to perform time-series compensation and real-time adjustment of engine load parameters based on the location of the components.

[0057] It should be noted that the intelligent fault diagnosis system for an agricultural tractor engine provided in this embodiment of the invention is used to execute all the process steps of the intelligent fault diagnosis method for an agricultural tractor engine in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0058] It should be noted that the device 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 device 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.

[0059] 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 intelligent fault diagnosis of an agricultural tractor engine, characterized in that, include: The original operating parameters of the engine are obtained, and noise removal processing is performed on the original operating parameters to obtain a noise removal parameter sequence; Temperature fluctuation sequence is extracted from the denoised parameter sequence, and the temperature fluctuation sequence is input into a preset time series prediction model to obtain the parameter change trend; By comparing the parameter change trend with a preset trend judgment threshold, anomaly detection is performed on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components; Based on the local abnormal components and the preset historical fault database, a potential correlation analysis of component faults is performed to obtain a cause distribution map corresponding to the abnormal fluctuations. Damage distribution matrix is ​​obtained by comparing the damage distribution map with the denoising parameter sequence. Then, the damage concentration areas are grouped according to the damage distribution matrix to obtain the location of the components with hidden engine damage. Based on the location of the components, the engine load parameters are compensated and adjusted in real time according to timing.

2. The intelligent fault diagnosis method for agricultural tractor engines according to claim 1, characterized in that, The process involves acquiring the engine's original operating parameters, performing noise removal processing on these parameters to obtain a denoised parameter sequence, including: By deploying a sensor network on the engine, vibration signals, temperature data and speed indicators are collected synchronously from multiple points to obtain raw operating parameters; The original operating parameters are subjected to noise removal using a Kalman filter algorithm to obtain a denoised parameter sequence; wherein the original operating parameters include vibration signals, temperature data, and rotational speed indicators.

3. The intelligent fault diagnosis method for agricultural tractor engines according to claim 1, characterized in that, The step of extracting the temperature fluctuation sequence from the denoised parameter sequence and inputting the temperature fluctuation sequence into a preset time series prediction model to obtain the parameter change trend includes: The denoised parameter sequence is subjected to wavelet packet decomposition to extract frequency domain energy features and construct a time-series feature set; A local data segment of the time series feature set is extracted by a preset sliding time window, and the variance of the temperature data within the local data segment is calculated to obtain the temperature fluctuation sequence. The predicted weights of the temperature fluctuation sequence are corrected by combining a preset fluctuation increment threshold to obtain a corrected temperature sequence. The corrected temperature sequence is input into a preset time series prediction model to obtain the parameter change trend.

4. The intelligent fault diagnosis method for agricultural tractor engines according to claim 3, characterized in that, The process involves comparing the parameter change trend with a preset trend judgment threshold, and performing anomaly detection on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components, including: Compare the parameter change trend with a preset trend judgment threshold. If the trend judgment threshold is exceeded, trigger the anomaly detection process and acquire the collected data of each sensor node that is aligned with the time period of the anomaly trend. The collected data is grouped according to node location and synchronously compared to obtain abnormal fluctuation characteristics. The correlation of the abnormal fluctuation characteristics is calculated to obtain overload hazard distribution data. By comparing the collected data corresponding to the distribution of overload hazards with the collected data of adjacent normal nodes, local abnormal components are obtained. If the trend determination threshold is not exceeded, return to the original operation parameter acquisition step.

5. The intelligent fault diagnosis method for agricultural tractor engines according to claim 1, characterized in that, The step of performing a potential correlation analysis of the faults based on the local abnormal components and a preset historical fault database to obtain a distribution map of the causes of abnormal fluctuations includes: Extract the fluctuation data corresponding to the local abnormal components and construct a fluctuation feature vector; The candidate fault dataset is obtained by retrieving the preset historical fault database based on the fluctuation feature vector. Extract potential fault correlations and transmission links between components from the candidate fault dataset; The fault contribution value of each component is obtained by weighted calculation based on the potential correlation of the fault and the transmission link. Combined with the engine topology and the fault contribution value, a cause distribution map corresponding to the abnormal fluctuation is generated.

6. The intelligent fault diagnosis method for agricultural tractor engines according to claim 1, characterized in that, The damage distribution matrix is ​​obtained by comparing the damage distribution map with the denoising parameter sequence. The damage distribution matrix is ​​then grouped into damage concentration regions to determine the location of components with latent engine damage, including: Based on the causal distribution diagram, obtain the abnormal association fields of potential components; The abnormal correlation field and the denoising parameter sequence are dynamically time-normalized and compared to obtain the damaged sequence set; Damage link analysis is performed on the damage sequence set to generate a damage distribution matrix; Based on the damage distribution matrix and the preset damage judgment threshold, the matrix elements that exceed the damage judgment threshold are clustered and grouped to obtain the core region of the damage concentration. The core region is mapped to the engine equipment topology to obtain the location of the components with hidden engine damage.

7. The intelligent fault diagnosis method for agricultural tractor engines according to claim 6, characterized in that, The step of performing time-series compensation and real-time adjustment of engine load parameters based on the location of the components includes: Based on the component location, extract the full-link transmission parameters of the corresponding sensor node, data transmission bus and engine actuator; Calculate the transmission delay time of the full-link transmission parameters; The load parameters of the engine are obtained, and the load parameters are compensated for timing advance based on the transmission delay time to obtain the load control command; The load control command is executed to adjust the engine's operating load.

8. The intelligent fault diagnosis method for agricultural tractor engines according to claim 7, characterized in that, After executing the load control command, the method further includes: Real-time fluctuation data during load adjustment is collected, and the fluctuation data is structurally concatenated with the anomaly association field and written into the historical fault database to obtain updated fault records. Extract the updated fault records, generate a structured processing log, and parse the processing log to obtain an environmental feature vector; Based on the environmental feature vector, the preset fault matching rules are weighted and redistributed to obtain an updated set of rule weights. The fault matching logic is reconstructed based on the set of rule weights to obtain optimized matching rules; The processing logs are subjected to feature matching using the optimized matching rules to generate fault diagnosis criteria adapted to the current operating environment.

9. An intelligent fault diagnosis system for an agricultural tractor engine, characterized in that, include: The data acquisition module is used to acquire the engine's raw operating parameters, perform noise removal processing on the raw operating parameters, and obtain a denoised parameter sequence. The trend analysis module is used to extract the temperature fluctuation sequence from the denoised parameter sequence, input the temperature fluctuation sequence into a preset time series prediction model, and obtain the parameter change trend. An anomaly detection module is used to compare the parameter change trend with a preset trend judgment threshold, and to perform anomaly detection on the original operating parameters that exceed the trend judgment threshold to obtain local abnormal components. The fault analysis module is used to perform potential correlation analysis of the faults based on the local abnormal components and the preset historical fault database to obtain the cause distribution map corresponding to the abnormal fluctuations. The damage localization module is used to perform damage comparison analysis based on the cause distribution map and the denoising parameter sequence to obtain a damage distribution matrix, and to group the damage distribution matrix into damage concentration areas to obtain the component locations of hidden engine damage. The load control module is used to perform time-series compensation and real-time adjustment of engine load parameters based on the location of the components.

Images