Digital processing system for adaptive depth of tillage machine
By combining edge controllers and sensors, frequency domain feature vectors are extracted and a nonlinear programming model is constructed, which solves the problem of soil loosening depth control for adaptive tillage machines in uneven soil environments, achieving stable and efficient soil loosening effect and reducing frame vibration and energy consumption.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAANXI JINWOLAN CONSTR ENG CO LTD
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-07
AI Technical Summary
Existing adaptive tillage machines fail to effectively combine soil physical and mechanical properties with the depth adjustment process in controlling soil loosening depth. This results in frequent operation of the lifting mechanism in environments with uneven soil compaction and alternating soft and hard soil conditions, leading to frame vibration and increased energy consumption of the hydraulic system.
An edge controller is used in conjunction with a torque sensor and a triaxial accelerometer. The torque and vibration frequency domain feature vectors are extracted by a sliding window fast Fourier transform. The soil dynamic parameters are identified by a support vector data description classifier. A nonlinear programming model is constructed with the objectives of minimizing the torque fluctuation variance and maximizing the soil fragmentation volume. Combined with the constraints of the hydraulic system, a smooth hydraulic proportional valve current control sequence is generated.
It achieves adaptive adjustment of soil loosening depth, avoids frequent operation of the lifting mechanism caused by fixed step length adjustment, reduces frame vibration and transient impact load on the transmission system, and improves the operation stability and efficiency of the tillage machine.
Smart Images
Figure CN122345997A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electronic digital data processing technology, specifically relating to a digital processing system for adaptive soil loosening depth of tillage machines. Background Technology
[0002] Existing adaptive tillage machines typically rely on torque sensors mounted on the cutter spindle and accelerometers on the frame to collect operational signals during soil loosening depth control. Upon receiving these sensor signals, the digital processing system directly calculates the average torque or vibration amplitude in the time domain and compares it with a preset fixed traction resistance threshold within the controller. When the current torque signal exceeds the preset threshold, the controller generates a lifting command with a fixed step size to reduce the loosening depth; conversely, when the current torque signal falls below the threshold, the controller generates a pressing command with a fixed step size to increase the loosening depth. This control logic, based on time-domain threshold comparison and fixed step size adjustments, maintains its judgment rules and adjustment range unchanged throughout the entire field operation cycle.
[0003] The aforementioned existing technology uses a fixed threshold and a fixed step size for linear adjustment, without performing dynamic coupling calculations on the physical and mechanical properties of the soil and the depth adjustment process. As a result, in field environments with uneven soil compaction and alternating soft and hard soil, the lifting mechanism will repeatedly perform lifting actions due to frequent fluctuations in resistance, causing severe vibration of the frame and increasing the energy consumption of the hydraulic system. Summary of the Invention
[0004] The purpose of this invention is to provide a digital processing system for the soil loosening depth of an adaptive tillage machine, which can solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A digital processing system for adaptive soil loosening depth of a tillage machine includes an edge controller, a torque sensor, a triaxial accelerometer, and a hydraulic proportional valve. The edge controller is electrically connected to the torque sensor, the triaxial accelerometer, and the hydraulic proportional valve, respectively. The edge controller is configured to perform the following steps: real-time acquisition of the tool spindle torque signal sequence collected by the torque sensor and the frame triaxial acceleration signal sequence collected by the triaxial accelerometer. Fast Fourier transform is performed on the torque signal sequence and the frame triaxial acceleration signal sequence using a sliding window to extract the torque frequency domain feature vector and the vibration frequency domain feature vector. The torque frequency domain feature vector and the vibration frequency domain feature vector are input into the soil type identification classifier, which outputs the set of dynamic parameters of the current cutting soil. Substituting the set of dynamic parameters into the state-space equation of soil-cutter interaction, with minimizing the torque fluctuation variance and maximizing the effective soil breaking volume as the dual objective functions, and with the maximum response rate of the hydraulic system and the maximum allowable structural stress as constraints, a nonlinear programming model for depth adjustment is constructed. Within each control cycle, the sequential quadratic programming algorithm is invoked to iteratively solve the nonlinear programming model and output the optimal soil loosening depth setting value at the current moment. The optimal loosening depth setting is converted into a smooth hydraulic proportional valve current control sequence of the digital processing system for the adaptive tillage machine's loosening depth using a trajectory interpolation algorithm, and then sent to the hydraulic proportional valve.
[0007] Preferably, the specific method for performing fast Fourier transform on the torque signal sequence and the frame triaxial acceleration signal sequence through a sliding window includes: dividing the sliding window with a fixed number of sampling points, and applying a Hanning window function to the torque signal sequence and the frame triaxial acceleration signal sequence respectively within the sliding window; Perform a Fast Fourier Transform on the windowed torque signal sequence and the frame triaxial acceleration signal sequence to obtain initial torque spectrum data and initial vibration spectrum data; The initial torque spectrum data and the initial vibration spectrum data are subjected to amplitude squared processing to extract the spectral peaks and corresponding frequency values within a preset frequency range. The spectral peaks and the frequency values are then combined to generate the torque frequency domain feature vector and the vibration frequency domain feature vector.
[0008] Preferably, the soil type identification classifier is a support vector data description classifier. The process of inputting the torque frequency domain feature vector and the vibration frequency domain feature vector into the soil type identification classifier includes: concatenating the torque frequency domain feature vector and the vibration frequency domain feature vector to generate a high-dimensional joint feature vector. The high-dimensional joint feature vector is input into the pre-trained support vector data description classifier, and the high-dimensional joint feature vector is mapped to a high-dimensional feature space through a Gaussian kernel function; In the high-dimensional feature space, calculate the distance from the high-dimensional joint feature vector to the center of the hypersphere, select the soil category label corresponding to the hypersphere with the smallest distance, retrieve the internal friction angle, cohesion and soil bulk density parameters associated with the soil category label, and combine the internal friction angle, the cohesion and the soil bulk density parameters into the dynamic parameter set.
[0009] Preferably, the method of constructing a nonlinear programming model for depth adjustment with the dual objective functions of minimizing torque fluctuation variance and maximizing effective soil breaking volume includes: establishing a discretized state-space equation containing tool cutting depth state variables based on the set of dynamic parameters, and using the tool cutting depth state variables as decision variables; The first objective function term is the variance calculation formula of the predicted torque sequence output by the discretized state space equation, and the second objective function term is the product integral calculation formula of the tool cutting depth state variable, tool width and grounding machine forward speed. The first objective function term and the second objective function term are assigned fixed weight coefficients and then linearly weighted and summed to generate the dual objective function.
[0010] Preferably, the method of using the maximum response rate of the hydraulic system and the maximum allowable structural stress as constraints includes: obtaining the rodless chamber area and rod chamber area of the hydraulic cylinder and the rated displacement of the hydraulic pump, and combining the rated flow of the hydraulic proportional valve to calculate a first upper limit value of the change in the tool spindle depth between adjacent control cycles, and constructing a first inequality constraint; Obtain the moment of inertia of the section of the tillage machine cutter arm and the ultimate yield strength of the material. Combine the predicted soil resistance value in the set of dynamic parameters to calculate the bending moment borne by the root of the cutter arm. Set a second upper limit value for the bending moment and construct a second inequality constraint. The first inequality constraint and the second inequality constraint are used as the boundary constraints of the nonlinear programming model.
[0011] Preferably, the method of using the sequential quadratic programming algorithm to iteratively solve the nonlinear programming model includes: using the historical optimal soil loosening depth setpoint output by the control cycle as the initial iteration point of the current control cycle; In each iteration, the biobjective function in the nonlinear programming model is expanded in second order Taylor at the current iteration point to generate a quadratic programming subproblem; Solve the quadratic programming subproblem to obtain the search direction and step size, update the current iteration point, and stop iterating when the absolute value of the difference between the two objective functions in two adjacent iterations is less than the preset convergence threshold, and output the current iteration point as the optimal soil loosening depth setting value.
[0012] Preferably, before dividing the sliding window with a fixed number of sampling points, the method further includes preprocessing the torque signal sequence and the frame triaxial acceleration signal sequence: extracting each data point in the torque signal sequence within the sliding window, and calculating the first difference between each data point and the mean torque value within the sliding window; Data points whose absolute value of the first difference is greater than a first preset multiple of the standard deviation are removed, and data at the positions of the removed data points are supplemented using a linear interpolation algorithm to generate a first preprocessed torque signal sequence. The frame triaxial acceleration signal sequence within the sliding window is processed using the same elimination criteria and interpolation algorithm as the first preprocessed torque signal sequence to generate the first preprocessed acceleration signal sequence.
[0013] Preferably, after retrieving the internal friction angle, cohesion, and soil bulk density parameters associated with the soil category label, the method further includes online updating of the support vector data description classifier: constructing online sample pairs containing the high-dimensional joint feature vector and the soil category label, and storing the online sample pairs in a circular buffer queue. When the number of online sample pairs in the circular buffer queue reaches a preset batch number, the preset batch number of online sample pairs is extracted. Using the coordinates and radius of the hypersphere's center as the initial values, the Lagrange multiplier solution process is re-executed using the preset batch number of online samples to update the coordinates and radius of the hypersphere's center, and the updated parameters of the Gaussian kernel function are saved.
[0014] Preferably, the method of assigning fixed weight coefficients to the first objective function term and the second objective function term and performing linear weighted summation is replaced by: extracting the historical predicted torque sequence output by the discretized state-space equation in the previous control cycle, and calculating the variance of the historical predicted torque sequence as the historical variance index; The historical variance index is input into a preset piecewise mapping function, wherein the piecewise mapping function uses the historical variance index as the independent variable and the first weight coefficient corresponding to the first objective function term and the second weight coefficient corresponding to the second objective function term as the dependent variables. When the historical variance index is greater than the preset segmentation threshold, the first weight coefficient is increased and the second weight coefficient is decreased; when the historical variance index is less than the preset segmentation threshold, the first weight coefficient is decreased and the second weight coefficient is increased.
[0015] Preferably, after converting the optimal loosening depth setting value into a smooth hydraulic proportional valve current control sequence through trajectory interpolation algorithm, the method further includes performing dead zone compensation processing on the hydraulic proportional valve current control sequence: obtaining the current-flow characteristic curve of the hydraulic proportional valve, and extracting the current interval in the current-flow characteristic curve where the flow rate remains zero as the dead zone current interval. Calculate the second difference between each current control point in the hydraulic proportional valve current control sequence and the center point of the dead zone current interval; When the absolute value of the second difference is less than half the width of the dead zone current interval, the value of the current control point is forcibly modified to the upper limit boundary value of the dead zone current interval. When the absolute value of the second difference is greater than or equal to half the width of the dead zone current range, the value of the current control point remains unchanged.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts torque and vibration frequency domain feature vectors using a sliding window fast Fourier transform. It then utilizes support vector data description to output a set of dynamic parameters including internal friction angle, cohesion, and soil bulk density from the classifier's output. This transforms depth control from time-domain threshold comparison to a nonlinear programming solution based on state-space equations. Using minimizing torque fluctuation variance and maximizing effective soil breaking volume as dual objective functions, and combining the maximum response rate of the hydraulic system and cantilever bending moment as constraints, the optimal loosening depth setpoint is solved using a sequential quadratic programming algorithm. A smooth current control sequence is then generated through trajectory interpolation. This process directly incorporates the dynamic nonlinear physical properties of the soil into the depth decision model, avoiding frequent lifting mechanism movements caused by fixed step size adjustments, eliminating vibration of the frame under alternating soft and hard soil conditions, and reducing the transient impact load on the transmission system.
[0017] 2. Before frequency domain feature extraction, data points deviating from the mean by multiples of the standard deviation were removed, and linear interpolation was performed to eliminate interference from abnormal jump signals during sensor acquisition on the classifier's judgment. A circular buffer queue was used to store online sample pairs, and the Lagrange multiplier solution process was re-executed, allowing the classifier's hypersphere parameters to adaptively adjust according to changes in the actual working soil. The weight coefficients of the dual objective function were piecewise mapped and adjusted based on the variance index of historical predicted torque sequences, enabling the control strategy to switch its optimization focus between highly fluctuating and low-fluctuating soil conditions. Dead-zone compensation was introduced for the current-flow characteristic curve of the hydraulic proportional valve, forcibly modifying control points falling into the dead-zone current range to the upper limit boundary value, eliminating the hysteresis error of the hydraulic valve having no mechanical response under small current input. Attached Figure Description
[0018] Figure 1 Main flowchart of the digital processing system for adaptive soil loosening depth of tillage machine; Figure 2 Here is a flowchart of frequency domain feature vector extraction and signal preprocessing; Figure 3 Flowchart for soil type identification and online classifier update; Figure 4 Flowchart for constructing and adjusting weights in a nonlinear programming model; Figure 5 Flowchart for constructing constraints for nonlinear programming models; Figure 6 The flowchart for solving the quadratic programming problem and compensating for dead zones is provided. Detailed Implementation
[0019] Please refer to Figure 1 This embodiment provides a digital processing system for adaptive soil loosening depth of a tillage machine, comprising an edge controller, a torque sensor, a triaxial accelerometer, and a hydraulic proportional valve. The edge controller is electrically connected to the torque sensor, the triaxial accelerometer, and the hydraulic proportional valve. The torque sensor is installed at the transmission connection end of the tillage machine's cutter spindle to collect real-time torque signals generated by the spindle during soil cutting. The triaxial accelerometer is installed at the cutter cantilever mounting position on the tillage machine frame to collect real-time acceleration signals of the frame along the X, Y, and Z orthogonal axes during operation. The hydraulic proportional valve is installed in the hydraulic control circuit of the tillage machine's lifting mechanism to control the extension and retraction of the hydraulic cylinder, thereby adjusting the cutter's soil penetration depth.
[0020] The edge controller executes corresponding signal processing and control logic in a fixed control cycle. At the beginning of each control cycle, it acquires the tool spindle torque signal sequence from the torque sensor and the three-axis acceleration signal sequence from the three-axis accelerometer in real time. The torque signal sequence is a single-channel, equally timed discrete sampling sequence, while the three-axis acceleration signal sequence is a three-channel, equally timed discrete sampling sequence (X, Y, Z). The sampling clocks of the torque signal sequence and the three-axis acceleration signal sequence are synchronized. The torque sample value and the three-axis acceleration sample value at the same sampling moment correspond to the same tool cutting position and the same frame vibration state, ensuring the timing consistency of the signals.
[0021] The edge controller performs Fast Fourier Transform (FFT) on the acquired torque signal sequence and the triaxial acceleration signal sequence of the frame using a sliding window to extract torque frequency domain feature vectors and vibration frequency domain feature vectors. The sliding window uses a fixed-length time window, sliding across the continuous signal sequence with a step size consistent with the control cycle. Each window contains a fixed number of sampling points, and one window sliding and corresponding signal processing is completed within each control cycle. Within the sliding window, the edge controller applies a window function to each channel of the torque signal sequence and the triaxial acceleration signal sequence to suppress spectral leakage caused by time-domain signal truncation. Then, it performs an FFT on the windowed discrete signal sequence to convert the time-domain signal into a frequency-domain signal, obtaining the corresponding spectral data. The edge controller squares the amplitude of the spectral data to obtain the corresponding power spectral density data. Then, it extracts the spectral peaks and corresponding frequency values within a preset frequency range, and combines the extracted spectral peaks and frequency values in a preset order to generate the corresponding torque frequency domain feature vector and vibration frequency domain feature vector.
[0022] Table 1. Correspondence between preset frequency ranges and feature dimensions for frequency domain feature extraction.
[0023] Table 1 clarifies the frequency range division and feature dimension setting for different signal types during frequency domain feature extraction. The ranges of the tool cutting fundamental frequency range and harmonic frequency range are determined based on the rated speed range and number of cutting teeth of the tillage machine tool. The range of the frame vibration natural frequency range is determined based on the modal analysis results of the tillage machine frame. No features are extracted within the high-frequency noise suppression range to exclude the influence of high-frequency electrical interference and non-operational vibration signals on the feature extraction results. The feature extraction dimension within each frequency range corresponds to the number of spectral peaks and corresponding frequency values extracted within that range. The feature elements of all ranges are concatenated in a preset order to generate the corresponding signal's frequency domain feature vector.
[0024] The edge controller inputs the generated torque frequency domain feature vector and vibration frequency domain feature vector into the soil type identification classifier, and outputs a set of dynamic parameters for the soil being cut. The soil type identification classifier is a machine learning classification model that has been pre-trained offline. The model's input is the torque frequency domain feature vector and the vibration frequency domain feature vector, and the model's output is a set of dynamic parameters corresponding to the soil being cut. The set of dynamic parameters includes three core parameters: the soil's internal friction angle, cohesion, and soil bulk density. These three parameters are key parameters describing the dynamic characteristics of soil cutting, and can directly reflect the degree of soil compaction, cutting resistance characteristics, and difficulty of breaking up, providing a soil dynamic basis for subsequent depth adjustment decisions.
[0025] The edge controller substitutes the acquired set of dynamic parameters into the state-space equation of soil-tool interaction, using minimizing torque fluctuation variance and maximizing effective soil breaking volume as dual objective functions, and the maximum response rate of the hydraulic system and the maximum allowable structural stress as constraints, to construct a nonlinear programming model for depth adjustment. The state-space equation of soil-tool interaction is a discretized state equation describing the dynamic relationship between the tool cutting depth, spindle torque, and soil cutting resistance. The equation uses the tool cutting depth as the state variable, soil dynamic parameters, tillage machine forward speed, and tool structural parameters as model input parameters, and the tool spindle torque as the model output variable. Based on the current soil dynamic parameters, it can predict the changes in spindle torque and soil resistance at different cutting depths. The dual objective function includes two parallel optimization objectives: minimizing torque fluctuation variance is used to suppress drastic fluctuations in spindle torque, reducing frame vibration amplitude and transient impact loads on the transmission system; maximizing effective soil breaking volume is used to ensure the loosening effect of tillage operations and meet the agronomic requirements of field operations. The constraints include two types of inequality constraints: the constraint corresponding to the maximum response rate of the hydraulic system and the constraint corresponding to the maximum allowable structural stress. The constraint on the maximum response rate of the hydraulic system is used to limit the depth adjustment range within adjacent control cycles, ensuring that the hydraulic system can complete the corresponding action within the control cycle. The constraint on the maximum allowable structural stress is used to limit the maximum bending moment that the tool cantilever can withstand, avoiding plastic deformation or fatigue damage caused by structural overload.
[0026] Within each control cycle, the edge controller invokes a sequential quadratic programming algorithm to iteratively solve the constructed nonlinear programming model, outputting the optimal soil loosening depth setpoint for the current moment. The sequential quadratic programming algorithm is an iterative optimization algorithm for solving nonlinear programming problems. Within each control cycle, the edge controller uses the historically optimal soil loosening depth setpoint output from the previous control cycle as the initial iteration point for the current control cycle's iterative solution. This initial iteration point is close to the optimal solution under the current operating conditions, effectively reducing the number of iterations, improving the solution speed, and meeting the timing requirements of real-time control. During each iteration, the edge controller performs a second-order Taylor expansion of the biobjective function in the nonlinear programming model at the current iteration point, and performs a first-order Taylor expansion of the nonlinear inequality constraints at the current iteration point, generating the corresponding quadratic programming subproblems. The edge controller solves the quadratic programming subproblem to obtain the iterative search direction and optimal step size. It updates the current iteration point according to the search direction and step size. After each iteration update, it calculates the absolute value of the difference between the two objective functions in adjacent iterations. When this absolute value is less than a preset convergence threshold, the iteration process stops, and the current iteration point is output as the optimal soil loosening depth setting value for the current control cycle. The optimal soil loosening depth setting value is the tool cutting depth value that satisfies all constraints and makes the two objective functions optimal.
[0027] The edge controller uses a trajectory interpolation algorithm to convert the optimal loosening depth setpoint into a smooth hydraulic proportional valve current control sequence, and then sends the generated current control sequence to the hydraulic proportional valve. The trajectory interpolation algorithm transforms the discrete optimal loosening depth setpoint into a continuous and smooth depth control trajectory, which in turn is converted into the corresponding hydraulic proportional valve current control sequence, avoiding hydraulic system shocks and frame vibrations caused by abrupt changes in the depth setpoint. The current control sequence contains current values corresponding to multiple interpolation points at equal time intervals within the current control cycle. The edge controller outputs the current values sequentially to the hydraulic proportional valve according to a preset interpolation cycle, controlling the hydraulic cylinder to complete the corresponding depth adjustment action, thus achieving adaptive closed-loop control of the loosening depth.
[0028] In this embodiment, through a complete processing flow of frequency domain feature extraction, soil dynamic parameter identification, nonlinear programming model construction and solution, and smooth control sequence generation, the dynamic nonlinear physical properties of the soil are directly introduced into the deep decision-making process, realizing adaptive adjustment of the loosening depth. This avoids the frequent operation of the lifting mechanism caused by fixed step size adjustment and reduces frame vibration and transient impact load on the transmission system.
[0029] refer to Figure 2 and 3 In a preferred embodiment, before dividing the sliding window with a fixed number of sampling points, the edge controller preprocesses the acquired torque signal sequence and the rack triaxial acceleration signal sequence. During preprocessing, the edge controller extracts each data point from the torque signal sequence within the sliding window, calculates the mean of all torque data points within the sliding window, calculates the first difference between each data point and the mean, and simultaneously calculates the standard deviation of the torque data points within the sliding window. Data points whose absolute value of the first difference is greater than a first preset multiple of the standard deviation are removed. A linear interpolation algorithm is then used to supplement the data at the positions of the removed data points, generating a first preprocessed torque signal sequence. The first preset multiple of the standard deviation is an outlier threshold pre-set based on sensor noise characteristics and operating conditions, used to identify and remove abnormal jump data in the signal. The edge controller processes each channel of the rack triaxial acceleration signal sequence within the sliding window according to the same removal criteria and interpolation algorithm as the first preprocessed torque signal sequence, generating a first preprocessed acceleration signal sequence.
[0030] The outlier removal calculation process is described by the following formula:
[0031] in, For the first in the sliding window The signal value at each sampling point This is the mean of all signal sampling points within the sliding window. The first preset multiple, Let be the standard deviation of all signal sampling points within the sliding window. When a sampling point satisfies the above inequality, the sampling point is determined to be an outlier and is removed.
[0032] The edge controller divides the sliding window with a fixed number of sampling points. Within the sliding window, a Hanning window function is applied to the preprocessed first preprocessed torque signal sequence and the first preprocessed acceleration signal sequence, respectively. The time-domain expression of the Hanning window function is:
[0033] in, The fixed number of sampling points within the sliding window. The sampling point number within the window. For the first The edge controller multiplies the preprocessed signal sequence with the Hanning window function point by point to complete the windowing process and suppress spectral leakage caused by signal truncation.
[0034] The edge controller performs a Fast Fourier Transform (FFT) on the windowed torque signal sequence and the triaxial acceleration signal sequence to obtain initial torque spectrum data and initial vibration spectrum data. The calculation expression for the Discrete Fast Fourier Transform is:
[0035] in, For the first in the window Preprocessed signal values at each sampling point For the first The frequency domain complex values corresponding to each spectral line For spectral line numbers, This represents the fixed number of sampling points within the window. The edge controller performs amplitude square processing on the obtained initial torque spectrum data and initial vibration spectrum data to obtain the corresponding power spectral density data. It extracts the spectral peaks and corresponding frequency values within the preset frequency range, and combines the spectral peaks and frequency values in a preset order to generate torque frequency domain feature vectors and vibration frequency domain feature vectors.
[0036] The edge controller employs a Support Vector Data Description (SVM) classifier for soil type identification. When inputting the torque and vibration frequency domain feature vectors into the SVM, the edge controller first concatenates them to generate a high-dimensional joint feature vector. The concatenation is performed by sequentially arranging all elements of the torque frequency domain feature vector, followed by sequentially concatenating the feature elements along the X, Y, and Z axes of the vibration frequency domain feature vector to generate a one-dimensional high-dimensional joint feature vector. The edge controller then inputs this high-dimensional joint feature vector into a pre-trained offline SVM, which uses a Gaussian kernel function to map the high-dimensional joint feature vector into a high-dimensional feature space.
[0037] The expression for the Gaussian kernel function is:
[0038] in, and For two high-dimensional joint feature vectors, Let be the Euclidean distance between two eigenvectors. The width parameter of the Gaussian kernel function. These are the kernel function values corresponding to the two feature vectors.
[0039] The edge controller calculates the distance from the high-dimensional joint feature vector to the center of each predefined hypersphere in the high-dimensional feature space. The expression for the distance calculation is:
[0040] in, The input is a high-dimensional joint feature vector. For feature vectors The distance to the center of the corresponding hypersphere This represents the number of support vectors corresponding to the hypersphere. and For the Lagrange multipliers corresponding to the support vectors, and This is the support vector corresponding to the hypersphere. The edge controller selects the soil category label corresponding to the hypersphere with the smallest distance, retrieves the internal friction angle, cohesion, and soil bulk density parameters associated with that soil category label, and combines these parameters into a set of dynamic parameters for the current soil cutting.
[0041] Table 2. Correlation table of dynamic parameters for different soil types
[0042] Table 2 is used to store the dynamic parameter ranges corresponding to different soil category labels. The parameter range corresponding to each soil category label is obtained through offline soil cutting test and geotechnical test calibration. After the classifier outputs the soil category label, the edge controller can directly retrieve the dynamic parameter range of the corresponding category through this table, and take the midpoint of the range as the current set of dynamic parameters for cutting soil, so as to realize the rapid matching and retrieval of parameters.
[0043] After retrieving the internal friction angle, cohesion, and soil bulk density parameters associated with the soil category label, the edge controller updates the support vector data description classifier online. The edge controller constructs online sample pairs containing the current high-dimensional joint feature vector and the corresponding soil category label, and stores these online sample pairs in a circular buffer queue. The circular buffer queue uses a first-in, first-out (FIFO) storage mechanism, and its maximum storage capacity is twice the preset batch size. When the number of online sample pairs in the circular buffer queue reaches the preset batch size, the edge controller extracts this batch of online sample pairs, uses the coordinates and radius of the current hypersphere center of the support vector data description classifier as initial values, re-executes the Lagrange multiplier solution process using these batches of online sample pairs, updates the coordinates and radius of the hypersphere center, and saves the updated Gaussian kernel function parameters.
[0044] The optimization objective expression for solving the Lagrange multipliers is:
[0045]
[0046]
[0047] in, The number of online sample pairs within a batch. The penalty coefficient for the classifier is... For the first The Lagrange multipliers corresponding to each sample For the first The high-dimensional joint feature vector corresponds to each sample. The edge controller obtains the updated Lagrange multipliers by solving the above optimization problem, and then calculates the updated coordinates and radius of the hypersphere center, thus completing the online update of the classifier.
[0048] In this embodiment, abnormal jump data in the sensor-acquired signals are eliminated through signal preprocessing, reducing the interference of abnormal signals on frequency domain feature extraction and classifier identification results. Accurate extraction of frequency domain features is achieved through Fast Fourier Transform with Hanning window and windowing. Soil type and dynamic parameters are accurately identified through Support Vector Data Description Classifier. Through online sample pair cyclic buffering and model updating, the classifier can adaptively correct itself to follow changes in soil properties during operation, improving the accuracy and adaptability of soil dynamic parameter identification.
[0049] In another preferred embodiment, the edge controller establishes a discretized state-space equation containing the tool cutting depth state variable based on the acquired set of dynamic parameters. The sampling period of the discretized state-space equation is consistent with the control period of the system. The expression of the discretized state-space equation is:
[0050] in, This is the sequence number of the control cycle. For the first The state variable for each control cycle is specifically the tool cutting depth in the current control cycle. For the first The input variables for each control cycle include the internal friction angle, cohesion, and soil bulk density from the soil dynamics parameter set, as well as the tillage machine's forward speed and cutter width parameters. For the first The output variable for each control cycle is specifically the predicted torque value of the tool spindle. The state matrix, For the input matrix, For the output matrix, To directly transmit the matrix, the elements of the state matrix and the input matrix are calibrated through the soil cutting dynamics model. They are directly related to the soil's internal friction angle, cohesion, and unit weight parameters, and can reflect the influence of the cutting depth variation on the spindle torque under different soil properties.
[0051] The edge controller uses the tool cutting depth state variable as the decision variable and constructs a first objective function term as the variance calculation formula for the predicted torque sequence output by the discretized state-space equation. The expression for the first objective function term is:
[0052] in, For the first objective function term, To predict the length of the time domain, For the first The first control cycle prediction Torque value per cycle, To predict the mean of the predicted torque sequence in the time domain, the value of the first objective function term corresponds to the variance of the predicted torque sequence fluctuation; the smaller the value, the smaller the torque fluctuation.
[0053] The edge controller constructs a second objective function term that is the integral of the product of the tool cutting depth state variable, the tool width, and the grounding machine's forward speed. The expression for the second objective function term is:
[0054] in, For the second objective function term, The depth of cut is the tool cutting depth in the continuous time domain. For the tool width, The speed of the tillage machine is denoted as , the integration interval is the time range corresponding to the prediction time domain, and the absolute value of the second objective function term corresponds to the effective soil breaking volume within the prediction time domain. The smaller the value, the larger the effective breaking volume.
[0055] The edge controller assigns fixed weight coefficients to the first objective function term and the second objective function term, performs a linear weighted summation, and generates a bi-objective function. The expression for the bi-objective function is as follows:
[0056] in, It is a dual objective function. The first weight coefficient corresponds to the first objective function term. This is the second weighting coefficient corresponding to the second objective function term. The weighting coefficient is used to balance the priority of the two optimization objectives. The fixed weighting coefficient can be preset according to the agronomic requirements of the operation scenario.
[0057] The edge controller constructs corresponding inequality constraints based on the maximum response rate of the hydraulic system and the maximum allowable structural stress. The edge controller obtains the rodless chamber area and rod chamber area of the hydraulic cylinder, as well as the rated displacement of the hydraulic pump. Combined with the rated flow rate of the hydraulic proportional valve, it calculates the first upper limit value of the change in tool spindle depth between adjacent control cycles, constructing the first inequality constraint. The expression for the first inequality constraint is:
[0058] in, The first upper limit value is the change in tool spindle depth between adjacent control cycles. This value is calculated based on the rated flow rate of the hydraulic system, the working area of the cylinder, and the duration of the control cycle. The first inequality constraint is used to limit the change in depth within adjacent control cycles to not exceed the maximum response capability of the hydraulic system, ensuring that the hydraulic system can complete the corresponding action within the control cycle.
[0059] The edge controller acquires the moment of inertia of the section of the tillage machine cutter cantilever and the ultimate yield strength of the material. Combined with the predicted soil resistance values from the dynamic parameter set, it calculates the bending moment borne at the root of the cutter cantilever, sets a second upper limit for the bending moment, and constructs a second inequality constraint. The expression for the second inequality constraint is as follows:
[0060] in, This represents the bending moment borne at the root of the tool cantilever at the current cutting depth. This value is calculated based on the predicted soil cutting resistance and the tool cantilever length. The second upper limit of the bending moment is calculated based on the material's yield strength limit, the moment of inertia of the cantilever section, and a preset safety factor. The second inequality constraint is used to limit the bending moment at the root of the tool cantilever to not exceed the allowable value of the structure, thus preventing plastic deformation or fatigue damage due to structural overload. The edge controller uses the first and second inequality constraints as boundary constraints for the nonlinear programming model.
[0061] refer to Figure 4 In a preferred embodiment, the edge controller replaces the linear weighted summation method with a fixed weight coefficient with an adaptive weight adjustment method based on the historical variance index. The edge controller extracts the historical predicted torque sequence output from the discretized state-space equation in the previous control cycle, calculates the variance of the historical predicted torque sequence as the historical variance index, and inputs the historical variance index into a preset piecewise mapping function. The piecewise mapping function uses the historical variance index as the independent variable and the first weight coefficient corresponding to the first objective function term and the second weight coefficient corresponding to the second objective function term as the dependent variable. When the historical variance index is greater than a preset piecewise threshold, the edge controller increases the first weight coefficient and decreases the second weight coefficient, increasing the optimization priority of torque fluctuation suppression; when the historical variance index is less than the preset piecewise threshold, the edge controller decreases the first weight coefficient and increases the second weight coefficient, increasing the optimization priority of effective soil fragmentation volume.
[0062] The expression for the piecewise mapping function is:
[0063] in, This is a historical variance indicator. The preset high segment threshold, The preset low segment threshold, , , These represent the maximum, median, and minimum values of the first weighting coefficient, respectively. , , These are the maximum, median, and minimum values of the second weighting coefficient, respectively, and they satisfy the following conditions: The constraints.
[0064] Table 3. Weighting coefficients and piecewise mapping parameters for the biobjective function.
[0065] Table 3 is used to clarify the weight coefficient values corresponding to different historical variance index intervals, so as to realize the segmented adaptive adjustment of the weight coefficients. The weight coefficient values in the table satisfy the constraint that the sum of the two weight coefficients is 1, so that the control strategy can switch the optimization focus according to the actual torque fluctuation state, prioritizing the operation stability under high fluctuation conditions and prioritizing the operation effect under low fluctuation conditions.
[0066] In this embodiment, a discretized state-space equation was established based on the dynamic characteristics of soil-tool interaction, and a dual objective function that takes into account both torque fluctuation suppression and operational effectiveness was constructed. Constraints were established by combining the hydraulic system response capability and structural strength limitations, forming a complete depth-adjusted nonlinear programming model. By adaptively adjusting the weight coefficients based on the historical torque fluctuation variance, the optimization model can dynamically adjust the priority of the optimization objectives according to the changes in actual working conditions, thus taking into account both the stability of the operation process and the operational effectiveness.
[0067] refer to Figure 5 and 6 In another preferred embodiment, when the edge controller calls the sequential quadratic programming algorithm to iteratively solve the nonlinear programming model in each control cycle, it uses the historically optimal soil loosening depth setpoint output from the previous control cycle as the initial iteration point for the current control cycle. This initial iteration point is close to the optimal solution under the current operating conditions, which can effectively reduce the number of iterations, improve the solution speed, and meet the timing requirements of real-time control. During each iteration, the edge controller performs a second-order Taylor expansion of the biobjective function in the nonlinear programming model at the current iteration point to generate the objective function of the quadratic programming subproblem. The expression of the second-order Taylor expansion is:
[0068] in, At the current iteration point, For the direction of iterative search, Let be the gradient vector of the biobjective function at the current iteration point. Let be the Hessian matrix of the bi-objective function at the current iteration point. The quadratic programming subproblem uses this second-order Taylor expansion as the objective function and the constraints after the first-order Taylor expansion as the constraint boundary.
[0069] The edge controller solves the quadratic programming subproblem to obtain the search direction and optimal step size for iteration. A line search algorithm is then used to determine the optimal step size that satisfies the constraints, and the current iteration point is updated according to the search direction and step size. After each iteration update, the edge controller calculates the absolute value of the difference between the two objective functions in two adjacent iterations. When this absolute value is less than a preset convergence threshold, the iteration process stops, and the current iteration point is output as the optimal soil loosening depth setting value for the current moment. The edge controller presets a maximum number of iterations. When the preset maximum number of iterations is reached, the iteration process stops regardless of whether the convergence threshold is met, and the current iteration point is output as the optimal soil loosening depth setting value. This prevents the iteration process from exceeding the control cycle time limit and ensures real-time control.
[0070] The edge controller converts the optimal soil loosening depth setpoint into a smooth hydraulic proportional valve current control sequence using a trajectory interpolation algorithm. The trajectory interpolation algorithm employs cubic spline interpolation. The edge controller uses the optimal soil loosening depth setpoint of the current control cycle, the optimal soil loosening depth setpoint of the previous control cycle, and the optimal soil loosening depth setpoints of the two previous control cycles as spline nodes to construct a cubic spline interpolation function. This function is a cubic polynomial within each segmented interval, and its first and second derivatives are continuous at the nodes, ensuring a smooth depth trajectory without abrupt changes. The expression for the cubic spline interpolation function is:
[0071] in, For cubic spline interpolation function, corresponding to time... The target depth value, For the first The time corresponding to each spline node , , , For the first The polynomial coefficients for each segmented interval are obtained by solving for the depth values of the spline nodes, the continuity conditions of the first and second derivatives. The edge controller divides a single control cycle into multiple interpolation cycles with equal time intervals. It calculates the target depth value corresponding to each interpolation cycle using a cubic spline interpolation function. Based on the pre-calibrated hydraulic proportional valve current-displacement characteristic curve, the target depth value is converted into the corresponding hydraulic proportional valve current value, generating a hydraulic proportional valve current control sequence.
[0072] After generating the hydraulic proportional valve current control sequence, the edge controller performs dead-zone compensation on the current control sequence. The edge controller pre-acquires the current-flow characteristic curve of the hydraulic proportional valve and extracts the current interval where the flow rate remains zero as the dead-zone current interval. The dead-zone current interval includes a forward dead zone and a reverse dead zone, corresponding to the unresponsive current intervals during the extension and retraction of the hydraulic cylinder, respectively. The edge controller calculates the second difference between each current control point in the hydraulic proportional valve current control sequence and the center point of the dead-zone current interval. Based on the relationship between the absolute value of the second difference and half the width of the dead-zone current interval, the current control points are corrected. The piecewise function expression for dead-zone compensation is:
[0073] in, These are the original current control point values in the current control sequence. This represents the current control point value after dead-zone compensation. The value is the value at the center point of the dead zone current range. This represents the upper boundary value of the dead zone current range. This represents the width of the dead zone current interval. When the absolute value of the second difference is less than half the width of the dead zone current interval, it indicates that the current control point falls within the dead zone interval, and the edge controller forcibly modifies the value of the current control point to the upper boundary value of the dead zone current interval; when the absolute value of the second difference is greater than or equal to half the width of the dead zone current interval, it indicates that the current control point is outside the dead zone interval, and the edge controller keeps the value of the current control point unchanged.
[0074] Table 4. Dead Zone Compensation Parameters and Current Correction Rules for Hydraulic Proportional Valves
[0075] Table 4 is used to clarify the dead zone determination results and correction rules corresponding to different current control points. The dead zone interval parameters in the table are obtained through the current-flow characteristic test calibration of the hydraulic proportional valve to ensure the accurate execution of dead zone compensation, eliminate the non-response hysteresis error of the hydraulic valve when a small current input is applied, and improve the control accuracy of depth adjustment.
[0076] The edge controller outputs each current value in the current control sequence after dead zone compensation to the hydraulic proportional valve according to the interpolation cycle. The hydraulic proportional valve adjusts the valve opening according to the input current value, controls the extension and retraction speed and displacement of the hydraulic cylinder, and realizes smooth adaptive adjustment of the soil loosening depth of the tool.
[0077] In this embodiment, a sequential quadratic programming algorithm is used to achieve rapid iterative solution of the nonlinear programming model, ensuring the real-time performance and optimality of the depth adjustment decision; a smooth depth control trajectory is generated by a cubic spline trajectory interpolation algorithm, avoiding system shocks caused by sudden depth changes; and dead zone compensation processing eliminates the hysteresis error of the hydraulic proportional valve, improving the control accuracy and response speed of the depth adjustment.
Claims
1. A digital processing system for adaptive soil loosening depth of a tillage machine, characterized in that, The device includes an edge controller, a torque sensor, a triaxial accelerometer, and a hydraulic proportional valve. The edge controller is electrically connected to the torque sensor, the triaxial accelerometer, and the hydraulic proportional valve, respectively. The edge controller is configured to perform the following steps: real-time acquisition of the tool spindle torque signal sequence collected by the torque sensor and the frame triaxial acceleration signal sequence collected by the triaxial accelerometer. Fast Fourier transform is performed on the torque signal sequence and the frame triaxial acceleration signal sequence using a sliding window to extract the torque frequency domain feature vector and the vibration frequency domain feature vector. The torque frequency domain feature vector and the vibration frequency domain feature vector are input into the soil type identification classifier, which outputs the set of dynamic parameters of the current cutting soil. Substituting the set of dynamic parameters into the state-space equation of soil-cutter interaction, with minimizing the torque fluctuation variance and maximizing the effective soil breaking volume as the dual objective functions, and with the maximum response rate of the hydraulic system and the maximum allowable structural stress as constraints, a nonlinear programming model for depth adjustment is constructed. Within each control cycle, the sequential quadratic programming algorithm is invoked to iteratively solve the nonlinear programming model and output the optimal soil loosening depth setting value at the current moment. The optimal loosening depth setting is converted into a smooth hydraulic proportional valve current control sequence of the digital processing system for the adaptive tillage machine's loosening depth using a trajectory interpolation algorithm, and then sent to the hydraulic proportional valve.
2. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 1, characterized in that, The specific method for performing Fast Fourier Transform on the torque signal sequence and the rack triaxial acceleration signal sequence through a sliding window includes: dividing the sliding window with a fixed number of sampling points, and applying Hanning window functions to the torque signal sequence and the rack triaxial acceleration signal sequence respectively within the sliding window; Perform a Fast Fourier Transform on the windowed torque signal sequence and the frame triaxial acceleration signal sequence to obtain initial torque spectrum data and initial vibration spectrum data; The initial torque spectrum data and the initial vibration spectrum data are subjected to amplitude squared processing to extract the spectral peaks and corresponding frequency values within a preset frequency range. The spectral peaks and the frequency values are then combined to generate the torque frequency domain feature vector and the vibration frequency domain feature vector.
3. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 1, characterized in that, The soil type identification classifier is a support vector data description classifier. The process of inputting the torque frequency domain feature vector and the vibration frequency domain feature vector into the soil type identification classifier includes: concatenating the torque frequency domain feature vector and the vibration frequency domain feature vector to generate a high-dimensional joint feature vector. The high-dimensional joint feature vector is input into the pre-trained support vector data description classifier, and the high-dimensional joint feature vector is mapped to a high-dimensional feature space through a Gaussian kernel function; In the high-dimensional feature space, calculate the distance from the high-dimensional joint feature vector to the center of the hypersphere, select the soil category label corresponding to the hypersphere with the smallest distance, retrieve the internal friction angle, cohesion and soil bulk density parameters associated with the soil category label, and combine the internal friction angle, the cohesion and the soil bulk density parameters into the dynamic parameter set.
4. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 1, characterized in that, The method of constructing a nonlinear programming model for depth regulation with the dual objective functions of minimizing torque fluctuation variance and maximizing effective soil breaking volume includes: establishing a discretized state-space equation containing tool cutting depth state variables based on the set of dynamic parameters, and using the tool cutting depth state variables as decision variables; The first objective function term is the variance calculation formula of the predicted torque sequence output by the discretized state space equation, and the second objective function term is the product integral calculation formula of the tool cutting depth state variable, tool width and grounding machine forward speed. The first objective function term and the second objective function term are assigned fixed weight coefficients and then linearly weighted and summed to generate the dual objective function.
5. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 1, characterized in that, The method of using the maximum response rate of the hydraulic system and the maximum allowable structural stress as constraints includes: obtaining the rodless chamber area and rod chamber area of the hydraulic cylinder and the rated displacement of the hydraulic pump, and combining the rated flow of the hydraulic proportional valve to calculate the first upper limit value of the change in the tool spindle depth between adjacent control cycles, and constructing the first inequality constraint; Obtain the moment of inertia of the section of the tillage machine cutter arm and the ultimate yield strength of the material. Combine the predicted soil resistance value in the set of dynamic parameters to calculate the bending moment borne by the root of the cutter arm. Set a second upper limit value for the bending moment and construct a second inequality constraint. The first inequality constraint and the second inequality constraint are used as the boundary constraints of the nonlinear programming model.
6. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 1, characterized in that, The method of using the sequential quadratic programming algorithm to iteratively solve the nonlinear programming model includes: using the historical optimal soil loosening depth setpoint output by the control cycle as the initial iteration point of the current control cycle; In each iteration, the biobjective function in the nonlinear programming model is expanded in second order Taylor at the current iteration point to generate a quadratic programming subproblem; Solve the quadratic programming subproblem to obtain the search direction and step size, update the current iteration point, and stop iterating when the absolute value of the difference between the two objective functions in two adjacent iterations is less than the preset convergence threshold, and output the current iteration point as the optimal soil loosening depth setting value.
7. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 2, characterized in that, Before dividing the sliding window with a fixed number of sampling points, the process further includes preprocessing the torque signal sequence and the frame triaxial acceleration signal sequence: extracting each data point in the torque signal sequence within the sliding window, and calculating the first difference between each data point and the mean torque within the sliding window. Data points whose absolute value of the first difference is greater than a first preset multiple of the standard deviation are removed, and data at the positions of the removed data points are supplemented using a linear interpolation algorithm to generate a first preprocessed torque signal sequence. The frame triaxial acceleration signal sequence within the sliding window is processed using the same elimination criteria and interpolation algorithm as the first preprocessed torque signal sequence to generate the first preprocessed acceleration signal sequence.
8. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 3, characterized in that, After retrieving the internal friction angle, cohesion, and soil bulk density parameters associated with the soil category label, the method further includes online updating of the support vector data description classifier: constructing online sample pairs containing the high-dimensional joint feature vector and the soil category label, and storing the online sample pairs in a circular buffer queue. When the number of online sample pairs in the circular buffer queue reaches a preset batch number, the preset batch number of online sample pairs is extracted. Using the coordinates and radius of the hypersphere's center as the initial values, the Lagrange multiplier solution process is re-executed using the preset batch number of online samples to update the coordinates and radius of the hypersphere's center, and the updated parameters of the Gaussian kernel function are saved.
9. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 4, characterized in that, The method of linearly weighting and summing the first objective function term and the second objective function term by assigning fixed weight coefficients is replaced by: extracting the historical predicted torque sequence output by the discretized state space equation in the previous control cycle, and calculating the variance of the historical predicted torque sequence as the historical variance index; The historical variance index is input into a preset piecewise mapping function, wherein the piecewise mapping function uses the historical variance index as the independent variable and the first weight coefficient corresponding to the first objective function term and the second weight coefficient corresponding to the second objective function term as the dependent variables. When the historical variance index is greater than the preset segmentation threshold, the first weight coefficient is increased and the second weight coefficient is decreased; when the historical variance index is less than the preset segmentation threshold, the first weight coefficient is decreased and the second weight coefficient is increased.
10. The digital processing system for the soil loosening depth of the adaptive tillage machine according to claim 6, characterized in that, After converting the optimal soil loosening depth setting value into a smooth hydraulic proportional valve current control sequence using a trajectory interpolation algorithm, the method further includes dead zone compensation processing of the hydraulic proportional valve current control sequence: obtaining the current-flow characteristic curve of the hydraulic proportional valve, and extracting the current interval in the current-flow characteristic curve where the flow rate remains zero as the dead zone current interval. Calculate the second difference between each current control point in the hydraulic proportional valve current control sequence and the center point of the dead zone current interval; When the absolute value of the second difference is less than half the width of the dead zone current interval, the value of the current control point is forcibly modified to the upper limit boundary value of the dead zone current interval. When the absolute value of the second difference is greater than or equal to half the width of the dead zone current range, the value of the current control point remains unchanged.