A method for controlling a combine harvester cleaning system based on trend prediction

Abstract

The application discloses a kind of based on trend prediction's combined harvester cleaning system control method, comprising the following steps: construct fuzzy predictor, according to expert experience library, predict the future prediction trend of the control action of harvester cleaning system to cleaning performance;The language trend of the fuzzy inference output of fuzzy predictor is equalized into real utility value;According to real utility value, construct total utility target function J (U), control action sequence when J (U) maximization is solved.The application advantage is to predict the influence trend of " control action to impurity rate / loss rate" using fuzzy rule base, and the trend is converted into symbolic reward / penalty in MPC optimization target, so that robust, smooth cleaning parameter adaptive adjustment is realized under the condition of model and sensing limitation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for agricultural machinery, and specifically relates to a control method for a combine harvester cleaning system based on trend prediction. Background Technology

[0002] In agricultural harvesting machinery, the core objective of the winnowing cleaning system is to minimize grain loss while ensuring a low impurity content. Current mainstream control strategies include: (1) Fixed parameter control: relies on manual adjustment and cannot adapt to changes in crops, humidity and feed amount; (2) PID feedback control: requires high-precision online sensors (such as near-infrared impurity detection), which are costly and susceptible to dust interference; (3) Pure fuzzy control: Although it can incorporate expert experience, it lacks multi-step forward-looking capability and is prone to control oscillation; (4) Data-driven MPC (such as LSTM-MPC): relies on a large amount of high-quality training data, and has poor generalization in agricultural machinery with small sample size and strong noise.

[0003] Therefore, there is an urgent need for a new control architecture that does not rely on high-precision models and sensors but can achieve multi-step optimization. Summary of the Invention

[0004] This application provides a control method for a combine harvester cleaning system based on trend prediction. Its advantage is that it uses a fuzzy rule base to predict the "influence trend of control actions on impurity / loss rate" (rather than the absolute value) and transforms the trend into a symbolic reward / penalty in the MPC optimization objective, thereby achieving robust and stable adaptive adjustment of cleaning parameters under model and sensor limit conditions.

[0005] The above-mentioned objective of this application is achieved through the following technical solution: A control method for a combine harvester cleaning system based on trend prediction includes the following steps: S1: Construct a fuzzy predictor to predict the future trend of the control actions of the harvester cleaning system on the cleaning performance based on the expert experience base; S2: Construct a trend-utility mapping function Φ(·) to convert the linguistic trend hierarchy output of the fuzzy predictor into real utility values; S3: Let the prediction time domain be N steps. In each control period t, enumerate all physically feasible candidate control action sequences U = [u(t), u(t+1), ..., u(t+N-1)] of length N. For each sequence U, according to the fuzzy predictor, obtain the trend sequence of changes in impurity rate C and loss rate L in the next N steps. Construct the total utility objective function J(U) based on the real utility value, and find the control action sequence that maximizes J(U). S4: Decode the control action sequence when J(U) is maximized and send it to the harvester cleaning system for execution.

[0006] Furthermore, in step S1: The control actions of the harvester cleaning system include fan operation, vibrating screen operation, and vibrating screen opening and closing angle adjustment. The cleaning performance of the harvester cleaning system includes impurity content level and loss rate level, wherein: Fan actions include decreasing, maintaining, and increasing; The actions of a vibrating screen include decreasing, maintaining, and increasing the frequency; The opening and closing angle adjustment of the vibrating screen includes decreasing, maintaining, and increasing. The impurity level and loss level use a five-level language set: {significant decrease, slight decrease, essentially unchanged, slight increase, significant increase}.

[0007] Furthermore, in step S2, the specific mapping relationship of the trend-utility mapping function Φ(·) is set as follows: "Significant decrease" → +2.0; "Slight decrease" → +1.0; "Basically unchanged" → 0.0; "Slight increase" → -1.0; "Significant increase" → -2.0.

[0008] Furthermore, in step S3, the control action sequence U is dynamically trimmed to exclude actions that are invalid or even harmful under certain operating conditions.

[0009] Furthermore, in step S3, the total utility objective function J(U) is:

[0010] in: and These are the utility values ​​corresponding to the trends in impurity and loss rates predicted in step k, respectively. and Weighting coefficients that can be configured by the user; γ is the discount factor, 0 < γ ≤ 1.

[0011] Furthermore, in step S3, a control smoothness reward term is introduced into the total utility objective function: for any candidate action sequence, if the control instructions of its two adjacent steps are the same, a positive utility reward is given; otherwise, if an unnecessary switch occurs, a negative utility penalty is imposed.

[0012] Furthermore, in step S4: The fan operation is decoded into fan control signals: the "increase / decrease / hold" command is mapped to the target speed setpoint of the fan inverter.

[0013] The vibration screen action is decoded into a vibration screen control signal: the "increase screen frequency" command is converted into the target frequency setting value of the vibration motor; The opening and closing angle adjustment of the vibrating screen is decoded into a servo motor for adjusting the screen body tilt angle, which drives the screen to move towards the preset angle.

[0014] Furthermore, in step S4, a closed-loop execution logic in the rolling time domain is used to control the harvester cleaning system: In each control cycle t, the five-dimensional status of the cleaning operation is collected in real time: impurity content, loss rate, air volume, screening frequency, and screening angle, and these are quantified into the levels required for fuzzy inference. Based on the fuzzy predictor and the total utility objective function, the optimal action sequence U* for the next N steps is searched within the clipped action space. Only the first action u*(t) in the optimal sequence U* is extracted and executed, decoded, and then sent to the execution mechanism; Wait until the next control cycle t+1, and then repeat the above steps; using the latest status information, update the expert database based on the actual measured loss and mixed data.

[0015] In summary, the beneficial effects of this application are as follows: 1. Utilize a fuzzy rule base to predict the "trend of the impact of control actions on impurity rate / loss rate" (rather than the absolute value), and transform this trend into a symbolic reward / penalty in the MPC optimization objective, thereby achieving robust and stable adaptive adjustment of cleaning parameters under model and sensor limit conditions; 2. Through a rolling implementation and continuous correction mechanism, the controller can continuously use the latest field feedback to correct its predictions and decisions, effectively overcoming the impact of factors such as model mismatch and external disturbances (such as sudden changes in crop humidity and feed amount), thereby ensuring the long-term stability and optimal cleaning performance.

[0016] 3. Through the above three key technical features—trend prediction, utility optimization, and rolling execution—this invention constructs a complete intelligent control solution suitable for strongly nonlinear agricultural equipment, which significantly improves the automation level and operation quality of the wind-screen cleaning system. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0018] The specific embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] Example: A trend-prediction-based control method for a combine harvester cleaning system, used for the coordinated adjustment of actuators such as wind speed, vibrating screen frequency, and screen opening angle in the combine harvester cleaning system. This method encodes expert experience into a fuzzy rule base to predict the future trend of cleaning performance (impurity and loss rate) affected by control actions (e.g., "significant decrease" or "slight increase"), rather than predicting absolute values. Regarding control action execution, the control actions for the fan, vibrating screen frequency, and screen opening are decomposed into three types: increase, maintain, and decrease, all of which are directional actions rather than specific target values. Therefore, the system has only 26 action options (actions where all are maintain actions should be excluded, as there is no action in this case). In each control cycle, the system first acquires and fuzzifies the current cleaning state; then, for the discretized candidate action sequence, it calls a fuzzy inference engine to predict the trend of impurity and loss rate changes after multiple steps; next, it clarifies the linguistic trend into numerical rewards or penalties through a preset cost function; finally, under the premise of satisfying the physical constraints of the actuators, it solves for the control sequence that optimizes the overall trend utility, and only executes the first step of the control. The above process is repeated in the next cycle to achieve rolling optimization and closed-loop control. After the action is completed, before the start of the next cycle, the loss sensor and impurity sensor detect the changes in impurity and loss caused by the previous action. After normalizing the data, it is updated to the expert database at a certain update rate to achieve online optimization.

[0020] This architecture effectively integrates the robustness of fuzzy logic with the foresight of model predictive control, and utilizes various action comparison optimization methods of model predictive control to effectively suppress time delay and oscillation problems, significantly improving the adaptive capability and control stability of the cleaning system under complex field conditions.

[0021] like Figure 1 The trend-prediction-based control method for the combine harvester cleaning system includes the following steps: S1: Construct a fuzzy predictor to predict the future trend of the control actions of the harvester cleaning system on the cleaning performance based on the expert experience base.

[0022] The control actions of the harvester cleaning system include fan operation, vibrating screen operation, and vibrating screen opening and closing angle adjustment. The cleaning performance of the harvester cleaning system includes impurity content level and loss rate level, wherein: Fan actions include decreasing, maintaining, and increasing; The actions of a vibrating screen include decreasing, maintaining, and increasing the frequency; The opening and closing angle adjustment of the vibrating screen includes decreasing, maintaining, and increasing. The impurity level and loss level use a five-level language set: {significant decrease, slight decrease, essentially unchanged, slight increase, significant increase}.

[0023] The fuzzy inference system of this invention is specifically designed to predict the future trends of cleaning performance indicators (impurity rate, loss rate) after applying specific control actions. Its inputs fully encompass the five-dimensional current state of the cleaning system: Impurity level (low / medium / high) Loss rate level (low / medium / high) Current airflow level (low / medium / high) Current vibrating screen frequency (low / medium / high) Current vibrating screen angle (small / medium / large) The candidate control actions to be evaluated are composed of independent adjustments and combinations of the fan and the vibrating screen, forming a total of 3 × 3 = 9 basic action pairs.

[0024] However, considering the strong nonlinear characteristics of the system, some actions may be ineffective or even harmful under certain operating conditions. Therefore, in actual reasoning, the action space is dynamically trimmed, that is, the control action sequence U is dynamically trimmed to exclude actions that are ineffective or even harmful under certain operating conditions. For example, when the air volume is already at a "high" level, the "fan increase" action will be disabled; when the screen angle has reached a "large" level, the "screen angle increase" action will also be excluded, thereby avoiding ineffective or dangerous operations.

[0025] The core of the fuzzy rule base lies in establishing a mapping relationship of "five-dimensional state + two-dimensional action → two-indicator trend". The rule form is as follows: Rule 1 (Suppressing airflow in high-volume areas) IF (Impurity = High) AND (Loss = High) AND (Air Volume = High) AND (Screening Frequency = Medium) AND (Screening Angle = Medium) AND (Fan Action = Increase) AND (Screening Frequency Action = Maintain) AND (Screening Angle Action = Maintain) THEN (Δ Impurity = Slightly Increased) AND (Δ Loss = Significantly Increased) Note: Continuing to increase the airflow under high air volume will not have a cleaning effect, but will instead exacerbate air loss.

[0026] Rule 2 (Actively increase airflow in low-airflow areas) IF (Impurity = High) AND (Loss = Low) AND (Air Volume = Low) AND (Screening Frequency = Medium) AND (Screening Angle = Medium) AND (Fan Operation = Increase) AND (Screening Frequency Operation = Maintain) AND (Screening Angle Operation = Maintain) THEN (Δ impurity = significant decrease) AND (Δ loss = slight increase) Note: Increasing the airflow at low air volume can effectively separate impurities, although losses will increase slightly but remain manageable.

[0027] All output trends use a five-level language set: {significant decrease, slight decrease, basically unchanged, slight increase, significant increase}. This mechanism does not generate control instructions; it only serves as a multi-step trend prediction engine, providing MPC with the expected performance evaluation basis for each candidate action sequence, thereby achieving intelligent and robust rolling optimization in complex nonlinear environments.

[0028] Building upon fuzzy trend predictors, this invention further proposes a novel model predictive control (MPC) rolling optimization framework. The core innovation of this framework lies in the fact that its optimization objective is not based on precise numerical predictions of the system state, but rather entirely on the utility value mapped by verbal trend predictions. This enables the control strategy to achieve intelligent and robust optimization of cleaning performance even in the absence of high-precision sensors and complex mathematical models.

[0029] S2: Construct a trend-utility mapping function Φ(·) to convert the linguistic trend hierarchy output of the fuzzy predictor into real-valued utility values; to reflect the contribution of this change to system performance. The specific mapping relationship is set as follows: "Significant decrease" → +2.0; "Slight decrease" → +1.0; "Basically unchanged" → 0.0; "Slight increase" → -1.0; "Significant increase" → -2.0.

[0030] This mapping reflects a clear preference for control objectives: a reduction in impurity and loss rates is a positive gain, and the greater the reduction, the higher the utility; conversely, a deterioration in these indicators brings negative utility.

[0031] S3: Let the prediction time domain be N steps. In each control period t, enumerate all physically feasible candidate control action sequences U = [u(t), u(t+1), ..., u(t+N-1)] of length N. For each sequence U, according to the fuzzy predictor, obtain the trend sequence of changes in impurity rate C and loss rate L in the next N steps. Construct the total utility objective function J(U) based on the real utility value. Its design aims to maximize the comprehensive performance gain in the entire prediction time domain. Find the control action sequence when J(U) is maximized.

[0032] The total utility objective function J(U) is:

[0033] in: and These are the utility values ​​corresponding to the trends in impurity and loss rates predicted in step k, respectively. and This is a user-configurable weighting factor used to dynamically balance cleaning quality (low impurity content) and harvesting efficiency (low loss) based on operational needs. For example, when harvesting high-value seed crops, a weighting factor of > can be set to prioritize low loss. γ is a discount factor (0 < γ ≤ 1), used to reflect the characteristic that prediction uncertainty increases over time. By setting γ ≤ 1 (e.g., 0.9), the optimization process can focus more on recent and more reliable prediction results, thus enhancing the robustness of the system.

[0034] To avoid mechanical wear or system oscillations caused by frequent and abrupt action switching in the actuator, a control smoothness reward term is introduced into the overall utility objective function: for any candidate action sequence, a positive utility reward is given if the control commands of its two adjacent steps are the same; conversely, a negative utility penalty is imposed if unnecessary switching occurs. Unnecessary switching includes "increase" followed immediately by "decrease," and "decrease" followed immediately by "increase." This design encourages the generation of smooth and stable control commands.

[0035] More importantly, the generation of all candidate action sequences strictly follows the aforementioned nonlinear operating condition action pruning rules. For example, when the current airflow is at a "high" level, any sequence containing the "fan increase" action will be directly excluded from the candidate set. This hard constraint ensures that all evaluated and ultimately executed control commands are physically safe and effective, fundamentally avoiding the risk of unnecessary operations in the system saturation or ineffective regions.

[0036] In summary, the proposed method successfully combines expert knowledge-driven qualitative prediction with the rolling optimization concept of modern MPC, constructing an intelligent control framework that is independent of precise models, robust to sensor noise, and adaptable to system nonlinearity. The designed total utility objective function J(U) provides the controller with a clear optimization direction, enabling it to make optimal decisions in complex multi-objective trade-offs. In each control cycle, this objective function drives a rolling optimization process, selecting the sequence with the highest overall utility from all physically feasible action sequences and executing its first action, thereby achieving closed-loop intelligent control.

[0037] S4: Decode the control action sequence when J(U) is maximized and send it to the harvester cleaning system for execution. That is, transform the abstract "optimal action" output by the MPC optimizer into specific physical control instructions and complete the timing logic of the entire control closed loop to ensure the stable and reliable operation of the system.

[0038] The optimal action output by the MPC optimizer is a symbolic pair of instructions, such as {fan: increase, screen frequency: maintain, screen angle: increase}. The control system needs to convert this into signals that can be recognized by the underlying actuators through an instruction decoding module. The fan action is decoded into fan control signals: "Increase / Decrease / Hold" commands are mapped to the target speed setting value of the fan inverter; for example, "Increase" corresponds to adding a preset increment ΔRPM_fan to the current speed.

[0039] The vibration screen action is decoded into vibration screen control signals: the "increase screen frequency" command is converted into the target frequency setting value of the vibration motor; the "increase screen angle" command is sent to the screen body tilt angle adjustment servo motor, driving it to move in the direction of the preset angle increment Δθ.

[0040] The harvester cleaning system is controlled using a rolling time-domain closed-loop execution logic. In each control cycle t, the five-dimensional status of the cleaning operation is collected in real time: impurity content, loss rate, air volume, screening frequency, and screening angle, and these are quantified into the levels required for fuzzy inference. Based on the fuzzy predictor and the total utility objective function, the optimal action sequence U* for the next N steps is searched within the clipped action space. Only the first action u*(t) in the optimal sequence U* is extracted and executed, decoded, and then sent to the execution mechanism; Wait until the next control cycle t+1, and then repeat the above steps; using the latest status information, update the expert database based on the actual measured loss and mixed data.

[0041] In this specific embodiment, the method is deployed in a combine harvester cleaning system. The hardware includes: a rear-mounted machine vision camera for estimating impurity and loss rates; a wind speed sensor installed in the air duct (or indirectly obtaining airflow through fan speed); and an encoder for real-time monitoring of the vibrating screen's operating frequency and screen tilt angle. All sensor signals are connected to an embedded main controller (such as an ARM Cortex-A53 industrial control module) via a CAN bus. The controller outputs control commands to the fan inverter, vibrating motor driver, and screen angle adjustment servo motor, forming a closed loop. The signal flow is: sensor acquisition → controller fuzzification and prediction → MPC optimization → command decoding → actuator action → next cycle feedback, constituting a complete rolling time-domain control loop.

[0042] In this embodiment, the fuzzy trend predictor is designed as a feedforward inference engine. Its function is to predict the qualitative trends of impurity rate and loss rate within a short future time window, given the current system state and a candidate control action to be evaluated. This prediction result is not used to directly drive the actuator, but rather serves as the basis for evaluating the objective function of multi-step rolling optimization (MPC).

[0043] The predictor's input is a seven-dimensional combination: • Five-dimensional current status: Impurity level (low / medium / high), Loss rate level (low / medium / high), Air volume level (low / medium / high), Vibrating screen frequency (low / medium / high), Screen body angle (small / medium / large); • Three-dimensional candidate actions: fan action (decrease / hold / increase), vibrating screen compound action (composed of screen frequency action {decrease / hold / increase} and screen angle action {decrease / hold / increase}, a total of 26 combinations (all actions that are held are considered no action and should be deleted).

[0044] All linguistic variables are fuzzified using a triangular membership function, and the universe of discourse parameters are determined based on field calibration data from combine harvesters. For example, the sieve angle "small" corresponds to [0°, 0°, 8°], "medium" to [6°, 10°, 14°], and "large" to [12°, 18°, 18°].

[0045] The fuzzy rule base is constructed in the form of "IF (state 1 is X) AND ... AND (action 2 is Y) THEN (Δ impurity is T1) AND (Δ loss is T2)", covering typical working conditions and boundary scenarios. The output trend adopts a five-level language set: {significant decrease, slight decrease, basically unchanged, slight increase, significant increase}, and is mapped to the numerical values ​​{-2, -1, 0, +1, +2} for subsequent calculations.

[0046] To address the system's strong nonlinearity, a dynamic action pruning mechanism is introduced: before generating candidate action sequences, invalid or dangerous actions are disabled in real time based on the current state. For example, when the airflow is high, "increase fan speed" is prohibited; when the sieve angle is large, "increase sieve angle" is prohibited. This mechanism ensures that all actions fed into the fuzzy predictor are physically feasible and safe operations.

[0047] Through this design, the fuzzy trend predictor can effectively capture the nonlinear dynamic characteristics of the cleaning system in the absence of a precise mathematical model, providing reliable trend prediction capabilities for MPC.

[0048] In this embodiment, the MPC rolling optimizer uses a fuzzy trend predictor as the core prediction model to construct a finite-time open-loop optimization problem. In each control cycle, the optimizer first generates a set of physically feasible candidate control sequences from the action space, where each U is a three-dimensional action vector (fan action, vibrating screen frequency, screen opening action), and invalid terms have been removed through the aforementioned dynamic pruning mechanism.

[0049] For each candidate sequence, MPC calls a fuzzy trend predictor to perform multi-step forward simulation. Specifically, using the current state X[k] and the first action u[k] as input, it predicts the performance index change at the next time step (Δc[k+1], Δl[k+1]); then, using the updated state x[k+1] and u[k+1] as input, it predicts (Δc[k+2], Δl[k+2]), and so on, until the prediction time domain endpoint N[p]. Finally, the predicted values ​​of impurity rate and loss rate trajectory for the entire sequence are accumulated.

[0050] The optimization objective function J(Uk) is defined as the weighted sum of squared deviations between the predicted trajectory and the ideal target (e.g., impurity <3%, loss rate <1.5%), with a penalty term for the rate of change of the control variable added to ensure smooth action.

[0051] Where w c ,w l ρ is the weighting coefficient, and ρ is the smoothing factor.

[0052] The MPC solver traverses all feasible candidate sequences (or uses a heuristic search), selects the sequence that minimizes J(Uk), and sends only the first control action u[k] of that sequence to the execution layer. Subsequently, in the next cycle k+1, the above process is repeated based on the newly acquired state information to achieve rolling time-domain optimization.

[0053] The closed-loop workflow of this system is executed cyclically according to a fixed sampling period (e.g., 8-10 times per second). The specific steps are as follows: 1. Status awareness: The current impurity content ck and loss rate lk are estimated in real time through the rear vision camera; at the same time, data from the wind speed sensor, screen frequency encoder and tilt sensor are read to obtain vk, fk and θk.

[0054] 2. State classification: The above five-dimensional continuous quantities are mapped to the corresponding fuzzy language levels (low / medium / high or small / medium / large) to form the current system state description xk.

[0055] 3. Motion generation and trimming: Based on xk, all possible combinations of fan and vibrating screen motions are dynamically generated, and motions that are invalid or dangerous due to nonlinear characteristics are immediately trimmed (such as prohibiting air increase under high air volume).

[0056] 4. MPC Rolling Optimization: The MPC optimizer uses a fuzzy trend predictor to perform multi-step effect simulations on all possible action sequences and selects the optimal first action uk* based on the objective function.

[0057] 5. Command Decoding and Execution: The symbolic optimal action uk∗ (such as "increase air volume", "maintain vibrating screen frequency", "maintain screen opening") is decoded into specific physical control signals (such as the target speed of the fan inverter, the drive frequency of the vibrating motor, and the target angle of the servo motor), and sent to each actuator via the CAN bus.

[0058] 6. System Response and Waiting: After the actuator takes action, the system enters a waiting period until the next sampling cycle begins, at which point it re-perceives the state, forming a complete "perception-decision-execution-feedback" closed loop.

[0059] This process ensures that the control system can make full use of expert knowledge and field experience (through fuzzy rules) and achieve forward-looking and global intelligent decision-making through rolling optimization, effectively dealing with strong nonlinearity and time-varying disturbances in the combine harvester cleaning process.

[0060] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the protection scope of this application.

Claims

1. A control method for a combine harvester cleaning system based on trend prediction, characterized in that, Includes the following steps: S1: Construct a fuzzy predictor to predict the future trend of the control actions of the harvester cleaning system on the cleaning performance based on the expert experience base; S2: Construct a trend-utility mapping function Φ(·) to convert the linguistic trend hierarchy output of the fuzzy predictor into real utility values; S3: Let the prediction time domain be N steps. In each control period t, enumerate all physically feasible candidate control action sequences U = [u(t), u(t+1), ..., u(t+N-1)] of length N. For each sequence U, according to the fuzzy predictor, obtain the trend sequence of changes in impurity rate C and loss rate L in the next N steps. Construct the total utility objective function J(U) based on the real utility value, and find the control action sequence that maximizes J(U). S4: Decode the control action sequence when J(U) is maximized and send it to the harvester cleaning system for execution.

2. The control method for a combine harvester cleaning system based on trend prediction according to claim 1, characterized in that, In step S1: The control actions of the harvester cleaning system include fan operation, vibrating screen operation, and vibrating screen opening and closing angle adjustment. The cleaning performance of the harvester cleaning system includes impurity content level and loss rate level, wherein: Fan actions include decreasing, maintaining, and increasing; The actions of a vibrating screen include decreasing, maintaining, and increasing the frequency; The opening and closing angle adjustment of the vibrating screen includes decreasing, maintaining, and increasing. The impurity level and loss level use a five-level language set: {significant decrease, slight decrease, essentially unchanged, slight increase, significant increase}.

3. The control method for a combine harvester cleaning system based on trend prediction according to claim 2, characterized in that, In step S2, the specific mapping relationship of the trend-utility mapping function Φ(·) is set as follows: "Significantly decreased" → +2.0; "Slight decrease" → +1.0; "Basically unchanged" → 0.0; "Slight increase" → -1.0; "Significantly increased" → -2.

0.

4. The control method for a combine harvester cleaning system based on trend prediction according to claim 3, characterized in that, In step S3, the control action sequence U is dynamically trimmed to exclude actions that are invalid or even harmful under certain operating conditions.

5. The control method for a combine harvester cleaning system based on trend prediction according to claim 4, characterized in that, In step S3, the total utility objective function J(U) is: in: and These are the utility values ​​corresponding to the trends in impurity and loss rates predicted in step k, respectively. and Weighting coefficients that can be configured by the user; γ is the discount factor, 0 < γ ≤ 1.

6. The control method for a combine harvester cleaning system based on trend prediction according to claim 5, characterized in that, In step S3, a control smoothness reward term is introduced into the total utility objective function: for any candidate action sequence, if the control instructions of its two adjacent steps are the same, a positive utility reward is given; otherwise, if an unnecessary switch occurs, a negative utility penalty is imposed.

7. The control method for a combine harvester cleaning system based on trend prediction according to claim 4, characterized in that, In step S4: The fan operation is decoded into fan control signals: "increase / decrease / hold" commands are mapped to the target speed setpoint of the fan inverter; The vibration screen action is decoded into a vibration screen control signal: the "increase screen frequency" command is converted into the target frequency setting value of the vibration motor; The opening and closing angle adjustment of the vibrating screen is decoded into a servo motor for adjusting the screen body tilt angle, which drives the screen to move towards the preset angle.

8. The control method for a combine harvester cleaning system based on trend prediction according to claim 4, characterized in that, In step S4, the harvester cleaning system is controlled using closed-loop execution logic in the rolling time domain: In each control cycle t, the five-dimensional status of the cleaning operation is collected in real time: impurity content, loss rate, air volume, screening frequency, and screening angle, and these are quantified into the levels required for fuzzy inference. Based on the fuzzy predictor and the total utility objective function, the optimal action sequence U* for the next N steps is searched within the clipped action space. Only the first action u*(t) in the optimal sequence U* is extracted and executed, decoded, and then sent to the execution mechanism; Wait until the next control cycle t+1, and then repeat the above steps; using the latest status information, update the expert database based on the actual measured loss and mixed data.

Images