Multi-dimensional feature perception based experimental huller parameter adaptive matching method and system

Abstract

This invention relates to the field of automation control technology and discloses a multi-dimensional feature perception-based adaptive parameter matching method and system for experimental rice hullers. It aims to solve the problems of poor parameter matching accuracy, large transient processing errors, and the inability to calculate optimal parameters in advance based on rice characteristics in existing methods. The solution mainly includes: collecting rice data when the feed gate is locked and constructing a five-dimensional feature vector through orthogonal dimensionality reduction; instantaneously solving the optimal gap, rotational speed, and airflow using a fixed set of equations; calculating the wear compensation correction gap based on the cumulative processing volume; opening the gate after adjusting to the command value under no-load conditions and confirming stability; and closing the feedback loop during processing, allowing the rice to complete hulling in a constant parameter field. This invention improves parameter matching accuracy, saves experimental samples, and significantly increases the hulling integrity rate and head rice rate, making it particularly suitable for agricultural breeding, grain storage inspection, and quality assessment.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, specifically to a method and system for adaptive matching of parameters of an experimental rice huller based on multi-dimensional feature perception. Background Technology

[0002] Rice hulling is the first and crucial step in rice processing, and its effectiveness directly impacts subsequent milling, quality evaluation, and rice yield. Laboratory rice hullers, serving as benchmark equipment for agricultural breeding, grain storage inspection, and quality assessment, rely on a core hulling principle based on a pair of elastic rollers rotating in opposite directions at different speeds. Normal crushing pressure is generated by adjusting the physical gap between the rollers, while tangential tearing shearing force is generated by the difference in linear velocity between the fast and slow rollers, thus peeling away the rice husk. Therefore, accurately matching processing parameters such as roller gap, roller speed, and airflow according to the physical characteristics of the rice (e.g., geometric dimensions, moisture content) is key to ensuring high hulling and head rice yield. Unlike large-scale continuous industrial production, laboratory rice hullers are characterized by small sample sizes, short processing times, and irreversible destructive processes. If processing parameters are incorrect, the current batch of samples is wasted and cannot be reprocessed.

[0003] Existing experimental rice huller processing parameter matching methods typically employ the following three technical solutions:

[0004] The first method is the traditional manual parameter adjustment scheme based on experience. Most laboratory rice hullers rely on operators visually observing the rice variety and morphology, and manually adjusting the roller gap and speed by rotating handwheels based on experience. A few devices (such as the experimental rice huller disclosed in CN223404988U) have introduced electric quantitative adjustment mechanisms, but other parameters still rely on manual settings. This method cannot quantitatively perceive hidden physical properties such as moisture content and internal mechanical characteristics. Different batches of rice have significant differences in microscopic toughness, and it is difficult to establish a stable mapping relationship based on experience, resulting in poor repeatability and accuracy of parameter matching.

[0005] The second type is the conventional mechanical online dynamic closed-loop matching scheme. Some experimental rice hullers adopt the closed-loop control logic of industrial continuous production lines: after the material enters the rolling zone, the system monitors the motor feedback current or broken rice rate in real time. When an anomaly is detected, the PID algorithm is used to reversely adjust the roller gap or speed (such as the predictive control method, device, equipment and storage medium for rice hullers disclosed in CN117205989A). This type of scheme is subject to physical constraints in laboratory sample preparation scenarios, such as small sample size, short processing time, and irreversible damage. The actuators (such as stepper motors and frequency converters) have inherent physical hysteresis. When the system detects the error and completes parameter adjustment, most of the sample has already passed through the rolling zone under non-optimal transition parameters, resulting in sample crushing or incomplete hulling, leading to distorted test data. This scheme is essentially suitable for continuous large-scale production, but not for rapid testing of small samples.

[0006] The third type is an online control scheme based on artificial intelligence. Its main principle is to use algorithms such as reinforcement learning for closed-loop control (e.g., a reinforcement learning-based milling process control method disclosed in CN121232595A), allowing the machine to continuously fine-tune neural network parameters based on feedback errors during actual processing, achieving online self-learning. This type of scheme requires a large number of online trial-and-error iterations to converge. In a laboratory setting, each batch of rice samples is independent and irreversible, making repeated trial and error impossible. Within a processing window of 15-20 seconds, reinforcement learning algorithms cannot effectively explore and converge; instead, parameter oscillations may cause complete sample destruction. Summary of the Invention

[0007] This invention aims to solve the problems of poor parameter matching accuracy, large transition state processing error, and inability to calculate the optimal parameters in advance based on the characteristics of rice in existing experimental rice hulling machine parameter matching methods. It proposes a multi-dimensional feature perception-based experimental rice hulling machine parameter adaptive matching method and system.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] In a first aspect, the present invention provides a multi-dimensional feature-aware experimental rice huller parameter adaptive matching method, the method comprising:

[0010] Step 1: Static pre-detection and feature extraction:

[0011] With the feed gate physically locked, raw sensing data of the rice sample to be processed is collected in the feed detection chamber, and the raw sensing data is subjected to orthogonal dimensionality reduction and statistical purification to construct a multidimensional orthogonal feature vector matrix.

[0012] Step 2, Online forward solution:

[0013] The multidimensional orthogonal feature vector matrix is ​​input into a set of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. This set of equations is pre-obtained through offline optimization training and includes fixed membership function parameters, conclusion parameters, and feature importance weight parameters. The lower-level execution unit solves the set of parallel multi-channel forward continuous equations through algebraic operations and outputs the theoretically optimal instruction matrix for the current batch of rice samples. ;in, The target rubber roller displacement clearance. The target linear velocity of the fast roller, The target linear velocity of the slow roller. For the target linear velocity difference, To separate air volume for the target;

[0014] Step 3, Cross-batch wear compensation:

[0015] Obtain the total processing power equivalent of the experimental rice huller since the last zero-point calibration. The wear compensation increment is calculated based on pre-calibrated empirical wear fitting coefficients. The theoretically optimal instruction matrix The target rubber roller displacement gap With the wear compensation increment Superimpose the data to generate the final physical displacement command. ;

[0016] Step 4: Execute the no-load command:

[0017] The drive roller gap adjustment mechanism executes the final physical displacement command. The frequency converters driving the fast and slow rollers respectively execute the target linear speed of the fast roller. and the target linear velocity of the slow roller Drive the separator fan to execute the target separation air volume ;

[0018] Step 5: No-load steady-state closed-loop confirmation:

[0019] In the no-load state, continuously poll the status register of the underlying driver to obtain the actual physical displacement. Actual rotational speed of the fast roller and the actual speed of the slow roller When determining the actual physical displacement With final physical displacement command The deviation, the actual speed of the fast roller Deviation from the corresponding target speed, actual speed of the slow roller The deviation from the corresponding target speed simultaneously enters the preset steady-state tolerance dead zone, and the duration exceeds the preset anti-shake filtering time threshold. At that time, output a global steady-state flag signal;

[0020] Step 6, Feeding and Processing:

[0021] In response to the global steady-state flag signal, the feed gate is opened, allowing the rice sample to be processed to be moved by the final physical displacement command. Target linear velocity of the fast roller target linear velocity of slow roller and target separation air volume The shell removal process is completed under a defined constant physical field; within the processing window, all parameter correction loops based on real-time feedback are forcibly shut down.

[0022] Furthermore, the multidimensional orthogonal eigenvector matrix is ​​a five-dimensional orthogonal eigenvector matrix. ;in, The average aspect ratio of the batch. This represents the average projected area of ​​the batch. This represents the average grain thickness distribution. The standard deviation of the grain thickness distribution. The initial moisture content, Indicates matrix transpose;

[0023] Constructing a multidimensional orthogonal eigenvector matrix specifically includes:

[0024] The vibration device was activated to lay the rice sample in a single layer, simultaneously triggering the image sensor and moisture sensor to collect data; the sample was then segmented. For each independent rice sample, the two-dimensional projection major axis, two-dimensional projection minor axis, contour projection area, single grain thickness, and dielectric capacitance variation were extracted.

[0025] Calculate the average length-to-width ratio of the batch using the following formula. :

[0026] ;

[0027] in, This represents the total number of independent rice samples. For the first The major axis of the two-dimensional projection of an individual rice sample. For the first The minor axis of the two-dimensional projection of an individual rice sample;

[0028] Calculate the average projected area of ​​the batch using the following formula. :

[0029] ;

[0030] in, For the first Binarized images of individual rice samples. For the first Two-dimensional closed contour region of an independent rice sample;

[0031] The mean grain thickness distribution is calculated using the following formula. ;

[0032] ;

[0033] in, For the first Single grain thickness of an independent rice sample;

[0034] Calculate the standard deviation of grain thickness distribution using the following formula. :

[0035] ;

[0036] Calculate the initial moisture content using the following formula. :

[0037] ;

[0038] in, This is the variation in dielectric capacitance. and This is the sensor environmental compensation calibration constant. It represents the natural logarithm.

[0039] Furthermore, the parallel multi-channel forward continuous equation system obtains and solidifies its internal parameters through an offline optimization training process, which includes:

[0040] Step 21: Select feature-instruction mapping pairs from the orthogonal physical de-shelling experimental data that maximize the comprehensive de-shelling quality evaluation index, and use the set of multiple feature-instruction mapping pairs as a structured benchmark dataset; in the feature-instruction mapping pair, the feature represents the feature component in the five-dimensional orthogonal feature vector matrix, and the instruction represents the corresponding optimal processing instruction.

[0041] Step 22: Input the structured benchmark dataset into the decision tree model, calculate the information gain of each input feature, obtain the feature importance weight distribution, and solidify it;

[0042] Step 23: Construct a parallel multi-channel ANFIS containing a fuzzing layer, a conclusion layer, and a defuzzing layer. Inject the feature importance weight distribution into the rule excitation intensity calculation. Train the ANFIS using the structured benchmark dataset. After training converges, solidify the membership function parameters and conclusion parameters.

[0043] Step 24: Use the particle swarm optimization algorithm to globally optimize the membership function parameters and conclusion parameters of ANFIS to minimize the preset fitness function. Replace the initial parameters with the optimal parameters obtained by optimization and finally solidify them to form the parallel multi-channel forward continuous equation system.

[0044] Furthermore, the comprehensive shelling quality evaluation index is defined as follows:

[0045] ;

[0046] in, To comprehensively evaluate the quality of shelling, and The preset penalty weight constant, For shelling rate, The percentage of broken rice. This represents the maximum value function.

[0047] Furthermore, step 22 specifically includes:

[0048] The structured benchmark dataset is input into the decision tree model, and the input features are automatically calculated using the node splitting mechanism of the decision tree based on impurity measurement. Feature Importance Weight Distribution The input features The corresponding average aspect ratios of the batches Batch average projected area Mean of grain thickness distribution Standard deviation of particle thickness distribution Initial moisture content ;

[0049] The impurity measure in the regression task is the mean squared error. Calculate the information gain using the following formula:

[0050] ;

[0051] in, Input features Information gain This represents the mean square error of the parent node before the split. This represents the total number of training samples contained in the parent node. For child node indexes, The set of all child nodes. For the first after the split The total number of training samples for each child node. For the first The mean square error of each child node;

[0052] Mean square error Calculate using the following formula:

[0053] ;

[0054] in, This represents the total number of training samples for the current node. For the first The true value of each training sample. This is the average value of all training samples at the current node;

[0055] The information gain The larger the value, the more significant the input features. The more important the prediction of the output processing command, the better; after training, the information gain of all input features is normalized to obtain the feature importance weight distribution. And solidify it, the solidified feature importance weight distribution This refers to the feature importance weight parameters contained in the parallel multi-channel forward continuous equation system.

[0056] Furthermore, step 23 specifically includes:

[0057] Step 231: Decompose the optimal processing instructions in the structured benchmark dataset into gaps. High roller speed Slow roller speed and separation air volume Four independent baseline action vectors are used, and a parallel ANFIS subnetwork is deployed for each baseline action vector.

[0058] Step 232: In the blurring layer of each ANFIS sub-network, the input features are processed according to the following formula. Membership degree converted into fuzzy linguistic variable:

[0059] ;

[0060] in, Input features The membership degree of the corresponding fuzzy linguistic variable. The center of the Gaussian membership function, The width of the Gaussian membership function. Represents the natural exponential function;

[0061] In the conclusion layer, each ANFIS subnetwork has an independent Takagi-Sugeno linear consequent coefficient matrix, and its output node equations are as follows:

[0062] For gaps The corresponding ANFIS subnetwork's input feature is the mean of the grain thickness distribution. and standard deviation of particle thickness distribution , No. Output of the mapping rule for:

[0063] ;

[0064] For the separation air volume The corresponding ANFIS subnetwork has batch average projected area as its input feature. and batch average aspect ratio , No. Output of the mapping rule for:

[0065] ;

[0066] For the speed of the fast roller The corresponding ANFIS subnetwork takes initial water content as its input feature. and batch average aspect ratio , No. Output of the mapping rule for:

[0067] ;

[0068] For slow roller speed The corresponding ANFIS subnetwork takes initial water content as its input feature. and batch average aspect ratio , No. Output of the mapping rule for:

[0069] ;

[0070] in, , , These are the conclusion parameters for the ANFIS subnetwork corresponding to the gap. , , These are the conclusion parameters of the ANFIS subnetwork corresponding to the air volume. , , The parameters for the ANFIS subnetwork corresponding to the high-speed roller rotation are: , , The conclusion parameters of the ANFIS subnetwork are the corresponding values ​​for the slow roller speed;

[0071] In the deblurring layer, each ANFIS sub-network outputs smooth and continuous processing instructions using the weighted average centroid method:

[0072] ;

[0073] in, , , , These represent the final outputs of the ANFIS sub-networks for the gap, air volume, fast roller speed, and slow roller speed, respectively. For the first The excitation intensity of the mapping rule, the excitation intensity The calculation process incorporates feature importance weight distribution. ;

[0074] Step 233: Train each ANFIS sub-network using a structured benchmark dataset. Once offline training converges, the center of the Gaussian membership function... Width of Gaussian membership function And the conclusion parameters are fixed, the center of the fixed Gaussian membership function Width of Gaussian membership function These are the membership function parameters contained in the parallel multi-channel forward continuous equation system, and the fixed conclusion parameters are the conclusion parameters contained in the parallel multi-channel forward continuous equation system.

[0075] Furthermore, step 24 specifically includes:

[0076] Constructing the fitness function :

[0077] ;

[0078] in, , , The preset weighting coefficients, For shelling rate, The percentage of broken rice. For specific energy consumption;

[0079] The particle swarm optimization (PSO) algorithm is used to perform a global search in the offline domain on the center, width, and conclusion parameters of the Gaussian membership function of the ANFIS subnetwork, in order to minimize the fitness function. After the search converges, the center, width, and conclusion parameters of the optimal Gaussian membership function are obtained. These parameters are then incorporated into the parallel multi-channel forward continuous equation system, replacing the initial parameters. This allows the lower-level execution unit to output the theoretically optimal instruction matrix through algebraic operations based on these incorporated optimal parameters during online execution. .

[0080] Furthermore, step 3 specifically includes:

[0081] Calculate the total machining power equivalent using the following formula. :

[0082] ;

[0083] in, This represents the total number of batches processed since the last zero-point calibration. For the first The quality of batches of rice samples For the first The difference in linear velocity used during the processing of batches of rice samples. For the first Processing time of batch rice samples;

[0084] The wear compensation increment is calculated using the following formula. :

[0085] ;

[0086] in, The pre-calibrated empirical fitting coefficients for radial wear of the rubber roller are... This is a wear nonlinearity correction factor. This is a manual bias calibration value;

[0087] The final physical displacement command is generated according to the following formula. :

[0088] ;

[0089] Furthermore, based on the same wear compensation increment The radial wear of the rubber roller is reflected by the following formula relative to the angular velocity of the fast and slow rollers. Compensation will be provided.

[0090] ;

[0091] in, For the target linear velocity, The initial radius of the rubber roller is given.

[0092] Furthermore, in step 5, the actual rotational speed of the fast roller... The deviation from the corresponding target speed and the actual speed of the slow roller The condition for the deviation from the corresponding target speed to enter the preset steady-state tolerance dead zone is:

[0093] ;

[0094] ;

[0095] in, The target frequency of the fast roller corresponds to the target rotational speed of the fast roller. The target frequency of the slow roll corresponding to the target speed of the slow roll. To allow for the speed tolerance, For a moment The real-time stator electrical frequency output by the high-speed roller inverter. For a moment The real-time stator frequency output by the slow roller inverter;

[0096] The actual physical displacement With final physical displacement command The condition for the deviation to enter the preset steady-state tolerance dead zone is:

[0097] ;

[0098] in, This represents the maximum allowable mechanical tolerance of the system.

[0099] The global steady-state flag signal is defined as When all three conditions above are met simultaneously and the duration exceeds the anti-shake filtering time threshold, hour, ,otherwise ;

[0100] The global steady-state flag signal is used to generate the opening command for the feed gate. :

[0101] ;

[0102] in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; The start time of the calibration. The preset safe dead time, This is a command signal indicating that the feed gate is fully open. This indicates that the opening level is fully open.

[0103] In a second aspect, the present invention provides a multi-dimensional feature-aware experimental rice huller parameter adaptive matching system for implementing the multi-dimensional feature-aware experimental rice huller parameter adaptive matching method as described in the first aspect, the system comprising:

[0104] The static pre-inspection and feature extraction module is used to collect raw sensing data of rice samples to be processed in the feeding detection chamber when the feeding gate is in a physically locked state, and to perform orthogonal dimensionality reduction and statistical purification on the raw sensing data to construct a multidimensional orthogonal feature vector matrix.

[0105] The online forward computation module is used to input the multidimensional orthogonal feature vector matrix into a system of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. This system of parallel multi-channel forward continuous equations is pre-obtained through offline optimization training and includes fixed membership function parameters, conclusion parameters, and feature importance weight parameters. The lower-level execution unit solves the system of parallel multi-channel forward continuous equations through algebraic operations and outputs the theoretically optimal instruction matrix for the current batch of rice samples. ;in, The target rubber roller displacement clearance. The target linear velocity of the fast roller, The target linear velocity of the slow roller. For the target linear velocity difference, To separate air volume for the target;

[0106] The cross-batch wear compensation module is used to obtain the total processing power equivalent accumulated by the experimental rice huller since the last zero-position calibration. The wear compensation increment is calculated based on pre-calibrated empirical wear fitting coefficients. The theoretically optimal instruction matrix The target rubber roller displacement gap With the wear compensation increment Superimpose the data to generate the final physical displacement command. ;

[0107] The no-load command execution module is used to drive the rubber roller gap adjustment mechanism to execute the final physical displacement command. The frequency converters driving the fast and slow rollers respectively execute the target linear speed of the fast roller. and the target linear velocity of the slow roller Drive the separator fan to execute the target separation air volume ;

[0108] The no-load steady-state closed-loop confirmation module is used to continuously poll the status register of the underlying driver in the no-load state to obtain the actual physical displacement. Actual rotational speed of the fast roller and the actual speed of the slow roller When determining the actual physical displacement With final physical displacement command The deviation, the actual speed of the fast roller Deviation from the corresponding target speed, actual speed of the slow roller The deviation from the corresponding target speed simultaneously enters the preset steady-state tolerance dead zone, and the duration exceeds the preset anti-shake filtering time threshold. At that time, output a global steady-state flag signal;

[0109] The feeding and processing module is used to control the opening of the feeding gate in response to the global steady-state flag signal, so that the rice sample to be processed is subject to the final physical displacement command. Target linear velocity of the fast roller target linear velocity of slow roller and target separation air volume The shell removal process is completed under a defined constant physical field; within the processing window, all parameter correction loops based on real-time feedback are forcibly shut down.

[0110] The beneficial effects of this invention are as follows: The multi-dimensional feature perception-based experimental rice huller parameter adaptive matching method and system provided by this invention collects multiple independent physical characteristics of rice grains when the feed gate is closed through a static pre-inspection process. This allows for accurate prior knowledge of rice grain characteristics and instantaneous calculation of optimal processing parameters using a pre-fixed forward equation set, completely replacing manual experience and significantly improving the accuracy of parameter matching. The use of a no-load stabilization before feeding ensures that the roller gap, speed, and airflow are stable at target values ​​before opening the feed gate. This ensures that all rice samples are hulled under constant optimal parameters from start to finish, completely eliminating the transitional processing errors caused by traditional on-the-fly adjustments and avoiding sample waste. A cross-batch wear compensation mechanism is introduced, automatically correcting the gap command based on the cumulative processing volume since the last calibration. This maintains processing accuracy even after long-term use, eliminating the need for frequent manual calibration. Within a short processing window, all adjustment loops based on real-time feedback are forcibly closed, ensuring the hulling process is completed in one go. This saves experimental samples and significantly improves the hulling integrity rate and head rice rate. Attached Figure Description

[0111] Figure 1 A flowchart illustrating the multi-dimensional feature perception-based experimental rice huller parameter adaptive matching method provided in this embodiment;

[0112] Figure 2 This is a schematic diagram of the structure of the experimental rice huller parameter adaptive matching system with multi-dimensional feature perception provided in the embodiment. Detailed Implementation

[0113] The present invention provides a method and system for adaptive matching of parameters of an experimental rice huller based on multidimensional feature perception. First, with the feed gate physically locked, raw sensing data of the rice sample to be processed is collected in the feed detection chamber. A multidimensional orthogonal feature vector matrix is ​​constructed through orthogonal dimensionality reduction and statistical purification. Then, the multidimensional orthogonal feature vector matrix is ​​input into a set of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. Through algebraic operations, the theoretically optimal command matrix for the current batch of rice samples is instantaneously solved and output. This matrix includes the target rubber roller displacement gap, the target linear velocity of the fast roller, the target linear velocity of the slow roller, the difference in target linear velocities, and the target separation air volume. Next, the total processing power equivalent accumulated by the experimental rice huller since the last zero-position calibration is obtained. The wear compensation increment is calculated based on a pre-calibrated wear empirical fitting coefficient, and the target rubber roller displacement gap is superimposed with the wear compensation increment to generate the final physical displacement command. Finally, the rubber roller gap adjustment mechanism is driven to execute the final physical displacement command. The drive inverters for the fast and slow rollers execute the target linear speeds of the fast and slow rollers respectively, and drive the separation fan to execute the target separation air volume. Under no-load conditions, the status register of the underlying driver is continuously polled to obtain the actual physical displacement, the actual speed of the fast roller, and the actual speed of the slow roller. When the deviations of the actual physical displacement from the final physical displacement command, the deviations of the actual speed of the fast roller from the corresponding target speed, and the deviations of the actual speed of the slow roller from the corresponding target speed all enter the preset steady-state tolerance dead zone and the duration exceeds the preset anti-shake filtering time threshold, a global steady-state flag signal is output. Finally, in response to the global steady-state flag signal, the feed gate is opened, so that the rice sample to be processed completes the dehulling process under a constant physical field defined by the final physical displacement command, the target linear speed of the fast roller, the target linear speed of the slow roller, and the target separation air volume. Within the processing window, all parameter correction loops based on real-time feedback are forcibly closed, thereby achieving zero-transition adaptive matching of the experimental rice huller parameters.

[0114] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0115] Figure 1 A flowchart illustrating a multi-dimensional feature-aware adaptive matching method for experimental rice huller parameters is shown. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0116] Step 1: Static pre-detection and feature extraction:

[0117] With the feed gate physically locked, raw sensing data of the rice sample to be processed is collected in the feed detection chamber, and the raw sensing data is subjected to orthogonal dimensionality reduction and statistical purification to construct a multidimensional orthogonal feature vector matrix.

[0118] In this embodiment, the multidimensional orthogonal eigenvector matrix is ​​a five-dimensional orthogonal eigenvector matrix. ;in, The average aspect ratio of the batch. This represents the average projected area of ​​the batch. This represents the average grain thickness distribution. The standard deviation of the grain thickness distribution. The initial moisture content, This indicates the matrix transpose.

[0119] Constructing a multidimensional orthogonal eigenvector matrix specifically includes:

[0120] The vibration device was activated to lay the rice sample in a single layer, simultaneously triggering the image sensor and moisture sensor to collect data; the sample was then segmented. For each independent rice sample, the two-dimensional projection major axis, two-dimensional projection minor axis, contour projection area, single grain thickness, and dielectric capacitance variation were extracted.

[0121] Calculate the average length-to-width ratio of the batch using the following formula. :

[0122] ;

[0123] in, This represents the total number of independent rice samples. For the first The major axis of the two-dimensional projection of an individual rice sample. For the first The minor axis of the two-dimensional projection of an individual rice sample;

[0124] Calculate the average projected area of ​​the batch using the following formula. :

[0125] ;

[0126] in, For the first Binarized images of individual rice samples. For the first Two-dimensional closed contour region of an independent rice sample;

[0127] The mean grain thickness distribution is calculated using the following formula. ;

[0128] ;

[0129] in, For the first Single grain thickness of an independent rice sample;

[0130] Calculate the standard deviation of grain thickness distribution using the following formula. :

[0131] ;

[0132] Calculate the initial moisture content using the following formula. :

[0133] ;

[0134] in, This is the variation in dielectric capacitance. and This is the sensor environmental compensation calibration constant. It represents the natural logarithm.

[0135] In practical applications, the feed gate is first closed, creating an independent static detection space in the feed detection chamber. A micro-vibration device is then activated to evenly distribute the rice samples into a single layer within the chamber, preventing particle stacking. Simultaneously, an image sensor captures images, and a moisture sensor collects signals indicating changes in dielectric constant.

[0136] The image captured by the image sensor is binarized to segment the background. Each rice sample is analyzed individually. For each grain, an image processing algorithm extracts the two-dimensional projection major axis (the longest distance along the grain's longest direction) and the two-dimensional projection minor axis (the maximum width perpendicular to the major axis), and calculates the pixel area enclosed by the contour as the projected area. Simultaneously, the thickness of each grain is measured using structured light or stereo vision methods. A moisture sensor acquires the dielectric capacitance variation of the entire rice sample.

[0137] Based on the above measurements, multiple features are calculated. This embodiment includes five features: the first is the average aspect ratio of the batch. First, calculate the ratio of the major axis to the minor axis of each grain, then average the values ​​of all grains. This characteristic reflects the tumbling trend of the rice within the roller zone. The second is the batch average projected area. The first is the arithmetic mean of the projected areas of all particles, representing the overall size of the rice grain and providing an aerodynamic basis for subsequently setting the separation airflow. The third is the average particle thickness distribution. The average thickness of all individual grains determines the degree of normal pressure that the rollers should apply during peeling. The fourth is the standard deviation of the grain thickness distribution. The first indicator reflects the consistency or dispersion of the thickness of the rice grains in that batch. When the standard deviation is large, the system will automatically limit the minimum gap between the rollers to prevent thick grains from being crushed. The fifth indicator is the initial moisture content. The capacitance change was obtained through linear calibration using the natural logarithm of the capacitance change. This value affects the bonding toughness between the rice husk and brown rice, thus influencing the required speed difference between the fast and slow rollers. Finally, a five-dimensional orthogonal feature vector matrix was constructed based on the five extracted, uncorrelated orthogonal features. .

[0138] Step 2, Online forward solution:

[0139] The multidimensional orthogonal feature vector matrix is ​​input into a set of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. This set of equations is pre-obtained through offline optimization training and includes fixed membership function parameters, conclusion parameters, and feature importance weight parameters. The lower-level execution unit solves the set of parallel multi-channel forward continuous equations through algebraic operations and outputs the theoretically optimal instruction matrix for the current batch of rice samples. ;in, The target rubber roller displacement clearance. The target linear velocity of the fast roller, The target linear velocity of the slow roller. For the target linear velocity difference, Separate the air volume for the target.

[0140] Specifically, the five-dimensional orthogonal eigenvector matrix obtained in step 1 will be... The pre-defined parallel multi-channel forward continuous equation system is fed into the lower-level execution unit. This system was obtained through extensive orthogonal experiments and offline optimization training using artificial intelligence algorithms before the equipment leaves the factory. Internally, the membership function parameters, conclusion parameters, and importance weight parameters for each feature are already fixed. The lower-level execution unit receives the five-dimensional orthogonal feature vector matrix. Then, a simple algebraic operation is performed, that is, the eigenvalues ​​are substituted into a pre-prepared system of polynomial equations to instantly solve for the optimal processing instructions for the current batch of rice. These optimal processing instructions constitute the theoretically optimal instruction matrix. Among them, the target rubber roller displacement gap The target linear speed of the high-speed roller is used to determine the tightness with which the rubber roller grips the rice grains. and the target linear velocity of the slow roller Used to determine the rotation speed of the two rollers, the target linear velocity difference. Used to determine the tearing force, by and The difference is automatically obtained, and the target separation air volume is determined. Used to determine the wind force used to blow away rice husks and brown rice.

[0141] In this embodiment, the offline optimization training of the parallel multi-channel forward continuous equation system includes steps 21 to 24:

[0142] Step 21: Select feature-instruction mapping pairs from the orthogonal physical de-shelling experimental data that maximize the comprehensive de-shelling quality evaluation index, and use the set of multiple feature-instruction mapping pairs as a structured benchmark dataset; in the feature-instruction mapping pair, the feature represents the feature component in the five-dimensional orthogonal feature vector matrix, and the instruction represents the corresponding optimal processing instruction.

[0143] Specifically, an experimental rice huller was first used to conduct orthogonal physical dehulling experiments on a large number of rice samples with different moisture contents and grain shapes. The five-dimensional orthogonal eigenvector values ​​corresponding to each group of experiments were recorded, namely, the average aspect ratio of the batch, the average projected area of ​​the batch, the mean of the grain thickness distribution, the standard deviation of the grain thickness distribution, and the initial moisture content. At the same time, the dehulling rate and broken rice rate of the group of experiments were recorded.

[0144] Then, a comprehensive evaluation index for shelling quality is defined:

[0145] ;

[0146] in, To comprehensively evaluate the quality of shelling, and The preset penalty weight constant, For shelling rate, The percentage of broken rice. This represents the maximum value function.

[0147] From all possible combinations of processing instructions, select the comprehensive dehulling quality evaluation index. The set of parameters with the highest value is the optimal processing instruction, which includes the target rubber roller displacement gap, the target linear velocity of the fast roller, the target linear velocity of the slow roller, the target linear velocity difference, and the target separation air volume.

[0148] Finally, the five-dimensional orthogonal feature vector of each set of experiments is paired with the optimal processing instruction corresponding to that set of experiments to form a feature-instruction mapping pair. All feature-instruction mapping pairs are collected to form a structured benchmark dataset.

[0149] Step 22: Input the structured benchmark dataset into the decision tree model, calculate the information gain of each input feature, obtain the feature importance weight distribution, and solidify it.

[0150] In this embodiment, step 22 specifically includes:

[0151] The structured benchmark dataset is input into the decision tree model, and the input features are automatically calculated using the node splitting mechanism of the decision tree based on impurity measurement. Feature Importance Weight Distribution The input features The corresponding average aspect ratios of the batches Batch average projected area Mean of grain thickness distribution Standard deviation of particle thickness distribution Initial moisture content ;

[0152] The impurity measure in the regression task is the mean squared error. Calculate the information gain using the following formula:

[0153] ;

[0154] in, Input features Information gain This represents the mean square error of the parent node before the split. This represents the total number of training samples contained in the parent node. For child node indexes, The set of all child nodes. For the first after the split The total number of training samples for each child node. For the first The mean square error of each child node;

[0155] Mean square error Calculate using the following formula:

[0156] ;

[0157] in, This represents the total number of training samples for the current node. For the first The true value of each training sample. This is the average value of all training samples at the current node;

[0158] The information gain The larger the value, the more significant the input features. The more important the prediction of the output processing command, the better; after training, the information gain of all input features is normalized to obtain the feature importance weight distribution. And solidify it, the solidified feature importance weight distribution This refers to the feature importance weight parameters contained in the parallel multi-channel forward continuous equation system.

[0159] Specifically, in step 22, the structured benchmark dataset obtained in step 21 is fed into the decision tree model for training. The decision tree model is a supervised learning algorithm that repeatedly splits the dataset according to a certain feature value, making the output value (i.e., the optimal processing instruction) of each child node purer after the split. To measure the improvement in performance brought by each split, an impurity metric is introduced. Since this task is a regression problem (predicting continuous values, such as the gap between rubber rollers), the impurity metric uses mean squared error.

[0160] The specific process is as follows: The decision tree starts from the root node and contains all training samples. For each input feature (five in total: batch average aspect ratio, batch average projected area, mean grain thickness distribution, standard deviation of grain thickness distribution, and initial moisture content), all possible split points are tried. The difference between the mean square error of the parent node before splitting and the weighted mean square error of each child node after splitting is calculated. This difference is the information gain. The larger the information gain, the more effectively the input feature can reduce the prediction error when splitting at the current node. Therefore, the more important the input feature is for predicting the output processing instructions.

[0161] After the decision tree is trained, the information gain of each input feature in all split nodes is calculated and normalized to between 0 and 1 to obtain the feature importance weight distribution. ( (Each corresponds to one of the five input features). The feature importance weight distribution reflects the relative importance of the five input features in predicting the optimal processing command. For example, the mean and standard deviation of the particle thickness distribution are often more important for predicting the roller gap, while the projected area and aspect ratio are more important for predicting the air volume.

[0162] Finally, the feature importance weights are distributed. The solidified features are incorporated into the parallel multi-channel forward continuous equation system, serving as prior weights in the subsequent ANFIS model calculation of fuzzy rule excitation intensity, thus giving important features a higher weight in fuzzy inference. That is, the feature importance weight parameters contained within the parallel multi-channel forward continuous equation system.

[0163] Step 23: Construct a parallel multi-channel ANFIS containing a fuzzing layer, a conclusion layer, and a defuzzing layer. Inject the feature importance weight distribution into the rule excitation intensity calculation. Train the ANFIS using the structured benchmark dataset. After training converges, solidify the membership function parameters and conclusion parameters.

[0164] In this embodiment, step 23 specifically includes steps 231 to 233:

[0165] Step 231: Decompose the optimal processing instructions in the structured benchmark dataset into gaps. High roller speed Slow roller speed and separation air volume Four independent baseline action vectors are used, and a parallel ANFIS subnetwork is deployed for each baseline action vector.

[0166] Step 232: In the blurring layer of each ANFIS sub-network, the input features are processed according to the following formula. Membership degree converted into fuzzy linguistic variable:

[0167] ;

[0168] in, Input features The membership degree of the corresponding fuzzy linguistic variable. The center of the Gaussian membership function, The width of the Gaussian membership function. Represents the natural exponential function;

[0169] In the conclusion layer, each ANFIS subnetwork has an independent Takagi-Sugeno linear consequent coefficient matrix, and its output node equations are as follows:

[0170] For gaps The corresponding ANFIS subnetwork's input feature is the mean of the grain thickness distribution. and standard deviation of particle thickness distribution , No. Output of the mapping rule for:

[0171] ;

[0172] For the separation air volume The corresponding ANFIS subnetwork has batch average projected area as its input feature. and batch average aspect ratio , No. Output of the mapping rule for:

[0173] ;

[0174] For the speed of the fast roller The corresponding ANFIS subnetwork takes initial water content as its input feature. and batch average aspect ratio , No. Output of the mapping rule for:

[0175] ;

[0176] For slow roller speed The corresponding ANFIS subnetwork takes initial water content as its input feature. and batch average aspect ratio , No. Output of the mapping rule for:

[0177] ;

[0178] in, , , These are the conclusion parameters for the ANFIS subnetwork corresponding to the gap. , , These are the conclusion parameters of the ANFIS subnetwork corresponding to the air volume. , , The parameters for the ANFIS subnetwork corresponding to the high-speed roller rotation are: , , The conclusion parameters of the ANFIS subnetwork are the corresponding values ​​for the slow roller speed;

[0179] In the deblurring layer, each ANFIS sub-network outputs smooth and continuous processing instructions using the weighted average centroid method:

[0180] ;

[0181] in, , , , These represent the final outputs of the ANFIS sub-networks for the gap, air volume, fast roller speed, and slow roller speed, respectively. For the first The excitation intensity of the mapping rule, the excitation intensity The calculation process incorporates feature importance weight distribution. ;

[0182] Step 233: Train each ANFIS sub-network using a structured benchmark dataset. Once offline training converges, the center of the Gaussian membership function... Width of Gaussian membership function And the conclusion parameters are fixed, the center of the fixed Gaussian membership function Width of Gaussian membership function These are the membership function parameters contained in the parallel multi-channel forward continuous equation system, and the fixed conclusion parameters are the conclusion parameters contained in the parallel multi-channel forward continuous equation system.

[0183] Specifically, in step 23, a parallel multi-channel adaptive neurofuzzy inference system (ANFIS) is constructed to establish a continuous nonlinear mapping between five-dimensional orthogonal feature vectors and optimal processing instructions. The entire training process is as follows:

[0184] First, the optimal processing instructions in the structured benchmark dataset are decomposed into four independent benchmark action vectors: roller gap, fast roller speed, slow roller speed, and separation air volume. An ANFIS sub-network is then deployed for each of these four benchmark action vectors.

[0185] Then, in the fuzzification layer of each ANFIS sub-network, five orthogonal features are input: batch average aspect ratio, batch average projected area, mean grain thickness distribution, standard deviation of grain thickness distribution, and initial moisture content. Each input feature is transformed into the membership degree of the fuzzy linguistic variable through a Gaussian membership function, where the Gaussian function is centered at... and width Decide.

[0186] In the conclusion layer, each ANFIS subnetwork uses a Takagi-Sugeno type linear equation. The input features of the ANFIS subnetwork corresponding to the gap are the mean and standard deviation of the particle thickness distribution; the input features of the ANFIS subnetwork corresponding to the air volume are the projected area and aspect ratio; and the input features of the ANFIS subnetwork corresponding to the fast and slow roller speeds are both moisture content and aspect ratio. Each linear equation contains a constant term and two feature terms, for a total of three conclusion parameters.

[0187] In the deblurring layer, the excitation intensity of each mapping rule is first calculated. Excitation intensity The calculation process incorporates feature importance weight distribution. This makes the important features contribute more to the excitation of the rules. Then, the weighted average centroid method is used, that is, the output of each rule is multiplied by the excitation intensity, the sum is obtained, and then the result is divided by the sum of the excitation intensities to obtain the continuous output processing instructions of the four ANFIS subnetworks.

[0188] Finally, four ANFIS sub-networks were trained using a structured benchmark dataset, adjusting the center and width of the membership functions and the conclusion parameters. After training converged, these parameters were solidified to form an offline-trained parallel multi-channel forward continuous equation system. The solidified membership function parameters and conclusion parameters constitute the core coefficients of this equation system.

[0189] Step 24: Use the particle swarm optimization algorithm to globally optimize the membership function parameters and conclusion parameters of ANFIS to minimize the preset fitness function. Replace the initial parameters with the optimal parameters obtained by optimization and finally solidify them to form the parallel multi-channel forward continuous equation system.

[0190] In this embodiment, step 24 specifically includes:

[0191] Constructing the fitness function :

[0192] ;

[0193] in, , , The preset weighting coefficients, For shelling rate, The percentage of broken rice. For specific energy consumption;

[0194] The particle swarm optimization (PSO) algorithm is used to perform a global search in the offline domain on the center, width, and conclusion parameters of the Gaussian membership function of the ANFIS subnetwork, in order to minimize the fitness function. After the search converges, the center, width, and conclusion parameters of the optimal Gaussian membership function are obtained. These parameters are then incorporated into the parallel multi-channel forward continuous equation system, replacing the initial parameters. This allows the lower-level execution unit to output the theoretically optimal instruction matrix through algebraic operations based on these incorporated optimal parameters during online execution. .

[0195] Specifically, in step 24, in order to further improve the prediction accuracy of the ANFIS model and avoid the traditional gradient descent method from getting stuck in local optima, the particle swarm optimization algorithm is used to globally optimize the parameters of the four ANFIS sub-networks.

[0196] In practical applications, the first step is to construct a fitness function. This function is used to evaluate the overall performance of a set of parameters. It consists of three terms: the first term is 1 minus the shelling rate. The higher the hulling rate, the lower this item; the second item is the broken rice rate. The lower the broken rice rate, the better; the third item is energy consumption. This refers to the energy consumed in processing a unit mass of rice. The three terms are multiplied by preset weighting coefficients and then summed. The objective is to minimize the fitness function. This means pursuing a high hulling rate while suppressing broken rice rate and reducing energy consumption.

[0197] The particle swarm optimization algorithm simulates the foraging behavior of bird flocks by randomly initializing a swarm of particles in the parameter space. Each particle represents a set of ANFIS parameters (including the center of all Gaussian membership functions). and width (and all conclusion parameters). Each particle continuously updates its velocity and position based on its current local optimal position and the global optimal position of the entire swarm, moving towards a better solution. The algorithm iterates multiple times until the fitness function is satisfied. It converges to a stable value or reaches the preset number of iterations.

[0198] After the search converges, a globally optimal set of ANFIS parameters is obtained, including the optimal Gaussian membership function center, width, and conclusion parameters. These optimal parameters replace the original initial parameters and are permanently embedded into the parallel multi-channel forward continuous equation system. Thus, during online execution, the lower-level execution unit can quickly output the theoretically optimal instruction matrix based on the globally optimized parameters through simple algebraic operations. ,

[0199] Step 3, Cross-batch wear compensation:

[0200] Obtain the total processing power equivalent of the experimental rice huller since the last zero-point calibration. The wear compensation increment is calculated based on pre-calibrated empirical wear fitting coefficients. The theoretically optimal instruction matrix The target rubber roller displacement gap With the wear compensation increment Superimpose the data to generate the final physical displacement command. .

[0201] In this embodiment, step 3 specifically includes:

[0202] Calculate the total machining power equivalent using the following formula. :

[0203] ;

[0204] in, This represents the total number of batches processed since the last zero-point calibration. For the first The quality of batches of rice samples For the first The difference in linear velocity used during the processing of batches of rice samples. For the first Processing time of batch rice samples;

[0205] The wear compensation increment is calculated using the following formula. :

[0206] ;

[0207] in, The pre-calibrated empirical fitting coefficients for radial wear of the rubber roller are... This is a wear nonlinearity correction factor. This is a manual bias calibration value;

[0208] The final physical displacement command is generated according to the following formula. :

[0209] ;

[0210] Furthermore, based on the same wear compensation increment The radial wear of the rubber roller is reflected by the following formula relative to the angular velocity of the fast and slow rollers. Compensation will be provided.

[0211] ;

[0212] in, For the target linear velocity, The initial radius of the rubber roller is given.

[0213] Specifically, in step 3, the machining work equivalent of all batches since the last zero-position calibration is first accumulated. That is, the quality of each batch of rice samples The linear velocity difference used in this batch and processing time The products are summed. Then, the empirical fitting coefficients for radial wear of the rubber rollers, which have been pre-calibrated, are used. Wear nonlinearity correction factor and manual bias calibration amount Calculate the wear compensation increment The target roller displacement gap in the theoretically optimal command. Adding this compensation increment yields the final physical displacement command. This is to offset the effect of the reduced radius and increased actual gap of the rubber roller caused by wear. Simultaneously, based on the same wear compensation increment... The radial wear amount is used to compensate for the angular velocity of the fast and slow rollers, thus obtaining the angular velocity. This ensures that the target linear velocity can be maintained even after wear, thereby maintaining a constant peeling and tearing force.

[0214] Step 4: Execute the no-load command:

[0215] The drive roller gap adjustment mechanism executes the final physical displacement command. The frequency converters driving the fast and slow rollers respectively execute the target linear speed of the fast roller. and the target linear velocity of the slow roller Drive the separator fan to execute the target separation air volume .

[0216] Specifically, in step 4, under no-load conditions with the feed gate closed, the roller gap adjustment mechanism is first driven according to the final physical displacement command. The movement brings the actual gap between the two rubber rollers to the compensated target value; simultaneously, the high-speed roller frequency converter and the low-speed roller frequency converter output the target linear velocity of the high-speed roller respectively. and the target linear velocity of the slow roller This causes the two rubber rollers to rotate at a set speed difference; in addition, it drives the separator fan to output the target separation air volume. This prepares for subsequent grain-based air separation. During this stage, all actuators operate under no-load conditions, meaning only processing parameters are adjusted without material feeding, ensuring that the parameters are stable for subsequent processing.

[0217] Step 5: No-load steady-state closed-loop confirmation:

[0218] In the no-load state, continuously poll the status register of the underlying driver to obtain the actual physical displacement. Actual rotational speed of the fast roller and the actual speed of the slow roller When determining the actual physical displacement With final physical displacement command The deviation, the actual speed of the fast roller Deviation from the corresponding target speed, actual speed of the slow roller The deviation from the corresponding target speed simultaneously enters the preset steady-state tolerance dead zone, and the duration exceeds the preset anti-shake filtering time threshold. At that time, a global steady-state flag signal is output.

[0219] In this embodiment, the actual rotational speed of the fast roller is... The deviation from the corresponding target speed and the actual speed of the slow roller The condition for the deviation from the corresponding target speed to enter the preset steady-state tolerance dead zone is:

[0220] ;

[0221] ;

[0222] in, The target frequency of the fast roller corresponds to the target rotational speed of the fast roller. The target frequency of the slow roll corresponding to the target speed of the slow roll. To allow for the speed tolerance, For a moment The real-time stator electrical frequency output by the high-speed roller inverter. For a moment The real-time stator frequency output by the slow roller inverter;

[0223] The actual physical displacement With final physical displacement command The condition for the deviation to enter the preset steady-state tolerance dead zone is:

[0224] ;

[0225] in, This represents the maximum allowable mechanical tolerance of the system.

[0226] The global steady-state flag signal is defined as When all three conditions above are met simultaneously and the duration exceeds the anti-shake filtering time threshold, hour, ,otherwise ;

[0227] The global steady-state flag signal is used to generate the opening command for the feed gate. :

[0228] ;

[0229] in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; The start time of the calibration. The preset safe dead time, This is a command signal indicating that the feed gate is fully open. This indicates that the opening level is fully open.

[0230] Specifically, in step 5, the status register of the underlying driver is continuously read under no-load conditions to obtain the actual physical displacement of the rubber roller. Real-time stator frequency of the fast roller inverter and the real-time stator frequency of the slow roller inverter When the absolute value of the deviation between the actual physical displacement and the final physical displacement command does not exceed the maximum allowable mechanical tolerance of the system. Simultaneously, the real-time frequency of the fast roller and the target frequency of the fast roller The absolute value of the deviation does not exceed the allowable speed tolerance. Furthermore, the real-time frequency of the slow roll and the target frequency of the slow roll The absolute value of the deviation does not exceed These three conditions must be met simultaneously and maintained for a duration exceeding the preset anti-shake filtering time threshold. At this point, confirm that the equipment has entered a stable state and output a global steady-state flag signal. Then, check the time since the start of the calibration. Has the preset safe dead zone time not been exceeded? If both conditions are met, then the indicator function will be used. Generate a command to fully open the feed gate. This means that material feeding is allowed. If any condition is not met, the gate remains closed and a timeout alarm is triggered. This ensures that all actuators only begin processing after reaching a fully steady state, achieving zero-transition unloading.

[0231] Step 6, Feeding and Processing:

[0232] In response to the global steady-state flag signal, the feed gate is opened, allowing the rice sample to be processed to be moved by the final physical displacement command. Target linear velocity of the fast roller target linear velocity of slow roller and target separation air volume The shell removal process is completed under a defined constant physical field; within the processing window, all parameter correction loops based on real-time feedback are forcibly shut down.

[0233] Specifically, in step 6, upon receiving the global steady-state flag signal output in step 5, the feed gate is immediately opened. At this time, the rice sample to be processed falls into the rubber roller zone and completes dehulling within a pre-adjusted constant physical field. This constant physical field is defined by three parts: first, the final physical displacement command... The determined roller gap provides a stable normal rolling pressure; secondly, the target linear velocity of the fast roller. and the target linear velocity of the slow roller The determined speed difference It provides a constant tangential tearing force; thirdly, it targets the separation air volume. This provides a stable airflow field for hulling. Throughout the entire processing window, typically 15 to 20 seconds, all parameter correction loops based on real-time feedback are forcibly shut down, without any online adjustments to gaps, speed, or airflow. This ensures that all rice samples are hulled under identical, theoretically optimal processing conditions from start to finish, avoiding transient processing errors caused by adjustment lag in traditional closed-loop control, thus maximizing the accuracy of hulling and head rice yields.

[0234] In summary, the multi-dimensional feature-aware adaptive matching method for experimental rice hullers provided in this embodiment, during the static pre-inspection stage of the feed gate locking, extracts multiple independent features through orthogonal dimensionality reduction, completely eliminating collinearity interference between geometric parameters such as length and width in conventional methods, making subsequent parameter matching more accurate and reliable. The parallel multi-channel forward continuous equations, converged after offline training, are solidified in the lower-level execution unit, requiring only one algebraic operation to instantly output the optimal processing command during online operation, avoiding the real-time bottleneck of complex iterative algorithms within a short processing window. Wear compensation increments are calculated by accumulating processing work equivalent across batches, correcting the target gap and simultaneously compensating for angular velocity, ensuring that the rubber rollers maintain initial processing accuracy even after long-term use, eliminating the need for frequent manual calibration. The timing sequence of first adjusting under no-load conditions and confirming steady state before opening the feed gate ensures that both the actual gap and rotational speed enter the steady-state tolerance dead zone and remain under anti-shake conditions for a sustained period before material release, fundamentally eliminating the transient processing errors caused by traditional on-the-fly adjustments and protecting valuable experimental samples. By forcibly shutting down all real-time feedback loops within a processing window of 15 to 20 seconds, all rice samples are dehulled in one go under a pre-adjusted constant gap, constant speed difference, and constant air volume field, which significantly improves the accuracy of the measurement of dehulling integrity rate and head rice rate.

[0235] Based on the above technical solution, this embodiment also provides a multi-dimensional feature-aware experimental rice huller parameter adaptive matching system, used to implement the multi-dimensional feature-aware experimental rice huller parameter adaptive matching method as described in the embodiment. Please refer to [link to relevant documentation]. Figure 2 The system includes:

[0236] The static pre-inspection and feature extraction module is used to collect raw sensing data of rice samples to be processed in the feeding detection chamber when the feeding gate is in a physically locked state, and to perform orthogonal dimensionality reduction and statistical purification on the raw sensing data to construct a multidimensional orthogonal feature vector matrix.

[0237] The online forward computation module is used to input the multidimensional orthogonal feature vector matrix into a system of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. This system of parallel multi-channel forward continuous equations is pre-obtained through offline optimization training and includes fixed membership function parameters, conclusion parameters, and feature importance weight parameters. The lower-level execution unit solves the system of parallel multi-channel forward continuous equations through algebraic operations and outputs the theoretically optimal instruction matrix for the current batch of rice samples. ;in, The target rubber roller displacement clearance. The target linear velocity of the fast roller, The target linear velocity of the slow roller. For the target linear velocity difference, To separate air volume for the target;

[0238] The cross-batch wear compensation module is used to obtain the total processing power equivalent accumulated by the experimental rice huller since the last zero-position calibration. The wear compensation increment is calculated based on pre-calibrated empirical wear fitting coefficients. The theoretically optimal instruction matrix The target rubber roller displacement gap With the wear compensation increment Superimpose the data to generate the final physical displacement command. ;

[0239] The no-load command execution module is used to drive the rubber roller gap adjustment mechanism to execute the final physical displacement command. The frequency converters driving the fast and slow rollers respectively execute the target linear speed of the fast roller. and the target linear velocity of the slow roller Drive the separator fan to execute the target separation air volume ;

[0240] The no-load steady-state closed-loop confirmation module is used to continuously poll the status register of the underlying driver in the no-load state to obtain the actual physical displacement. Actual rotational speed of the fast roller and the actual speed of the slow roller When determining the actual physical displacement With final physical displacement command The deviation, the actual speed of the fast roller Deviation from the corresponding target speed, actual speed of the slow roller The deviation from the corresponding target speed simultaneously enters the preset steady-state tolerance dead zone, and the duration exceeds the preset anti-shake filtering time threshold. At that time, output a global steady-state flag signal;

[0241] The feeding and processing module is used to control the opening of the feeding gate in response to the global steady-state flag signal, so that the rice sample to be processed is subject to the final physical displacement command. Target linear velocity of the fast roller target linear velocity of slow roller and target separation air volume The shell removal process is completed under a defined constant physical field; within the processing window, all parameter correction loops based on real-time feedback are forcibly shut down.

[0242] It is understood that the multi-dimensional feature-aware experimental rice huller parameter adaptive matching system described in this embodiment is a system for implementing the multi-dimensional feature-aware experimental rice huller parameter adaptive matching method described in the embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant parts, please refer to the description of the method. It will not be repeated here.

Claims

1. A multi-dimensional feature-aware experimental rice huller parameter adaptive matching method, characterized in that, The method includes: Step 1: Static pre-detection and feature extraction: With the feed gate physically locked, raw sensing data of the rice sample to be processed is collected in the feed detection chamber, and the raw sensing data is subjected to orthogonal dimensionality reduction and statistical purification to construct a multidimensional orthogonal feature vector matrix. Step 2, Online forward solution: The multidimensional orthogonal feature vector matrix is ​​input into a set of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. This set of equations is pre-obtained through offline optimization training and includes fixed membership function parameters, conclusion parameters, and feature importance weight parameters. The lower-level execution unit solves the set of parallel multi-channel forward continuous equations through algebraic operations and outputs the theoretically optimal instruction matrix for the current batch of rice samples. ;in, The target rubber roller displacement clearance. The target linear velocity of the fast roller, The target linear velocity of the slow roller. For the target linear velocity difference, Separate air volume for target purpose; Step 3, Cross-batch wear compensation: Obtain the total processing power equivalent of the experimental rice huller since the last zero-point calibration. The wear compensation increment is calculated based on pre-calibrated empirical wear fitting coefficients. The theoretically optimal instruction matrix The target rubber roller displacement gap With the wear compensation increment Superimpose the data to generate the final physical displacement command. ; Step 4: Execute the no-load command: The drive roller gap adjustment mechanism executes the final physical displacement command. The frequency converters driving the fast and slow rollers respectively execute the target linear speed of the fast roller. and the target linear velocity of the slow roller Drive the separator fan to execute the target separation air volume ; Step 5: No-load steady-state closed-loop confirmation: In the no-load state, continuously poll the status register of the underlying driver to obtain the actual physical displacement. Actual rotational speed of the fast roller and the actual speed of the slow roller When determining the actual physical displacement With final physical displacement command The deviation, the actual speed of the fast roller Deviation from the corresponding target speed, actual speed of the slow roller The deviation from the corresponding target speed simultaneously enters the preset steady-state tolerance dead zone, and the duration exceeds the preset anti-shake filtering time threshold. At that time, output a global steady-state flag signal; Step 6, Feeding and Processing: In response to the global steady-state flag signal, the feed gate is opened, allowing the rice sample to be processed to be moved by the final physical displacement command. Target linear velocity of the fast roller target linear velocity of slow roller and target separation air volume The shell removal process is completed under a defined constant physical field; within the processing window, all parameter correction loops based on real-time feedback are forcibly shut down.

2. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 1, characterized in that, The multidimensional orthogonal eigenvector matrix is ​​a five-dimensional orthogonal eigenvector matrix. ;in, The average aspect ratio of the batch. This represents the average projected area of ​​the batch. This represents the average grain thickness distribution. The standard deviation of the grain thickness distribution. The initial moisture content, Indicates matrix transpose; Constructing a multidimensional orthogonal eigenvector matrix specifically includes: The vibration device was activated to lay the rice sample in a single layer, simultaneously triggering the image sensor and moisture sensor to collect data; the sample was then segmented. For each independent rice sample, the two-dimensional projection major axis, two-dimensional projection minor axis, contour projection area, single grain thickness, and dielectric capacitance variation were extracted. Calculate the average length-to-width ratio of the batch using the following formula. : ； in, This represents the total number of independent rice samples. For the first The major axis of the two-dimensional projection of an individual rice sample. For the first The minor axis of the two-dimensional projection of an individual rice sample; Calculate the average projected area of ​​the batch using the following formula. : ； in, For the first Binarized images of individual rice samples. For the first Two-dimensional closed contour region of an independent rice sample; The mean grain thickness distribution is calculated using the following formula. ; ； in, For the first Single grain thickness of an independent rice sample; Calculate the standard deviation of grain thickness distribution using the following formula. : ； Calculate the initial moisture content using the following formula. : ； in, This is the variation in dielectric capacitance. and This is the sensor environmental compensation calibration constant. It represents the natural logarithm.

3. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 2, characterized in that, The parallel multi-channel forward continuous equation system obtains its internal parameters and solidifies them through an offline optimization training process, which includes: Step 21: Select feature-instruction mapping pairs from the orthogonal physical de-shelling experimental data that maximize the comprehensive de-shelling quality evaluation index, and use the set of multiple feature-instruction mapping pairs as a structured benchmark dataset; in the feature-instruction mapping pair, the feature represents the feature component in the five-dimensional orthogonal feature vector matrix, and the instruction represents the corresponding optimal processing instruction. Step 22: Input the structured benchmark dataset into the decision tree model, calculate the information gain of each input feature, obtain the feature importance weight distribution, and solidify it; Step 23: Construct a parallel multi-channel ANFIS containing a fuzzing layer, a conclusion layer, and a defuzzing layer. Inject the feature importance weight distribution into the rule excitation intensity calculation. Train the ANFIS using the structured benchmark dataset. After training converges, solidify the membership function parameters and conclusion parameters. Step 24: Use the particle swarm optimization algorithm to globally optimize the membership function parameters and conclusion parameters of ANFIS to minimize the preset fitness function. Replace the initial parameters with the optimal parameters obtained by optimization and finally solidify them to form the parallel multi-channel forward continuous equation system.

4. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 3, characterized in that, The comprehensive shelling quality evaluation index is defined as follows: ； in, To comprehensively evaluate the quality of shelling, and The preset penalty weight constant, For the shelling rate, The percentage of broken rice. This represents the maximum value function.

5. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 3, characterized in that, Step 22 specifically includes: The structured benchmark dataset is input into the decision tree model, and the input features are automatically calculated using the node splitting mechanism of the decision tree based on impurity measurement. Feature Importance Weight Distribution The input features The corresponding average aspect ratios of the batches Batch average projected area Mean of grain thickness distribution Standard deviation of particle thickness distribution Initial moisture content ; The impurity measure in the regression task is the mean squared error. Calculate the information gain using the following formula: ； in, Input features Information gain This represents the mean square error of the parent node before the split. This represents the total number of training samples contained in the parent node. For child node indexes, The set of all child nodes. For the first after the split The total number of training samples for each child node. For the first The mean square error of each child node; Mean square error Calculate using the following formula: ； in, This represents the total number of training samples for the current node. For the first The true value of each training sample. This is the average value of all training samples at the current node; The information gain The larger the value, the more significant the input features. The more important the prediction of the output processing command, the better; after training, the information gain of all input features is normalized to obtain the feature importance weight distribution. And solidify it, the solidified feature importance weight distribution This refers to the feature importance weight parameters contained in the parallel multi-channel forward continuous equation system.

6. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 3, characterized in that, Step 23 specifically includes: Step 231: Decompose the optimal processing instructions in the structured benchmark dataset into gaps. High roller speed Slow roller speed and separation air volume Four independent baseline action vectors are used, and a parallel ANFIS subnetwork is deployed for each baseline action vector. Step 232: In the blurring layer of each ANFIS sub-network, the input features are processed according to the following formula. Membership degree converted into fuzzy linguistic variable: ； in, Input features The membership degree of the corresponding fuzzy linguistic variable. The center of the Gaussian membership function, The width of the Gaussian membership function. Represents the natural exponential function; In the conclusion layer, each ANFIS subnetwork has an independent Takagi-Sugeno linear consequent coefficient matrix, and its output node equations are as follows: For gaps The corresponding ANFIS subnetwork's input feature is the mean of the grain thickness distribution. and standard deviation of particle thickness distribution , No. Output of the mapping rule for: ； For the separation air volume The corresponding ANFIS subnetwork has batch average projected area as its input feature. and batch average aspect ratio , No. Output of the mapping rule for: ； For the speed of the fast roller The corresponding ANFIS subnetwork takes initial water content as its input feature. and batch average aspect ratio , No. Output of the mapping rule for: ； For slow roller speed The corresponding ANFIS subnetwork takes initial water content as its input feature. and batch average aspect ratio , No. Output of the mapping rule for: ； in, , , The gap corresponds to the conclusion parameters of the ANFIS subnetwork. , , These are the conclusion parameters of the ANFIS subnetwork corresponding to the air volume. , , The parameters for the ANFIS subnetwork corresponding to the high-speed roller rotation are: , , The conclusion parameters of the ANFIS subnetwork are the corresponding values ​​for the slow roller speed; In the deblurring layer, each ANFIS sub-network outputs smooth and continuous processing instructions using the weighted average centroid method: ； in, , , , These represent the final outputs of the ANFIS sub-networks for the gap, air volume, fast roller speed, and slow roller speed, respectively. For the first The excitation intensity of the mapping rule, the excitation intensity The calculation process incorporates feature importance weight distribution. ; Step 233: Train each ANFIS sub-network using a structured benchmark dataset. Once offline training converges, the center of the Gaussian membership function... Width of Gaussian membership function And the conclusion parameters are fixed, the center of the fixed Gaussian membership function Width of Gaussian membership function These are the membership function parameters contained in the parallel multi-channel forward continuous equation system, and the fixed conclusion parameters are the conclusion parameters contained in the parallel multi-channel forward continuous equation system.

7. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 3, characterized in that, Step 24 specifically includes: Constructing the fitness function : ； in, , , The preset weighting coefficients, For the shelling rate, The percentage of broken rice. For specific energy consumption; The particle swarm optimization (PSO) algorithm is used to perform a global search in the offline domain on the center, width, and conclusion parameters of the Gaussian membership function of the ANFIS subnetwork, in order to minimize the fitness function. After the search converges, the center, width, and conclusion parameters of the optimal Gaussian membership function are obtained. These parameters are then incorporated into the parallel multi-channel forward continuous equation system, replacing the initial parameters. This allows the lower-level execution unit to output the theoretically optimal instruction matrix through algebraic operations based on these incorporated optimal parameters during online execution. .

8. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 1, characterized in that, Step 3 specifically includes: Calculate the total machining power equivalent using the following formula. : ； in, This represents the total number of batches processed since the last zero-point calibration. For the first The quality of batch rice samples For the first The difference in linear velocity used during the processing of batches of rice samples. For the first Processing time of batch rice samples; The wear compensation increment is calculated using the following formula. : ； in, The pre-calibrated empirical fitting coefficients for radial wear of the rubber roller are... This is a wear nonlinearity correction factor. This is a manual bias calibration value; The final physical displacement command is generated according to the following formula. : ； Furthermore, based on the same wear compensation increment The radial wear of the rubber roller is reflected by the following formula relative to the angular velocity of the fast and slow rollers. Compensation will be provided. ； in, For the target linear velocity, The initial radius of the rubber roller is given.

9. The multi-dimensional feature-aware experimental rice huller parameter adaptive matching method according to claim 1, characterized in that, In step 5, the actual rotational speed of the fast roller The deviation from the corresponding target speed and the actual speed of the slow roller The condition for the deviation from the corresponding target speed to enter the preset steady-state tolerance dead zone is: ； ； in, The target frequency of the fast roller corresponds to the target rotational speed of the fast roller. The target frequency of the slow roll corresponding to the target speed of the slow roll. To allow for the speed tolerance, For a moment The real-time stator electrical frequency output by the high-speed roller inverter. For a moment The real-time stator frequency output by the slow roller inverter; The actual physical displacement With final physical displacement command The condition for the deviation to enter the preset steady-state tolerance dead zone is: ； in, This represents the maximum allowable mechanical tolerance of the system. The global steady-state flag signal is defined as When all three conditions above are met simultaneously and the duration exceeds the anti-shake filtering time threshold, hour, ,otherwise ; The global steady-state flag signal is used to generate the opening command for the feed gate. : ； in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; The start time of the calibration. The preset safe dead time, This is a command signal indicating that the feed gate is fully open. This indicates that the screen is fully open.

10. A multi-dimensional feature-aware experimental rice huller parameter adaptive matching system, characterized in that, The system is used to implement the experimental rice huller parameter adaptive matching method for multi-dimensional feature perception as described in any one of claims 1 to 9, the system comprising: The static pre-inspection and feature extraction module is used to collect raw sensing data of rice samples to be processed in the feeding detection chamber when the feeding gate is in a physically locked state, and to perform orthogonal dimensionality reduction and statistical purification on the raw sensing data to construct a multidimensional orthogonal feature vector matrix. The online forward computation module is used to input the multidimensional orthogonal feature vector matrix into a system of parallel multi-channel forward continuous equations fixed in the lower-level execution unit. This system of parallel multi-channel forward continuous equations is pre-obtained through offline optimization training and includes fixed membership function parameters, conclusion parameters, and feature importance weight parameters. The lower-level execution unit solves the system of parallel multi-channel forward continuous equations through algebraic operations and outputs the theoretically optimal instruction matrix for the current batch of rice samples. ;in, The target rubber roller displacement clearance. The target linear velocity of the fast roller, The target linear velocity of the slow roller. For the target linear velocity difference, Separate air volume for target purpose; The cross-batch wear compensation module is used to obtain the total processing power equivalent accumulated by the experimental rice huller since the last zero-position calibration. The wear compensation increment is calculated based on pre-calibrated empirical wear fitting coefficients. The theoretically optimal instruction matrix The target rubber roller displacement gap With the wear compensation increment Superimpose the data to generate the final physical displacement command. ; The no-load command execution module is used to drive the rubber roller gap adjustment mechanism to execute the final physical displacement command. The frequency converters driving the fast and slow rollers respectively execute the target linear speed of the fast roller. and the target linear velocity of the slow roller Drive the separator fan to execute the target separation air volume ; The no-load steady-state closed-loop confirmation module is used to continuously poll the status register of the underlying driver in the no-load state to obtain the actual physical displacement. Actual rotational speed of the fast roller and the actual speed of the slow roller When determining the actual physical displacement With final physical displacement command The deviation, the actual speed of the fast roller Deviation from the corresponding target speed, actual speed of the slow roller The deviation from the corresponding target speed simultaneously enters the preset steady-state tolerance dead zone, and the duration exceeds the preset anti-shake filtering time threshold. At that time, output a global steady-state flag signal; The feeding and processing module is used to control the opening of the feeding gate in response to the global steady-state flag signal, so that the rice sample to be processed is subject to the final physical displacement command. Target linear velocity of the fast roller target linear velocity of slow roller and target separation air volume The shell removal process is completed under a defined constant physical field; within the processing window, all parameter correction loops based on real-time feedback are forcibly shut down.

Images