A machine learning-based tillage machine operation parameter optimization processing method

Abstract

This invention relates to the field of agricultural machinery operation parameter optimization technology, and discloses a machine learning-based method for optimizing tillage machine operation parameters. The method involves acquiring location data of soil hardness sampling points and conic index profile data, and collecting continuous operation data of the tillage machine. The position of the tillage component is determined based on the installation offset and heading of the positioning device and tillage component. A sweeping action domain is constructed based on the effective working width, and the effective working interval and spatial matching parameters corresponding to the sampling points are determined. The depth-based impact weight is determined based on the actual tillage depth, and the contribution of hardness and depth matching parameters are obtained, with time delay compensation applied to the tillage response. Valid training samples are selected based on registration reliability, and a machine learning algorithm is used to establish a mapping relationship between soil hardness, operation parameters, and tillage response, thereby determining the target combination of operation parameters. This invention can reduce the spatiotemporal and depth-based mismatch between soil hardness and tillage response data, improving the reliability of operation parameter optimization.

Description

Technical Field

[0001] This invention relates to the field of tillage machine data processing technology, specifically to a method for optimizing tillage machine operation parameters based on machine learning. Background Technology

[0002] In agricultural tillage, soil hardness is a significant environmental factor affecting tillage resistance, operational energy consumption, tillage depth stability, and the load on tillage components. Existing methods for optimizing tillage machine operating parameters typically involve obtaining the conic index at several sampling locations within a plot using a conic hardness tester. This data is then combined with information on the tillage machine's forward speed, tillage depth, tillage component rotation speed, and tillage response to create a dataset. Machine learning and other methods are then employed to analyze the relationship between soil condition, operating parameters, and tillage response, ultimately determining suitable operating parameters.

[0003] However, the conic index is typically obtained using discrete point sampling. Tillers move continuously along their work paths in the field, and soil hardness exhibits significant spatial heterogeneity, potentially changing considerably near compacted layers, stubble, or locally hard soil. When soil hardness data obtained from discrete sampling is directly assigned to adjacent continuous work sections, it can easily lead to inconsistencies between the soil hardness sampling location and the actual application location of the tillage components, resulting in a spatial mismatch between soil hardness data and the corresponding tillage response. The conic index is usually collected within a preset depth range, while the actual tillage depth of the tiller is dynamically affected by soil resistance, implement posture, and workload. If the conic index or its average value within a fixed depth range is directly used to characterize soil hardness, information about soil layers not actually affected by the tillage components may be introduced into the work sample, causing the depth range represented by the soil hardness characteristics to be inconsistent with the actual cut soil layer. Furthermore, positioning data, tillage depth data, and tillage response data are typically obtained from different sensors or acquisition links, and these data types differ in sampling frequency, signal transmission, and response time. Especially during continuous operation of the tillage machine, a temporal offset may occur between the actual location where the tillage components are subjected to soil action and the location where the response signal is collected, further distorting the data correspondence between soil conditions, operating parameters, and tillage responses. These spatial, depth, and temporal data mismatches reduce the reliability of machine learning training samples, causing the established relationship between soil hardness, operating parameters, and tillage responses to be affected by erroneous samples. In areas with significant variations in soil hardness, parameter optimization results obtained based on this type of data are prone to bias, making it difficult to stably determine tillage parameters that match actual operating conditions based on different soil states. Summary of the Invention

[0004] The purpose of this invention is to provide a method for optimizing tillage machine operation parameters based on machine learning, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing tillage machine operation parameters based on machine learning, comprising: S1. Obtain the location data of multiple soil hardness sampling points and conic index profile data to form a depth sequence and hardness feature vector; S2. Continuously acquire the tiller's operation data to form a first positioning sequence, a heading sequence, a first tillage depth sequence, an operation parameter sequence, and a first response sequence; S3. Based on the installation offset and heading sequence of the positioning device and the tillage component, perform coordinate transformation on the first positioning sequence to obtain the second positioning sequence of the tillage component; S4. Construct the sweeping action area based on the second positioning sequence and the effective working width of the tillage component, and determine the effective working interval and spatial matching parameters based on the positional relationship between each sampling point and the sweeping action area. S5. Extract the second tillage depth sequence corresponding to the effective working area, compare it with the depth sequence, determine the depth effect weight according to the proportion of actual tillage depth reaching each sampling depth, and weight the hardness feature vector to obtain the hardness effect contribution vector and depth matching parameters. S6. Perform time compensation on the first response sequence according to the response delay parameter to obtain the second response sequence, and extract the response data corresponding to the effective working area to form a tillage response feature vector. S7. Determine the registration confidence level based on the spatial matching parameters and depth matching parameters. Use the hardness contribution vector with a registration confidence level not less than the preset threshold, the corresponding operation parameters, and the tillage response feature vector to form an effective training sample. Use machine learning algorithms to establish the mapping relationship between soil hardness, operation parameters, and tillage response.

[0006] As can be seen from the above technical solution, the present invention has the following beneficial effects: This invention correlates soil hardness sampling data with continuous tillage data, obtains the actual movement position of the tillage component based on the installation relationship between the positioning device and the tillage component, and determines the effective operating range corresponding to the soil hardness sampling point by combining the effective operating width of the tillage component. This reduces the spatial mismatch between discrete soil hardness sampling positions and continuous tillage positions. Furthermore, it determines the weight of different sampling depths based on the actual tillage depth within the effective operating range, ensuring that the hardness contribution reflects the actual soil layer affected by the tillage component and reduces the interference of untilled soil layers on hardness characteristics. Simultaneously, it improves the temporal correspondence between response data and actual operating positions by performing time delay compensation on the tillage response sequence.

[0007] Based on this, the credibility of sample registration is evaluated by spatial matching degree and deep matching degree, and machine learning training samples are constructed using data that meet the credibility requirements. This helps to reduce the impact of erroneous registration data on the mapping relationship between soil hardness, operation parameters and tillage response, and makes the predicted tillage response closer to the actual operation state, thereby improving the reliability of the determination of the target operation parameter combination and the adaptability to spatial changes in soil hardness. Attached Figure Description

[0008] Figure 1 This is a flowchart of the tillage machine operation parameter optimization method based on machine learning of the present invention; Figure 2 This is a timing diagram of the machine learning-based tillage machine operation parameter optimization method of the present invention. Detailed Implementation

[0009] 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.

[0010] Example 1:

[0011] like Figure 1 and Figure 2 As shown, the present invention provides a technical solution: a method for optimizing tillage machine operation parameters based on machine learning, comprising: S1. Obtain the location data of multiple soil hardness sampling points and conic index profile data to form a depth sequence and hardness feature vector; S2. Continuously acquire the tiller's operation data to form a first positioning sequence, a heading sequence, a first tillage depth sequence, an operation parameter sequence, and a first response sequence; S3. Based on the installation offset and heading sequence of the positioning device and the tillage component, perform coordinate transformation on the first positioning sequence to obtain the second positioning sequence of the tillage component; S4. Construct the sweeping action area based on the second positioning sequence and the effective working width of the tillage component, and determine the effective working interval and spatial matching parameters based on the positional relationship between each sampling point and the sweeping action area. S5. Extract the second tillage depth sequence corresponding to the effective working area, compare it with the depth sequence, determine the depth effect weight according to the proportion of actual tillage depth reaching each sampling depth, and weight the hardness feature vector to obtain the hardness effect contribution vector and depth matching parameters. S6. Perform time compensation on the first response sequence according to the response delay parameter to obtain the second response sequence, and extract the response data corresponding to the effective working area to form a tillage response feature vector. S7. Determine the registration confidence level based on the spatial matching parameters and depth matching parameters. Use the hardness contribution vector with a registration confidence level not less than the preset threshold, the corresponding operation parameters, and the tillage response feature vector to form an effective training sample. Use machine learning algorithms to establish the mapping relationship between soil hardness, operation parameters, and tillage response. S8. Based on the mapping relationship, predict and calculate the soil hardness and candidate operation parameter combinations of the operation area to be optimized, and determine the target operation parameter combination according to the predicted tillage response and the preset optimization target.

[0012] In the above embodiments, this application first establishes training samples that reflect the actual tillage relationship by using soil hardness sampling data and continuous tillage machine operation data. Then, it uses these training samples to establish a machine learning mapping relationship between soil hardness, operation parameters, and tillage response. Finally, it uses the trained machine learning model to predict the possible tillage responses of the area to be optimized under different candidate operation parameters and determines the target operation parameters based on a preset optimization objective. The entire process does not simply involve directly inputting soil hardness data into the machine learning model. Instead, spatial location correction, operation range matching, action depth matching, and response time matching are performed sequentially before model training, ensuring that the soil hardness characteristics, operation parameters, and tillage responses in the training samples correspond to the same actual tillage process.

[0013] In the above implementation, multiple soil hardness sampling points are first selected within the cultivated area, and the spatial location of each sampling point is recorded. A soil cone penetration tester is then used to acquire cone index profile data along the soil depth direction for each sampling point. The cone index reflects the resistance of the soil to cone penetration at a corresponding depth, thus characterizing the hardness of different soil layers. During sampling, the cone index can be recorded at predetermined depth intervals, so that each soil hardness sampling point corresponds to multiple sampling depths arranged from shallow to deep, along with the corresponding cone index for each sampling depth. The sampling depths are arranged in chronological order to form a depth sequence; the corresponding cone indices are arranged in the same order to form a hardness feature vector.

[0014] In the above embodiments, different soil hardness sampling points preferably use the same sampling depth and depth interval to ensure that the hardness features corresponding to different sampling points have a consistent data structure. When the actual sampling depth of different sampling points differs due to detection conditions, the data can be organized according to a unified target sampling depth. For example, if the target sampling depth is located between two actual detection depths, interpolation can be performed based on the conic index corresponding to two adjacent actual detection depths to obtain the conic index corresponding to the target sampling depth. Through the above processing, different soil hardness sampling points form hardness features with the same depth level, providing a foundation for subsequent machine learning models to receive data of a unified dimension.

[0015] In the above embodiment, after soil hardness sampling is completed, the tiller performs continuous tillage according to a predetermined working route, and continuously acquires positioning data, heading data, actual tillage depth data, operating parameters, and tillage response data of the tiller during the actual operation. Positioning data characterizes the position of the positioning device installed on the tiller at each sampling time; heading data characterizes the direction of travel of the tiller at each sampling time; actual tillage depth data characterizes the actual depth of the tillage components entering the soil at the corresponding time; operating parameters characterize the adjustable operating state of the tiller at the corresponding time; and tillage response data characterizes the operating response of the tiller after being subjected to soil action. The operating parameters may include one or more of the following: tiller speed, target tillage depth, tillage component speed, power output state, or other operating parameters that can change the tillage process. The tillage response may include one or more of the following: traction resistance, drive torque, power consumption, fuel consumption, electrical energy consumption, wheel slip ratio, engine load, tillage component load, or other response data that can reflect the tillage effect and mechanical load state. The actual operating parameters and tillage response to be used can be determined based on the type of tiller, the adjustable actuator, and the optimization objectives.

[0016] In the above embodiments, all types of operational data are accompanied by sampling time information, enabling data obtained from different sensors to establish a correspondence based on time. When positioning data, heading data, tillage depth data, operational parameters, and response data have different sampling frequencies, a unified time interval can be selected as a benchmark to resample each data sequence. For data for which there is no direct measurement value at a unified sampling time, values ​​from nearby times can be used or interpolation can be performed based on adjacent measurement values. This forms a first positioning sequence, heading sequence, first tillage depth sequence, operational parameter sequence, and first response sequence that correspond to each other in time.

[0017] Specifically, the first positioning sequence cannot be directly used as the basis for matching soil hardness sampling points with the tillage action location. This is because the positioning device is usually installed on the tiller body, the top of the cab, or other suitable locations for receiving positioning signals, while the tillage component that actually interacts with the soil is usually located behind or to the side of the positioning device. There is a fixed or calibrable installation offset distance between the two. When the tiller's heading changes, the actual offset direction of the same installation offset distance in the geographic coordinate system will also change with the heading. Therefore, this application uses the installation location of the positioning device, the installation location of the tillage component, and the real-time heading of the tiller together for calculating the position of the tillage component.

[0018] Specifically, the distances between the positioning device and the reference position of the tillage component along the longitudinal direction and the lateral direction of the tiller are measured in advance. At each positioning sampling moment, based on the tiller's heading at that moment, the longitudinal and lateral distances are converted from the machine coordinate direction (which rotates with the machine body) to the planar coordinate direction used in the current working area. The converted lateral and longitudinal displacements are then superimposed on the position of the positioning device at that moment, thereby obtaining the actual reference position of the tillage component at the same moment. The above conversion is performed sequentially on each positioning point in the first positioning sequence to form the second positioning sequence.

[0019] Through the aforementioned position transformation, the subsequent location data used for spatial matching reflects the actual locations traversed by the tillage component, rather than the locations traversed by the positioning device. This position transformation algorithm has a direct functional connection with subsequent machine learning training: without performing installation offset and heading correction, soil hardness data located near the positioning device's trajectory but not affected by the tillage component may be incorrectly paired with the tillage response, while soil sampling points actually affected by the tillage component may be missed, thus creating an incorrect correspondence between input and output in the training samples. Therefore, coordinate transformation processing is used to improve the spatial realism of machine learning training samples.

[0020] In this process, a sweeping action zone is constructed based on the second positioning sequence and the effective working width of the tillage component. Since the tillage component has a certain lateral working width, its actual action area is not a zero-width trajectory line, but rather a continuous strip-shaped area formed as the tillage component moves. When constructing the sweeping action zone, the movement trajectory of the tillage component formed by the second positioning sequence can be used as the central trajectory, and the sweeping action zone can be expanded to both sides of the trajectory according to the effective working width, thus forming the spatial area where the tillage component may actually interact with the soil throughout the entire operation.

[0021] In the above implementation, the positional relationship between each soil hardness sampling point and the sweeping action area is determined sequentially. When a soil hardness sampling point is located outside the sweeping action area, it indicates that the soil corresponding to that sampling point is not directly affected by the tillage component in the current tillage trajectory. Therefore, this sampling point is not paired with the response data in the current operation trajectory as training samples. When a soil hardness sampling point is located within the sweeping action area, the time when the tillage component enters the local action area corresponding to the sampling point and the time when it leaves the local action area are further determined based on the process of the tillage component moving along the second positioning sequence. The data interval between the two times is taken as the effective operation interval corresponding to the sampling point. The spatial matching parameter is used to characterize the degree of matching between the soil hardness sampling point and the actual action area of ​​the tillage component in the horizontal direction. Specifically, the shortest distance from the soil hardness sampling point to the center trajectory corresponding to the second positioning sequence can be obtained, and the spatial matching degree can be determined in combination with the effective operation width of the tillage component. When a sampling point approaches the center trajectory of the tillage component, it indicates that it is in a relatively stable area within the tillage component's effective range, corresponding to a high degree of spatial matching. Conversely, when a sampling point approaches the edge of the effective working width, it suggests that the sampling point may only be affected by the edge area of ​​the tillage component, resulting in a corresponding decrease in the degree of spatial matching. In practical processing, the proportion of the shortest distance to half the effective working width can be converted into a spatial matching level or normalized matching value from low to high, allowing the spatial matching parameters to be directly used for subsequent registration reliability calculations.

[0022] In the above implementation, after determining the effective working area, the actual tillage depth data corresponding to the effective working area is extracted from the first tillage depth sequence to form a second tillage depth sequence. The second tillage depth sequence is used to determine the actual soil depth reached by the tillage component during its passage through the current soil hardness sampling point. This process plays an important role in the accurate construction of machine learning input features because soil hardness sampling yields a complete conic index profile from shallow to deep layers. However, in a specific tillage process, the tillage component may not necessarily reach the entire depth of the conic index profile. If the conic index of the entire depth is directly input into the machine learning model with the same weight, the deep hardness that has not been affected by the tillage component will also participate in the interpretation of the tillage response, easily forming a training relationship inconsistent with the actual physical process.

[0023] Therefore, in the above implementation, for each sampling depth in the depth sequence, the actual tillage depth within the effective working interval is statistically analyzed to see if it reaches or exceeds that sampling depth. Specifically, for each actual tillage depth data in the second tillage depth sequence, when the actual tillage depth reaches or exceeds the current sampling depth, that moment is recorded as the current depth layer being actually affected by the tillage component; when the actual tillage depth is less than the current sampling depth, that moment is recorded as the current depth layer not being affected by the tillage component.

[0024] For any given sampling depth, the number of times that depth is actually affected within the effective working range is counted, along with the total number of samplings for all effective tillage depths within the effective working range. The ratio of the former to the latter is then used as the depth effect weight corresponding to that sampling depth. For example, when a certain depth layer is covered by the actual tillage depth throughout the entire effective working range, the effect weight corresponding to that depth layer is assigned a higher value; when a certain depth layer is only reached by the actual tillage depth for a portion of the time, its effect weight is correspondingly reduced; and when the actual tillage depth never reaches a certain depth layer, the effect weight corresponding to that depth layer is assigned a minimum value.

[0025] Subsequently, the conic index corresponding to each sampling depth is adjusted according to the depth-related weight of that sampling depth. Specifically, soil layers with higher weights retain a higher contribution of their conic index to the hardness contribution characteristics; soil layers with lower weights have their corresponding conic index contributions weakened; and soil layers not affected by actual tillage depth are not considered as key information regarding the current tillage response. The hardness data from each sampling depth, processed as described above, are arranged in ascending order to form a hardness contribution vector.

[0026] Through the above processing, the hardness contribution vector does not reflect all the objectively existing hardness at the soil sampling point, but rather the soil hardness information that may actually interact with tillage components under the current actual tillage depth conditions. This deep-effect weight calculation is directly related to the construction of input features for the machine learning model, enabling the model to learn the relationship between soil hardness actually involved in tillage and the tillage response, thereby avoiding interference from deep hardness features that have no actual interaction with the model parameters.

[0027] In the above implementation, a depth matching parameter is also determined based on the depth effect weights corresponding to multiple sampling depths. This depth matching parameter is used to comprehensively evaluate the degree of matching between the current soil hardness profile and the actual tillage process in the depth direction. Specifically, the effect weights corresponding to all sampling depths can be averaged, or a weighted average can be performed after setting higher statistical weights for the main tillage layers. Alternatively, the proportion of depth layers with effective effects can be calculated out of the total number of hardness sampling depth layers. A higher depth matching parameter indicates that a significant portion of the collected soil hardness profile actually participated in the current tillage process; a lower depth matching parameter indicates a larger difference between the hardness profile and the actual tillage depth.

[0028] In the above implementation, the first response sequence also requires response delay compensation. This is because after the tillage component enters a soil area and generates mechanical action, the traction resistance, driving torque, power load, or other responses, after being transmitted through the mechanical structure, detected by sensors, acquired by signals, and processed by the controller, are usually not recorded at exactly the same moment. Therefore, a certain response data in the first response sequence may correspond in time to a soil action process that has already occurred.

[0029] Response delay parameters can be obtained through pre-calibration. For example, by subjecting the tiller to significant load changes during operation, the time it takes for the tilling components to enter the load change zone is compared with the time it takes for the response sensor to detect the significant response change. The time difference between these two times can be used as the delay parameter of the corresponding response signal. Response delays can be calibrated separately for different types of response data. When the response delay is significantly affected by the operating speed, the response delay can also be calibrated separately for different speed ranges.

[0030] During time compensation, based on the response delay parameter, the response data recorded at later times in the first response sequence are shifted to correspond to the actual tillage time that produced the response. For example, if a response signal is known to occur a certain time later than the actual tillage action, the response data is re-correlated to the operation time that is the same time earlier. After processing all the response data in sequence, a second response sequence is formed. The response data in the second response sequence is closer in time to the actual soil action process that produced the response. Response delay compensation works in conjunction with the aforementioned spatial registration. If the effective operating range corresponding to the soil sampling point is determined solely based on the location of the tillage component, and the response data without time delay compensation is directly extracted, the extracted response data may actually originate from the next soil area acted upon after the tillage component has left the current sampling point. Therefore, time delay compensation is used to ensure that the response output in the training samples is consistent with the corresponding soil hardness input in time.

[0031] In the above implementation, after the second response sequence is formed, response data is extracted based on the effective operating range corresponding to each soil hardness sampling point, and a tillage response feature vector is formed. For continuously changing response data, the average level, peak level, minimum level, fluctuation degree, range of change, cumulative response amount, and other statistical characteristics that can reflect the tillage machine load or operating status within the effective operating range can be calculated. When multiple response data are collected simultaneously, corresponding statistical features can be extracted from different response data and combined in a fixed order to form a tillage response feature vector.

[0032] In the above implementation, the registration reliability is determined based on spatial matching parameters and depth matching parameters. Specifically, the contribution ratios of spatial matching and depth matching to the overall reliability can be preset, and the spatial matching parameters and depth matching parameters are weighted according to their respective contribution ratios and then summed to obtain the overall registration reliability. Alternatively, the spatial matching degree and the depth matching degree can be combined by multiplication, so that if the matching degree in either direction is significantly low, the overall registration reliability decreases accordingly.

[0033] The spatial matching parameter constrains whether the soil hardness sampling location is truly within the actual sweeping range of the tillage component, while the depth matching parameter constrains whether the soil hardness sampling profile matches the actual depth of action of the tillage component. Both parameters work together to improve the accuracy of the correspondence between soil hardness information in the training samples and the actual tillage process. Therefore, the spatial matching processing, depth weighting processing, and registration credibility processing are not independent data preparation operations of the machine learning model, but rather participate together in the definition of the training samples and directly affect the correspondence between soil hardness and tillage response learned by the model.

[0034] In the above implementation, the registration confidence level is compared with a preset confidence threshold. When the registration confidence level corresponding to a soil hardness sampling point reaches or exceeds the preset confidence threshold, it is considered that the hardness contribution characteristics, operation parameters, and tillage response corresponding to that sampling point have a sufficiently reliable correspondence, and the three types of data are combined into a set of effective training samples. When the registration confidence level is lower than the preset confidence threshold, this set of data is not used for model training to reduce erroneous samples caused by positioning errors, operation edge effects, insufficient actual tillage depth, or other registration errors.

[0035] In a specific implementation of a machine learning model, a multi-layer feedforward neural network can be used to establish a mapping model between soil hardness, operational parameters and tillage response. This neural network includes at least an input layer, a soil hardness feature processing layer, an operational parameter feature processing layer, a feature fusion layer and a response output layer.

[0036] The input layer includes a soil hardness input terminal and an operational parameter input terminal. The soil hardness input terminal receives the hardness contribution vector corresponding to the current valid training sample, and its number of input nodes corresponds to the number of unified soil sampling depth layers. The operational parameter input terminal receives operational parameters corresponding to the same valid operational interval, and its number of input nodes corresponds to the number of selected operational parameter types. This ensures that the soil hardness and operational parameters entering the model originate from the same spatial location and the same actual operational process. The soil hardness feature processing layer is connected to the soil hardness input terminal and is used to extract soil resistance distribution characteristics from the hardness contribution data of different soil layers. As one possible implementation, the soil hardness feature processing layer can be set with two fully connected layers. The first layer has 64 neurons, and the second layer has 32 neurons. Each layer uses a nonlinear activation function, enabling the model to characterize the combined changes of soil hardness at different depths and the possible nonlinear relationship between them and tillage load.

[0037] The operation parameter feature processing layer is connected to the operation parameter input terminal and is used to extract the operating status features of the current combination of operation parameters. As one possible implementation, the operation parameter feature processing layer can be configured with a fully connected layer of 32 neurons and a fully connected layer of 16 neurons, so that the driving speed, actual or set tillage depth, operating status of tillage components and other operation parameters can be mapped to unified operation status features.

[0038] The feature fusion layer is connected to both the soil hardness feature processing layer and the operational parameter feature processing layer, concatenating the processed soil hardness and operational parameter features to form a joint feature. This joint feature describes the specific tillage state under the current soil hardness conditions and the combination of operational parameters. The concatenated joint feature can be further processed sequentially through fully connected layers with 64 and 32 neurons to learn the interaction between soil hardness and operational parameters. The response output layer is connected to the feature fusion layer and outputs the corresponding predicted tillage response. When only one tillage response is selected as the optimization basis, the response output layer sets the output node corresponding to that response. When simultaneously predicting traction resistance, power consumption, wheel slippage rate, or multiple other responses, the response output layer sets output nodes corresponding to the number of response types, so that each output node corresponds to a tillage response or a pre-constructed response feature.

[0039] Before model training, input data of different dimensions can be normalized. For data on the contribution of soil hardness and operational parameters, scaling can be performed based on the maximum and minimum values ​​in the training samples, or standardization can be performed based on the mean and standard deviation of the training samples. The same scaling transformation can be applied to the response output data. The same data transformation parameters are used in both the model training and actual prediction phases to avoid inconsistencies in the data scale between the training input and the actual application input.

[0040] In the above implementation, all valid training samples can be divided into training samples, validation samples, and test samples according to the proportions of 70%, 15%, and 15%, respectively. Training samples are used to update the internal connection weights and bias parameters of the neural network; validation samples are used to judge the prediction performance of the model on data that are not directly involved in parameter updates, and are used to select training epochs or model structures; test samples are used to evaluate the model's prediction ability for unknown samples after the model training is completed.

[0041] During model training, the hardness contribution features and corresponding operation parameters from a batch of valid training samples are simultaneously input into the neural network to obtain the predicted tillage response output by the model. The predicted tillage response is compared with the actual tillage response from the same valid training sample, and the training loss is calculated based on the difference between the two. For continuous numerical tillage responses, mean squared error can be used as the training loss, resulting in a larger loss for samples where the predicted response differs significantly from the actual response. Subsequently, an adaptive moment estimation optimization method is used to iteratively adjust the internal connection weights and bias parameters of the model.

[0042] As an feasible training parameter configuration, the initial learning rate can be set to 0.001. Each training batch uses 32 or 64 sets of effective training samples. Each training cycle is counted as one training epoch, with a maximum of 500 training epochs. After each training epoch, the validation error is calculated on the validation samples. If the validation error does not decrease further after 20 consecutive training epochs, training can be stopped early, and the model parameters corresponding to the lowest validation error can be saved. This method can reduce the model's overfitting to the training samples. When the number of training samples is small, random forests, gradient boosting decision trees, or support vector regression can also be used to establish the mapping relationship. Regardless of the machine learning algorithm used, the model's input data must include at least the soil hardness characteristics corrected for actual tillage depth and the corresponding operational parameters. The model's supervised output comes from the actual tillage response after response delay compensation and corresponding to the same effective operational interval. This ensures a clear data correspondence between the machine learning algorithm and the physical processes in the tillage field.

[0043] In the above implementation, after model training is completed, test samples can be used to calculate the error between the predicted response and the actual response. Evaluation metrics such as mean absolute error, root mean square error, or coefficient of determination can be used to determine whether the model has reached the preset prediction accuracy. When the model's prediction accuracy does not meet the application requirements, the number of effective training samples can be increased, the number of neurons in the hidden layer of the neural network can be adjusted, the learning rate can be adjusted, the input features can be modified, or a new confidence threshold can be selected, and model training can be performed again.

[0044] In the above implementation, when the trained machine learning model is used in the area to be optimized, the input data processing method for the area to be optimized remains consistent with that used in the model training phase. First, a soil hardness profile of the area to be optimized is obtained and processed according to the same depth level used in the model training phase. When the area to be optimized already has a projected tillage depth or a candidate tillage depth, the expected effect of each soil layer under the corresponding candidate parameter combination is determined based on the candidate tillage depth, and a soil hardness effect characteristic consistent with the model training input structure is formed accordingly.

[0045] Subsequently, multiple candidate operation parameter combinations are generated within the parameter range allowed by the tiller. Each set of parameters in the candidate operation parameter combination, along with the soil hardness characteristics corresponding to the area to be optimized, is input into the trained machine learning model. The soil hardness feature processing part of the model first extracts the hardness characteristics of the soil at different depths in the current area; the operation parameter feature processing part extracts the operating state characteristics corresponding to the current candidate operation parameters; the feature fusion part further calculates the combination relationship between soil hardness conditions and candidate operation parameters; and the response output part outputs the expected tillage response of the candidate operation parameter under the current soil conditions. Therefore, there is a clear domain correlation between the model input and output: the soil hardness input represents the soil resistance conditions that the tillage components will overcome; the operation parameter input represents the operating mode adopted by the tiller for these soil conditions; and the model output represents the mechanical load, energy consumption, or operating state result generated after the combined effect of the above soil conditions and operating mode. This mapping relationship ensures that the machine learning model does not make pure data predictions detached from the tillage process, but rather establishes predictive relationships around the actual action chain composed of soil conditions, operation parameters, and tillage response.

[0046] In the above implementation, after generating predicted tillage responses corresponding to multiple candidate combinations of operating parameters, the prediction results are compared according to preset optimization objectives. When the optimization objective is to reduce energy consumption, the parameter combination with lower predicted energy consumption can be selected from the candidate parameter combinations that meet preset tillage depth or operation quality conditions. When the optimization objective is to reduce tillage load, the parameter combination with lower predicted traction resistance or driving torque that meets the operation requirements can be selected. When energy consumption, operation efficiency, and mechanical load are considered simultaneously, different response indicators can first be converted to a unified evaluation scale, and then corresponding evaluation weights can be set according to the importance of each response indicator. The evaluation results of each indicator are then integrated to obtain a comprehensive evaluation result of the candidate operating parameter combinations, and the candidate operating parameter combination whose comprehensive evaluation result meets the preset optimization objective is selected as the target operating parameter combination.

[0047] In the above implementation, the coordinate transformation between soil hardness sampling points and the trajectory of the tillage component, the judgment of the sweeping action domain, the calculation of the depth action weight corresponding to the actual tillage depth, the response delay compensation, and the registration credibility screening work together to determine the input and output correspondence of the training samples used by the machine learning model. Specifically, coordinate transformation addresses the inconsistency between the positioning device position and the actual tillage position; the sweeping action domain addresses the issue of the tillage component having an actual operating width; the depth action weight addresses the inconsistency between the complete soil hardness profile and the actual tillage depth; the response delay compensation addresses the inconsistency between the soil action time and the detection response time; and the registration credibility further comprehensively filters the spatial and depth matching results. The combined effect of these algorithms ensures that the soil hardness features, operating parameters, and response features ultimately entering the model have a consistent correspondence in spatial location, action depth, and time process. This allows the machine learning model to learn the correlation between soil hardness, operating parameters, and tillage response based on training samples with actual physical meaning, and to optimize the tillage machine's operating parameters accordingly.

[0048] Example 2:

[0049] In this embodiment, S1 includes: The spatial coordinates and cone index profile data of each soil hardness sampling point were obtained, and the cone indices were sorted according to the sampling depth to form an initial hardness sequence and a depth sequence. The initial hardness sequence is resampled according to a preset depth interval to obtain a standard hardness sequence. Invalid data are identified and neighborhood correction is performed based on the effective range and repeated sampling consistency threshold of the conic index acquisition device, while hardness mutation data within the effective range are retained. Calculate the rate of change of the conic index at adjacent depths in the standard hardness sequence to form a hardness gradient sequence, and combine it with the standard hardness sequence to form a hardness feature matrix. The hardness feature matrix is ​​encoded in depth order to form a hardness feature vector associated with the spatial coordinates of the sampling points.

[0050] In the above implementation, the spatial coordinates of each soil hardness sampling point are recorded, and a cone index acquisition device is continuously inserted into the soil along the depth direction, simultaneously recording the penetration depth and the corresponding cone index. After data acquisition, the cone indices are arranged from shallowest to deep according to their corresponding sampling depth, using the penetration depth as the sorting criterion. Sampling depths and cone indices at the same sorting positions are saved separately, thus forming an initial hardness sequence and a depth sequence. For duplicate detection values ​​at the same sampling depth, the duplicate detection values ​​are retained until the subsequent consistency determination step and are not directly deleted during the sorting stage.

[0051] Because the actual data acquisition location during continuous penetration is affected by sampling frequency, penetration speed, and depth detection resolution, the actual sampling depth corresponding to each soil hardness sampling point is not completely consistent. Therefore, the initial hardness sequence is resampled according to a preset depth interval. In this embodiment, the starting and ending depths of resampling are first determined based on the effective operating depth range of the tillage component, and then standard sampling depths are generated within this depth range according to fixed depth intervals. The preset depth interval is taken as an integer multiple of the depth detection resolution of the conic index acquisition device and is less than the minimum soil layer thickness to be identified, so that the variation information of soil hardness along the depth direction can be retained between two adjacent standard sampling depths.

[0052] When a standard sampling depth coincides with an original sampling depth, the conic index corresponding to that original sampling depth is directly taken. When the standard sampling depth lies between two adjacent original sampling depths, the distances between the standard sampling depth and both the shallower and deeper original sampling depths are first calculated, and then the sum of these two distances is calculated. The conic index at the shallower original sampling depth is weighted by the proportion of the distance from the standard sampling depth to the deeper original sampling depth to the sum of these distances, and the conic index at the deeper original sampling depth is weighted by the same proportion. The two weighted results are then added together, and the resulting value is taken as the conic index corresponding to that standard sampling depth. The conic indices corresponding to all standard sampling depths are obtained sequentially in the same manner and arranged from shallowest to deepest to obtain a standard hardness sequence. This ensures that different soil hardness sampling points correspond to the same soil depth at the same sequence position.

[0053] Before performing validity processing on the standard hardness sequence, the effective range of the cone index acquisition device should be determined first.

[0054] In this embodiment, the cone index acquisition device is calibrated before field sampling. Multiple standard load points, ranging from low to high, are set within the device's rated range. Each standard load point is tested 10 times consecutively, and the test results are compared with the corresponding benchmark values. Using the indication error limit specified in the force sensor calibration document as the allowable error, within the load range where consecutive test results do not exceed the allowable error, the corresponding lowest and highest cone indices are determined as the lower and upper limits of the effective range, respectively. Data in the standard hardness sequence that are below the lower limit or above the upper limit of the effective range are marked as invalid data.

[0055] The consistency threshold for repeated sampling is determined through repeatability tests before sampling. At least 20 sampling locations are selected in a test area with soil type consistent with the area to be sampled. Three consecutive cone penetration tests are performed at each sampling location, and the results of the three tests are resampled at the same depth intervals as the initial sampling. For each standard sampling depth, the average of the three cone indices is calculated first, and then the absolute difference between each cone indices and the average is calculated. The largest absolute difference is divided by the average to obtain the relative deviation of repeated sampling at that depth. The relative deviations of repeated sampling obtained from all test locations and all standard sampling depths are summarized, sorted from smallest to largest, and the upper limit covering 95% of the data is taken as the consistency threshold for repeated sampling. Thus, the consistency threshold is jointly determined by the repeated detection error of the actual device and the current soil conditions.

[0056] Specifically, during the formal sampling process, three repeated penetration tests were performed at each soil hardness sampling point, and each test was resampled to the same standard sampling depth. For the same standard sampling depth, the average of the three conic indices and the absolute deviation of each conic indice relative to the average were calculated, and the ratio of the maximum absolute deviation to the average was taken as the relative deviation of the repeated sampling at that depth. When the relative deviation was not greater than the repeated sampling consistency threshold, the repeated sampling results at that depth were considered consistent, and the arithmetic mean of the three conic indices was taken as the valid conic indices at that depth. When the relative deviation was greater than the repeated sampling consistency threshold, the three test results were further compared, and the test value with the largest deviation from the other two test results was deleted; if the relative deviation between the remaining two test results was not greater than the repeated sampling consistency threshold, the average of the remaining two test results was taken as the valid conic indices at that depth; otherwise, the data at that depth was marked as invalid data.

[0057] For data within the effective measurement range and that pass the consistency test for repeated sampling, even if there are significant differences between the conic index and that of adjacent depths, they are retained without smoothing or deletion based on the amplitude of adjacent data. This processing ensures that sudden increases or decreases in hardness that occur in consecutive repeated measurements are retained in the standard hardness sequence, thereby preventing the true hardness changes at the soil compaction layer or the interface between different soil layers from being treated as outliers.

[0058] Specifically, neighborhood correction is applied to identified invalid data. When both the preceding and following standard sampling depths of a single invalid data point have valid conic indices, linear correction is performed based on the distances from the invalid data's depth to both valid sampling depths. Specifically, the distances between the invalid data's depth and both the preceding and following valid sampling depths are calculated. Based on the proportion of each distance in the total distance, the two valid conic indices are weighted inversely, and the two weighted results are summed. The resulting value is used as the correction value for the invalid data. Valid conic indices closer to the invalid depth have a larger weight in the correction result.

[0059] When two or more consecutive invalid data points appear, the nearest valid conic index is searched both towards the shallower and deeper layers from the invalid region. Then, linear interpolation is performed point-by-point based on the distance from each standard sampling depth within the invalid region to the valid data points on both sides. For data located at the beginning or end of the standard hardness sequence and having only a single-sided valid neighborhood, the average of the three valid conic indices closest to the invalid location is calculated, and this average is used as a correction value. After completing the above processing, the standard hardness sequence used for feature extraction is obtained.

[0060] The rate of change of the conic index between adjacent depths is then calculated. For each depth position starting from the second standard sampling depth, the conic index of the current depth is subtracted from the conic index of the previous standard sampling depth to obtain the change in the conic index between two adjacent depths. This change is then divided by the actual depth interval between the two standard sampling depths, and the result is taken as the rate of change of the conic index corresponding to the current depth. The rate of change of the conic index is positive as the depth increases and negative as the depth decreases. Since there is no data for the previous depth at the first standard sampling depth, its rate of change is set to 0. The rates of change are arranged from shallow to deep to form a hardness gradient sequence.

[0061] The standard hardness sequence and hardness gradient sequence are combined according to the same standard sampling depth to form a hardness feature matrix. Each row in the hardness feature matrix corresponds to one standard sampling depth. The first column records the conic index corresponding to that depth, and the second column records the rate of change of the conic index corresponding to that depth. Thus, each depth location simultaneously retains the soil hardness value at that location and the degree of hardness change relative to the previous depth location.

[0062] Specifically, after completing the hardness feature matrix, encoding is performed from shallow to deep according to the standard sampling depth. During encoding, the conic index and rate of change of the conic index corresponding to the first standard sampling depth are read first, then the conic index and rate of change of the conic index corresponding to the second standard sampling depth are read, and so on, until the last standard sampling depth. These are then arranged consecutively according to the reading order to form a hardness feature vector. All soil hardness sampling points use the same standard depth, the same feature arrangement, and the same encoding order; therefore, data at the same position in the hardness feature vector has the same soil layer meaning across different sampling points. Each hardness feature vector is bound and stored to the spatial coordinates of the soil hardness sampling point that generated it. When performing sweeping domain matching, the correspondence between the sampling point and the actual operating trajectory of the tillage component is first determined based on these spatial coordinates, and then the hardness feature vector associated with these spatial coordinates is called. Subsequently, when determining the depth-based weighting based on the actual tillage depth, the degree of influence of the corresponding hardness features is determined layer by layer according to the standard sampling depth. This ensures that the absolute value of the conic index, the variation characteristics of the conic index along the depth direction, and the actual tillage depth form a consistent depth correspondence. The processing results are then used as input data in the machine learning model to reflect the soil hardness status.

[0063] Example 3:

[0064] Specifically, S2 includes: Acquire positioning data, heading data, actual tillage depth data, operation parameter data, and tillage response data according to the corresponding sampling frequency, and add time markers; A unified timeline is constructed based on time identifiers, mapping data with different sampling frequencies to corresponding time nodes; Based on the effective range and data continuity threshold of each data acquisition device, missing or invalid data is identified and compensated using adjacent valid data. Construct a multi-source operation data matrix according to a unified timeline, and extract the corresponding data columns respectively; The first positioning sequence, heading sequence, first tillage depth sequence, operation parameter sequence, and first response sequence are formed in chronological order.

[0065] In the above embodiments, positioning data, heading data, actual tillage depth data, operation parameter data, and tillage response data are acquired by corresponding data acquisition devices. Positioning data is acquired by the positioning device, heading data by the heading sensor, actual tillage depth data by the tillage depth detection device, operation parameter data by the tillage machine controller, and tillage response data by the response sensor installed on the power output end, traction mechanism, or tillage component. Each data acquisition device operates according to its set sampling frequency, and each time a set of data is acquired, it is simultaneously written to the current sampling time. The sampling time is uniformly based on the time reference output by the data acquisition controller.

[0066] Before formal data collection, the clocks of all data acquisition devices are synchronized. The data acquisition controller sends a time synchronization signal to each acquisition device at the same moment, and each acquisition device uses the moment it receives the time synchronization signal as a common timing starting point. For data with its own output timestamp, its original timestamp is converted to the timing reference of the data acquisition controller. After synchronization is completed, each set of positioning data, heading data, actual tillage depth data, operation parameter data, and tillage response data has a definite time identifier.

[0067] Specifically, when establishing a unified timeline based on the time identifiers of each data point, the time interval between two adjacent time identifiers under continuous normal sampling conditions of each data source is first calculated. Then, the maximum value among the normal sampling intervals of each data source is taken as the time step of the unified timeline. The starting time of the unified timeline is the earliest common time after all data types have produced their first valid value, and the ending time is the latest common time before all data types still have valid values. Starting from the starting time, one time node is set for each time step, continuously set until the ending time, thereby forming the unified timeline.

[0068] When mapping data from different sampling frequencies to a unified time node, the time signatures of the original data are first compared with the current time node. If the time signature of the original data matches the current time node, the data is directly written to the current time node. If no data with completely matching time signatures exists, the nearest valid data before and after the current time node is found, and the mapping is completed based on the position of the current time node between the two sampling times.

[0069] For location coordinates, actual tillage depth, and continuously changing tillage response data, first calculate the time length from the current time node to the previous valid data point and the time length from the next valid data point, then calculate the total time length between the two valid data points. The closer the current time node is to the previous valid data point, the greater the contribution of the previous valid data point to the current time node's value; the closer it is to the next valid data point, the greater the contribution of the next valid data point's value. In specific processing, the proportion of the time length from the current time node to the next valid data point to the total time length is used as the weight of the previous valid data point, and the proportion of the time length from the current time node to the previous valid data point to the total time length is used as the weight of the next valid data point. The two valid data points are weighted separately and then summed. The result is used as the data value of the current time node.

[0070] The heading data is mapped according to the direction of the minimum turning angle. First, the two heading angles before and after the current time point are obtained, and the angle differences between the two heading angles in the clockwise and counterclockwise directions are calculated respectively. The angle difference with the smaller absolute value is selected as the actual heading change. Then, based on the proportion of time occupied by the current time point between the two sampling times before and after it, the actual heading change is allocated to the previous heading angle to obtain the heading angle corresponding to the current time point. When the heading data before and after are located at positions close to 360 degrees and close to 0 degrees respectively, the calculation is based on the smaller angle difference across 0 degrees to avoid abrupt changes in the interpolation results due to the periodicity of the heading angle.

[0071] For operational parameter data, mapping is performed according to parameter attributes. Continuously adjusted driving speed, RPM, and power output parameters are processed using the aforementioned time-weighted method for adjacent valid data. Setting parameters issued by the controller and remaining unchanged until the next control command arrives are used, with the actual effective time of the parameter command as the dividing line; the current time node uses the most recently effective parameter value. This ensures that the operational parameters on a unified time axis are consistent with the actual operating state of the tiller at the same time.

[0072] After all data mappings are completed, invalid data is identified based on the effective range of the corresponding acquisition device. The effective range is determined by the calibration results of the acquisition device. During calibration, at least five standard detection points are evenly set within the rated measurement range of the acquisition device, and each standard detection point is continuously measured 10 times. Each detection result is compared with the standard detection value. When all 10 detection results are within the allowable error range specified by the acquisition device, the standard detection point is recorded as a valid detection point. The lowest and highest measured values ​​covered by consecutive valid detection points are respectively used as the lower and upper limits of the effective range of the acquisition device. During formal operation, data below the lower limit or above the upper limit of the effective range are marked as invalid data.

[0073] Specifically, the data continuity threshold is determined using pre-sampled data from each acquisition device under normal operating conditions. Before formal operation, each data acquisition device continuously collects no less than 1000 sets of data, and the interval between the time markers of two adjacent sets of data is calculated in chronological order. First, all time intervals are added together and divided by the number of time intervals to obtain the average normal sampling interval. Then, the difference between each sampling interval and the average value is calculated, and the square of the squared difference is taken to obtain the discrete value of the sampling interval. The average normal sampling interval is added to three times the discrete value, and the resulting time length is used as the data continuity threshold for the corresponding data acquisition device.

[0074] During the formal operation, if the time interval between two adjacent sets of raw data is no greater than the data continuity threshold, the data collection for that period is considered continuous; if the time interval between two adjacent sets of raw data is greater than the data continuity threshold, data gaps are considered to exist between the two sets of data. If no valid mapped data is obtained at a unified time node, or if the mapped data exceeds the effective range of the corresponding acquisition device, it is also treated as data to be compensated. When processing the data to be compensated, the nearest valid data point is searched backwards from the time node to be compensated, and then the nearest valid data point is searched forwards. When valid data exists on both sides, the time length from the time node to be compensated to the valid data points before and after it is calculated, and then compensation is performed according to the aforementioned reverse weighting method based on time distance. Valid data closer to the time node to be compensated has a higher weight, and valid data farther away has a lower weight. Positioning coordinates, actual tillage depth, and continuous tillage response data are all compensated in this way.

[0075] When heading data is missing, the effective heading angles before and after the missing node are first obtained. The actual change between the two effective heading angles is determined according to the direction of the minimum turning angle. Then, the data is allocated according to the time position of the missing node between the preceding and following effective data. When maintenance parameters are missing, instead of averaging the preceding and following values, the most recent effective parameter value before the missing node is read and carried over to the time when the next parameter command takes effect. After compensation is completed, the effective range is checked again on the compensation result. If the compensation result is within the effective range, it is retained; if the compensation result exceeds the upper limit of the effective range, it is corrected to the upper limit of the effective range; if it is below the lower limit of the effective range, it is corrected to the lower limit of the effective range. This avoids interpolation compensation generating data that exceeds the physical measurement range of the acquisition device.

[0076] After all unified time nodes have been processed, a multi-source operation data matrix is ​​constructed according to a unified timeline. Each time node corresponds to one row in the matrix, and the same row contains the time identifier, positioning coordinates, heading data, actual tillage depth data, operation parameter data, and tillage response data for that time node in sequence. The two directional coordinates of the positioning coordinates are set up with independent data columns; when there are multiple operation parameters or multiple tillage responses, each parameter and each response occupies an independent data column. All rows of the matrix use the exact same data arrangement order, ensuring that the same column always corresponds to the same type of operation data.

[0077] Subsequently, data is extracted from the multi-source operation data matrix according to the data column attributes. Positioning data corresponding to each time node is read sequentially and arranged chronologically to form the first positioning sequence; heading data is extracted to form the heading sequence; actual tillage depth data is extracted to form the first tillage depth sequence; operation parameter data is extracted to form the operation parameter sequence; and tillage response data is extracted to form the first response sequence. Since all the above sequences are extracted from the same multi-source operation data matrix along the same time axis, data at the same time position in different sequences correspond to the same actual operation time. After the above processing, the first positioning sequence and heading sequence can be directly used for subsequent coordinate transformation from the positioning device position to the tillage component position; the first tillage depth sequence can extract corresponding tillage depth data according to the determined effective operation range; the first response sequence can implement time compensation based on the response delay parameter; and the operation parameter sequence can establish a correspondence with the soil action state and tillage response at the same operation time, thereby providing time-consistent data for the construction of subsequent effective training samples.

[0078] Example 4:

[0079] Specifically, S3 includes: An installation offset vector is constructed based on the installation distance and direction between the positioning device and the tillage component; Extract the heading angles at each time point from the heading sequence, and construct the corresponding rotation matrix based on the heading angles; By using the rotation matrix transformation to install the offset vector, the spatial offset vector corresponding to each time node is obtained; The spatial offset vector is superimposed with the corresponding position vector in the first positioning sequence, and invalid positions are corrected based on the displacement of adjacent positions and a preset displacement threshold. Arrange the corrected position vectors in chronological order to form a second positioning sequence that represents the actual movement trajectory of the tillage component.

[0080] In the above embodiment, the installation distance between the positioning reference point of the positioning device is measured, and the center position of the effective working width of the tillage component is measured, with the positioning reference point of the positioning device as the starting point and the center position of the tillage component as the ending point. The installation angle from the longitudinal forward direction of the tiller to the ending point direction is also measured. A machine body plane coordinate system is established that moves with the tiller, with the longitudinal forward direction as the first coordinate axis and the lateral direction perpendicular to the longitudinal forward direction as the second coordinate axis, and forward and rightward are defined as positive directions.

[0081] When constructing the installation offset vector, the installation distance is multiplied by the cosine of the installation angle to obtain the longitudinal offset between the positioning device and the tillage component along the first coordinate axis; the installation distance is multiplied by the sine of the installation angle to obtain the lateral offset along the second coordinate axis. When the tillage component is located behind the positioning device, the longitudinal offset is recorded as a negative value; when it is located to the left of the positioning device, the lateral offset is recorded as a negative value. The installation offset vector is formed by arranging the longitudinal offset first, followed by the lateral offset. The installation distance and installation angle are determined by actual installation dimensions before the tiller is put into operation. When the installation positions of the same positioning device and tillage component remain unchanged, the same installation offset vector is used throughout the entire operation.

[0082] Specifically, from the heading sequence formed above, the heading angles at each time point are read sequentially according to a unified time axis. When the heading angle output by the positioning device uses true north as 0 degrees and increases clockwise, the heading angle is first converted to the angle reference used by the coordinate system of the working area, so that the converted heading angle represents the angle from the positive direction of the first coordinate axis to the longitudinal forward direction of the tiller. Before participating in trigonometric function calculations, the angle is divided by 180 and then multiplied by pi to obtain the corresponding radian value.

[0083] For each time point, a 2x2 rotation matrix is ​​constructed based on the transformed heading angle. The first value in the first row of the rotation matrix is ​​the cosine of the heading angle, and the second value in the first row is the negative of the sine of the heading angle; the first value in the second row is the sine of the heading angle, and the second value in the second row is the cosine of the heading angle. This transforms the installation offset vector from the body coordinate system that rotates with the vehicle body to the fixed-plane coordinate system used in the first positioning sequence.

[0084] During the offset transformation, the two values ​​in the first row of the rotation matrix are multiplied by the longitudinal and lateral offsets in the installation offset vector, respectively, and then the two products are added together. The result is the spatial offset of the current time node along the first coordinate axis of the fixed plane coordinate system. The two values ​​in the second row of the rotation matrix are multiplied by the longitudinal and lateral offsets, respectively, and then the two products are added together to obtain the spatial offset along the second coordinate axis. These spatial offsets are arranged in the order of first coordinate axis spatial offset first, then second coordinate axis spatial offset, forming the spatial offset vector for the current time node. When the heading changes, the above transformation is re-executed according to the new heading angle to obtain the spatial offset vectors corresponding to each time node.

[0085] After obtaining the spatial offset vector, the position vector of the positioning device at the same time node is read from the first positioning sequence. The position coordinate of the first coordinate axis of the positioning device is added to the offset of the first coordinate axis of the spatial offset vector, and the position coordinate of the second coordinate axis of the positioning device is added to the offset of the second coordinate axis of the spatial offset vector. The resulting two coordinates constitute the candidate position vector of the tillage component at that time node. All position data in the first positioning sequence are processed in the same way to convert the positioning trajectory of the positioning device into the candidate motion trajectory corresponding to the reference position of the tillage component.

[0086] When validating candidate position vectors, the displacement between two adjacent candidate positions is calculated in chronological order. First, the first coordinate value of the current candidate position is subtracted from the first coordinate value of the previous candidate position. Then, the second coordinate value of the current candidate position is subtracted from the second coordinate value of the previous candidate position. The squares of the two coordinate differences are then added together, and the square root of the sum is taken. The resulting distance is the actual displacement between two adjacent time points. The preset displacement threshold is determined by the maximum displacement that the tiller can produce under normal operating conditions and the positioning error of the positioning device. First, the maximum operating speed limited by the tiller controller is read, and this speed is multiplied by the time length between two adjacent time points to obtain the maximum theoretical displacement allowed for normal operation of the tiller within that time period.

[0087] The positioning error is calibrated before formal operation. The positioning device is fixed at a reference position with known planar coordinates, and at least 300 sets of positioning coordinates are continuously acquired under stable positioning signal conditions. For each set of positioning coordinates, the difference between its two coordinate values ​​and the corresponding coordinate values ​​of the reference position is calculated. The two differences are squared, added together, and the square root is taken to obtain the planar position error of that positioning. The 300 sets of planar position errors are arranged in ascending order, and the largest error corresponding to 95% of the detection data after arrangement is taken as the positioning error limit.

[0088] Specifically, the maximum theoretical displacement corresponding to adjacent time nodes is added to twice the positioning error limit, and the resulting distance is used as the preset displacement threshold between the two time nodes. The reason for using twice the positioning error limit is that there are positioning deviations between two adjacent positioning points, and the error range of adjacent displacements is simultaneously affected by both the preceding and following positioning points.

[0089] When the actual displacement between adjacent candidate positions is not greater than the corresponding preset displacement threshold, the current candidate position is determined as a valid position; when the actual displacement is greater than the preset displacement threshold, the current candidate position is marked as an invalid position. After an invalid position occurs, the search continues backward based on the previous valid position, and the corresponding maximum theoretical displacement and displacement threshold are recalculated according to the actual time interval between the previous valid position and the position to be detected, until the next valid position whose displacement meets the corresponding threshold is found. For invalid positions located between two valid positions, position correction is performed according to the time ratio. First, the time length between the time node of the invalid position and the previous valid position is calculated, and then the total time length between the two valid positions is calculated. The former is divided by the latter to obtain the time ratio of the invalid time node in this time interval. Then, the coordinate difference between the next valid position and the previous valid position on the two coordinate axes is calculated respectively. Each coordinate difference is multiplied by the time ratio, and then added to the corresponding coordinate of the previous valid position. The two coordinate values ​​obtained are used as the correction result of the invalid position. When multiple invalid positions exist consecutively, the time ratio is calculated according to their respective time nodes, and the correction is completed point by point.

[0090] Specifically, for invalid positions located at the end of the sequence and lacking a subsequent valid position, the direction of movement is determined by the two most recent valid positions before the abnormal segment at the end. The coordinate changes of these two valid positions on the two coordinate axes and the time length between the two positions are calculated respectively. The coordinate changes are divided by the time length to obtain the displacement per unit time in the two coordinate directions. These displacements are then multiplied by the time length between the position to be corrected and the nearest valid position, and then superimposed on the coordinates of the nearest valid position to obtain the corrected coordinates of the invalid position at the end. The obtained corrected position is still checked according to a preset displacement threshold for the corresponding time interval, and is retained if the displacement threshold is met.

[0091] After all position corrections are completed, the position vectors corresponding to each time node are arranged from first to last along a unified timeline to form a second positioning sequence. The position data in the second positioning sequence represents the actual movement position of the tillage component reference point and corresponds to the heading data, actual tillage depth data, operational parameter data, and tillage response data at the same time node. When constructing the sweeping action area based on the second positioning sequence, the trajectory of the tillage component after installation position compensation and abnormal displacement correction is used, thereby avoiding position mismatch between soil hardness sampling points and the actual tillage area caused by installation position deviation of the positioning device and instantaneous positioning jumps.

[0092] Example 5:

[0093] S4 includes: Connect adjacent position vectors in the second positioning sequence according to time order to form a continuous operation trajectory segment; The effective working width of the tillage component is extended laterally along each working trajectory segment to form a continuous sweeping action area; Calculate the lateral distance from each soil hardness sampling point to the corresponding operation trajectory segment, and form a spatial matching sequence based on the relationship between the lateral distance and the effective operation half-width; Extract time nodes that are consecutively within the sweeping domain from the spatial matching sequence to determine the effective working interval; Spatial weights are determined based on the ratio of each lateral distance within the effective work area to the effective work half-width, and then aggregated to obtain spatial matching parameters.

[0094] In the above implementation, each position vector is read sequentially according to the time identifier of the second positioning sequence. The position vector of the previous time node is connected with the position vector of the next time node to form a work trajectory segment. The start time of the work trajectory segment is recorded as the time corresponding to the previous time node, and the end time is recorded as the time corresponding to the next time node. The same processing is performed on adjacent positions in the second positioning sequence until the last position vector, thereby forming work trajectory segments arranged continuously according to the actual work time.

[0095] The effective working width of the tillage component is calibrated before formal operation. The tillage component completes at least 10 straight, stable tillage operations at the set depth. Within the disturbance zone created by each operation, a cross-section perpendicular to the direction of travel is selected, and the width of the continuously tilled portion of this cross-section is measured. The measured widths are summed and divided by the number of measurements; the average value is taken as the effective working width. The effective working width is then divided by 2 to obtain the effective working half-width.

[0096] Specifically, for each work trajectory segment, first subtract the first coordinate of the starting point from the first coordinate of the endpoint to obtain the displacement in the first coordinate direction. Then, subtract the second coordinate of the starting point from the second coordinate of the endpoint to obtain the displacement in the second coordinate direction. Square the two displacements respectively, add them together, and take the square root to obtain the length of the trajectory segment. Then, divide the two displacements by the length of the trajectory segment to obtain the two unit directional components along the trajectory travel direction.

[0097] Based on the stated direction of travel, the two unit directional components are converted into lateral unit directions by rotating them 90 degrees. Moving one effective working half-width along both the forward and reverse directions of the lateral unit direction yields the boundary positions on both sides of the trajectory segment. The same expansion is performed on all working trajectory segments, with the expansion areas of adjacent trajectory segments continuously connected in chronological order, thus forming the sweeping action domain corresponding to the actual area traversed by the tillage component.

[0098] When calculating the lateral distance from a soil hardness sampling point to the operational trajectory segment, the two coordinates of the sampling point are subtracted from the corresponding coordinates of the trajectory segment's starting point to obtain the two coordinate differences between the sampling point and the trajectory segment's starting point. The first coordinate difference is multiplied by the first direction component of the trajectory segment, and the second coordinate difference is multiplied by the second direction component of the trajectory segment. The two products are then added together. The result represents the sampling point's position relative to the starting point along the trajectory direction. When the result is not less than 0 and not greater than the length of the trajectory segment, it indicates that the foot of the perpendicular from the sampling point to the starting point of the trajectory segment is located between the starting and ending points of the trajectory segment. In this case, the two coordinate differences between the sampling point and the trajectory segment's starting point are multiplied by their corresponding lateral unit direction components, and the two products are added together. The absolute value of the result is then taken to obtain the lateral distance from the sampling point to the current operational trajectory segment.

[0099] When the position of a sampling point along the trajectory direction is less than 0, it indicates that its perpendicular foot is outside the starting point of the trajectory segment, and the current trajectory segment is not considered as the orthogonal corresponding trajectory segment for that sampling point. When the position is greater than the length of the trajectory segment, it indicates that its perpendicular foot is outside the ending point of the trajectory segment, and the lateral distance of that trajectory segment is also not used as the effective matching distance. For sampling points at the boundary of adjacent trajectory segments, the effective lateral distance to the preceding and following adjacent trajectory segments is calculated, and the smaller value is taken as the lateral distance corresponding to that time position to avoid incorrect matching in turning areas due to trajectory extension lines. After obtaining the lateral distance, it is directly compared with the effective working half-width. When the lateral distance is not greater than the effective working half-width, it is determined that the sampling point is located within the sweeping action area corresponding to the working trajectory segment; when the lateral distance is greater than the effective working half-width, it is determined that the sampling point is located outside the sweeping action area. According to the time sequence of the working trajectory segments, the lateral distance of each time position and the corresponding in-domain or out-of-domain state are recorded sequentially to form the spatial matching sequence of the soil hardness sampling point.

[0100] When determining the effective work interval from the spatial matching sequence, the search starts continuously from the earliest node. When the current node is outside the sweep scope or enters the sweep scope, the current node is determined as the starting node of the effective work interval. The search continues forward until the first node outside the sweep scope appears; the node most recently inside the sweep scope before that node is determined as the ending node. The continuous time range from the starting node to the ending node constitutes one effective work interval.

[0101] When a single soil hardness sampling point is divided into multiple separate sections due to repeated sampling operations, each consecutive section forms an independent effective operating range, and time periods outside the sweeping area are not merged. Therefore, subsequent actual tillage depth, operational parameters, and tillage response are extracted separately according to the time range corresponding to each actual tillage operation.

[0102] After determining the effective working range, the spatial weight is calculated for each time node within the range. First, the lateral distance corresponding to the time node is divided by the effective working half-width to obtain the lateral deviation ratio. Then, the lateral deviation ratio is subtracted from 1, and the result is used as the spatial weight of that time node. When the lateral distance is 0, the sampling point is located at the center of the working trajectory, and the spatial weight is 1; when the lateral distance is equal to the effective working half-width, the sampling point is located at the boundary of the sweeping action domain, and the spatial weight is 0; when the lateral distance is between the two, the spatial weight decreases as the lateral distance increases.

[0103] Spatial matching parameters are obtained by aggregating all spatial weights within the effective operating interval. When adjacent time intervals on a unified time axis are consistent, all spatial weights within the effective operating interval are summed and then divided by the number of effective time nodes contained in that interval; the average value is used as the spatial matching parameter. When the time intervals between adjacent time nodes are inconsistent, the spatial weight of each time node is multiplied by the time length from that node to the next node to obtain the corresponding time-weighted result; all time-weighted results within the interval are summed and then divided by the total duration of the effective operating interval; the result is used as the spatial matching parameter. The spatial matching parameter also reflects whether the soil hardness sampling point is within the actual sweeping range of the tillage component and the degree to which it deviates from the center trajectory within that range. This parameter is further used to determine the registration reliability mentioned above, while the effective operating interval serves as the time range for subsequently extracting the second tillage depth sequence, corresponding operating parameters, and tillage response data.

[0104] Example 6:

[0105] Specifically, S5 includes: The second tillage depth sequence is extracted from the first tillage depth sequence based on the effective operating range; The actual tillage depth in the second tillage depth sequence is compared with each sampling depth in the depth sequence to construct a depth coverage matrix that characterizes whether the corresponding sampling depth has been reached at each time node. The depth effect weight is determined based on the effective coverage ratio corresponding to each sampling depth in the depth coverage matrix, forming a depth effect weight vector. A depth weight matrix is ​​constructed with the weights of each depth as diagonal elements, and the hardness feature vector is reconstructed into a matrix according to the sampling depth. The hardness contribution vector is obtained by weighting and expanding the depth weight matrix. The depth matching parameters are obtained by aggregating the depth effects based on the weights of each depth and the degree of continuous coverage.

[0106] In the above implementation, based on the determined effective operating interval, actual tillage depth data within the corresponding time range are extracted from the first tillage depth sequence. Specifically, using the start and end time nodes of the effective operating interval as boundaries, all actual tillage depth data with time markers between the two boundaries are retained and arranged in chronological order to form a second tillage depth sequence. Each actual tillage depth in the second tillage depth sequence corresponds to one time node within the effective operating interval.

[0107] After obtaining the second tillage depth sequence, the actual tillage depth at each time point is compared with each sampled depth in the depth sequence. A depth coverage matrix is ​​constructed using the time points within the effective working interval as rows and the sampled depths in the depth sequence as columns. For any given time point, if the actual tillage depth at that time point is greater than or equal to a certain sampled depth, the position corresponding to that sampling depth at that time point is marked as 1; if the actual tillage depth is less than that sampling depth, the corresponding position is marked as 0. After comparing all time points with all sampled depths, the depth coverage matrix is ​​obtained.

[0108] Each column in the depth coverage matrix corresponds to a fixed sampling depth. Therefore, the effective coverage ratio of a sampling depth is determined by statistically analyzing each column. For any column, the number of time nodes with a value of 1 is accumulated, and this number is then divided by the total number of effective time nodes corresponding to that column. The result is taken as the effective coverage ratio of that sampling depth, and the depth effect weight is determined accordingly. When all time nodes reach the sampling depth, the depth effect weight is 1; when none of the time nodes reach the sampling depth, the depth effect weight is 0; when only some time nodes reach the sampling depth, the depth effect weight is the ratio of the number of covered time nodes to the total number of effective time nodes. The depth effect weights are arranged from shallowest to deepest sampling depth to form a depth effect weight vector.

[0109] Specifically, for time nodes where only a portion of the unified sampling period is covered at the beginning or end of the effective working interval, the calculation is based on the actual duration of that time node. Specifically, the actual duration of that node within the effective working interval is divided by the unified sampling period, and the resulting proportion is used as the effective count for that node's participation in coverage statistics. All other complete time nodes are counted as 1, thus ensuring that the depth-based weighting corresponds to the actual coverage duration.

[0110] After forming the depth-based weight vector, a depth-based weight matrix is ​​constructed according to the order of the sampling depths. The number of rows and columns in the matrix equals the number of sampling depths. The first row and first column contain the depth-based weight corresponding to the first sampling depth, the second row and second column contain the depth-based weight corresponding to the second sampling depth, and so on until the last sampling depth. Except for the main diagonal positions, all other positions are set to 0. This ensures that each depth-based weight only applies to the hardness feature of the corresponding sampling depth, without changing the original arrangement relationship between different depths.

[0111] When reconstructing the hardness feature vector formed above into a matrix, the hardness features corresponding to each sampling depth are restored according to their original encoding order. Each sampling depth sequentially records the standard conic index and the corresponding hardness gradient. Therefore, starting from the beginning of the hardness feature vector, every two consecutive feature values ​​are read as one row. The first feature value is the standard conic index of that sampling depth, and the second feature value is the hardness gradient of that sampling depth. These are arranged row by row according to the sampling depth from shallow to deep, forming a hardness feature matrix. When using the depth weight matrix for weighting, the depth-based weight corresponding to each sampling depth is applied to all hardness features in the row containing that depth. Specifically, the depth-based weight of the first sampling depth is multiplied by the standard conic index and hardness gradient of the first row to obtain two weighted features for the first depth layer; then, the depth-based weight of the second sampling depth is multiplied by the two hardness features of the second row, and so on for all sampling depths. When the depth effect weight is 1, the hardness feature at that depth remains unchanged; when the depth effect weight is between 0 and 1, the hardness feature is reduced according to the actual coverage ratio; when the depth effect weight is 0, the hardness feature corresponding to that depth is adjusted to 0.

[0112] Specifically, after weighting at each depth, the features are re-expanded according to the same depth and feature order as described above. First, the standard conic index and hardness gradient after weighting at the first sampling depth are arranged, then the two weighted features at the second sampling depth are arranged, and so on until the last sampling depth, to obtain the hardness contribution vector. This maintains the same dimension and feature positions as the original hardness feature vector, only changing the contribution level of hardness features at different depths based on the actual tillage depth.

[0113] The degree of continuous coverage is determined by the continuous coverage status of adjacent time nodes in the depth coverage matrix. For each matrix column corresponding to a sampling depth, data is read sequentially starting from the beginning time node of the effective work interval. When consecutive values ​​of 1 appear, the number of continuously covered nodes is accumulated; when a value of 0 appears, the accumulation ends and the count restarts. After traversing the entire column, the segment with the most consecutive values ​​of 1 is selected, and the number of nodes in that segment is divided by the total number of effective time nodes in that column. The resulting ratio is the degree of continuous coverage for that sampling depth. Therefore, under the same effective coverage ratio, a work process that continuously maintains the target tillage depth achieves a higher degree of continuous coverage, while intermittent coverage caused by the actual tillage depth frequently exceeding the target depth achieves a lower degree of continuous coverage.

[0114] The depth matching parameters are aggregated based on the depth effect weight and the degree of continuous coverage. For each sampling depth, the depth effect weight of that depth is first added to the degree of continuous coverage, and then divided by 2 to obtain the comprehensive depth matching value of that sampling depth. Then, the comprehensive depth matching values ​​of all sampling depths are added together and divided by the total number of sampling depths. The result is used as the depth matching parameter of the current effective working area.

[0115] The depth weighting reflects the cumulative extent to which each sampling depth is covered by the actual tillage depth within the effective working area, while the continuous coverage reflects the duration of this coverage over time. Both factors contribute to the depth matching parameter calculation, distinguishing between working areas that briefly reach a certain depth but exhibit poor continuity and those that consistently and stably reach that depth.

[0116] The resulting hardness contribution vector is used as the soil hardness input in subsequent effective training samples, and the depth matching parameter is used together with the spatial matching parameter to determine the registration reliability. Thus, the soil hardness features formed above, after being corrected for actual tillage depth, establish a depth-consistent correspondence with the operational parameters and tillage responses within the same effective operating interval.

[0117] Example 7:

[0118] Specifically, S6 includes: Obtain the time identifier of the first response sequence and the pre-calibrated response delay parameters to form a time offset sequence; The time signature of the first response sequence is corrected according to the time offset sequence, and the second response sequence is obtained by resampling according to a unified time axis; Extract the effective response subsequence from the second response sequence based on the effective operating interval; Invalid response data are removed based on the effective range of the response acquisition device and the valid data identifier, while the transient response peak value within the effective range is retained, and the mean, peak value, root mean square value and fluctuation characteristics are extracted. The extracted response features are combined in a preset order to form a tillage response feature vector.

[0119] In the above implementation, each response data in the first response sequence carries a time identifier under the unified timing reference formed above. The response delay parameter is calibrated before the formal operation data acquisition. During calibration, the tiller is continuously operated with constant operating parameters, and a load change zone with known boundary positions is set in the operation path; the time when the reference point of the tillage component crosses the boundary is determined according to the second positioning sequence, and the time when the response acquisition device detects a continuous change in the response quantity is recorded. The former is subtracted from the latter to obtain one response delay. Ten consecutive calibrations are completed, and the ten response delays are arranged in ascending order of value. One maximum value and one minimum value are removed, and the remaining eight delays are added together and divided by 8. The resulting time is used as the response delay parameter of that response channel. When there are multiple response channels, each response channel performs the above calibration separately.

[0120] In forming the time offset sequence, response data is read item by item according to the time order of the first response sequence. For data belonging to the same response channel, the response delay parameter of that response channel is configured to the corresponding time node; when multiple response channels have different response delays, they are configured according to their respective calibration results. This ensures that each response data has a corresponding time offset, and they are arranged in the original time identifier order to form a time offset sequence.

[0121] When correcting the time stamp based on the time offset sequence, the original time stamp of each response data point in the first response sequence is subtracted from its corresponding response delay, and the resulting time is used as the corrected time stamp. This process shifts the actual time of the sensor-recorded response forward to the tillage time that produced the response. After all response data has been time-corrected, they are reordered according to the corrected time stamp.

[0122] Subsequently, resampling is performed based on the unified timeline established above. When the timestamp of a certain corrected response data is consistent with the unified time node, the response value is directly assigned to the corresponding time node; when there is no completely consistent timestamp, the response value that is most recent before the time node and the response value that is most recent after the time node are read.

[0123] Specifically, when interpolating the two response values, first calculate the time length between the current unified time node and the previous response data, then calculate the total time length between the two response data. Divide the former by the total time length to obtain the time ratio of the current time node between the two response data. Then, subtract the previous response value from the latter response value to obtain the response change. Multiply the response change by the time ratio and add it to the former response value. The result is written to the current time node. All nodes in the unified time axis are processed in the same way to form the second response sequence. After multiple response channels have completed time correction and resampling respectively, they are arranged according to the unified time nodes.

[0124] When extracting the second response sequence based on the aforementioned valid operational intervals, the start and end time nodes of the valid operational intervals are used as extraction boundaries. Only response data with time markers within these intervals are retained, maintaining the original time order, thus forming a valid response subsequence. In cases where multiple separate valid operational intervals exist for the same soil hardness sampling point, valid response subsequences are extracted for each interval separately, without merging response data across intervals.

[0125] Before the effective response subsequence enters feature calculation, it is first screened based on the effective range of the response acquisition device and the data validity identifier. The effective range of the response acquisition device is determined by calibration data. At least five standard load points are set within the rated range, and each standard load point is continuously measured 10 times. Each measured value is compared with the corresponding standard load value. If all 10 measured results meet the indication error requirements specified by the device, the load point is determined as an effective calibration point. The minimum load value among the continuous effective calibration points is taken as the lower limit of the effective range, and the maximum load value is taken as the upper limit of the effective range. During the formal acquisition process, the response acquisition device synchronously records the data validity identifier. When the sensor self-test is normal, no overflow occurs in the acquisition channel, the data frame verification passes, and the communication data is complete, the corresponding data is recorded as valid; if any one of these conditions is not met, the corresponding data is recorded as invalid. Response data in the effective response subsequence that is below the lower limit of the effective range, above the upper limit of the effective range, or has an invalid data validity identifier is discarded, and the remaining response data enters feature calculation.

[0126] Transient response peaks within the effective measurement range and with valid data identification are retained. During peak identification, comparisons are performed item by item starting from the second data point of the effective response subsequence. If the current response value is greater than the previous response value and greater than or equal to the next response value, the current response value is identified as a local peak. If multiple consecutive time points maintain the same maximum response value, the maximum response value in that consecutive segment is considered a single peak. Peaks confirmed by the effective measurement range and valid data identification are not subject to smoothing replacement, ensuring that short-term loads generated when the tillage component passes through locally high-resistance soil layers are retained in the response characteristics.

[0127] The mean is calculated from all valid response data within the valid response subsequence. All valid response values ​​are summed one by one, and then divided by the number of valid response data to obtain the mean response for the current valid work interval, which characterizes the average level of response volume within that interval.

[0128] The peak value is the maximum value of all valid response data in the valid response subsequence. During the calculation, the first valid response value is used as the initial maximum value, and comparisons are made sequentially starting from the second valid response value. When the current response value is greater than the recorded maximum value, the current response value replaces the original maximum value, until the valid response subsequence is traversed, and the final retained value is taken as the peak value.

[0129] The root mean square (RMS) value is determined by averaging the squares of all valid response values. First, each valid response value is squared individually. Then, all the squared values ​​are summed. The sum is divided by the number of valid response data points. Finally, the square root of the resulting average is taken to obtain the RMS value. This value characterizes the overall amplitude level of the response signal within the valid operating range.

[0130] Specifically, the fluctuation feature is determined using the standard deviation. First, the mean response is obtained. Then, each valid response value is subtracted from the mean response value to obtain the deviation of each response value relative to the average level. Each deviation is squared and summed, then divided by the number of valid response data points to obtain the mean squared deviation. The square root of the mean squared deviation is used as the fluctuation feature. When multiple response channels exist, the mean, peak value, root mean square value, and fluctuation feature are extracted for each channel. The feature combination order is fixed before the training samples are established. The response channels are processed according to their predetermined order, and then the mean, peak value, root mean square value, and fluctuation feature are arranged sequentially within each response channel. All valid work intervals use the same arrangement rule, thus forming a tillage response feature vector with consistent dimensions and feature positions.

[0131] The obtained tillage response feature vector is correlated with the hardness contribution vector and operating parameters corresponding to the same effective operating interval. Response delay correction maps the response data to the actual tillage time that produces the response. Effective range and data validity identifiers are used to exclude collection anomalies, while the retained transient peak reflects the short-term load caused by local soil resistance changes. The four types of extracted response features characterize the average response level, maximum transient response, overall effective amplitude, and response fluctuation degree, respectively, providing response-side feature data for the construction of subsequent effective training samples.

[0132] Example 8:

[0133] Specifically, S7 includes: The spatial matching parameters and depth matching parameters are normalized and weighted and fused to obtain the registration confidence level; The registration confidence level is compared with a preset registration confidence threshold, and valid data that meets the threshold is filtered out. The training samples are formed by combining the hardness contribution vector, operation parameters and tillage response feature vectors corresponding to the valid data. A training feature matrix and a response target matrix are constructed based on multiple training samples, and a sample weight vector is generated by a weight function that monotonically increases with the registration confidence. A supervised machine learning algorithm is used to perform weighted learning on the training feature matrix and the response target matrix based on the sample weight vector, thereby establishing a mapping relationship between soil hardness, operation parameters and tillage response.

[0134] In the above implementation, the spatial matching parameters and depth matching parameters corresponding to each valid work interval are first read. Since the two types of parameters differ in their numerical ranges, normalization is performed separately before fusion. For all data to be processed, the maximum and minimum values ​​of the spatial matching parameters are first determined. For any spatial matching parameter, the minimum value of the spatial matching parameter is subtracted from the parameter, and the difference is divided by the difference between the maximum and minimum values ​​of the spatial matching parameter to obtain a spatial normalization result within the range of 0 to 1. The depth matching parameters are processed in the same way to obtain the depth normalization result. When the maximum and minimum values ​​of a certain type of matching parameter in the current batch are the same, this type of parameter does not have a distinguishing effect between samples, and the corresponding normalization result is uniformly recorded as 1.

[0135] After normalization, the spatial normalization result and the depth normalization result are each assigned a fusion weight of 0.5. Specifically, the spatial normalization result is multiplied by 0.5, the depth normalization result is multiplied by 0.5, and the two calculation results are then added together. The resulting value is used as the registration confidence level for the corresponding effective working area. The registration confidence level is maintained between 0 and 1. The closer the spatial location is to the actual operating area of ​​the tillage component and the more fully the actual tillage depth covers the hardness sampling depth, the higher the registration confidence level.

[0136] The preset registration confidence threshold is determined before formal training based on historical operation data that has already been registered. At least 300 sets of historical data with obtained hardness contribution vectors, operation parameters, and tillage response characteristics are selected, and values ​​of 0.50, 0.55, 0.60 up to 0.90 are successively used as candidate registration confidence thresholds. For each candidate threshold, only historical data with a registration confidence not lower than that threshold are retained, and a temporary prediction model is established using the same training data partitioning method and the same model training parameters.

[0137] After each temporary prediction model is trained, the mean absolute error (MAO) between the predicted and actual tillage responses is calculated on the same validation dataset. The MAO is calculated by taking the absolute difference between the predicted and actual responses for each validation sample, summing all the differences, and then dividing by the number of response data points used in the validation. The MAOs corresponding to different candidate thresholds are compared, and the number of remaining training samples after each threshold selection is counted. With the constraint that the number of retained training samples should not be less than 30% of the original historical sample size, the candidate threshold with the smallest MAO is selected as the preset registration confidence threshold. Once determined, the threshold remains unchanged within the same model version.

[0138] Specifically, when formally constructing training data, the registration confidence level corresponding to each data point is compared with the preset registration confidence level threshold item by item. Data with a registration confidence level greater than or equal to the preset threshold is retained as valid data; data with a registration confidence level lower than the preset threshold is not included in model training. This excludes data with insufficient spatial correspondence or depth effect, avoiding the participation of erroneously registered data in the establishment of supervised objectives. For each set of valid data obtained through screening, the hardness contribution vector, the operation parameters within the same valid operation interval, and the tillage response feature vector are read. When the operation parameters are continuously collected data, all valid values ​​of the same operation parameter within the valid operation interval are added together and divided by the number of valid data points; the average value obtained is used as the operation parameter for that interval. When the operation parameter is a set value that remains unchanged after being issued by the controller, the actual effective parameter value within the valid operation interval is used.

[0139] The hardness contribution vector is concatenated with the corresponding operational parameters in a fixed order to form a set of training input data. The tillage response feature vector corresponding to the same effective operational interval is used as the supervision target of this training input data, and the registration confidence of this set of data is saved at the same time, thus forming a training sample. All training samples adopt the same hardness feature arrangement order, the same operational parameter arrangement order, and the same response feature arrangement order.

[0140] When constructing the training feature matrix based on multiple training samples, each row corresponds to one training sample. The first part of each row arranges the conic exponential features and hardness gradient features in the hardness contribution vector according to the sampling depth, and the second part arranges the tiller speed, tillage depth, tillage component rotation speed, and other operational parameters included in the training in that order. The same data columns maintain the same physical meaning in all samples.

[0141] The response target matrix is ​​constructed according to the same sample row order as the training feature matrix, with each row recording the tillage response features of the corresponding training sample. When the response target contains multiple response channels, the mean, peak, root mean square value, and fluctuation features of each response channel are arranged in the aforementioned feature order, so that the first row in the training feature matrix corresponds to the same effective working interval as the first row in the response target matrix, and subsequent rows maintain the same correspondence.

[0142] Before model training, each input feature in the training feature matrix is ​​standardized. For a given feature column, all training data in that column are summed and divided by the number of samples to obtain the mean of that feature. Then, the difference between each data point and the mean is calculated, the squares of these differences are summed, divided by the number of samples, and the square root of the result is taken to obtain the standard deviation of that feature. Subsequently, each original feature value is subtracted from the mean of that column and then divided by the standard deviation of that column to obtain the standardized input feature. The mean and standard deviation of each column obtained during training are saved and used to process the input data of the region to be optimized during subsequent prediction.

[0143] Specifically, the sample weights are generated according to the monotonically increasing relationship of registration confidence. In this embodiment, the registration confidence of each training sample is first multiplied by 0.5, and then added to 0.5. The result is used as the sample weight of that training sample. When the registration confidence is 0, the sample weight is 0.5; when the registration confidence is 1, the sample weight is 1. As the registration confidence increases, the sample weight continues to increase. The sample weights are arranged according to the sample row order of the training feature matrix to form a sample weight vector.

[0144] After the above processing, all samples that reach the registration confidence threshold participate in model training, but different samples have different degrees of influence on model parameter updates. Samples with higher registration confidence are given larger sample weights, while samples close to the registration confidence threshold are given smaller sample weights, thus retaining effective data while reducing the impact of low-registration quality samples on training results.

[0145] This embodiment uses a multi-layer feedforward neural network as a supervised machine learning model. The network consists of an input layer, a 128-node first fully connected layer, a 64-node second fully connected layer, a 32-node third fully connected layer, and a response output layer. The number of nodes in the input layer matches the number of columns in the training feature matrix, and each input node receives one standardized feature or operational parameter related to the contribution of stiffness. The first fully connected layer is connected to all input nodes, the second fully connected layer is connected to the first fully connected layer, and the third fully connected layer is connected to the second fully connected layer. All three fully connected layers use linear rectified activation, setting the computation result of neurons with a value less than 0 to 0, while retaining the original value for neurons with a value not less than 0.

[0146] Specifically, the response output layer is connected to the third fully connected layer. The number of output nodes is the same as the number of columns in the response target matrix. Each output node corresponds to one tillage response feature, and continuous numerical output is used. Thus, the model input receives soil hardness features and corresponding operational parameters within the actual tillage depth range, and the output provides predicted tillage response values ​​under the combined effects of the same soil conditions and operational parameters.

[0147] During model training, the effective training samples are divided into training, validation, and test sets at ratios of 70%, 15%, and 15%, respectively. Each time, 32 training samples are read from the training set, and the corresponding training features are input into the neural network. The predicted response is obtained after passing through three fully connected layers.

[0148] For each training sample, subtract the corresponding actual response in the response target matrix from each predicted response, square the difference, sum the squared differences of all response terms for that sample, and divide by the number of response terms to obtain the prediction error for that sample. Then, multiply this prediction error by the sample weight corresponding to that sample to obtain the weighted prediction error for that sample.

[0149] Within a single training batch, the weighted prediction errors of 32 training samples are summed, and the sum is then divided by the sum of the weights of the 32 samples in that batch. The result is used as the training loss for that batch. Since the training loss directly includes the sample weights, data with higher registration reliability contributes more to the training loss when the prediction bias is the same, thus increasing the adjustment effect on the neural network connection parameters.

[0150] An adaptive moment estimation optimization method is used to update the connection weights and bias parameters of each fully connected layer based on the training loss. The initial learning rate is set to 0.001, 32 training samples are used per batch, and the maximum number of training epochs is set to 300. After each training epoch, the validation loss is calculated using the validation set. Training stops when the validation loss does not fall below the previously obtained minimum validation loss for 20 consecutive epochs, and the network parameters saved in the epoch with the lowest validation loss are restored.

[0151] After training, the test set is input into the saved model, and the mean absolute error and root mean square error (RMSE) between the predicted and actual responses are calculated. The RMSE is calculated by squaring the difference between each predicted and actual response, averaging the squares, and then taking the square root. The resulting test results are used as a record of the model's prediction performance, and the trained network structure, connection weights, bias parameters, and standardized input feature parameters are saved. The established mapping relationship uses hardness contribution features and operational parameters as inputs, and corresponding tillage response features as outputs. Hardness contribution features originate from the soil layer reached by the actual tillage depth, operational parameters correspond to the actual operating state that produced the tillage action, and the response target originates from the same effective operational interval after time delay compensation. Spatial matching, depth matching, and the resulting registration reliability further participate in sample selection and training weight allocation, ensuring that the influence of each sample during training corresponds to its actual registration reliability. The resulting mapping relationship is used for subsequent tillage response prediction for candidate operational parameters.

[0152] Example 9:

[0153] Specifically, S8 is as follows: Obtain the conic index profile features and allowable range of operation parameters for the operation area to be optimized, and generate a candidate operation parameter matrix containing candidate tillage depths; For each candidate operation parameter, the weight of the candidate depth effect is determined based on the candidate tillage depth and the corresponding depth sequence, and the conic index profile features are weighted to form the candidate hardness effect contribution vector. The contribution vectors of each candidate hardness effect are combined with the corresponding candidate operation parameters to form a prediction feature matrix, and the prediction tillage response vector is obtained based on the mapping relationship. Based on at least one preset optimization objective among tillage power consumption, peak load and tillage depth deviation, the predicted tillage response vector is normalized and evaluated, and the target evaluation value is calculated based on the response constraint threshold. Candidate job parameters that do not meet the response constraint threshold are eliminated, and the target job parameter combination is determined based on the target evaluation value.

[0154] In the above implementation, firstly, the conic index profile data of the work area to be optimized is acquired, and then depth resampling, invalid data correction, hardness gradient extraction, and depth sequence encoding are performed according to the data processing method described above to obtain conic index profile features with the same depth level and feature arrangement order as those in the model training stage. The allowable range of work parameters is directly read from the minimum allowable value, maximum allowable value, and minimum adjustment step size of each controlled parameter in the tillage machine controller. The work parameters include at least tillage depth and other work parameters actually involved in the training during the model training stage. When generating the candidate work parameter matrix, starting from the minimum allowable value of each work parameter, one candidate value is formed for each additional minimum adjustment step size until the maximum allowable value is reached. For the tillage depth parameter, the depth sequence interval used by the model is used as the value interval for the candidate tillage depth, so that each candidate tillage depth can establish a definite depth relationship with the corresponding sampling depth. Each candidate value of the first work parameter is combined with all candidate values ​​of the second work parameter, and then the candidate values ​​of the remaining work parameters are added sequentially to obtain all parameter combinations. Each parameter combination occupies one row of the candidate job parameter matrix, and the columns are set according to the order of the job parameters during model training.

[0155] Specifically, for each row of the candidate task parameter matrix, candidate tillage depths are read and compared item by item from shallowest to deepest according to the depth sequence. When a candidate tillage depth is greater than or equal to a certain sampling depth, the corresponding candidate depth contribution weight is recorded as 1; when the candidate tillage depth is less than the sampling depth, the corresponding candidate depth contribution weight is recorded as 0. After comparing all sampling depths, the weights are arranged in depth order to obtain the candidate depth contribution weights corresponding to the current candidate task parameter. Based on the candidate depth contribution weights, the conic section features of the task area to be optimized are weighted depth by depth, and the standard conic index and hardness gradient corresponding to each sampling depth are multiplied by the candidate depth contribution weight corresponding to that depth. When the weight is 1, the standard conic index and hardness gradient of that depth are retained; when the weight is 0, both features corresponding to that depth are set to 0. After processing all sampling depths, they are arranged in the order of depth and feature used in the model training phase to form the candidate hardness contribution vector corresponding to the current candidate task parameter.

[0156] Therefore, the same soil hardness profile generates different candidate hardness contribution vectors under different candidate tillage depths. As the candidate tillage depth increases, the deep soil hardness features involved in the prediction increase accordingly; as the candidate tillage depth decreases, hardness features below the candidate tillage depth do not enter the actual effect features corresponding to the current parameter combination, thus ensuring that the predicted input is consistent with the soil effect range corresponding to the candidate tillage depth. When forming the prediction feature matrix, each candidate hardness contribution vector is connected to the candidate operation parameters that generated it in a fixed order. First, the hardness contribution features corresponding to each sampling depth are arranged, then the candidate tillage depths and other candidate operation parameters are arranged to form one row of prediction features; the prediction features corresponding to all candidate operation parameters are arranged row by row to obtain the prediction feature matrix. The number of columns, column order, and physical meaning of each column in the prediction feature matrix are consistent with the training feature matrix mentioned above. Before prediction, the prediction feature matrix is ​​processed using standardized parameters saved during the model training phase. For each feature column, the current predicted feature value is subtracted from the average value of that column during the training phase, and then divided by the standard deviation of that column during the training phase. The result is used as the actual prediction input received by the model. All candidate job parameters use the same set of standardized parameters from the training phase, without recalculating the mean and standard deviation based on the prediction data.

[0157] Specifically, the standardized prediction feature matrix is ​​input row by row into the trained supervised machine learning model. Each row of prediction features passes through the input layer to the first fully connected layer (128 nodes), then through the second fully connected layer (64 nodes) and the third fully connected layer (32 nodes), and finally outputs the corresponding tillage response prediction. The output data is arranged in the same order as the target response during model training. The prediction results are arranged according to the row order of the candidate task parameter matrix to form the predicted tillage response vector. When evaluating the predicted tillage response, the tillage power consumption, peak load, and tillage depth deviation corresponding to the current optimization task are obtained first. Tillage power consumption and peak load are directly taken from the corresponding predicted response terms in the model output. The tillage depth deviation is determined by subtracting the candidate tillage depth from the target tillage depth required by the current task, and the absolute value of the difference is used as the tillage depth deviation for that candidate task parameter.

[0158] Specifically, the response constraint thresholds are determined before the evaluation of candidate parameters. The tillage power consumption constraint threshold is taken as the upper limit of the allowable power for continuous operation specified in the technical parameters of the tiller's power source. The peak load constraint threshold is determined by reading the allowable peak loads of the engine, power transmission components, traction connection components, and tillage components, and taking the minimum allowable value in the power transmission path corresponding to the current response quantity as the peak load constraint threshold. The tillage depth deviation constraint threshold is taken as the allowable tillage depth deviation specified for the current tillage task. The above thresholds are determined before the start of this operation parameter optimization and remain unchanged during the candidate parameter screening process for the same operation area to be optimized.

[0159] Subsequently, each response index involved in the optimization is normalized. Taking tillage power consumption as an example, the maximum and minimum values ​​of the predicted power consumption corresponding to all candidate operation parameters are first obtained. The predicted power consumption of the current candidate parameter is then subtracted from the minimum value, and the difference is divided by the difference between the maximum and minimum values ​​to obtain the normalized power consumption value of the current candidate parameter. Peak load and tillage depth deviation are processed in the same way to obtain their corresponding normalized values. When the maximum and minimum values ​​of a certain evaluation index are equal among all candidate parameters, the index has no distinguishing effect on the candidate parameters, and the corresponding normalized value is uniformly recorded as 0.

[0160] When performing constraint judgment, the predicted tillage power consumption of each candidate parameter is compared with the tillage power consumption constraint threshold, the predicted peak load is compared with the peak load constraint threshold, and the tillage depth deviation is compared with the tillage depth deviation constraint threshold. If any response exceeds the corresponding constraint threshold, the candidate operation parameter is determined to not meet the response constraint and is eliminated; if none of the three responses exceed the corresponding constraint threshold, the candidate operation parameter enters the target evaluation. The target evaluation value is calculated according to the pre-determined evaluation weights of the current operation task. In this embodiment, the evaluation weights of tillage power consumption, peak load, and tillage depth deviation are 0.4, 0.4, and 0.2, respectively. The normalized power consumption value is multiplied by 0.4, the normalized peak load value is multiplied by 0.4, and the normalized tillage depth deviation value is multiplied by 0.2. The three results are then added together to obtain the target evaluation value corresponding to the current candidate operation parameter. When only one optimization objective is set, the normalized result of that response is directly used as the target evaluation value.

[0161] The evaluation weights are determined based on the optimization objectives during the task setting phase. When reducing power consumption is the primary objective, tillage power consumption is set to the highest weight; when limiting mechanical load is the primary objective, peak load is set to the highest weight. All evaluation weights are summed to 1 and remain unchanged throughout the parameter selection process for the area to be optimized. After calculating all candidate parameters, candidate operation parameters exceeding any response constraint threshold are first deleted. Then, the remaining candidate parameters are sorted from smallest to largest according to their target evaluation values, and the parameter combination with the smallest target evaluation value is determined as the target operation parameter combination. If there are two or more identical minimum target evaluation values, peak load, tillage power consumption, and tillage depth deviation are compared sequentially, prioritizing the option with the smaller peak load; if peak loads are the same, the option with the smaller tillage power consumption is selected; if the first two are the same, the option with the smaller tillage depth deviation is selected. The final determined target operation parameter combination includes the target tillage depth and other corresponding controlled operation parameters and is sent to the tiller controller for execution. In this process, the candidate tillage depth first changes the contribution of the corresponding soil layer's hardness, and then enters the above-established mapping relationship together with the candidate operation parameters. Therefore, the model prediction results correspond to the actual tillage state formed by specific soil hardness conditions and specific candidate parameters, rather than being evaluated independently based solely on the operation parameters.

[0162] Example 10:

[0163] Furthermore, it also includes: Control the tiller to perform operations according to the target combination of operating parameters and obtain the corresponding actual tillage response vector; Calculate the response residuals between the actual tillage response vector and the corresponding predicted tillage response vector, normalize and weight the response residuals to obtain the response deviation value; When the response deviation value exceeds the preset deviation threshold, the current hardness contribution vector is combined with the target operation parameters to form the current operation feature vector; Calculate the feature distance between each candidate operation feature vector and the current operation feature vector, filter local candidate parameters whose feature distance is not greater than a preset neighborhood threshold, and use the response residual to correct the corresponding predicted tillage response based on the correction weight that decreases as the feature distance increases. Based on the corrected predicted tillage response, the target evaluation values ​​of the local candidate parameters are recalculated, and the updated target operation parameter combination is determined.

[0164] In the above embodiment, the determined target operation parameter combination is sent to the tillage machine controller. The controller controls the tillage machine to perform operations according to the target tillage depth, travel speed, tillage component rotation speed, and corresponding operation parameters. After each operation parameter reaches the target set value, the effective operation interval corresponding to the current soil hardness sampling point is used as the response acquisition interval, and the actual response data within this interval is extracted. The actual response data is processed according to the response delay compensation, effective range selection, and feature extraction process described above, respectively obtaining the response mean, peak value, root mean square value, and fluctuation characteristics corresponding to the model output, and forming the actual tillage response vector according to the response feature arrangement order determined during model training. The actual tillage response vector has the same number of features and arrangement position as the predicted tillage response vector obtained for the target operation parameter combination.

[0165] The actual tillage response and the predicted tillage response are compared item by item. For the same response feature, the actual response value is subtracted from the corresponding predicted response value, and the difference is used as the response residual for that feature. A positive response residual indicates that the actual operation response is higher than the predicted result; a negative response residual indicates that the actual operation response is lower than the predicted result. Each response residual retains its positive or negative direction and is used to subsequently correct the predicted response of candidate parameters.

[0166] To eliminate the influence of different response feature dimensions and numerical ranges, the response residuals are normalized. For each response feature, the maximum and minimum values ​​of that feature are obtained from the training response data used to build the model. The training value range of that response feature is obtained by subtracting the minimum value from the maximum value. Then, the absolute value of the current response residual is taken and divided by the training value range. The resulting value is used as the normalized residual of that response feature. When the training value range is 0, the response feature does not change in the training data and is not included in the calculation of the response bias value.

[0167] In this process, after normalization, the response evaluation weights used for the current optimization objective are aggregated. Each normalized residual is multiplied by its corresponding evaluation weight, and all products are summed. The sum is then divided by the sum of the evaluation weights actually used in the calculation to obtain the response deviation value. Thus, the response deviation value reflects both the degree of deviation between the actual and predicted responses and the importance of different response indicators in the current operational objective. The preset deviation threshold is determined by the prediction error during the model validation phase. After model training is completed, the response residuals between the predicted tillage response and the corresponding actual tillage response are calculated for each set of samples in the validation set. Then, the response deviation value of each validation sample is calculated using the same normalization and weighted aggregation process as in the formal operation phase. The response deviation values ​​of all validation samples are arranged in ascending order, and the response deviation value at the 95th percentile is taken as the preset deviation threshold. This threshold is saved along with the current model parameters and remains unchanged unless the model is retrained.

[0168] If the current response deviation is not greater than the preset deviation threshold, the actual operation response is determined to be within the model verification error range, and the current target operation parameter combination is continued to be used; if the current response deviation is greater than the preset deviation threshold, local parameter correction is initiated.

[0169] When forming the current operation feature vector, the hardness contribution vector corresponding to the current effective operation interval is read. Following the column order of the training feature matrix, the target tillage depth, driving speed, tillage component rotation speed, and other operation parameters are sequentially arranged after the hardness contribution vector to obtain the current operation feature vector. The current hardness contribution vector is obtained in the manner described above, that is, the depth contribution weight of each sampling depth is determined based on the current actual tillage depth, and the corresponding conic index and hardness gradient are weighted using each depth contribution weight.

[0170] The feature vectors of each candidate operation are obtained using the same processing method described above when forming the prediction feature matrix. For each candidate operation parameter, a corresponding candidate hardness contribution vector is generated based on its candidate tillage depth. This vector is then concatenated with the corresponding candidate tillage depth, driving speed, tillage component rotation speed, and other candidate operation parameters in a fixed order to form a single candidate operation feature vector. Before calculating the feature distance, the current operation feature vector and all candidate operation feature vectors are standardized using the average and standard deviation of the training features saved above. For each feature, the current feature value is subtracted from the average value of that feature during model training, and then divided by the corresponding standard deviation to obtain the standardized feature value. This avoids the distance calculation being dominated by a single numerical scale due to the dimensional differences between the conic index, hardness gradient, tillage depth, and driving speed.

[0171] Specifically, for each candidate operation feature vector, the difference between its standardized eigenvalues ​​and the corresponding standardized eigenvalues ​​of the current operation feature vector is calculated sequentially. Each difference is squared, and the sum of all squared results is taken as the square root of the sum. The resulting value is used as the feature distance between the candidate operation feature vector and the current operation feature vector. The smaller the feature distance, the closer the soil hardness state and operation parameter state corresponding to the candidate parameter are to the current actual operation state. The preset neighborhood threshold is determined based on the distance distribution between training feature samples. After model training is complete, for each group of samples in the training feature matrix, the feature distance between that sample and the other training samples is calculated, and the smallest feature distance is taken as the nearest neighbor distance of that sample. The nearest neighbor distances of all training samples are arranged from smallest to largest, and the nearest neighbor distance corresponding to the 95th percentile position is taken as the preset neighborhood threshold. This threshold is saved together with the model and the standardized parameters of the training features to ensure that the local correction range matches the feature range actually covered by the model training data.

[0172] The feature distances corresponding to the feature vectors of each candidate job are compared with a preset neighborhood threshold item by item. Candidate job parameters whose feature distances are not greater than the preset neighborhood threshold are retained as local candidate parameters; candidate job parameters whose feature distances are greater than the preset neighborhood threshold are not corrected using the current response residual. A correction weight is calculated for each local candidate parameter. First, the feature distance of the candidate parameter is divided by the preset neighborhood threshold to obtain the distance ratio of the candidate parameter relative to the neighborhood boundary. Then, the distance ratio is subtracted from 1 to obtain the correction weight. When the candidate job feature overlaps with the current job feature, the feature distance is 0 and the correction weight is 1; when the feature distance reaches the preset neighborhood threshold, the correction weight is 0; when it is between the two, the correction weight continuously decreases as the feature distance increases.

[0173] When using response residual correction to predict tillage response, for each predicted response of a local candidate parameter, the residual of the same type of response corresponding to the current actual operation is multiplied by the correction weight of that local candidate parameter to obtain the correction amount for that response term. This correction amount is then added to the original predicted response of that local candidate parameter to obtain the corrected predicted response. When the current actual response is higher than the model prediction, the corresponding predicted responses of neighboring candidate parameters are corrected upwards; when the current actual response is lower than the model prediction, it is corrected downwards according to the negative value of the response residual. As the feature distance between the candidate parameter and the current operation state increases, the correction magnitude of the response residual for that candidate parameter decreases accordingly.

[0174] After correcting the predicted response of the local candidate parameters, the response constraint judgment described above is re-executed. The corrected tillage power consumption is compared with the power consumption constraint threshold, the corrected peak load is compared with the peak load constraint threshold, and the candidate tillage depth deviation is compared with the tillage depth deviation constraint threshold. Local candidate parameters whose indicators exceed the corresponding constraint threshold are eliminated. The target evaluation value is recalculated for the remaining local candidate parameters. The corrected predicted results are used for tillage power consumption and peak load, while the absolute difference between the candidate tillage depth and the target tillage depth is still used for tillage depth deviation. The normalized values ​​of each evaluation indicator are obtained according to the normalization method determined above, and then multiplied by the evaluation weight corresponding to the current optimization task. The weighted results are then summed to obtain the new target evaluation value for the local candidate parameter.

[0175] Arrange the local candidate parameters that satisfy the constraints according to the new target evaluation value in ascending order, and determine the candidate parameter with the smallest target evaluation value as the updated target operation parameter combination. When there are two or more sets of the same minimum target evaluation value, compare the corrected peak load, corrected tillage power consumption, and tillage depth deviation in sequence, and determine the final parameter combination in order of smaller values. The updated target operation parameter combination is sent to the tiller controller for parameter execution in subsequent operation intervals. When the actual operation response exceeds the normal prediction deviation range during the model validation stage, only candidate parameters near the current operation state are used for local correction, and the degree of correction of the candidate prediction results by the response residual decreases with the increase of feature distance. Thus, while retaining the original mapping relationship prediction results, the target operation parameters are updated based on the current soil conditions and measured operation response.

[0176] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing tillage machine operating parameters based on machine learning, characterized in that, include: S1. Obtain the location data of multiple soil hardness sampling points and conic index profile data to form a depth sequence and hardness feature vector; S2. Continuously acquire the tiller's operation data to form a first positioning sequence, a heading sequence, a first tillage depth sequence, an operation parameter sequence, and a first response sequence; S3. Based on the installation offset and heading sequence of the positioning device and the tillage component, perform coordinate transformation on the first positioning sequence to obtain the second positioning sequence of the tillage component; S4. Construct the sweeping action area based on the second positioning sequence and the effective working width of the tillage component, and determine the effective working interval and spatial matching parameters based on the positional relationship between each sampling point and the sweeping action area. S5. Extract the second tillage depth sequence corresponding to the effective working area, compare it with the depth sequence, determine the depth effect weight according to the proportion of actual tillage depth reaching each sampling depth, and weight the hardness feature vector to obtain the hardness effect contribution vector and depth matching parameters. S6. Perform time compensation on the first response sequence according to the response delay parameter to obtain the second response sequence, and extract the response data corresponding to the effective working area to form a tillage response feature vector. S7. Determine the registration confidence level based on the spatial matching parameters and depth matching parameters. Use the hardness contribution vector with a registration confidence level not less than the preset threshold, the corresponding operation parameters, and the tillage response feature vector to form an effective training sample. Use machine learning algorithms to establish the mapping relationship between soil hardness, operation parameters, and tillage response.

2. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S1 includes: The spatial coordinates and cone index profile data of each soil hardness sampling point were obtained, and the cone indices were sorted according to the sampling depth to form an initial hardness sequence and a depth sequence. The initial hardness sequence is resampled according to a preset depth interval to obtain a standard hardness sequence. Invalid data are identified and neighborhood correction is performed based on the effective range and repeated sampling consistency threshold of the conic index acquisition device, while hardness mutation data within the effective range are retained. Calculate the rate of change of the conic exponent at adjacent depths in the standard hardness sequence to form a hardness gradient sequence, and combine it with the standard hardness sequence to form a hardness feature matrix. The hardness feature matrix is ​​encoded in depth order to form a hardness feature vector associated with the spatial coordinates of the sampling points.

3. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S2 includes: Acquire positioning data, heading data, actual tillage depth data, operation parameter data, and tillage response data according to the corresponding sampling frequency, and add time markers; A unified timeline is constructed based on time identifiers, mapping data with different sampling frequencies to corresponding time nodes; Based on the effective range and data continuity threshold of each data acquisition device, missing or invalid data is identified and compensated using adjacent valid data. Construct a multi-source operation data matrix according to a unified timeline, and extract the corresponding data columns respectively; The first positioning sequence, heading sequence, first tillage depth sequence, operation parameter sequence, and first response sequence are formed in chronological order.

4. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S3 includes: An installation offset vector is constructed based on the installation distance and direction between the positioning device and the tillage component; Extract the heading angles at each time point from the heading sequence, and construct the corresponding rotation matrix based on the heading angles; By using the rotation matrix transformation to install the offset vector, the spatial offset vector corresponding to each time node is obtained; The spatial offset vector is superimposed with the corresponding position vector in the first positioning sequence, and invalid positions are corrected based on the displacement of adjacent positions and a preset displacement threshold. Arrange the corrected position vectors in chronological order to form a second positioning sequence that represents the actual movement trajectory of the tillage component.

5. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S4 includes: Connect adjacent position vectors in the second positioning sequence according to time order to form a continuous operation trajectory segment; The effective working width of the tillage component is extended laterally along each working trajectory segment to form a continuous sweeping action area; Calculate the lateral distance from each soil hardness sampling point to the corresponding operation trajectory segment, and form a spatial matching sequence based on the relationship between the lateral distance and the effective operation half-width; Extract time nodes that are consecutively within the sweeping domain from the spatial matching sequence to determine the effective working interval; Spatial weights are determined based on the ratio of each lateral distance within the effective work area to the effective work half-width, and then aggregated to obtain spatial matching parameters.

6. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S5 includes: The second tillage depth sequence is extracted from the first tillage depth sequence based on the effective operating range; The actual tillage depth in the second tillage depth sequence is compared with each sampling depth in the depth sequence to construct a depth coverage matrix that characterizes whether the corresponding sampling depth has been reached at each time node. The depth effect weight is determined based on the effective coverage ratio corresponding to each sampling depth in the depth coverage matrix, forming a depth effect weight vector. A depth weight matrix is ​​constructed with the weights of each depth as diagonal elements, and the hardness feature vector is reconstructed into a matrix according to the sampling depth. The hardness contribution vector is obtained by weighting and expanding the depth weight matrix. The depth matching parameters are obtained by aggregating the depth effects based on the weights of each depth and the degree of continuous coverage.

7. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S6 includes: Obtain the time identifier of the first response sequence and the pre-calibrated response delay parameters to form a time offset sequence; The time signature of the first response sequence is corrected based on the time offset sequence, and the second response sequence is obtained by resampling according to a unified time axis; Extract the effective response subsequence from the second response sequence based on the effective operating interval; Invalid response data are removed based on the effective range of the response acquisition device and the valid data identifier, while the transient response peak value within the effective range is retained, and the mean, peak value, root mean square value and fluctuation characteristics are extracted. The extracted response features are combined in a preset order to form a tillage response feature vector.

8. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, S7 includes: The spatial matching parameters and depth matching parameters are normalized and weighted and fused to obtain the registration confidence level; The registration confidence level is compared with a preset registration confidence threshold, and valid data that meets the threshold is filtered out. The training samples are formed by combining the hardness contribution vector, operation parameters and tillage response feature vectors corresponding to the valid data. A training feature matrix and a response target matrix are constructed based on multiple training samples, and a sample weight vector is generated by a weight function that monotonically increases with the registration confidence. A supervised machine learning algorithm is used to learn the training feature matrix and the response target matrix by weighting the sample weight vector, and to establish the mapping relationship between soil hardness, operation parameters and tillage response.

9. The method for optimizing tillage machine operation parameters based on machine learning according to claim 1, characterized in that, It also includes S8, predicting and calculating the soil hardness and candidate operation parameter combinations of the area to be optimized based on the mapping relationship, and determining the target operation parameter combination based on the predicted tillage response and the preset optimization target, specifically: Obtain the conic index profile features and allowable range of operation parameters for the operation area to be optimized, and generate a candidate operation parameter matrix containing candidate tillage depths; For each candidate operation parameter, the weight of the candidate depth effect is determined based on the candidate tillage depth and the corresponding depth sequence, and the conic index profile features are weighted to form the candidate hardness effect contribution vector. The contribution vectors of each candidate hardness effect are combined with the corresponding candidate operation parameters to form a prediction feature matrix, and the prediction tillage response vector is obtained based on the mapping relationship. Based on at least one preset optimization objective among tillage power consumption, peak load and tillage depth deviation, the predicted tillage response vector is normalized and evaluated, and the target evaluation value is calculated based on the response constraint threshold. Candidate job parameters that do not meet the response constraint threshold are eliminated, and the target job parameter combination is determined based on the target evaluation value.

10. The method for optimizing tillage machine operation parameters based on machine learning according to claim 9, characterized in that, Also includes: Control the tiller to perform operations according to the target combination of operating parameters and obtain the corresponding actual tillage response vector; Calculate the response residuals between the actual tillage response vector and the corresponding predicted tillage response vector, normalize and weight the response residuals to obtain the response deviation value; When the response deviation value exceeds the preset deviation threshold, the current hardness contribution vector is combined with the target operation parameters to form the current operation feature vector; Calculate the feature distance between each candidate operation feature vector and the current operation feature vector, filter local candidate parameters whose feature distance is not greater than a preset neighborhood threshold, and use the response residual to correct the corresponding predicted tillage response based on the correction weight that decreases as the feature distance increases. Based on the corrected predicted tillage response, the target evaluation values ​​of the local candidate parameters are recalculated, and the updated target operation parameter combination is determined.

Images