Corn bionic powder taking end parameter online optimization method and system

Abstract

The present application relates to the field of agricultural automation, and particularly relates to a corn bionic powder taking end parameter online optimization method and system. The method fuses multi-modal perception data obtained by real-time collection of vibration, sound and image signals, and extracts a comprehensive state feature vector. Based on the vector, a bionic optimization algorithm is used to iteratively calculate the optimal working parameter combination, and then the speed of the end driving motor, the vibration amplitude and the spatial posture angle are dynamically adjusted. The present application solves the problems of poor powder taking quality, low efficiency and high energy consumption caused by the inability of existing powder taking parameters to adapt to working condition changes in real time and low optimization precision, realizes intelligent real-time optimization of powder taking parameters, significantly improves powder taking quality and operation efficiency, reduces energy consumption and grain breakage rate, and adapts to different working condition requirements.

Description

Technical Field

[0001] This invention relates to the field of agricultural automation, and in particular to a method and system for online optimization of parameters at the end of a corn biomimetic powder extraction device. Background Technology

[0002] In corn breeding, deep processing, and quality testing, corn flour extraction is a fundamental and crucial process. The quality and efficiency of flour extraction directly impact the accuracy of subsequent experimental analysis, product quality, and production efficiency. During corn flour extraction, the operating parameters at the end of the extraction device (such as drive motor speed, vibration amplitude, and spatial attitude angle) have a decisive influence on the extraction effect. Different corn varieties, kernel maturity, and ear morphology vary significantly, and dynamic factors such as device wear and load fluctuations continuously change the operating conditions, making it difficult for fixed operating parameters to adapt to complex and changing extraction scenarios. Currently, the end-operating parameters of existing corn flour extraction devices are mostly determined through offline presets or manual experience adjustment. Offline presets cannot respond to changes in operating conditions in real time. When corn varieties are changed, ear morphology is uneven, or the device experiences slight wear, problems such as incomplete flour extraction, excessive powder splashing, and severe kernel damage easily occur, thus reducing flour extraction quality and efficiency. Manual experience adjustment relies on the operator's professional skills, resulting in low adjustment accuracy, slow response speed, and high labor intensity, making it difficult to achieve automated and intelligent control of the flour extraction process. Furthermore, some existing methods for optimizing millet extraction parameters rely solely on a single type of sensing data (such as monitoring only vibration signals), which cannot fully characterize the complex state of the millet extraction process. This results in insufficient accuracy in parameter optimization, making it difficult to achieve coordinated optimization of millet extraction quality, efficiency, and energy consumption. Consequently, these methods fail to meet the demands of modern corn milling operations for high precision, high efficiency, low energy consumption, and intelligence. Therefore, developing a method that can adapt to changes in operating conditions in real time, fully perceive the milling state, and achieve precise online parameter optimization has become an urgent problem to be solved in the field of corn milling technology. Summary of the Invention

[0003] This invention provides a method and system for online optimization of parameters at the corn biomimetic powder extraction end, aiming to solve the problems in the prior art where the powder extraction parameters cannot adapt to changes in working conditions in real time and have low optimization accuracy, resulting in poor powder extraction quality, low efficiency, and high energy consumption.

[0004] To achieve the above objectives, the following technical solution is adopted.

[0005] The method for online optimization of parameters at the powder extraction end of corn using biomimetic technology includes the following steps: Vibration signals, sound signals, and image data are collected in real time at the end of the corn flour extraction device to obtain multimodal sensing data of the flour extraction process; Feature extraction and fusion processing are performed on the multimodal sensing data of the powder extraction process to obtain a comprehensive state feature vector representing the current powder extraction state and quality; Based on the comprehensive state feature vector, a biomimetic optimization algorithm that simulates the intelligent cooperation mechanism of biological groups is used to iteratively calculate the optimal combination of working parameters at the end of the powder extraction device online. Based on the optimal combination of operating parameters, the rotational speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder extraction device are dynamically adjusted.

[0006] Optionally, the step of performing feature extraction and fusion processing on the multimodal sensing data of the powder extraction process to obtain a comprehensive state feature vector representing the current powder extraction state and quality specifically includes: Empirical mode decomposition is performed on the vibration signal to obtain multiple intrinsic mode function components. The sample entropy and energy ratio of each component are calculated to form a subset of vibration time-frequency domain features. Mel frequency cepstral coefficients are extracted from the sound signal, and its zero-crossing rate and short-time energy are calculated to form a feature subset in the sound frequency domain; Image data is preprocessed and region of interest segmented to extract surface texture roughness of corn ears, area ratio of powder splashing region, and contour features of unthreshed kernels, forming a subset of visual morphological features; The vibration time-frequency domain feature subset, the acoustic audio domain feature subset, and the visual morphology feature subset are processed by an attention-based feature fusion network. The feature fusion network dynamically assigns weights to each feature subset and performs weighted concatenation. Then, a fully connected layer is used for dimensionality reduction and information compression, and finally outputs a low-dimensional and information-dense comprehensive state feature vector.

[0007] Optionally, the step of using a biomimetic optimization algorithm that simulates the intelligent cooperative mechanism of biological groups to iteratively calculate the optimal combination of operating parameters at the end of the powder extraction device online specifically includes: Initialize an optimization population consisting of multiple virtual agents, each virtual agent representing a set of candidate operating parameters, including the speed of the drive motor, vibration amplitude, spatial pitch angle and spatial yaw angle; Define a fitness function for the virtual agent. The input of the fitness function is the comprehensive state feature vector, and the output is a comprehensive evaluation score. The comprehensive evaluation score is obtained by weighted summation of the powder collection efficiency score, the powder collection integrity score, and the energy consumption score. Simulating bird flock foraging behavior, each virtual agent updates its candidate working parameters for the next iteration based on its own historical best parameter position, the population's global best parameter position, and a random perturbation term. A simulated ant colony pheromone diffusion mechanism is introduced to assign virtual pheromones to the parameter regions represented by virtual agents with higher fitness, guiding other virtual agents to conduct regional fine-grained searches in high-quality parameter regions; When the optimal fitness value in the population changes less than the convergence threshold within several consecutive iterations, or when the maximum number of iterations is reached, the iteration is terminated, and the candidate working parameters corresponding to the position of the global optimal parameter in the population at this time are determined as the optimal working parameter combination.

[0008] Optionally, it also includes parameter initialization and constraint setting steps based on maize variety characteristics: Before starting online optimization, the variety identification features of the corn ear to be processed are obtained through the image recognition unit, or the corn variety code information input from the upstream process is received. Based on the variety identification features or the maize variety coding information, query the pre-established maize variety-working parameter knowledge base to obtain the recommended initial working parameter values ​​and parameter safety boundary constraints that match the variety. The recommended values ​​of the initial working parameters are used as the initial center position of the optimization population in the biomimetic optimization algorithm, and the parameter safety boundary constraints are used to limit the movement range of each virtual agent in the parameter space. The parameter safety boundary constraints are embedded in the iterative update formula of the biomimetic optimization algorithm in the form of inequalities to ensure that the candidate working parameter combinations generated in each generation are within the safe operating range allowed by the mechanical structure and process requirements.

[0009] Optionally, online safety intervention procedures may also be included for situations such as stalling or abnormal operating conditions of the powder collection device: During the online operation of the biomimetic optimization algorithm, the real-time current value of the drive motor is monitored in parallel to see if it exceeds the first safety threshold, and the amplitude of the vibration signal is monitored to see if it remains below the second safety threshold within a preset time window. If the real-time current value exceeds the first safety threshold, it is determined that there is a risk of stalling, the current optimization iteration process is immediately interrupted, and the first safety strategy is invoked. The first safety strategy forces the speed of the drive motor to be reduced to a preset safe speed and reverses for a short period of time to try to eliminate the blockage. If the vibration signal amplitude remains below the second safety threshold, it is determined that there is an abnormal working condition where the powder collection device has lost contact with the corn cob or the efficiency is extremely low. The current optimization iteration process is immediately interrupted, and the second safety strategy is invoked. The second safety strategy controls the end of the powder collection device to perform a preset spatial posture search motion until the vibration signal amplitude returns to the normal range, and then the bionic optimization algorithm is restarted.

[0010] Optionally, it also includes a multi-objective dynamic weight adjustment step in the optimization process: During the iterative process of the biomimetic optimization algorithm, the average powder extraction efficiency, average powder extraction integrity, and average unit energy consumption in recent historical data are calculated in real time. A dynamic weight adjuster based on fuzzy logic rules is established. The input of the dynamic weight adjuster is the deviation of the average fan acquisition efficiency, average fan acquisition integrity and average unit energy consumption from their respective set targets. The output is the adjustment amount of the fan acquisition efficiency score weight, fan acquisition integrity score weight and energy consumption score weight in the fitness function. When the average fan acquisition efficiency is consistently lower than the target, the weight of the fan acquisition efficiency score is increased; when the average fan acquisition integrity is consistently lower than the target, the weight of the fan acquisition integrity score is increased; when the average unit energy consumption is consistently higher than the target, the weight of the energy consumption score is increased; the adjusted weights are applied to the fitness calculation of subsequent iteration cycles to guide the biomimetic optimization algorithm to search towards the current performance target that needs to be optimized.

[0011] Optionally, it also includes steps for verifying the optimal combination of working parameters and updating the knowledge base: After the biomimetic optimization algorithm outputs a set of optimal working parameters and applies them to the powder collection device, new multimodal sensing data of the powder collection process is continuously collected within a fixed verification time window. Based on the new multimodal sensing data of the powder extraction process, the actual powder extraction efficiency, powder extraction integrity and energy consumption indicators are calculated to obtain the actual performance index set; The actual performance index set is compared with the performance index predicted during the optimization process. If the overall improvement is higher than the verification threshold, the optimization is considered successful. The successful maize variety identification features, the final optimal working parameter combination, and the corresponding actual performance index set are added as a new sample pair to the maize variety-working parameter knowledge base to enrich the initial parameter recommendations for subsequent similar varieties.

[0012] Optionally, the step of dynamically adjusting the rotational speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder-collecting device relative to the corn cob specifically includes: The optimal combination of operating parameters is analyzed as the target speed of the drive motor, the target amplitude of the vibration device, the target pitch axis angle, and the target yaw axis angle. By adjusting the output of the power electronic converter of the drive motor through a speed closed-loop control circuit, the actual speed of the motor is made to track the target speed value. The excitation signal of the vibration generator is adjusted by the amplitude closed-loop control loop so that the actual vibration amplitude tracks the target amplitude. By using a dual closed-loop control loop for spatial attitude, the service motors of the pitch axis and yaw axis are adjusted respectively, so that the actual pitch angle and yaw angle at the end of the powder collection device track the target angle. During the adjustment process, the tracking error of each loop is monitored in real time. If the tracking error of any loop continues to exceed the allowable tolerance, a parameter fine-tuning command is triggered. The biomimetic optimization algorithm is used to perform local re-optimization in a small range near the optimal combination of working parameters to compensate for the effects of mechanical wear or load changes.

[0013] The online optimization system for parameters at the corn biomimetic powder extraction end includes: Multimodal sensing data acquisition module, comprehensive state feature extraction and fusion module, biomimetic online optimization calculation module, and end effector dynamic adjustment module; The multimodal sensing data acquisition module is used to collect vibration signals, sound signals and image data of the corn flour extraction device in real time during operation, obtain multimodal sensing data of the flour extraction process, and send the multimodal sensing data of the flour extraction process to the comprehensive state feature extraction and fusion module. The comprehensive state feature extraction and fusion module is used to perform feature extraction and fusion processing on the received multimodal perception data of the powder extraction process to obtain a comprehensive state feature vector that represents the current powder extraction state and quality, and send the comprehensive state feature vector to the biomimetic online optimization calculation module. The biomimetic online optimization calculation module is used to calculate the optimal combination of working parameters at the end of the powder extraction device online based on the received comprehensive state feature vector and a biomimetic optimization algorithm that simulates the intelligent cooperation mechanism of biological groups, and then send the optimal combination of working parameters to the end effector dynamic adjustment module. The end effector dynamic adjustment module is used to dynamically adjust the speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder collection device relative to the corn cob, based on the received optimal working parameter combination.

[0014] Optionally, the biomimetic online optimization calculation module includes: an optimization population management unit, a fitness evaluation unit, and a biomimetic collaborative iteration unit; The optimization population management unit is used to initialize and maintain an optimization population consisting of multiple virtual agents. Each virtual agent represents a set of candidate working parameters, including drive motor speed, vibration amplitude, spatial pitch angle and spatial yaw angle. The fitness evaluation unit is connected to the comprehensive state feature extraction and fusion module and the optimization population management unit. It is used to receive the comprehensive state feature vector and, according to the candidate working parameters provided by the optimization population management unit, call the preset fitness function to calculate the comprehensive evaluation score of each virtual agent. The comprehensive evaluation score is obtained by weighted summation of the powder collection efficiency score, the powder collection integrity score and the energy consumption score. The biomimetic collaborative iteration unit is connected to the fitness evaluation unit and the optimization population management unit. It is used to simulate the collaborative mechanism of bird flock foraging and ant colony pheromone diffusion. Based on the historical optimal parameters of each virtual agent, the global optimal parameters of the population, and the virtual pheromone distribution, it drives the optimization population management unit to update the candidate working parameters of all virtual agents. When the preset convergence condition is met, the final global optimal parameters of the population are output as the optimal working parameter combination.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This application achieves comprehensive acquisition of multimodal perception data during the corn milling process by real-time collection of vibration signals, sound signals, and image data from the end of the corn milling device. Feature extraction and fusion processing of this multimodal data accurately characterizes the current milling status and quality. Combined with a biomimetic optimization algorithm simulating intelligent collaborative mechanisms of biological groups, it enables online iterative calculation of the optimal combination of working parameters at the end of the milling device. Based on these optimal parameters, the drive motor speed, vibration amplitude, and spatial attitude angle are dynamically adjusted. This effectively solves the technical problems in existing technologies where milling parameters cannot adapt to changes in working conditions in real time, resulting in low optimization accuracy, poor milling quality, low efficiency, and high energy consumption. It achieves intelligent and real-time optimization of milling parameters, significantly improving milling quality and operational efficiency while reducing equipment energy consumption and grain breakage rate, adapting to corn milling needs under different working conditions. Targeted feature extraction methods are employed for multimodal sensing data, extracting features from the vibration time-frequency domain, acoustic domain, and visual morphology dimensions. A feature fusion network with an attention mechanism dynamically allocates weights and fuses features, highlighting key features and suppressing redundant information. This improves the accuracy and reliability of the comprehensive state feature vector, providing high-quality data support for subsequent parameter optimization. The biomimetic optimization algorithm combines bird flock foraging and ant colony pheromone diffusion mechanisms, ensuring both global search capability and improved local fine-grained search efficiency. It can quickly converge to the optimal parameter combination, shortening optimization time and improving optimization accuracy.

[0016] Based on the characteristics of different corn varieties, parameter initialization and constraint setting can utilize a pre-established knowledge base to provide suitable initial parameters and safety constraints for different corn varieties, accelerating the convergence speed of optimization while preventing parameters from exceeding the safety range required by the mechanical structure and process, thus improving the safety and stability of equipment operation. An online safety intervention mechanism during the milling process can monitor abnormal operating conditions such as stalling and device disengagement in real time, and promptly implement targeted handling strategies to avoid equipment damage and operation interruption, ensuring the continuity and safety of the milling process. Multi-objective dynamic weight adjustment during the optimization process can dynamically adjust the fitness function weights based on real-time performance index deviations, achieving synergistic optimization of milling efficiency, milling integrity, and energy consumption, flexibly adapting to different operational needs. The optimal parameter verification and knowledge base update mechanism can continuously enrich the knowledge base samples, improve the accuracy of subsequent initial parameter recommendations, and form a closed-loop optimization of "optimization-verification-update," further improving the stability and optimization effect of long-term operation. The multi-closed-loop control strategy of the end effector can ensure that each working parameter accurately tracks the target value, reduce parameter adjustment error, and at the same time compensate for the impact of mechanical wear and load changes through local re-optimization, maintaining the stability of long-term flour extraction effect. The corresponding online optimization system, through the collaborative work of various modules, realizes integrated management and control of flour extraction process perception, feature fusion, parameter optimization and dynamic adjustment, promoting the development of corn flour extraction operations towards automation, intelligence and high precision, reducing reliance on manual labor and improving the level of industrial scale operation. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart illustrating the steps of an embodiment of the online optimization method for corn biomimetic powder extraction parameters according to the present invention.

[0018] Figure 2 This is a schematic diagram of the system modules of an embodiment of the corn biomimetic powder extraction end parameter online optimization system of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0020] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0021] Example 1 like Figure 1As shown, this embodiment is used to achieve online optimization of the terminal working parameters of the corn flour extraction device, adapting to the flour extraction needs of different corn varieties and different working conditions. It solves the problems of fixed parameters, lag in adjustment, and poor adaptability in traditional flour extraction, improves flour extraction quality and operating efficiency, and reduces energy consumption. It can be widely used in automated flour extraction operation scenarios in corn breeding, deep processing, and quality inspection, and is suitable for various electric and pneumatic driven corn flour extraction devices, especially for small and medium-sized batch flour extraction operations with high requirements for flour extraction accuracy and automation.

[0022] The implementation of this embodiment is based on a conventional industrial control hardware platform and data processing module. No additional special equipment is required. The deployment and operation of the entire method can be completed simply by integrating the corresponding sensing device, control module and actuator at the end of the existing corn flour extraction device. The deployment cost is low and the compatibility is strong, which facilitates the upgrading and transformation of existing equipment.

[0023] The specific implementation process is as follows: First, prepare for the powder extraction operation by checking the mechanical structure of the powder extraction device to ensure that there is no jamming or abnormal wear of the drive motor, vibration generator, spatial attitude adjustment mechanism, and other actuators. Check that the installation of each sensing device is secure and that signal transmission is smooth to ensure stable acquisition of vibration signals, sound signals, and image data. Check that the power supply to the control module is stable and that the data processing program is loaded normally to ensure that subsequent steps can be executed smoothly. At the same time, ensure that the pre-established corn variety-working parameter knowledge base has been initialized. The knowledge base stores recommended initial working parameters and parameter safety boundary constraints for various common corn varieties. The knowledge base adopts an expandable storage format to facilitate the updating and retrieval of newly added corn variety samples and optimized parameter information.

[0024] After preparation, the corn harvesting device is started, and the various steps of the online optimization method are executed. First, real-time acquisition of vibration, sound, and image data from the end of the corn harvesting device is performed to obtain multimodal sensing data of the harvesting process. Vibration signal acquisition is achieved through a vibration sensor installed on the housing of the actuator at the end of the harvesting device. The vibration sensor is a conventional piezoelectric vibration sensor, installed close to the harvesting head and away from vibration interference from the drive motor. This ensures that the acquired vibration signal accurately reflects the vibration generated by the interaction between the harvesting head and the corn cob during the harvesting process, rather than interference from the drive motor itself. The sensor's acquisition direction is consistent with the vibration direction of the harvesting head, and the acquisition frequency is reasonably set according to the actual working conditions of the harvesting operation to ensure complete capture of the vibration characteristics during the harvesting process. The acquired vibration signal is an analog signal, which is filtered, amplified, and noise-reduced by the sensor's built-in signal conditioning module before being converted into a digital signal and transmitted to the data processing module, completing the vibration signal acquisition and preliminary preprocessing.

[0025] The sound signal is acquired through a sound acquisition device installed near the end of the powder-collecting device. The sound acquisition device uses a conventional omnidirectional microphone, which is installed at a certain distance from the powder-collecting head. This ensures that the sound generated during the powder-collecting process can be clearly acquired, including the sound generated by the friction between the powder-collecting head and the corn cob, the sound generated by powder splashing, and the sound generated by the threshing of kernels. At the same time, it is necessary to avoid the microphone being contaminated by powder or damaged by the vibration of the powder-collecting device if it is too close. The acquisition frequency of the microphone is matched with the acquisition frequency of the vibration signal to ensure the time synchronization of multimodal data. The acquired sound signal is also preprocessed by filtering, noise reduction, and amplification to remove environmental noise and noise interference generated by the operation of the drive motor, so as to obtain digital sound signal data that can truly reflect the powder-collecting status and transmit it to the data processing module.

[0026] Image data acquisition is achieved through an industrial camera installed above the end of the corn cob extraction device. The industrial camera is a conventional color industrial camera equipped with a lens of appropriate focal length. The shooting angle is adjusted to fully cover the interaction area between the extraction head and the corn cob, ensuring clear capture of the surface condition of the corn cob, the working status of the extraction head, powder splashing, and the distribution of unthreshed kernels. The shooting frame rate of the industrial camera is reasonably set according to the speed of the extraction operation to ensure real-time capture of dynamic changes during the extraction process and avoid the loss of key image frames due to a low frame rate. The captured image data is in color, and the resolution is set according to the actual extraction accuracy requirements. After the image data is transmitted to the data processing module, it undergoes preliminary preprocessing, including image denoising, image enhancement, and size normalization, to remove dust noise, light reflection interference, etc., and improve image clarity, laying the foundation for subsequent feature extraction.

[0027] During the data acquisition process, it is crucial to ensure the time synchronization of vibration signals, sound signals, and image data. This is achieved through a synchronization clock module within the control module, keeping the synchronization error within a reasonable range to prevent deviations in subsequent feature fusion due to time asynchrony, which could affect the accurate representation of the dust extraction status. The acquired data is uniformly stored in the cache unit of the data processing module, marked according to the acquisition timestamp, forming a multimodal sensing data set for the dust extraction process, which is then used for subsequent feature extraction and fusion processing.

[0028] Next, feature extraction and fusion processing are performed on the multimodal sensing data of the powder extraction process to obtain a comprehensive state feature vector representing the current powder extraction state and quality. First, feature extraction is performed on the vibration signal. Specifically, the empirical mode decomposition method is used to decompose the pre-processed vibration signal. The empirical mode decomposition process does not require any preset basis functions and can adaptively decompose according to the characteristics of the vibration signal itself, decomposing the complex vibration signal into multiple intrinsic mode function components and a residual component. The residual component reflects the overall trend of the vibration signal, and the multiple intrinsic mode function components reflect the local characteristics of the vibration signal in different frequency ranges. During the decomposition process, through steps such as continuously screening extreme points, fitting envelopes, and removing trend terms, it is ensured that each intrinsic mode function component meets the basic requirements of intrinsic mode functions, that is, the number of extreme points and the number of zero crossings in the component are equal or differ by no more than one, and at any time, the upper and lower envelopes of the signal are symmetrical about the time axis.

[0029] After decomposition, residual components are removed, retaining only multiple intrinsic mode function (IMF) components. For each IMF component, sample entropy and energy ratio are calculated. Sample entropy characterizes the complexity and irregularity of the vibration signal; a smaller sample entropy value indicates a more regular vibration signal and a more stable powder extraction process, while a larger value indicates a more complex vibration signal and potential anomalies in the powder extraction process. Energy ratio characterizes the proportion of energy of each IMF component in the entire vibration signal; a larger energy ratio indicates a greater influence of the frequency range of that component on the entire powder extraction process. When calculating sample entropy, appropriate embedding dimensions and similarity tolerances are set to ensure the accuracy and stability of the calculation results. The embedding dimension is reasonably set based on the acquisition frequency and complexity of the vibration signal, and the similarity tolerance is set based on the amplitude range of the vibration signal. When calculating the energy ratio, the ratio of the energy of each IMF component to the total energy of all IMF components is used to obtain the energy ratio of each component. The sample entropy and energy ratio of all intrinsic mode function components are integrated to form a vibration time-frequency domain feature subset. This feature subset can comprehensively reflect the time-frequency domain characteristics of the vibration signal during the powder extraction process, providing a basis for subsequent judgment of the powder extraction status based on the vibration dimension.

[0030] Then, feature extraction is performed on the audio signal. First, the pre-processed audio signal is divided into frames using an overlapping framing method. The frame length and frame shift are reasonably set according to the acquisition frequency of the audio signal and the changing characteristics of the sound during the extraction process, ensuring that the short-term features of the audio signal can be fully captured, while avoiding feature blurring due to excessively long frame lengths or excessive computational load due to excessively short frame lengths. After framing, Mel-frequency cepstral coefficients are extracted from each frame of the audio signal. Mel-frequency cepstral coefficients can simulate the auditory characteristics of the human ear and effectively represent the frequency domain features of the audio signal. The extraction process first maps the frequency domain response of the audio signal onto the Mel-frequency axis, filters the frequency domain signal through a Mel-filter bank to obtain the Mel-frequency spectrum, takes the logarithm of the Mel-frequency spectrum, and finally performs a discrete cosine transform to remove the correlation between features in each dimension, obtaining the Mel-frequency cepstral coefficients. The order of the extracted Mel-frequency cepstral coefficients is set according to actual needs to ensure that it can comprehensively reflect the frequency domain features of the audio signal.

[0031] Based on the extraction of Mel frequency cepstral coefficients, the zero-crossing rate and short-time energy of each frame of the sound signal are calculated simultaneously. The zero-crossing rate characterizes the number of times the sound signal crosses the zero level per unit time, reflecting the frequency characteristics and the degree of change of the sound signal. During the powder extraction process, powder splashing, kernel threshing, and other factors can cause significant changes in the zero-crossing rate of the sound signal. The short-time energy characterizes the energy level of each frame of the sound signal, reflecting the intensity of the sound. Higher powder extraction efficiency and more thorough kernel threshing may result in higher short-time energy, while the short-time energy will decrease significantly when the powder extraction device loses contact with the corn cob. The zero-crossing rate is calculated by counting the number of times the sound signal crosses the zero level in each frame and then normalizing the result. The short-time energy is calculated by summing the squares of the sound signals in each frame and then normalizing the result to ensure that the energy value is within a reasonable range. The extracted Mel frequency cepstral coefficients, zero-crossing rates of each frame, and short-time energy are integrated to form a subset of acoustic frequency domain features. This subset of features can comprehensively reflect the frequency domain features and short-time change features of the sound signal during the powder extraction process, providing a basis for judging the sound dimension of the powder extraction status.

[0032] Subsequently, feature extraction is performed on the image data. First, the color image, which has undergone preliminary preprocessing (denoising, enhancement, and size normalization), undergoes further preprocessing operations, including grayscale conversion and threshold segmentation. Grayscale conversion converts the color image into a grayscale image, reducing the amount of data and improving the efficiency of subsequent processing. The grayscale conversion uses a weighted average method, setting weights based on the sensitivity of the human eye to different color channels to ensure that the grayscale image can accurately reflect the brightness distribution of the original color image. Threshold segmentation is used to separate the background area from the target area (corn ears, powder extraction head, powder splash area, unthreshed kernels, etc.) in the grayscale image. By setting an appropriate threshold, areas with grayscale values ​​greater than the threshold are identified as target areas, and areas with grayscale values ​​less than or equal to the threshold are identified as background areas. The threshold setting adopts an adaptive threshold segmentation method, which can automatically adjust the threshold according to the local grayscale distribution of the image to avoid poor segmentation results caused by factors such as changes in light and dust interference.

[0033] After thresholding, region of interest (ROI) segmentation is performed. Image segmentation algorithms are used to segment the corn cob region, powder splash region, and unthreshed kernel region within the target area. ROI segmentation employs a combination of contour detection and region growing. First, a contour detection algorithm extracts all target contours in the image. Then, based on the morphological and size characteristics of the corn cob, powder splash region, and unthreshed kernel, the extracted contours are filtered and classified to determine the contour range of each ROI. Finally, a region growing algorithm is used to fill and optimize each ROI, ensuring that the segmented ROI is complete, accurate, and without omissions or missegments.

[0034] For the segmented surface region of the corn ear, surface texture roughness features are extracted. Texture roughness characterizes the roughness of the corn ear surface, reflecting the maturity and kernel density. The rougher the surface texture, the denser the kernel distribution and the higher the maturity, but the more difficult it is to extract pollen. Texture roughness is extracted by calculating the gray-level co-occurrence matrix of the corn ear surface region, and then extracting feature parameters such as contrast, correlation, and entropy from the gray-level co-occurrence matrix. These feature parameters are then fused to obtain feature values ​​characterizing the surface texture roughness of the corn ear. For the segmented powder splashing region, its area ratio is calculated, which is the ratio of the pixel area of ​​the powder splashing region to the pixel area of ​​the entire image. The smaller the area ratio of the powder splashing region, the smoother the pollen extraction process, the less powder waste, and the higher the pollen extraction quality. Conversely, a larger area ratio indicates that the pollen extraction process may have unreasonable parameters leading to excessive powder splashing. For the segmented unclamped kernel area, its contour features are extracted, including features such as perimeter, area, roundness, and rectangularity. These contour features characterize the size, shape, and distribution of the unclamped kernels. The contour features of the unclamped kernels reflect the integrity of flour extraction; fewer and smaller unclamped kernels indicate more complete flour extraction and higher extraction quality. The extracted surface texture roughness of the corn cob, the proportion of powder splash area, and the contour features of the unclamped kernels are integrated to form a visual morphological feature subset. This feature subset comprehensively reflects the state of the corn cob, powder splashing, and extraction integrity during the extraction process, providing a visual dimension for judging the extraction status and quality.

[0035] After extracting the three feature subsets, an attention-based feature fusion network is used to fuse the three feature subsets to obtain a comprehensive state feature vector. The feature fusion network adopts a deep learning network architecture, consisting of an input layer, an attention weight allocation layer, a weighted concatenation layer, a fully connected layer, and an output layer. The input layer receives the vibration time-frequency domain feature subset, the acoustic audio-visual domain feature subset, and the visual morphological feature subset, and converts the three feature subsets into feature matrix forms suitable for network processing. Each feature subset corresponds to an independent input channel, ensuring that the three feature subsets can be input into the network for processing in parallel.

[0036] The attention weight allocation layer dynamically assigns weights to each feature subset. The weight allocation adaptively adjusts based on the current powder extraction conditions and the representational ability of each feature subset. Feature subsets with stronger representational ability receive larger weights and have a greater impact on the overall state feature vector; conversely, feature subsets with weaker representational ability receive smaller weights and have a smaller impact on the overall state feature vector. The attention weight allocation layer calculates the feature importance score for each feature subset and then normalizes the score using a softmax function to obtain the weight coefficient for each feature subset. The feature importance score is calculated based on the correlation between the feature subset and the powder extraction state and quality; the higher the correlation, the higher the score and the larger the weight coefficient. For example, when excessive powder splatter occurs during extraction, the visual morphology feature subset has a stronger representational ability, and the attention weight allocation layer automatically increases the weight coefficient of the visual morphology feature subset. When abnormal vibration occurs during extraction, the vibration time-frequency domain feature subset has a stronger representational ability, and the weight coefficient of the vibration time-frequency domain feature subset is automatically increased.

[0037] The weighted concatenation layer multiplies each of the three feature subsets by its corresponding weight coefficient, resulting in weighted feature subsets. These weighted subsets are then concatenated to form a high-dimensional fusion feature matrix. This high-dimensional fusion feature matrix contains all key feature information from the vibration, sound, and visual dimensions. However, its high dimensionality and information redundancy necessitate further dimensionality reduction and information compression. The fully connected layer performs dimensionality reduction and information compression on the high-dimensional fusion feature matrix. This layer contains multiple hidden layers, each using an activation function to perform a non-linear transformation on the input features, gradually reducing the feature dimensionality and removing information redundancy. Finally, the output layer outputs a low-dimensional, information-dense comprehensive state feature vector. The dimensionality of this comprehensive state feature vector is set according to actual needs, ensuring it comprehensively represents the current powder extraction state and quality while reducing the computational load of subsequent optimization algorithms and improving optimization efficiency. Before use, the feature fusion network needs to be trained. The training data consists of multimodal sensing data collected under different corn varieties and different milling conditions, as well as the corresponding milling status and quality labels. The network parameters are iteratively optimized through the backpropagation algorithm to ensure that the network can accurately achieve feature fusion. The output comprehensive state feature vector can truly and accurately represent the current milling status and quality.

[0038] After obtaining the comprehensive state feature vector, a biomimetic optimization algorithm simulating the intelligent cooperation mechanism of biological swarms is used to iteratively calculate the optimal combination of operating parameters at the end of the pollen extraction device online. This biomimetic optimization algorithm combines the advantages of bird flock foraging behavior and ant colony pheromone diffusion mechanisms. It possesses both strong global search capabilities, avoiding getting trapped in local optima, and high efficiency in fine-grained local search, quickly converging to the global optimum. This makes it suitable for online real-time optimization of pollen extraction parameters.

[0039] First, the optimization population is initialized, consisting of multiple virtual agents. Each virtual agent corresponds to a set of candidate operating parameters, including the drive motor's rotational speed, vibration amplitude, spatial pitch angle, and spatial yaw angle. These four parameters collectively determine the working state of the dust collection device's end effector, directly impacting dust collection quality, efficiency, and energy consumption. The number of virtual agents is reasonably set based on actual computing power and optimization accuracy requirements. Too many agents will increase computational load and reduce optimization efficiency, while too few agents will result in insufficient search range and difficulty in finding the global optimum. A balance must be struck between optimization accuracy and efficiency.

[0040] During initialization, the candidate working parameters for each virtual agent are not randomly set, but are initialized and constrained based on the characteristics of the corn variety. Specifically, before starting online optimization, the image recognition unit obtains the variety identification features of the corn ear to be processed. The image recognition unit shares hardware with the industrial camera that collects image data. The image recognition algorithm identifies the variety of the corn ear image and extracts the variety identification features of the corn ear. The variety identification features include the length, diameter, kernel arrangement, kernel color, and other characteristic parameters that can distinguish different corn varieties. If the flour extraction device is linked with the upstream process, it can also directly receive the corn variety code information input from the upstream process. The corn variety code information is a unique code assigned by the upstream process after classifying the corn variety, which can quickly and accurately identify the corn variety.

[0041] After obtaining the corn variety identification features or corn variety coding information, a pre-established corn variety-operating parameter knowledge base is queried. This knowledge base stores recommended initial operating parameter values ​​and parameter safety boundary constraints for various common corn varieties. The recommended initial operating parameter values ​​are determined based on the optimal parameter statistics and experimental data of the corn variety in historical milling operations. This provides an initial search starting point close to the optimal solution for the optimization algorithm, accelerating the optimization convergence speed. The parameter safety boundary constraints are used to limit the value range of each candidate operating parameter. They are set based on the mechanical structural strength, process requirements, and milling quality requirements of the milling device, ensuring that the candidate operating parameters do not exceed the equipment's tolerance range and the process's allowable range, avoiding equipment damage or a serious decline in milling quality due to unreasonable parameters.

[0042] After obtaining the recommended initial working parameters and parameter safety boundary constraints that match the current maize variety, the recommended initial working parameters are used as the initial center position of the optimization population in the biomimetic optimization algorithm. The candidate working parameters corresponding to each virtual agent are randomly generated around this initial center position to ensure that the candidate parameters of the initial population are all within a reasonable search range and close to the optimal solution region. At the same time, parameter safety boundary constraints are used to limit the movement range of each virtual agent in the parameter space. The parameter safety boundary constraints are embedded in the iterative update formula of the biomimetic optimization algorithm in the form of inequalities. When updating the candidate working parameters in each iteration, the updated parameters are checked for boundaries. If the parameters exceed the safety boundary constraints, they are pruned or adjusted to ensure that the candidate working parameter combinations generated in each generation are within the safe operating range allowed by the mechanical structure and process requirements, thus ensuring the safe operation of the equipment and the stability of the powder extraction operation.

[0043] After initializing the optimization population, a fitness function for the virtual agents is defined. This function evaluates the merits of the candidate working parameters for each virtual agent. The input to the fitness function is a comprehensive state feature vector, and the output is a comprehensive evaluation score. A higher comprehensive evaluation score indicates better candidate working parameters, making them more suitable for the current pollen harvesting conditions; a lower comprehensive evaluation score indicates worse candidate working parameters, requiring optimization. The comprehensive evaluation score is obtained by weighted summation of pollen harvesting efficiency, pollen harvesting integrity, and energy consumption scores. The weighting coefficients are set according to the actual pollen harvesting requirements and can be adjusted based on different operational scenarios. For example, in corn breeding pollen harvesting scenarios, pollen harvesting integrity and quality have higher priority, so the weighting coefficient for pollen harvesting integrity can be appropriately increased; in large-scale deep processing pollen harvesting scenarios, pollen harvesting efficiency and energy consumption have higher priority, so the weighting coefficients for pollen harvesting efficiency and energy consumption scores can be appropriately increased.

[0044] The flour extraction efficiency score is calculated based on relevant feature parameters in the comprehensive state feature vector, reflecting the flour extraction speed and operational efficiency corresponding to the candidate working parameters. The calculation basis includes feature parameters such as the proportion of the flour splash area and the regularity of the vibration signal. For example, a moderate proportion of the flour splash area and strong regularity of the vibration signal indicate a stable flour extraction process, higher flour extraction efficiency, and a higher flour extraction efficiency score. The flour extraction integrity score is also calculated based on relevant feature parameters in the comprehensive state feature vector, reflecting the thoroughness of flour extraction corresponding to the candidate working parameters. The calculation basis includes feature parameters such as the outline characteristics of unthreshed kernels and the surface texture roughness of the corn cob. For example, fewer unthreshed kernels and smaller outline sizes indicate more complete flour extraction, resulting in a higher flour extraction integrity score. The energy consumption score is indirectly calculated based on parameters such as the vibration signal energy characteristics and the short-time energy of the sound signal in the comprehensive state feature vector, reflecting the equipment energy consumption level corresponding to the candidate working parameters. A higher energy consumption score indicates lower equipment energy consumption. For example, moderate vibration signal energy and stable short-time sound signal energy indicate lower equipment operating energy consumption, resulting in a higher energy consumption score.

[0045] After the fitness function is defined, online iterative computation begins. During the iteration, bird flock foraging behavior is simulated. Each virtual agent updates its candidate working parameters for the next iteration based on its own historical best parameter position, the population's global best parameter position, and a random perturbation term. The historical best parameter position refers to the position of the candidate working parameter with the highest fitness score in all previous iterations for that virtual agent. The population's global best parameter position refers to the position of the candidate working parameter with the highest fitness score in the entire optimization population in all previous iterations. The random perturbation term increases the randomness of the optimization process, preventing it from getting trapped in local optima. The range of values ​​for the random perturbation term is dynamically adjusted according to the number of iterations. In the early stages of iteration, the range of values ​​for the random perturbation term is larger, increasing the global search scope and ensuring a comprehensive search of the parameter space. In the later stages of iteration, the range of values ​​for the random perturbation term gradually decreases, reducing search randomness, improving local search accuracy, and accelerating convergence.

[0046] The update formula for candidate working parameters for each virtual agent is constructed based on the group cooperation mechanism of bird flock foraging. It combines the guidance of its own historical best and the global best of the population, while introducing random perturbations to ensure that each virtual agent can move towards a better parameter position while maintaining a certain level of exploration capability, thus preventing the population from getting trapped in local optima. Specifically, when updating candidate working parameters, each virtual agent first calculates the movement direction and movement step size based on its own historical best parameter position and the global best parameter position of the population. The movement step size is dynamically adjusted according to changes in fitness score. When the fitness score increases rapidly, the movement step size is appropriately increased to accelerate the optimization speed; when the fitness score increases slowly, the movement step size is appropriately decreased to improve the optimization accuracy. Then, a random perturbation term is added to fine-tune the movement direction and movement step size, ensuring the randomness and comprehensiveness of the optimization process.

[0047] While simulating bird flock foraging behavior for a global search, a simulated ant colony pheromone diffusion mechanism is introduced. Virtual pheromones are assigned to parameter regions represented by virtual agents with higher fitness, guiding other virtual agents to conduct refined regional searches towards higher-quality parameter regions, further improving optimization accuracy and convergence speed. The concentration of virtual pheromones is positively correlated with the fitness score of the virtual agent; the higher the fitness score of the virtual agent, the higher the concentration of virtual pheromones in its corresponding parameter region; conversely, the lower the fitness score of the virtual agent, the lower the concentration of virtual pheromones in its corresponding parameter region, even gradually decaying to zero.

[0048] The diffusion process of virtual pheromones follows the natural diffusion law of ant colony pheromones. The pheromone gradually decays with the increase of iterations, avoiding interference from old pheromones in subsequent optimization processes. Simultaneously, after each iteration, virtual pheromones are replenished to the parameter regions corresponding to virtual agents with high fitness scores, maintaining a high pheromone concentration in those regions. When updating candidate working parameters, other virtual agents are guided by the virtual pheromone concentration, tending to move towards parameter regions with high virtual pheromone concentrations, i.e., towards parameter regions with higher fitness. This achieves regional fine-grained search, further optimizing parameter accuracy and shortening the optimization time based on the global search.

[0049] The iterative process also includes a multi-objective dynamic weight adjustment step in the optimization process. This step dynamically adjusts the weight coefficients of each scoring item in the fitness function based on changes in real-time follower acquisition performance indicators, guiding the biomimetic optimization algorithm to search towards the most pressing performance objective and improving its adaptability and flexibility. Specifically, during the iterative process of the biomimetic optimization algorithm, the average follower acquisition efficiency, average follower acquisition completeness, and average unit energy consumption in recent historical data are calculated in real time. The time window for recent historical data is reasonably set according to the iteration cycle and follower acquisition speed to ensure that it reflects the recent follower acquisition performance status and avoids excessive deviation in calculation results due to insufficient data or response lag due to excessive data.

[0050] A dynamic weight adjuster based on fuzzy logic rules is established. The inputs of the dynamic weight adjuster are the deviations of the average fan acquisition efficiency, average fan acquisition integrity, and average unit energy consumption from their respective set targets. The outputs are the adjustment amounts of the fan acquisition efficiency score weight, fan acquisition integrity score weight, and energy consumption score weight in the fitness function. The fuzzy logic rules are set based on the actual needs and experience of the toner collection operation, covering weight adjustment strategies under different performance deviations. For example, when the average toner collection efficiency is consistently lower than the set target, it indicates that the current toner collection efficiency is insufficient and needs to be optimized first. At this time, the dynamic weight adjuster outputs a positive adjustment amount to the toner collection efficiency score weight, increasing the weight coefficient of the toner collection efficiency score, while appropriately decreasing the weight coefficients of the toner collection integrity score and energy consumption score, guiding the optimization algorithm to prioritize searching for parameter combinations that can improve toner collection efficiency. When the average toner collection integrity is consistently lower than the set target, it indicates that the current toner collection is not thorough enough and needs to be optimized first. At this time, the dynamic weight adjuster outputs a positive adjustment amount to the toner collection integrity score weight, increasing the weight coefficient of the toner collection integrity score. When the average unit energy consumption is consistently higher than the set target, it indicates that the current equipment energy consumption is too high and needs to be optimized first. At this time, the dynamic weight adjuster outputs a positive adjustment amount to the energy consumption score weight, increasing the weight coefficient of the energy consumption score.

[0051] The adjustment amount output by the dynamic weight adjuster is applied to the fitness calculation in subsequent iterations, and the weight coefficients of the fitness function are updated in real time. This ensures that the optimization algorithm can dynamically adjust the optimization direction according to the changes in real-time powder picking performance, so as to achieve coordinated optimization of powder picking efficiency, powder picking integrity and energy consumption, and meet the performance requirements under different operating scenarios.

[0052] During the iteration process, the changes in the optimal fitness value in the population are continuously monitored to determine whether the convergence condition is met. The convergence condition is that the change in the optimal fitness value in the population is less than the convergence threshold over multiple consecutive iteration cycles, or the maximum number of iterations is reached. The convergence threshold is set according to the required optimization accuracy. The smaller the convergence threshold, the higher the optimization accuracy, but the longer the optimization time; the larger the convergence threshold, the lower the optimization accuracy, but the shorter the optimization time. A reasonable balance needs to be struck based on the actual powder extraction accuracy requirements. The maximum number of iterations is set according to the computational power and optimization efficiency requirements to avoid wasting computational resources due to the optimization process getting stuck in a loop or too slow convergence, ensuring that the optimization process can be completed within a reasonable time.

[0053] When the convergence condition is met, the iteration terminates, and the candidate working parameters corresponding to the current position of the globally optimal parameter in the population are determined as the optimal working parameter combination. If the convergence condition is not met, the next round of iteration continues until the convergence condition is met. During the iteration process, if a sudden change occurs in the pollination conditions, such as a change in corn variety or slight wear and tear on the pollination device, the initialization process of the optimization population can be restarted. The iteration optimization can then be performed again based on the new conditions to ensure that the optimal working parameter combination always adapts to the current pollination conditions.

[0054] Simultaneously, during the online operation of the biomimetic optimization algorithm, online safety intervention steps for stalling and abnormal operating conditions of the corn harvesting device are executed in parallel. This is used to monitor the operating status of the corn harvesting device in real time, promptly detect abnormal operating conditions such as stalling and detachment of the corn harvesting device from the corn cob, and take targeted handling strategies to avoid equipment damage, operation interruption, or serious degradation of corn harvesting quality, thus ensuring the continuity and safety of the corn harvesting operation. Specifically, the online safety intervention steps are executed in parallel with the optimization iteration process without affecting the normal operation of the optimization algorithm. The monitoring unit in the control module collects the real-time current value of the drive motor and the vibration signal amplitude collected by the vibration sensor.

[0055] The monitoring unit determines in real time whether the real-time current value of the drive motor exceeds the first safety threshold. The first safety threshold is set based on the rated current of the drive motor and the mechanical load capacity of the powder collection device, and is used to determine whether there is a risk of stalling. When stalling occurs at the end of the powder collection device, the load of the drive motor will increase sharply, causing the real-time current value to rise significantly and exceed the first safety threshold. At the same time, the monitoring unit determines in real time whether the amplitude of the vibration signal is continuously lower than the second safety threshold within a preset time window. The second safety threshold is set based on the amplitude range of the vibration signal under normal powder collection conditions, and is used to determine whether there are abnormal working conditions such as the powder collection device losing contact with the corn cob or extremely low efficiency. When the powder collection device loses contact with the corn cob, the powder collection head no longer interacts with the corn cob, the amplitude of the vibration signal will decrease significantly, and remain at a low level within the preset time window.

[0056] If the real-time current value exceeds the first safety threshold, a risk of stalling is identified. The current optimization iteration process is immediately interrupted to prevent further deterioration of the stalling situation, which could lead to burnout of the drive motor or damage to the mechanical structure. Simultaneously, the first safety strategy is invoked, which forcibly reduces the drive motor speed to a preset safe speed. The preset safe speed is a low speed that can prevent further stalling and will not damage the equipment. Then, the drive motor is controlled to reverse for a short cycle. The reverse torque generated by the reversal attempts to remove blockages, such as corn kernels stuck in the powder collection head or powder clumps. If the real-time current value of the drive motor returns to below the first safety threshold after a short reversal, it indicates that the blockage has been cleared, and the optimization iteration process is restarted to continue parameter optimization. If the real-time current value still exceeds the first safety threshold after reversal, it indicates that the blockage is more severe and cannot be cleared by reversal. At this time, an alarm signal is issued to remind the operator to intervene manually to investigate the blockage problem. After the blockage problem is resolved, the optimization iteration process and powder collection operation are restarted.

[0057] If the vibration signal amplitude remains below the second safety threshold, an abnormal condition is identified, indicating that the pollen-collecting device has lost contact with the corn cob or is operating at extremely low efficiency. The current optimization iteration process is immediately interrupted, and the second safety strategy is invoked. This strategy controls the end effector of the pollen-collecting device to perform a preset spatial attitude search motion. This preset spatial attitude search motion involves a series of attitude adjustment actions covering common effective pollen-collecting attitudes, including multi-angle adjustments to the pitch and yaw axes. Through this attitude search motion, the device attempts to re-establish effective contact with the corn cob. During the attitude search motion, the vibration signal amplitude is monitored in real time. If the vibration signal amplitude returns to the normal range (i.e., above the second safety threshold), it indicates that the pollen-collecting device has re-established effective contact with the corn cob, and the optimization iteration process is restarted to continue parameter optimization. If, after completing the preset spatial attitude search motion, the vibration signal amplitude remains below the second safety threshold, it indicates other abnormal conditions, such as excessive corn cob position deviation or mechanical failure of the pollen-collecting device. An alarm signal is issued to remind the operator to intervene manually, adjust the corn cob position, or troubleshoot the mechanical failure. Once the failure is resolved, the optimization iteration process and pollen collection operation are restarted.

[0058] After determining the optimal combination of operating parameters, the speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the flour extraction device are dynamically adjusted according to this optimal combination. This ensures that the end of the flour extraction device operates according to the optimal combination of operating parameters, achieving synergistic optimization of flour extraction quality, efficiency, and energy consumption. The specific adjustment process is as follows: First, the optimal combination of operating parameters is analyzed into target speed values ​​for the drive motor, target amplitude values ​​for the vibration device, target pitch axis angles, and target yaw axis angles. This analysis is achieved through a parameter analysis program in the control module, converting the digital parameters in the optimal combination of operating parameters into control signals that can be recognized and executed by each actuator.

[0059] For speed adjustment of the drive motor, a closed-loop speed control circuit is used. This circuit consists of the drive motor, a speed sensor, a control module, and a power electronic converter. The speed sensor is mounted on the output shaft of the drive motor to collect the actual speed value in real time and feed it back to the control module. The control module compares the actual speed value of the drive motor with the target speed value, calculates the speed tracking error, and uses a proportional-integral-derivative (PID) control algorithm to generate a control signal to adjust the output of the power electronic converter of the drive motor. For example, adjusting the output voltage and current of the power electronic converter changes the input power of the drive motor, enabling the actual speed of the motor to quickly and stably track the target speed value. This ensures the accuracy and stability of speed adjustment and reduces the impact of speed fluctuations on powder collection quality.

[0060] The adjustment of vibration amplitude is achieved through an amplitude closed-loop control loop, which consists of a vibration generator, a vibration sensor, and a control module. The vibration sensor collects the actual vibration amplitude at the end of the powder collection device in real time and feeds it back to the control module. The control module compares the actual vibration amplitude with the target amplitude, calculates the amplitude tracking error, and generates a control signal based on the amplitude tracking error to adjust the excitation signal of the vibration generator. For example, adjusting the frequency and amplitude of the excitation signal changes the vibration output of the vibration generator, ensuring that the actual vibration amplitude accurately tracks the target amplitude. This ensures the stability of the vibration amplitude during powder collection and avoids excessive powder splashing and grain damage due to excessive vibration amplitude, or incomplete powder collection due to insufficient vibration amplitude.

[0061] The adjustment of spatial attitude angles is achieved through a dual-loop spatial attitude control system. This dual-loop control system corresponds to the pitch and yaw axes respectively, and consists of a pitch axis service motor, a yaw axis service motor, an angle sensor, and a control module. The angle sensor is installed on the attitude adjustment mechanism at the end of the powder-collecting device to collect the actual pitch and yaw angles in real time and feeds these values ​​back to the control module. The control module compares the actual pitch and yaw angles with the target angle, calculates the angle tracking error, and generates control signals for the pitch and yaw axes based on this error. This adjusts the speed and direction of the pitch and yaw axis service motors, driving the attitude adjustment mechanism at the end of the powder-collecting device. This ensures that the actual pitch and yaw angles accurately track the target angle, guaranteeing that the powder-collecting head contacts the corn cob with optimal spatial attitude. This improves powder collection efficiency and integrity, preventing insufficient contact and uneven powder collection caused by unreasonable attitude angles.

[0062] During the adjustment of the above parameters, the tracking error of each closed-loop control loop is monitored in real time. If the tracking error of any loop continues to exceed the allowable tolerance (which is set according to the powder extraction accuracy requirements), it indicates that there is a deviation in the current parameter adjustment, which may be due to factors such as mechanical wear, load changes, and environmental interference. At this time, a parameter fine-tuning command is triggered, and a biomimetic optimization algorithm is used to perform local re-optimization in a small range near the optimal working parameter combination. During the local re-optimization process, the initial position of the optimization population is set to the current optimal working parameter combination, and the optimization range is limited to a small area near the optimal working parameter combination to reduce the amount of optimization calculation and speed up the fine-tuning. The optimal working parameter combination after fine-tuning is obtained through local re-optimization to compensate for the impact of mechanical wear or load changes, ensuring that the powder extraction device always operates according to the optimal parameters and maintains a stable powder extraction effect.

[0063] Furthermore, after the biomimetic optimization algorithm outputs a set of optimal operating parameter combinations and applies them to the powder extraction device, the process also includes verification and knowledge base update steps for these optimal operating parameter combinations. These steps verify the actual effectiveness of the current optimal operating parameter combination and add valid parameter samples to the knowledge base, enriching its content and improving the accuracy and reliability of subsequent initial parameter recommendations, thus forming a closed-loop optimization mechanism of "optimization-verification-update". Specifically, after the optimal operating parameter combination is applied to the powder extraction device, a fixed verification time window is set. This window is reasonably set based on the powder extraction speed and sample validity requirements to ensure that sufficient actual performance data can be collected to verify the merits of the parameter combination.

[0064] Within the verification time window, new multimodal sensing data of the toner extraction process are continuously collected. The collection method is consistent with the multimodal sensing data collection method described above to ensure that the collected data can truly reflect the toner extraction status and quality after the application of the optimal combination of working parameters. Based on the newly collected multimodal sensing data of the toner extraction process, the actual toner extraction efficiency, toner extraction integrity and energy consumption indicators are calculated to obtain the actual performance index set. The calculation method of the actual performance index set is consistent with the calculation method of each scoring item in the fitness function to ensure the consistency and comparability of the calculation results.

[0065] The actual performance index set is compared with the predicted performance index during the optimization process to calculate the overall improvement. The overall improvement is used to evaluate the actual optimization effect of the optimal working parameter combination. The calculation of the overall improvement is based on factors such as the deviation between the actual performance index and the predicted performance index, and the deviation between the actual performance index and the historical average performance index, comprehensively reflecting the improvement in flour extraction performance. If the overall improvement is higher than the verification threshold, which is set based on the optimization requirements of the actual flour extraction operation, it indicates that the optimization was successful and the optimal working parameter combination can effectively improve flour extraction performance. At this time, the successful corn variety identification feature, the final optimal working parameter combination, and its corresponding actual performance index set are added as a new sample pair to the corn variety-working parameter knowledge base. If the overall improvement is lower than or equal to the verification threshold, it indicates that the optimization effect was poor and the optimal working parameter combination failed to achieve the expected optimization goal. At this time, the knowledge base is not updated, but the relevant data of this optimization process is recorded for subsequent analysis of the reasons for the poor optimization effect and optimization of the parameter settings of the optimization algorithm or the training strategy of the feature fusion network.

[0066] After new sample pairs are added to the knowledge base, the knowledge base is organized and optimized to ensure that the sample data in the knowledge base is accurate and effective, making it easy to query and call later. When encountering corn milling operations of the same or similar varieties in the future, more accurate recommended values ​​of initial working parameters can be found from the knowledge base, which can further accelerate the optimization convergence speed, improve the optimization effect, and at the same time reduce the amount of calculation in the optimization process and reduce equipment energy consumption.

[0067] This embodiment of the corn biomimetic powder extraction end parameter online optimization method comprehensively captures the state information during the powder extraction process through real-time acquisition of multimodal sensing data, avoiding the problem of incomplete representation by a single type of data; it improves the accuracy of powder extraction state and quality representation through feature fusion based on attention mechanism; and it achieves rapid and accurate online parameter optimization by combining a biomimetic optimization algorithm based on bird flock foraging and ant colony pheromone diffusion. At the same time, through steps such as multi-objective dynamic weight adjustment, safety intervention, and knowledge base update, it further improves the adaptability, safety, and stability of the algorithm, ensuring that the optimal combination of working parameters can adapt to complex and ever-changing powder extraction conditions in real time, effectively improving powder extraction quality and operation efficiency, reducing equipment energy consumption and manual dependence, solving many problems existing in traditional powder extraction parameter adjustment methods, and is applicable to various automated corn powder extraction operation scenarios.

[0068] Example 2 like Figure 2 As shown, this embodiment provides a corn bionic powder extraction end parameter online optimization system, which is used to implement the corn bionic powder extraction end parameter online optimization method described in Embodiment 1. It can comprehensively and stably complete functions such as multimodal sensing data acquisition, feature extraction and fusion, online parameter optimization and end effector dynamic adjustment. It is compatible with different types of corn powder extraction devices, has a high degree of automation, stability and compatibility, and can be widely used in automated powder extraction operations in corn breeding, deep processing and quality inspection, effectively improving the intelligence level and operation performance of powder extraction operations.

[0069] This embodiment of the system adopts a modular design concept, dividing the entire system into multiple functionally independent yet collaborative modules. Each module is responsible for a specific function, and the modules exchange data and commands through standardized interfaces, facilitating system installation, debugging, maintenance, and upgrades. Furthermore, the parameters and configurations of each module can be flexibly adjusted according to the specifications and operational requirements of the actual powder extraction device, improving the system's compatibility and applicability. The overall hardware architecture of the system is built on an industrial control platform, while the software architecture adopts a layered design, including a sensing layer, a data processing layer, a control layer, and an execution layer. Each layer has a clear division of labor and works collaboratively to ensure the system operates efficiently and stably.

[0070] Specifically, the system in this embodiment includes a multimodal sensing data acquisition module, a comprehensive state feature extraction and fusion module, a biomimetic online optimization calculation module, and an end effector dynamic adjustment module. These four modules are interconnected and work together to form a complete online optimization system. The multimodal sensing data acquisition module is used to collect multimodal sensing data during the powder extraction process and transmit it to the comprehensive state feature extraction and fusion module. The comprehensive state feature extraction and fusion module is used to extract and fuse features from the multimodal sensing data to obtain a comprehensive state feature vector, and transmit it to the biomimetic online optimization calculation module. The biomimetic online optimization calculation module is used to calculate the optimal combination of working parameters based on the comprehensive state feature vector using a biomimetic optimization algorithm, and transmit it to the end effector dynamic adjustment module. The end effector dynamic adjustment module is used to dynamically adjust the relevant actuators at the end of the powder extraction device according to the optimal combination of working parameters, thereby achieving real-time optimization of the powder extraction parameters.

[0071] The specific implementation methods of each module are described in detail below. First, the multimodal sensing data acquisition module is fundamental to the system's acquisition of state information during the corn milling process. It is used to collect vibration signals, sound signals, and image data in real time from the end of the corn milling device during operation, obtaining multimodal sensing data of the milling process. This multimodal sensing data is then sent to the comprehensive state feature extraction and fusion module. This module consists of a vibration acquisition unit, a sound acquisition unit, an image acquisition unit, a data preprocessing unit, and a data transmission unit. These units cooperate to ensure the accuracy, stability, and synchronization of the multimodal sensing data acquisition.

[0072] The vibration acquisition unit is used to collect vibration signals during the operation of the corn cob extraction device. It consists of a vibration sensor and a signal conditioning circuit. The vibration sensor is a conventional piezoelectric vibration sensor, characterized by high sensitivity, fast response speed, and strong anti-interference ability. It can accurately capture the vibration signals generated by the interaction between the extraction head and the corn cob during the extraction process. The vibration sensor is installed on the housing of the actuator at the end of the extraction device, close to the extraction head and far from the drive motor to reduce interference from the drive motor's own vibration. The acquisition direction of the vibration sensor is consistent with the vibration direction of the extraction head to ensure that the acquired vibration signal can accurately reflect the vibration state of the extraction process. The signal conditioning circuit is connected to the vibration sensor and is used to filter, amplify, reduce noise, and perform analog-to-digital conversion on the analog vibration signal acquired by the sensor. Filtering removes environmental noise and high-frequency interference signals, amplification amplifies the weak vibration signal to a range suitable for analog-to-digital conversion, noise reduction further suppresses noise interference and improves signal quality, and analog-to-digital conversion converts the analog vibration signal into a digital vibration signal for subsequent data processing and transmission.

[0073] The sound acquisition unit is used to collect sound signals generated during the corn cob extraction process. It consists of a sound acquisition device and a sound preprocessing circuit. The sound acquisition device uses a conventional omnidirectional microphone, which features a wide frequency response range, high sensitivity, and strong anti-interference capabilities. It can clearly capture the sounds generated by the friction between the extraction head and the corn cob, the sound of powder splashing, and the sound of kernel threshing. The microphone is installed at a certain distance from the extraction head to ensure the clarity of the sound acquisition while avoiding contamination by powder or damage from vibrations of the extraction device. The sound preprocessing circuit is connected to the microphone and is used to filter, amplify, reduce noise, and perform analog-to-digital conversion on the analog sound signal acquired by the microphone. The processing is similar to the preprocessing of vibration signals, removing interference signals such as environmental noise and drive motor noise, and converting the analog sound signal into a digital sound signal to ensure the quality and usability of the sound signal.

[0074] The image acquisition unit is used to collect image data during the corn cob extraction process. It consists of an industrial camera, lens, light source, and image preprocessing circuitry. The industrial camera is a conventional color industrial camera, characterized by high resolution, high frame rate, and good image quality. It can capture real-time image information such as the interaction between the extraction head and the corn cob, powder splashing, and the distribution of unthreshed kernels. The resolution and frame rate of the industrial camera are reasonably set according to the actual extraction accuracy and operating speed to ensure clear and real-time capture of dynamic changes during the extraction process. The lens is equipped with an industrial lens with a suitable focal length, which can adjust the shooting range and image clarity according to the structure of the extraction device and the shooting distance to ensure that the captured image completely covers the extraction area. The light source uses a conventional industrial supplementary light source to compensate for the impact of insufficient ambient light on image quality, ensuring the clarity and stability of image acquisition. The brightness and illumination angle of the supplementary light source can be adjusted according to the actual ambient light conditions to avoid light reflection interfering with image quality. The image preprocessing circuit is connected to the industrial camera and is used to perform preliminary preprocessing on the color images captured by the industrial camera, including image denoising, image enhancement, size normalization, grayscale conversion, etc., to remove dust noise, light reflection interference, etc. in the image, improve image clarity, and lay the foundation for subsequent feature extraction and region of interest segmentation.

[0075] The data preprocessing unit further unifies the preprocessing of digital signals and image data output from the vibration acquisition unit, sound acquisition unit, and image acquisition unit. This includes data normalization, time synchronization, and data verification. Data normalization converts different types and magnitudes of sensor data into a unified magnitude range, preventing deviations in subsequent feature extraction and fusion due to differences in data magnitude. Time synchronization is achieved through a synchronization clock module, ensuring that the acquisition time of vibration signals, sound signals, and image data is synchronized, with synchronization errors controlled within a reasonable range to avoid affecting the accurate representation of the powder collection status due to time asynchrony. Data verification checks the validity of the acquired data, removing invalid data such as abnormal or missing data to ensure the accuracy and validity of data transmitted to subsequent modules. Data verification uses conventional data verification algorithms, judging and filtering the acquired data based on its normal range and variation patterns. Invalid data is marked and replaced or discarded.

[0076] The data transmission unit sends preprocessed multimodal sensing data to the integrated state feature extraction and fusion module. The data transmission unit employs either wired or wireless transmission. Wired transmission utilizes conventional industrial Ethernet, RS485 bus, and other transmission methods, offering advantages such as high transmission speed, high stability, and strong anti-interference capabilities, making it suitable for fixed-installation dust collection devices. Wireless transmission uses conventional WiFi, Bluetooth, LoRa, and other wireless transmission methods, offering flexible installation and no wiring required, making it suitable for mobile dust collection devices or scenarios where wiring is inconvenient. The data transmission unit uses standardized transmission protocols to ensure the reliability and compatibility of data transmission, while also featuring data encryption to prevent data tampering or leakage during transmission, ensuring system security.

[0077] Each unit of the multimodal sensing data acquisition module adopts low-power, high-reliability hardware devices, which can adapt to the harsh environment of powder extraction operations (such as dust, vibration, temperature changes, etc.), ensuring that the module can operate stably for a long time. At the same time, the module has a self-test function, which can monitor its own operating status in real time. When problems such as sensor failure or abnormal data transmission occur, it will promptly issue an alarm signal to remind operators to troubleshoot and maintain the system, ensuring continuous operation.

[0078] Next is the comprehensive state feature extraction and fusion module. This module is used to extract and fuse features from the received multimodal sensing data of the powder extraction process, obtaining a comprehensive state feature vector representing the current powder extraction state and quality. This comprehensive state feature vector is then sent to the biomimetic online optimization calculation module. This module is built on an embedded processor or industrial computer, possessing strong data processing capabilities and computing speed. It can process large amounts of multimodal sensing data in real time and efficiently complete feature extraction and fusion tasks. The hardware core of this module can be a high-performance embedded processor, such as an ARM Cortex-A series processor, or a small industrial computer, flexibly selected according to the actual data processing volume and real-time requirements. Embedded processors are characterized by small size, low power consumption, and strong anti-interference capabilities, making them suitable for installation in the control box of the powder extraction device and adaptable to the industrial installation environment. Industrial computers, on the other hand, have stronger data processing capabilities, suitable for processing complex data such as high-resolution image data and large-scale feature vectors, and are suitable for scenarios with high data processing accuracy and speed requirements. The module's hardware also includes auxiliary components such as a memory, data interface, and power module. The memory is used to temporarily store the acquired multimodal sensing data, extracted feature subsets, and fused comprehensive state feature vectors, ensuring the continuity of the data processing process. Storage devices such as SD cards and solid-state drives can be selected, with appropriate capacity chosen according to storage requirements. The data interface is used for data interaction with the multimodal sensing data acquisition module and the bionic online optimization calculation module. It adopts standardized serial ports, Ethernet interfaces, or CAN bus interfaces to ensure the stability and compatibility of data transmission. The power module provides a stable power supply for the entire module. It adopts a wide voltage input design to adapt to different power supply environments in industrial sites, and also has overvoltage and overcurrent protection functions to prevent the module from being damaged due to abnormal power supply.

[0079] The software component of the integrated state feature extraction and fusion module adopts a modular design and works in conjunction with the hardware. It mainly includes a data receiving module, a feature extraction module, a feature fusion module, and a data sending module. Each software module has a clear division of labor and operates collaboratively to ensure the accuracy and real-time performance of feature extraction and fusion. The data receiving module receives multimodal sensing data from the multimodal sensing data acquisition module during the powder extraction process via a data interface. It parses and verifies the received data to ensure that the data format is correct, without missing data or anomalies. If an anomaly is detected, it promptly sends a feedback signal to the multimodal sensing data acquisition module, requesting retransmission of the data. Simultaneously, it marks and temporarily stores the abnormal data for subsequent troubleshooting. The data receiving module also has a data caching function. When the multimodal sensing data acquisition speed is slightly faster than the feature extraction speed, the caching mechanism prevents data loss and ensures the continuity of data processing.

[0080] The feature extraction module is the core of the software, used to perform targeted feature extraction on verified multimodal sensing data, obtaining vibration time-frequency domain feature subsets, acoustic frequency domain feature subsets, and visual morphological feature subsets. Its extraction algorithm corresponds completely to the method described in Example 1, ensuring consistency and accuracy in feature extraction. For vibration signal feature extraction, the feature extraction module deploys an empirical mode decomposition algorithm, which adaptively decomposes the vibration signal into multiple intrinsic mode function components. Then, through preset calculation logic, it automatically calculates the sample entropy and energy ratio of each component, integrating them to form a vibration time-frequency domain feature subset. The algorithm embeds parameter adaptive adjustment logic, which automatically adjusts parameters such as the empirical mode decomposition screening threshold and iteration number based on real-time parameters such as the vibration signal acquisition frequency and amplitude range, ensuring the decomposition effect and the accuracy of feature calculation, and avoiding feature extraction deviations due to signal differences.

[0081] For feature extraction of audio signals, the feature extraction module deploys algorithms such as frame segmentation, Mel frequency cepstral coefficient extraction, zero-crossing rate calculation, and short-time energy calculation. The frame segmentation algorithm adopts overlapping frame segmentation logic, and the frame length and frame shift can be adaptively adjusted according to the real-time changes of the audio signal to ensure complete capture of the short-time features of the audio signal. The Mel frequency cepstral coefficient extraction algorithm integrates a Mel filter bank, which can automatically adjust the parameters of the filter bank according to the frequency domain distribution of the audio signal to improve the representation ability of frequency domain features. The zero-crossing rate and short-time energy calculation algorithms use a sliding window mechanism to calculate the relevant parameters of each frame of audio signal in real time and perform normalization processing to ensure the consistency of feature values ​​and integrate them to form a subset of audio-visual domain features.

[0082] For feature extraction from image data, the feature extraction module deploys algorithms such as image preprocessing, threshold segmentation, region of interest (ROI) segmentation, and feature parameter calculation. Image preprocessing algorithms include denoising, enhancement, size normalization, and grayscale conversion. The denoising algorithm uses a combination of Gaussian filtering and median filtering to effectively remove dust noise and light reflection interference from the image while preserving target details. The threshold segmentation algorithm employs adaptive threshold logic, adjusting the segmentation threshold in real time based on the local grayscale distribution of the image to ensure effective separation of the background and target regions. The ROI segmentation algorithm integrates contour detection and region growing logic, automatically extracting ROIs such as corn ears, powder splashes, and unthreshed kernels, and optimizing them to avoid missegmentation and missed segmentation. The feature parameter calculation algorithm automatically calculates parameters such as texture roughness, area ratio, and contour features for each ROI, integrating them to form a subset of visual morphological features.

[0083] The feature fusion module is used to fuse the three extracted feature subsets to obtain a comprehensive state feature vector. This module deploys a feature fusion network based on an attention mechanism. The network structure is consistent with that described in Example 1, including an input layer, an attention weight allocation layer, a weighted concatenation layer, a fully connected layer, and an output layer. After the network model is pre-trained, it is fixed and deployed in this module and can be directly called and run. The input layer of the feature fusion network receives three feature subsets, converts them into a feature matrix in a unified format, and processes them in parallel through independent input channels. The attention weight allocation layer analyzes the correlation between each feature subset and the current follower acquisition status and quality in real time through a preset feature importance calculation logic, and automatically assigns weight coefficients. The feature subset with higher correlation has a larger weight coefficient, ensuring that key feature information is highlighted. The weighted concatenation layer performs a weighted multiplication of the three feature subsets with their respective weight coefficients and then concatenates them to form a high-dimensional fusion feature matrix. The fully connected layer performs dimensionality reduction and information compression on the high-dimensional fusion feature matrix through multiple hidden layers and activation functions, removing redundant information and improving the density of the feature vector. The output layer finally outputs a low-dimensional comprehensive state feature vector. The dimension of the feature vector is set according to the actual processing requirements to ensure that it can comprehensively represent the follower acquisition status and quality while reducing the computational load of subsequent optimization algorithms.

[0084] To improve the real-time performance of feature fusion, the feature fusion module also incorporates algorithm optimization logic. This parallelizes feature extraction and fusion processes, allocating vibration, sound, and image feature extraction tasks to different processing threads for simultaneous processing, thus shortening feature extraction time. The feature fusion network is also optimized for lightweighting, simplifying its structure and reducing the number of parameters. This improves network speed while maintaining fusion accuracy, ensuring real-time output of the comprehensive state feature vector. The data transmission module sends the fused comprehensive state feature vector to the biomimetic online optimization computation module via a data interface. Data encryption and verification mechanisms are used during transmission to ensure the feature vector is not tampered with or lost. The module also provides feedback on the data transmission status to the biomimetic online optimization computation module. If a transmission anomaly occurs, the data is retransmitted promptly, ensuring smooth data interaction throughout the system.

[0085] The integrated status feature extraction and fusion module also has self-testing and fault diagnosis functions. It can monitor its own hardware operating status and software algorithm operation in real time, including parameters such as processor load, memory usage, data interface connection status, and algorithm runtime. When a hardware fault is detected, such as memory damage or data interface disconnection, or a software fault, such as abnormal algorithm operation or feature extraction failure, an alarm signal is issued in a timely manner to remind operators to troubleshoot and maintain the system. At the same time, the fault information is stored in the memory for subsequent traceability and analysis, ensuring that the module can operate stably for a long time and providing reliable support for the normal operation of the entire system.

[0086] After detailing the comprehensive state feature extraction and fusion module, the next step is to elaborate on the biomimetic online optimization calculation module. This module is the core control module of the entire system. Based on the received comprehensive state feature vector, it uses a biomimetic optimization algorithm that simulates the intelligent collaborative mechanism of biological swarms to iteratively calculate the optimal combination of operating parameters at the end of the powder extraction device online, and then sends the optimal combination of operating parameters to the end effector dynamic adjustment module. This module is also built on an industrial control hardware platform, possessing strong computing power, real-time performance, and stability. It can quickly complete the optimization iterative calculation, adapting to the real-time optimization needs of powder extraction operations. It also has good compatibility and scalability, allowing for flexible adjustment of the optimization algorithm parameter settings according to different powder extraction conditions.

[0087] The hardware structure of the biomimetic online optimization computing module is similar to that of the comprehensive state feature extraction and fusion module. It mainly includes components such as a computing core, memory, data interface, power supply module, and alarm module. The computing core is the hardware core of the module and can be a high-performance industrial processor, FPGA, or DSP chip. The choice is flexible based on the complexity and real-time requirements of the optimization algorithm. Industrial processors are characterized by strong versatility and easy programming, making them suitable for running complex biomimetic optimization algorithms and multi-objective dynamic weight adjustment logic. FPGA chips have strong parallel computing capabilities and high real-time performance, making them suitable for handling large-scale data calculations during the optimization iteration process, significantly improving the optimization speed. DSP chips focus on digital signal processing, suitable for processing digital signals such as comprehensive state feature vectors, improving the algorithm's operating efficiency. The performance of the computing core directly determines the iteration speed and optimization accuracy of the optimization algorithm; therefore, it is necessary to select a suitable core based on actual needs to ensure rapid convergence to the optimal parameter combination.

[0088] The memory is used to store data such as the optimization algorithm program, optimization population data, comprehensive state feature vector, optimal working parameter combination, maize variety-working parameter knowledge base, and parameter safety boundary constraints. It is divided into program memory and data memory. The program memory is used to permanently store the optimization algorithm program, control logic program, etc., to ensure that the program can run stably and is not tampered with. The data memory is used to temporarily store various types of data in the optimization process, including candidate working parameters, fitness scores, historical best parameters, and global best parameters of each generation of optimization population. At the same time, it stores the complete data of the maize variety-working parameter knowledge base for easy real-time query and update. Large-capacity solid-state drives or DDR memory can be selected to ensure the speed of data storage and retrieval and meet the real-time requirements of optimization iteration.

[0089] The data interface is used to interact with the integrated state feature extraction and fusion module, the end effector dynamic adjustment module, and the multimodal sensing data acquisition module. On one hand, it receives the integrated state feature vector from the integrated state feature extraction and fusion module and the corn variety identification features or corn variety coding information from the multimodal sensing data acquisition module. On the other hand, it sends optimal operating parameter combinations and parameter adjustment commands to the end effector dynamic adjustment module, while simultaneously feeding back optimization status information to the multimodal sensing data acquisition module. The data interface uses standardized Ethernet and CAN bus interfaces to ensure data transmission stability and compatibility, supporting simultaneous data interaction between multiple modules and avoiding data transmission bottlenecks. The power supply module provides stable power to the entire module. It adopts a wide voltage input design consistent with the integrated state feature extraction and fusion module, and features overvoltage, overcurrent, and undervoltage protection. It also has a low-power mode; when the powder extraction device is in standby mode, the module automatically enters low-power mode to reduce energy consumption and meet the energy-saving requirements of industrial sites.

[0090] The alarm module is used to issue alarm signals when abnormal situations occur during the optimization process, including both audible and visual alarms and signal alarms. The audible and visual alarms are implemented through indicator lights and buzzers. Different colors of indicator lights distinguish different types of abnormalities, such as red indicator lights indicating serious faults and yellow indicator lights indicating general abnormalities. The buzzers use different sound frequencies to remind operators to pay attention. The signal alarm sends alarm signals to the main control system of the powder collection device through the data interface, which facilitates unified management and fault diagnosis by the main control system. Abnormal situations include optimization iteration timeout, parameters exceeding safety boundary constraints, abnormal comprehensive state feature vectors, and data transmission failures, ensuring that operators can detect and handle abnormalities in a timely manner, and ensuring the continuity and safety of powder collection operations.

[0091] The software portion of the biomimetic online optimization calculation module adopts a layered design, mainly including a data receiving and parsing layer, a knowledge base management layer, an optimization algorithm layer, a parameter output layer, and a fault diagnosis layer. These layers work collaboratively to ensure the smooth operation of the optimization algorithm and the accurate output of optimal parameters. The data receiving and parsing layer receives data such as comprehensive state feature vectors, corn variety identification features, or corn variety coding information through a data interface. It performs format parsing, verification, and preprocessing on the received data to ensure its accuracy and validity. If data anomalies are detected, such as incorrect dimensions of the comprehensive state feature vector or missing corn variety information, an alarm module is triggered promptly, and feedback signals are sent to relevant modules requesting data retransmission. Simultaneously, this layer sends the parsed comprehensive state feature vector to the optimization algorithm layer and the corn variety information to the knowledge base management layer, providing data support for subsequent optimization calculations.

[0092] The knowledge base management layer manages the pre-established maize variety-working parameter knowledge base, enabling functions such as querying, updating, and maintaining the knowledge base. This layer stores complete data in the knowledge base, including recommended initial working parameter values, parameter safety boundary constraints, and historical successful optimization samples for various maize varieties. Upon receiving maize variety information from the data receiving and parsing layer, the knowledge base management layer automatically queries the knowledge base to retrieve recommended initial working parameter values ​​and parameter safety boundary constraints matching the maize variety. The query results are then sent to the optimization algorithm layer to provide a basis for initializing the optimization population. After the optimization process is completed and successfully verified, the knowledge base management layer receives new sample pairs from the optimization algorithm layer, including maize variety identification features, optimal working parameter combinations, and actual performance index sets. It automatically adds the new sample pairs to the knowledge base and organizes and optimizes the knowledge base, removing duplicate and invalid sample data to ensure the accuracy and effectiveness of the knowledge base. Simultaneously, the knowledge base management layer has knowledge base backup and recovery functions, regularly backing up the knowledge base data to prevent data loss. When the knowledge base is damaged, it can quickly recover the data, ensuring the normal operation of the system.

[0093] The optimization algorithm layer is the core of this module's software. It is used to deploy a biomimetic optimization algorithm that simulates the intelligent collaborative mechanism of biological groups. It performs functions such as initialization of the optimization population, fitness calculation, iterative updates, and convergence judgment. The algorithm logic completely corresponds to the method described in Example 1, ensuring the consistency and accuracy of the optimization process. This layer mainly includes an optimization population management unit, a fitness evaluation unit, and a biomimetic collaborative iteration unit. These three units cooperate to complete the online iterative optimization task.

[0094] The optimization population management unit initializes and maintains an optimization population consisting of multiple virtual agents. Each virtual agent represents a set of candidate operating parameters, including the drive motor's rotational speed, vibration amplitude, and spatial pitch and yaw angles. Upon receiving the recommended initial operating parameters and parameter safety boundary constraints from the knowledge base management layer, this unit uses the recommended initial operating parameters as the initial center position of the population. Around this center position, it randomly generates initial candidate operating parameters for multiple virtual agents. The number of virtual agents is reasonably set based on the performance of the computing core and the optimization accuracy requirements, achieving a balance between optimization accuracy and optimization speed. Simultaneously, this unit embeds the parameter safety boundary constraints into the generation and update logic of the candidate operating parameters in the form of inequalities. It performs boundary checks on the candidate operating parameters of each virtual agent; if a parameter exceeds the safety boundary constraint, it is automatically pruned or adjusted to ensure that the candidate operating parameter combinations generated in each generation are within the safe operating range allowed by the mechanical structure and process requirements. During the optimization iteration process, this unit is responsible for maintaining the historical optimal parameter position of each virtual agent, recording the candidate working parameters with the highest fitness score for each virtual agent in all iteration cycles, and maintaining the global optimal parameter position of the entire population, recording the candidate working parameters with the highest fitness score for all virtual agents in all iteration cycles, providing a basis for subsequent iteration updates; when the powder picking conditions change abruptly, this unit can receive external trigger signals, reinitialize the optimization population, and restart the iterative optimization based on the new conditions, ensuring that the optimal parameter combination can adapt to changes in conditions in real time.

[0095] The fitness evaluation unit is connected to the data receiving and parsing layer and the optimization population management unit. It receives the comprehensive state feature vector sent by the data receiving and parsing layer and, based on the candidate working parameters provided by the optimization population management unit, calls a preset fitness function to calculate the comprehensive evaluation score for each virtual agent. The logic of the fitness function is consistent with that in Example 1. The comprehensive evaluation score is obtained by weighted summation of the pollen collection efficiency score, pollen collection integrity score, and energy consumption score. This unit pre-stores the calculation logic of the fitness function and can automatically calculate the pollen collection efficiency score, pollen collection integrity score, and energy consumption score based on the relevant feature parameters in the comprehensive state feature vector. Then, it calculates the comprehensive evaluation score based on preset weight coefficients, which can be adjusted according to the actual pollen collection operation requirements. Simultaneously, this unit is also responsible for real-time calculation of the average pollen collection efficiency, average pollen collection integrity, and average unit energy consumption from recent historical data, sending the calculation results to the dynamic weight adjuster to provide a basis for adjusting the weights of the fitness function.

[0096] The biomimetic collaborative iterative unit is connected to the fitness evaluation unit and the optimization population management unit. It simulates the collaborative mechanism of bird flock foraging and ant colony pheromone diffusion, driving the optimization population management unit to update the candidate working parameters of all virtual agents and determine whether the convergence condition is met. This unit deploys bird flock foraging simulation logic and ant colony pheromone diffusion simulation logic. The bird flock foraging simulation logic guides virtual agents to move towards their own historical optimal parameter position and the global optimal parameter position of the population. The candidate working parameter update formula for each virtual agent is constructed based on the group cooperation mechanism of bird flock foraging, combined with the guidance of its own historical best and the global optimal of the population. A random perturbation term is also introduced. The value range of the random perturbation term is dynamically adjusted according to the number of iterations. In the early stages of iteration, the value range is larger, increasing the global search range; in the later stages of iteration, the value range gradually decreases, improving the local search accuracy. The ant colony pheromone diffusion simulation logic is used to assign virtual pheromones to parameter regions represented by virtual agents with higher fitness. The concentration of virtual pheromones is positively correlated with the fitness score of the virtual agent; the higher the fitness score, the higher the pheromone concentration. At the same time, the pheromone gradually decays with the increase of the number of iterations to avoid old pheromones interfering with the subsequent optimization process. This logic guides other virtual agents to move to parameter regions with high virtual pheromone concentrations, realizing regional fine-grained search and improving optimization accuracy and convergence speed on the basis of global search.

[0097] In addition, the optimization algorithm layer also deploys a multi-objective dynamic weight adjustment unit and a convergence judgment unit. The multi-objective dynamic weight adjustment unit is used to realize the multi-objective dynamic weight adjustment in the optimization process. This unit deploys a dynamic weight adjuster based on fuzzy logic rules. The inputs are the deviations of the average fan acquisition efficiency, average fan acquisition integrity, and average unit energy consumption from their respective set targets. The output is the weight adjustment amount of each scoring item in the fitness function. The weight coefficients of the fitness function are updated in real time according to the weight adjustment amount, guiding the optimization algorithm to search towards the current performance target that needs to be optimized. The convergence judgment unit is used to monitor the changes in the optimal fitness value in the population in real time and determine whether the convergence condition is met. The convergence condition is that the change in the optimal fitness value in the population is less than the convergence threshold in multiple consecutive iteration cycles, or the maximum number of iterations is reached. The convergence threshold and the maximum number of iterations are pre-stored in this unit and can be flexibly adjusted according to the actual optimization accuracy and efficiency requirements. When the convergence condition is met, the unit issues a convergence signal, terminates the iteration process, and determines the candidate working parameters corresponding to the current global optimal parameter position of the population as the optimal working parameter combination, which is then sent to the parameter output layer. If the convergence condition is not met, the biomimetic cooperative iteration unit is triggered to perform the next round of iteration update.

[0098] Meanwhile, the optimization algorithm layer also integrates an online safety intervention linkage unit, which works in conjunction with the safety monitoring unit of the corn harvesting device to achieve online safety intervention for stalling and abnormal operating conditions of the corn harvesting device. This unit receives real-time current values ​​and vibration signal amplitudes of the drive motor sent by the safety monitoring unit to determine whether there are abnormal operating conditions such as stalling or the corn harvesting device detaching from the corn cob. If an abnormal operating condition is detected, the current optimization iteration process is immediately interrupted, the corresponding safety strategy is invoked, and an alarm signal is sent to the alarm module to remind the operator to pay attention. After the abnormal operating condition is eliminated, the unit restarts the optimization iteration process to ensure the continuity and safety of the corn harvesting operation.

[0099] The parameter output layer receives the optimal operating parameter combination sent by the optimization algorithm layer, parses and converts this optimal combination into a control signal that the end effector dynamic adjustment module can recognize and execute. The format of the control signal is reasonably set according to the type and control method of the end effector to ensure the accuracy of parameter adjustment. Simultaneously, this layer sends the parsed control signal to the end effector dynamic adjustment module through a data interface and receives parameter adjustment status information from the end effector dynamic adjustment module in real time. If parameter adjustment failure or adjustment error exceeding the allowable tolerance is detected, the optimization algorithm layer is promptly triggered to perform a small-scale local re-optimization, generating a fine-tuned optimal operating parameter combination, which is then resent to the end effector dynamic adjustment module to ensure that the powder extraction device operates accurately according to the optimal parameters. Furthermore, this layer is also responsible for sending the optimal operating parameter combination and its corresponding set of actual performance indicators to the knowledge base management layer, providing data support for knowledge base updates.

[0100] The fault diagnosis layer is used to monitor the operational status of each layer, including the optimization algorithm layer, data receiving and parsing layer, and knowledge base management layer, in real time. It detects faults such as algorithm malfunctions, data transmission anomalies, and knowledge base query failures. Simultaneously, it monitors the module's hardware operation status, such as excessive core load, excessive memory usage, and power supply anomalies. When a fault is detected, this layer immediately triggers the alarm module, issuing a corresponding alarm signal and recording fault information, including fault type, time of occurrence, and cause, storing it in memory for later traceability and analysis. For minor faults, this layer can automatically attempt recovery; for example, it automatically resends data in case of data transmission anomalies or automatically restarts the algorithm program in case of algorithm malfunctions, ensuring the module can quickly return to normal operation. For serious faults, this layer issues an emergency alarm signal and stops the module's operation to prevent further deterioration and protect the module hardware and the entire system.

[0101] Finally, the dynamic adjustment module of the end effector is described in detail. This module is the system's execution terminal, used to dynamically adjust the speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder collection device according to the received optimal combination of operating parameters. This ensures that the end of the powder collection device operates according to the optimal combination of operating parameters, achieving synergistic optimization of powder collection quality, efficiency, and energy consumption. This module connects directly to the end effector of the powder collection device, possessing strong driving capability, control precision, and stability. It can adapt to different types of end effector mechanisms in powder collection devices and has good anti-interference capabilities, enabling it to adapt to harsh environments such as vibration and dust in industrial settings.

[0102] The hardware of the end effector dynamic adjustment module mainly consists of a parameter parsing unit, a drive unit, a closed-loop control unit, a status monitoring unit, a data interface, and a power supply module. These components work together to complete parameter adjustment and execution control tasks. The parameter parsing unit receives the optimal working parameter combination sent by the biomimetic online optimization calculation module, parses and verifies it, resolving the optimal working parameter combination into the target speed value of the drive motor, the target amplitude of the vibration device, and the target angles of the pitch and yaw axes. Simultaneously, it verifies whether the parsed target parameters are within the safety boundary constraints. If abnormal parameters are found, an alarm module is immediately triggered, and a feedback signal is sent to the biomimetic online optimization calculation module, requesting a retransmission of the optimal working parameter combination. The parameter parsing unit also has a parameter caching function. When the optimal working parameter combination is updated, the new target parameters are cached to ensure a smooth transition during the parameter adjustment process and avoid instability in the powder collection device due to sudden parameter changes.

[0103] The drive unit drives the various actuators at the end of the powder collection device according to the control signals sent by the closed-loop control unit. These actuators include drive motors, vibration generators, pitch axis service motors, and yaw axis service motors. This unit consists of multiple drive modules, each corresponding to one actuator. The type of drive module is selected appropriately based on the type of actuator; for example, the drive motor uses a DC motor drive module or an AC motor drive module, the vibration generator uses a vibration excitation drive module, and the service motor uses a servo drive module. The drive modules possess strong driving capability and control precision, and can quickly adjust the operating state of the actuators according to changes in control signals, such as the speed of the drive motor, the excitation signal of the vibration generator, and the speed and direction of the service motor. They also have overcurrent, overvoltage, and overload protection functions to prevent damage to the actuators due to drive malfunctions. The drive unit also includes a power regulation module, which automatically adjusts the output power according to changes in the load of the actuators, improving energy efficiency and reducing energy consumption.

[0104] The closed-loop control unit is the core control part of this module, used to achieve precise control of the operating status of each actuator, ensuring that the actual operating parameters can accurately track the target parameters. This unit includes a speed closed-loop control loop, an amplitude closed-loop control loop, and a spatial attitude dual closed-loop control loop. The three control loops are independent of each other and work together, respectively corresponding to the adjustment of the drive motor speed, vibration amplitude, and spatial attitude angle. The control logic is consistent with the method described in Example 1. The speed closed-loop control loop consists of a speed sensor, a control chip, and a drive module. The speed sensor is installed on the output shaft of the drive motor to collect the actual speed value of the drive motor in real time and feed the actual speed value back to the control chip. The control chip compares the actual speed value with the target speed value, calculates the speed tracking error, and uses a proportional-integral-derivative control algorithm to generate a control signal based on the tracking error. This signal is sent to the drive module corresponding to the drive motor to adjust the input power of the drive motor, so that the actual speed can quickly and stably track the target speed value, reducing the impact of speed fluctuations on powder collection quality. The control chip also embeds speed fluctuation suppression logic, which can effectively suppress speed fluctuations caused by load changes and improve the stability of speed control.

[0105] The amplitude closed-loop control loop consists of a vibration sensor, a control chip, and a drive module. The vibration sensor is shared with the vibration sensor in the multimodal sensing data acquisition module, or a dedicated vibration sensor can be added separately. It is used to collect the actual vibration amplitude at the end of the powder collection device in real time and feed the actual vibration amplitude back to the control chip. The control chip compares the actual vibration amplitude with the target amplitude, calculates the amplitude tracking error, generates a control signal based on the tracking error, and sends it to the drive module corresponding to the vibration generator. This adjusts the excitation signal frequency and amplitude of the vibration generator to ensure that the actual vibration amplitude accurately tracks the target amplitude. This ensures the stability of the vibration amplitude during the powder collection process and avoids excessive powder splashing and grain damage due to excessive vibration amplitude, or incomplete powder collection due to insufficient vibration amplitude.

[0106] The spatial attitude dual closed-loop control loop corresponds to the pitch axis and yaw axis respectively. Each control loop consists of an angle sensor, a control chip, and a drive module. The angle sensor is installed on the attitude adjustment mechanism at the end of the powder collection device to collect the actual pitch angle and yaw angle at the end of the powder collection device in real time and feed the actual angle value back to the control chip. The control chip compares the actual angle value with the target angle value, calculates the angle tracking error, and uses a proportional-integral-derivative control algorithm to generate control signals for the pitch axis and yaw axis according to the tracking error. These signals are sent to the corresponding service motor drive module to adjust the speed and direction of the service motor, thereby driving the attitude adjustment mechanism to move. This ensures that the actual pitch angle and yaw angle accurately track the target angle value, ensuring that the powder collection head can contact the corn cob with the optimal spatial attitude, improving powder collection efficiency and integrity.

[0107] To improve the accuracy and response speed of closed-loop control, the control chip of the closed-loop control unit adopts a high-performance microcontroller, such as the STM32 series microcontroller, which features fast computing speed, high control accuracy, and strong anti-interference capability, and can quickly process feedback signals and control signals. At the same time, the control chip embeds adaptive adjustment logic for control parameters, which can adjust the parameters of the proportional-integral-derivative control algorithm in real time according to the load changes and mechanical wear of the actuator, thereby optimizing the control effect and reducing tracking error.

[0108] The status monitoring unit is used to monitor the operating status and parameter adjustments of each actuator in real time, including the speed, current, and temperature of the drive motor, the vibration amplitude and excitation signal of the vibration generator, the speed, direction, and angle of the service motor, and the operating status of the attitude adjustment mechanism. It collects relevant status parameters through various sensors and feeds these parameters back to the closed-loop control unit and the fault diagnosis unit. When an abnormality is detected in the operation of an actuator, such as excessively high drive motor temperature, vibration amplitude adjustment error exceeding the allowable tolerance, or service motor jamming, a feedback signal is immediately sent to the closed-loop control unit to trigger a parameter fine-tuning command. Simultaneously, an abnormal signal is sent to the fault diagnosis unit to trigger the alarm module. The status monitoring unit also has a data logging function, recording the operating status data and parameter adjustment data of each actuator to facilitate subsequent maintenance and troubleshooting.

[0109] The data interface is used for data interaction with the biomimetic online optimization calculation module and the status monitoring unit. It receives the optimal combination of operating parameters and control commands from the biomimetic online optimization calculation module, receives the operating status data of the actuators from the status monitoring unit, and simultaneously feeds back information such as parameter adjustment status and tracking errors to the biomimetic online optimization calculation module, while sending monitoring commands to the status monitoring unit. The data interface adopts a standardized interface type to ensure the stability and compatibility of data transmission, possesses anti-interference capabilities, and adapts to the complex environment of industrial sites. The power supply module provides a stable power supply to the entire end effector dynamic adjustment module and each actuator. It provides different voltage levels of power output according to the power requirements of the actuators, and has overvoltage, overcurrent, overload, and overheat protection functions to prevent damage to the module and actuators due to abnormal power supply. The power supply module adopts an isolation design to reduce the impact of external electromagnetic interference on module operation and ensure module stability.

[0110] The software component of the end effector dynamic adjustment module mainly includes a parameter parsing module, a closed-loop control module, a drive control module, a status feedback module, and a fault diagnosis module. These modules work collaboratively to ensure the accuracy and stability of parameter adjustment. The parameter parsing module parses, verifies, and caches the received optimal operating parameter combination, sends the parsed target parameters to the closed-loop control module, and simultaneously feeds back the parameter verification results to the biomimetic online optimization calculation module. The closed-loop control module deploys the control logic and algorithms for the three closed-loop control loops. It receives the target parameters from the parameter parsing module and the actual operating parameters from the status monitoring unit, calculates the tracking error, generates control signals, and sends them to the drive control module. Simultaneously, this module receives abnormal signals from the status monitoring unit, triggers parameter fine-tuning commands, sends fine-tuning requests to the biomimetic online optimization calculation module, obtains the fine-tuned optimal operating parameter combination, and readjusts the parameters to compensate for the effects of mechanical wear or load changes.

[0111] The drive control module receives control signals from the closed-loop control module, converts these signals into drive signals that the drive modules can recognize, controls the operation of each drive module, drives the actuators, and monitors the operating status of the drive modules. If an abnormality occurs in a drive module, the fault diagnosis module is immediately triggered. The status feedback module collects the operating status data and parameter adjustment data of the actuators sent by the status monitoring unit, processes and analyzes the data, generates a status feedback report, and sends it to the biomimetic online optimization calculation module via a data interface, providing a basis for optimizing the optimization algorithm and adjusting parameters. Simultaneously, this module sends status data to the fault diagnosis module, providing data support for fault diagnosis. The fault diagnosis module monitors the hardware and software operating status of the module, receives abnormal signals from the status monitoring unit and the drive control module, determines the fault type and cause, triggers the alarm module to issue corresponding alarm signals, and records fault information to the memory. For minor faults, the module can automatically attempt recovery; for example, if the drive module operates abnormally, it will automatically restart the drive module; if the parameter adjustment error is too large, it will automatically trigger a fine-tuning command. For serious faults, the module will stop the operation of the relevant execution components to prevent the fault from worsening and protect the equipment safety.

[0112] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A method for online optimization of parameters at the powder extraction end of corn using biomimetic technology, characterized in that, Includes the following steps: Vibration signals, sound signals, and image data are collected in real time at the end of the corn flour extraction device to obtain multimodal sensing data of the flour extraction process; Feature extraction and fusion processing are performed on the multimodal sensing data of the powder extraction process to obtain a comprehensive state feature vector representing the current powder extraction state and quality; Based on the comprehensive state feature vector, a biomimetic optimization algorithm that simulates the intelligent cooperation mechanism of biological groups is used to iteratively calculate the optimal combination of working parameters at the end of the powder extraction device online. Based on the optimal combination of operating parameters, the rotational speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder extraction device are dynamically adjusted.

2. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 1, characterized in that, The step of extracting and fusing features from the multimodal sensing data of the powder extraction process to obtain a comprehensive state feature vector representing the current powder extraction state and quality specifically includes: Empirical mode decomposition is performed on the vibration signal to obtain multiple intrinsic mode function components. The sample entropy and energy ratio of each component are calculated to form a subset of vibration time-frequency domain features. Mel frequency cepstral coefficients are extracted from the sound signal, and its zero-crossing rate and short-time energy are calculated to form a feature subset in the sound frequency domain; Image data is preprocessed and region of interest segmented to extract surface texture roughness of corn ears, area ratio of powder splashing region, and contour features of unthreshed kernels, forming a subset of visual morphological features; The vibration time-frequency domain feature subset, the acoustic audio domain feature subset, and the visual morphology feature subset are processed by an attention-based feature fusion network. The feature fusion network dynamically assigns weights to each feature subset and performs weighted concatenation. Then, a fully connected layer is used for dimensionality reduction and information compression, and finally outputs a low-dimensional and information-dense comprehensive state feature vector.

3. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 1, characterized in that, The steps of using a biomimetic optimization algorithm that simulates the intelligent cooperation mechanism of biological groups to iteratively calculate the optimal combination of operating parameters at the end of the powder extraction device include: Initialize an optimization population consisting of multiple virtual agents, each virtual agent representing a set of candidate operating parameters, including the speed of the drive motor, vibration amplitude, spatial pitch angle and spatial yaw angle; Define a fitness function for the virtual agent. The input of the fitness function is the comprehensive state feature vector, and the output is a comprehensive evaluation score. The comprehensive evaluation score is obtained by weighted summation of the powder collection efficiency score, the powder collection integrity score, and the energy consumption score. Simulating bird flock foraging behavior, each virtual agent updates its candidate working parameters for the next iteration based on its own historical best parameter position, the population's global best parameter position, and a random perturbation term. A simulated ant colony pheromone diffusion mechanism is introduced to assign virtual pheromones to the parameter regions represented by virtual agents with higher fitness, guiding other virtual agents to conduct regional fine-grained searches in high-quality parameter regions; When the optimal fitness value in the population changes less than the convergence threshold within several consecutive iterations, or when the maximum number of iterations is reached, the iteration is terminated, and the candidate working parameters corresponding to the position of the global optimal parameter in the population at this time are determined as the optimal working parameter combination.

4. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 1, characterized in that, It also includes parameter initialization and constraint setting steps based on maize variety characteristics: Before starting online optimization, the variety identification features of the corn ear to be processed are obtained through the image recognition unit, or the corn variety code information input from the upstream process is received. Based on the variety identification features or the maize variety coding information, query the pre-established maize variety-working parameter knowledge base to obtain the recommended initial working parameter values ​​and parameter safety boundary constraints that match the variety. The recommended values ​​of the initial working parameters are used as the initial center position of the optimization population in the biomimetic optimization algorithm, and the parameter safety boundary constraints are used to limit the movement range of each virtual agent in the parameter space. The parameter safety boundary constraints are embedded in the iterative update formula of the biomimetic optimization algorithm in the form of inequalities to ensure that the candidate working parameter combinations generated in each generation are within the safe operating range allowed by the mechanical structure and process requirements.

5. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 1, characterized in that, It also includes online safety intervention procedures for powder collection device stall and abnormal operating conditions: During the online operation of the biomimetic optimization algorithm, the real-time current value of the drive motor is monitored in parallel to see if it exceeds the first safety threshold, and the amplitude of the vibration signal is monitored to see if it remains below the second safety threshold within a preset time window. If the real-time current value exceeds the first safety threshold, it is determined that there is a risk of stalling, the current optimization iteration process is immediately interrupted, and the first safety strategy is invoked. The first safety strategy forces the speed of the drive motor to be reduced to a preset safe speed and reverses for a short period of time to try to eliminate the blockage. If the vibration signal amplitude remains below the second safety threshold, it is determined that there is an abnormal working condition where the powder collection device has lost contact with the corn cob or the efficiency is extremely low. The current optimization iteration process is immediately interrupted, and the second safety strategy is invoked. The second safety strategy controls the end of the powder collection device to perform a preset spatial posture search motion until the vibration signal amplitude returns to the normal range, and then the bionic optimization algorithm is restarted.

6. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 3, characterized in that, It also includes the multi-objective dynamic weight adjustment step in the optimization process: During the iterative process of the biomimetic optimization algorithm, the average powder extraction efficiency, average powder extraction integrity, and average unit energy consumption in recent historical data are calculated in real time. A dynamic weight adjuster based on fuzzy logic rules is established. The input of the dynamic weight adjuster is the deviation of the average fan acquisition efficiency, average fan acquisition integrity and average unit energy consumption from their respective set targets. The output is the adjustment amount of the fan acquisition efficiency score weight, fan acquisition integrity score weight and energy consumption score weight in the fitness function. When the average fan acquisition efficiency is consistently lower than the target, the weight of the fan acquisition efficiency score is increased; when the average fan acquisition integrity is consistently lower than the target, the weight of the fan acquisition integrity score is increased; when the average unit energy consumption is consistently higher than the target, the weight of the energy consumption score is increased; the adjusted weights are applied to the fitness calculation of subsequent iteration cycles to guide the biomimetic optimization algorithm to search towards the current performance target that needs to be optimized.

7. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 1, characterized in that, It also includes steps for verifying the optimal combination of working parameters and updating the knowledge base: After the biomimetic optimization algorithm outputs a set of optimal working parameters and applies them to the powder collection device, new multimodal sensing data of the powder collection process is continuously collected within a fixed verification time window. Based on the new multimodal sensing data of the powder extraction process, the actual powder extraction efficiency, powder extraction integrity and energy consumption indicators are calculated to obtain the actual performance index set; The actual performance index set is compared with the performance index predicted during the optimization process. If the overall improvement is higher than the verification threshold, the optimization is considered successful. The successful maize variety identification features, the final optimal working parameter combination, and the corresponding actual performance index set are added as a new sample pair to the maize variety-working parameter knowledge base to enrich the initial parameter recommendations for subsequent similar varieties.

8. The method for online optimization of parameters at the corn biomimetic powder extraction end according to claim 1, characterized in that, The steps of dynamically adjusting the speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder extraction device relative to the corn cob specifically include: The optimal combination of operating parameters is analyzed as the target speed of the drive motor, the target amplitude of the vibration device, the target pitch axis angle, and the target yaw axis angle. By adjusting the output of the power electronic converter of the drive motor through a speed closed-loop control circuit, the actual speed of the motor is made to track the target speed value. The excitation signal of the vibration generator is adjusted by the amplitude closed-loop control loop so that the actual vibration amplitude tracks the target amplitude. By using a dual closed-loop control loop for spatial attitude, the service motors of the pitch axis and yaw axis are adjusted respectively, so that the actual pitch angle and yaw angle at the end of the powder collection device track the target angle. During the adjustment process, the tracking error of each loop is monitored in real time. If the tracking error of any loop continues to exceed the allowable tolerance, a parameter fine-tuning command is triggered. The biomimetic optimization algorithm is used to perform local re-optimization in a small range near the optimal combination of working parameters to compensate for the effects of mechanical wear or load changes.

9. A corn biomimetic powder extraction end parameter online optimization system, used to implement the corn biomimetic powder extraction end parameter online optimization method according to any one of claims 1 to 8, characterized in that, include: Multimodal sensing data acquisition module, comprehensive state feature extraction and fusion module, biomimetic online optimization calculation module, and end effector dynamic adjustment module; The multimodal sensing data acquisition module is used to collect vibration signals, sound signals and image data of the corn flour extraction device in real time during operation, obtain multimodal sensing data of the flour extraction process, and send the multimodal sensing data of the flour extraction process to the comprehensive state feature extraction and fusion module. The comprehensive state feature extraction and fusion module is used to perform feature extraction and fusion processing on the received multimodal perception data of the powder extraction process to obtain a comprehensive state feature vector that represents the current powder extraction state and quality, and send the comprehensive state feature vector to the biomimetic online optimization calculation module. The biomimetic online optimization calculation module is used to calculate the optimal combination of working parameters at the end of the powder extraction device online based on the received comprehensive state feature vector and a biomimetic optimization algorithm that simulates the intelligent cooperation mechanism of biological groups, and then send the optimal combination of working parameters to the end effector dynamic adjustment module. The end effector dynamic adjustment module is used to dynamically adjust the speed, vibration amplitude, and spatial attitude angle of the drive motor at the end of the powder collection device relative to the corn cob, based on the received optimal working parameter combination.

10. The online optimization system for corn biomimetic powder extraction end parameters according to claim 9, characterized in that, The biomimetic online optimization calculation module includes: an optimization population management unit, a fitness evaluation unit, and a biomimetic collaborative iteration unit; The optimization population management unit is used to initialize and maintain an optimization population consisting of multiple virtual agents. Each virtual agent represents a set of candidate working parameters, including drive motor speed, vibration amplitude, spatial pitch angle and spatial yaw angle. The fitness evaluation unit is connected to the comprehensive state feature extraction and fusion module and the optimization population management unit. It is used to receive the comprehensive state feature vector and, according to the candidate working parameters provided by the optimization population management unit, call the preset fitness function to calculate the comprehensive evaluation score of each virtual agent. The comprehensive evaluation score is obtained by weighted summation of the powder collection efficiency score, the powder collection integrity score and the energy consumption score. The biomimetic collaborative iteration unit is connected to the fitness evaluation unit and the optimization population management unit. It is used to simulate the collaborative mechanism of bird flock foraging and ant colony pheromone diffusion. Based on the historical optimal parameters of each virtual agent, the global optimal parameters of the population, and the virtual pheromone distribution, it drives the optimization population management unit to update the candidate working parameters of all virtual agents. When the preset convergence condition is met, the final global optimal parameters of the population are output as the optimal working parameter combination.

Images