Collaborative control method for dynamic grain quality grading and stone removal based on multimodal sensor fusion

Abstract

The present invention relates to the field of grain sorting and cleaning technology, and more particularly to a method for coordinated control of dynamic grain quality grading and stone removal based on multimodal sensor fusion. The method comprises the following steps: Step 1: synchronous acquisition using multimodal sensors, where a high-speed photoelectric trigger generates synchronous pulses as the grain falls, triggering a linear array camera to capture a sequence of grain surface morphology images, a near-infrared spectrometer to capture absorbance curves and moisture content data, and a piezoelectric weighing platform to capture impact pressure waveforms; Step 2: dynamic grading decision-making; Step 3: coordinated parameter generation; Step 4: execution and feedback optimization, where the generated parameters are sent to the vibrating screen, fan, and inclination adjustment mechanism for execution and control; and detecting the residual impurity ratio at the stone removal outlet. If the ratio exceeds the limit continuously, retraining of the cascade decision model is triggered. Through the above design, the present invention significantly improves the efficiency and accuracy of grain grading and stone removal, providing users with a high-quality product experience and value-added benefits.

Description

Technical Field

[0001] The present invention relates to the technical field of grain sorting and cleaning, and in particular to a method for coordinated control of dynamic grain quality grading and stone removal based on multimodal sensor fusion. Background Art

[0002] Grain cleaning and sorting are key steps in grain processing, directly affecting grain storage safety and processing quality. Existing technologies primarily use equipment such as vibrating screens and gravity stoners to remove impurities. However, existing technologies have the following inherent drawbacks:

[0003] The existing production line designs quality grading (such as identification of moldy kernels) and stone removal as independent work sections: moldy kernel sorting relies on color sorters or near-infrared sorting equipment, while stone removal is completed by gravity sorters.

[0004] Industry rationality: Due to the overlapping physical properties of moldy particles and gravel (such as similar density), a single device cannot meet the sorting accuracy requirements of both types of impurities at the same time, forcing companies to adopt a segmented processing mode.

[0005] In addition, mainstream equipment relies on a single sensing technology: optical sorters only collect surface morphology / color, and specific gravity sorters only detect weight signals.

[0006] The unavoidable overlap of physical properties: Moldy grains increase in density due to water loss (1.2-1.4 g / cm 3 ), compared with common quartz stone (1.3-1.5 g / cm 3 ) density ranges overlap, resulting in:

[0007] The optical sensor misidentifies dark stones as moldy particles, and the specific gravity sensor misidentifies high-density moldy particles as stones.

[0008] The misjudgment rate in actual production lines reaches 15%-20%, and in serious cases, rework is required.

[0009] Therefore, there is an urgent need for a collaborative control method of dynamic grain quality grading and stone removal based on multimodal sensor fusion to solve the above problems. Summary of the Invention

[0010] Based on the above objectives, the present invention provides a method for coordinated control of dynamic grain quality grading and stone removal based on multimodal sensor fusion, comprising the following steps:

[0011] Step 1: Synchronous acquisition by multimodal sensors: A high-speed photoelectric trigger generates synchronous pulses as the grain falls, triggering a linear array camera to capture sequential images of the grain surface morphology, a near-infrared spectrometer to collect absorbance curves and moisture content data, and a piezoelectric weighing platform to collect impact pressure waveforms.

[0012] Step 2: Dynamic hierarchical decision-making: extract morphological features, compositional features, and mechanical features from the collected data. The morphological features are calculated based on the major axis / minor axis ratio and surface depression depth of the morphological sequence images. The compositional features are calculated based on the peak height ratio of the absorbance curve at the preset mold characteristic peak to calculate the mold indication coefficient. The mechanical features are calculated based on the maximum slope of the impact pressure waveform during the rising phase. The morphological, compositional, and mechanical features are input into the cascade decision model to execute cascade decision-making: the primary decision is to mark the suspected moldy grain when the mold indication coefficient exceeds the dynamic threshold. The secondary decision is to match the density characterization factor of the suspected moldy grain with the stone feature library for waveform matching and determine whether it is a stone or moldy grain based on the matching degree, and output the classification label.

[0013] Step 3: Collaborative parameter generation: Calculate the stone incorporation rate and moldy grain aggregation in real time based on the classification labels. The stone incorporation rate is the percentage of stones in the total processed volume, and the moldy grain aggregation is the standard deviation of the moldy grain distribution across the width of the conveyor belt. Based on the calculated results and the moisture content data from Step 1, generate de-stoner equipment parameters: the vibration amplitude of the vibrating screen increases in a stepwise manner with increasing stone incorporation rate, the airflow velocity increases with increasing moldy grain aggregation, and the screen surface inclination is adjusted inversely based on the moisture content.

[0014] Step 4: Execute and feedback optimization, and send the generated parameters to the vibrating screen, fan, and inclination adjustment mechanism for execution and control; detect the residual impurity ratio at the stone removal outlet, and trigger retraining of the cascade decision model when it exceeds the limit continuously.

[0015] Preferably, the implementation of the synchronous acquisition of multimodal sensors in step 1 includes:

[0016] The high-speed photoelectric trigger is composed of an infrared transmitter and a receiver, and generates a square wave pulse with adjustable pulse width when grains fall and block the light beam.

[0017] The linear array camera starts exposure after a first fixed time window is delayed after the rising edge of the pulse. The first fixed time window is determined through a calibration test: after grains of different particle sizes are allowed to freely fall from the drop port, the average time it takes for them to reach the center of the detection area is calculated, and this average time is set as the delay window;

[0018] The near-infrared spectrometer triggers high-speed scanning on the rising edge of the pulse. The scanning band covers the spectral range where the characteristic absorption peak of mold is located. The scanning resolution is adaptively adjusted according to the curvature of the grain surface. The greater the curvature, the higher the resolution is to capture local mold characteristics.

[0019] The piezoelectric weighing platform starts waveform recording at the rising edge of the pulse, and the recording duration is determined by the impact dynamics test: multiple groups of grain impact waveforms are recorded, and the maximum duration of the waveform from the beginning to the decay to the baseline noise level is calculated, and this duration is set as the recording duration.

[0020] Preferably, the method for generating the dynamic threshold in step 2 includes:

[0021] Construct a nonlinear mapping library between moisture content and mold indication coefficient thresholds: During the equipment commissioning phase, collect representative grain samples of different moisture contents. For each sample, compile a distribution histogram of the mold indication coefficient for moldy grains. The coefficient value corresponding to the bottom of the histogram is taken as the optimal threshold for that moisture content.

[0022] Real-time acquisition of current moisture content: The absorbance value of the near-infrared spectrometer in the moisture characteristic band is substituted into the pre-calibrated absorbance-moisture content conversion model for calculation;

[0023] Dynamic threshold query: Input the current moisture content into the mapping library. If the moisture content is between two calibration points, linear interpolation is used to calculate the real-time threshold; if it exceeds the calibration range, the nearest calibration threshold is used.

[0024] Preferably, the specific process of waveform matching in step 2 is:

[0025] The stone feature database establishes a multidimensional index based on origin and mineral type, and calls the corresponding sub-database based on the current grain origin information;

[0026] Waveform normalization processing: extract the rising edge segment of the current pressure waveform, normalize the amplitude based on the time axis, and eliminate the amplitude fluctuation caused by the impact velocity difference;

[0027] Dynamic time warping distance calculation: The normalized waveform and the standard waveform are discretized into a sequence of points with equal time intervals. The Euclidean distance matrix between the two waveform point pairs is constructed. A dynamic programming algorithm is used to search for the path with the minimum cumulative distance from the upper left corner to the lower right corner of the matrix. The cumulative distance of this path is the warping distance.

[0028] Matching degree conversion: Establish a negative correlation mapping relationship between regularized distance and matching degree. This relationship is calibrated through the confusion matrix: input a sample set with known classification, calculate the correct matching probability under different regularized distances, and fit the distance-probability conversion curve.

[0029] Preferably, in step 3, the control logic for the stepwise increase of the vibration screen amplitude as the stone mixing rate increases includes: the higher the mixing rate, the higher the amplitude level, and the interval division is dynamically optimized according to the historical distribution density, and the amplitude switching needs to be triggered after a stable exceeding of the limit for multiple consecutive cycles, which specifically includes the following:

[0030] Amplitude level interval division: Measure the amplitude values ​​under different eccentric block weights during the equipment no-load test, and divide the amplitude range into several levels with the maximum safe amplitude as the upper limit;

[0031] Adaptive mixing rate range: Based on historical data, the distribution density of stone mixing rate is calculated. In high-density ranges, the grade span is reduced to improve control accuracy, and in low-density ranges, the span is increased to reduce the adjustment frequency.

[0032] Level switching rules: Amplitude level switching is triggered only when the real-time mixing rate remains stable in the new range for multiple consecutive detection cycles. During the switching process, the mass distribution of the eccentric block is gradually adjusted according to the preset acceleration curve to avoid instantaneous impact loads.

[0033] Preferably, the reverse adjustment of the screen surface inclination angle according to the moisture content in step 3 is implemented as follows: when the moisture content exceeds a preset critical value, the inclination angle is reduced according to the calibration curve; when the moisture content falls below the critical value, the slope is restored to the basic inclination angle, which specifically includes:

[0034] Moisture content critical value calibration: Conduct flowability tests on different types of grains, gradually increase the moisture content at a fixed inclination angle, and record the moisture content as the critical value when the grain flow rate drops to the safety threshold;

[0035] Compensation calculation model: A proportional relationship between the excess moisture content and the inclination compensation amount is established. The proportional coefficient is determined through a residence time calibration test. At different moisture contents exceeding the critical value, the inclination reduction required to achieve the target residence time is measured, and a moisture content-compensation curve is fitted.

[0036] Dynamic compensation execution: When the real-time moisture content exceeds the critical value, the compensation amount is calculated according to the curve, and the electric push rod drives the screen surface inclination to decrease; when the moisture content falls below the critical value, it is restored to the basic inclination at a slow slope.

[0037] Preferably, the triggering and execution of retraining in step 4 includes: when the X-ray measured impurity ratio continuously exceeds the model expected value and the weighted moving average exceeds the dynamic tolerance threshold, starting the incremental learning mechanism to optimize the cascade decision model, which specifically includes:

[0038] Calculation of impurity residual deviation: Taking the expected impurity ratio output by the cascade decision model as the benchmark, make the absolute difference with the actual value measured by the X-ray density meter, and then multiply it by the current flow weight coefficient to obtain the weighted deviation;

[0039] Trigger condition determination: Set a time window and calculate the moving average of the weighted deviation within the window. When the average value continuously exceeds the dynamic tolerance threshold, retraining is triggered. The tolerance threshold is set according to the impurity fluctuation range of the equipment's historical best operating conditions;

[0040] Incremental learning mechanism: All misjudgment sample data within the trigger period is collected, and its feature vectors are extracted to form a training set; with the reduction of weighted deviation as the optimization goal, the gradient descent method is used to iteratively update the weight matrix of the cascade decision model. The comprehensive recognition rate on the validation set is verified after each iteration, and training is terminated when the recognition rate reaches saturation.

[0041] Preferably, the method for constructing the stone feature library includes:

[0042] Sample collection and pre-processing: Collect typical mineral types of stones from the target production area, classify them according to geological composition, and select samples with particle sizes within the grain size range;

[0043] Impact waveform recording: Build a simulated free-fall device to make the stone fall from the actual production line drop height to impact the pressure station, and record the waveform of the entire impact process;

[0044] Feature extraction and optimization: After performing noise reduction filtering on the waveform, we extract three core features: rising edge slope, peak acceleration, and oscillation decay period. We perform cluster analysis on the feature sets of stones of the same type, remove samples that deviate from the main cluster, and calculate the feature mean and variance.

[0045] Feature library update mechanism: When the system is put into use in a new production area, the learning mode is first run to collect local stone samples, which are then added to the feature library after manual verification.

[0046] Preferably, the optimization process of the gradient descent method includes:

[0047] Loss function construction: The sum of weighted deviations of misjudged samples is used as the main loss term, and a model weight regularization term is added to prevent overfitting;

[0048] Adaptive learning rate setting: The initial learning rate is set according to the estimated value of the loss function surface curvature, and is dynamically adjusted according to the rate of change of the gradient direction during the iteration process: when the gradient direction of consecutive iterations is consistent, the learning rate is increased, and when the direction oscillates, the learning rate is reduced;

[0049] Early stopping mechanism: When the harmonic mean of the moldy grain recognition rate and the stone misidentification rate on the validation set no longer increases, the system rolls back to the optimal weight state.

[0050] Preferably, the specific method of increasing the air flow velocity as the aggregation degree of moldy particles increases includes:

[0051] Nozzle array zoning control: high-pressure airflow nozzles are arranged equidistantly along the width of the conveyor belt, with each nozzle covering a fixed area;

[0052] Aggregation distribution calculation: Real-time statistics are collected on the percentage of moldy particles in the nozzle partitions, and the standard deviation of the percentage of each partition and the average value is calculated as the local aggregation degree;

[0053] Spray strategy generation: For the zones where the local concentration exceeds the limit, the corresponding nozzles are opened, and the spray angle is dynamically adjusted according to the conveyor belt speed: when the speed is high, the spray angle is reduced to extend the action time, and when the speed is low, the angle is increased to expand the coverage area;

[0054] Spray parameter optimization: With the goal of uniform distribution of moldy particles, the optimal spray duration and pressure combination is determined through feedback adjustment.

[0055] Beneficial effects of the present invention:

[0056] 1. By incorporating concentration distribution calculation and spray strategy generation technology, this invention monitors the distribution of moldy grains in real time, analyzes the data, and dynamically adjusts the spray angle and area. This approach ensures that the spray device precisely targets areas where moldy grains are concentrated, effectively improving moldy grain removal efficiency and making the grain purification process more thorough.

[0057] 2. The present invention dynamically adjusts the spray angle according to the conveyor belt speed to ensure good spray coverage and action time under different speed conditions, thereby stabilizing the stone removal effect.

[0058] 3. The present invention optimizes injection parameters through feedback regulation, automatically adjusts injection duration and pressure, reduces manual intervention, and improves operation accuracy and response speed.

[0059] 4. The nozzle array zoning control of the present invention enables airflow control to be accurate in each zone, improving overall control accuracy and effectively reducing grain losses caused by over-spraying. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0061] Figure 1 is a flow chart of the steps of the method of the present invention;

[0062] Figure 2 Flow chart of the steps of the method for constructing a stone feature library in the method of the present invention;

[0063] Figure 3 The figure is a flow chart of the steps of the optimization process of the gradient descent method in the method of the present invention. DETAILED DESCRIPTION

[0064] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0065] See Figure 1-Figure 3 An embodiment of the present invention provides a method for dynamic grain quality grading and coordinated stone removal control based on multimodal sensor fusion. In step 1, the method uses multimodal sensors to synchronously collect multiple data points as the grain falls, thereby achieving dynamic grain quality grading and coordinated stone removal control. First, a high-speed photoelectric trigger generates synchronous pulses as the grain falls. This process triggers a linear array camera to capture a sequence of images of the grain surface morphology. Simultaneously, a near-infrared spectrometer records absorbance curves and moisture content data, while a piezoelectric weighing platform measures the impact pressure waveform. This data lays the foundation for subsequent dynamic grading decisions.

[0066] In step 2, during the dynamic classification process, detailed features are extracted from the aspects of form, composition, and mechanics. Specifically, morphological features are calculated by calculating the major axis / minor axis ratio and surface depression depth of the grain through morphological sequence images; compositional features are based on the absorbance curve, and the peak height ratio is calculated at the mold characteristic peak to obtain the mold indication coefficient; the acquisition of mechanical features relies on the impact pressure waveform, and the density characterization factor is calculated by calculating the maximum slope of the waveform rising phase. Through the cascade decision model, when the mold indication coefficient exceeds a specific threshold, the grain is marked as suspected moldy grain; the density characterization factor is further compared with the stone feature library to determine whether the grain is a stone or moldy grain, and finally form a classification label.

[0067] In step 3, based on this, the stone inclusion rate and moldy grain aggregation are calculated based on the real-time classification labels. These two indicators represent the percentage of stone in the total processing volume and the standard deviation of the moldy grain distribution on the conveyor belt, respectively. Based on these calculations and initial data, the parameters of the stone removal equipment were optimized: the vibration amplitude of the vibrating screen was increased based on the stone ratio, the airflow velocity was adjusted according to the moldy grain aggregation, and the screen surface inclination was adjusted inversely based on the moisture content.

[0068] Finally, by continuously monitoring the residual impurity ratio at the stone removal outlet, the cascade decision-making model's performance is improved and optimized through retraining when the ratio repeatedly exceeds the set value. This approach not only achieves efficient grain grading and stone removal, but also significantly improves processing efficiency and accuracy, reducing the probability of misjudgments and omissions.

[0069] In one possible implementation, a high-speed photoelectric trigger, consisting of an infrared transmitter and receiver, is designed to detect falling grain. When grain passes through the sensor's path, it interrupts the beam, generating a square wave pulse signal with an adjustable pulse width.

[0070] Specifically, after receiving the rising edge of a square wave pulse, the line array camera begins exposure according to a pre-set delay window. This delay window is determined through calibration testing. The goal is to calculate the average time required for grains of varying particle sizes to freely fall from the feed opening and reach the center of the inspection zone. This provides an optimal delay window, ensuring the camera can clearly capture the grain's surface morphology when it reaches the critical position.

[0071] The near-infrared spectrometer also uses the rising edge of a square wave pulse as a trigger, performing a high-speed scanning motion. The scanning band is concentrated in the spectral range of the characteristic absorption peaks of mold. Furthermore, to ensure detection accuracy, the spectrometer's scanning resolution is adaptively adjusted based on the curvature of the grain surface. Specifically, areas with greater curvature indicate a higher likelihood of mold, so increased resolution captures more detailed features.

[0072] For the piezoelectric weighing platform, the impact waveform of the grain begins recording after the rising edge of the square wave pulse is triggered. The recording duration is not arbitrarily set but rather determined through impact dynamics testing. This testing included statistical analysis of the impact waveforms of multiple groups of grains, with the maximum time from the waveform's onset to its decay to the baseline noise level being the upper limit of the recording. This setting ensures that the waveform recording fully captures the impact process of the grain.

[0073] This series of steps not only technically enhances the sensor fusion effect, but also improves the accuracy and real-time nature of data collection, providing a reliable guarantee for subsequent dynamic grading and stone removal processes, and significantly improving the response speed and classification accuracy of grain quality detection.

[0074] In one possible implementation, accurate grain quality grading is achieved through dynamic threshold generation, establishing a nonlinear mapping library between moisture content and mold indicator coefficients. During the equipment commissioning phase, representative grain samples with varying moisture contents are selected and statistically analyzed for mold indicator coefficients. This involves generating a histogram of the mold indicator coefficient distribution for each sample and then identifying the coefficient value corresponding to the valley bottom. This value is considered the optimal threshold for that moisture content. This nonlinear mapping method helps capture the complex dynamic relationship between mold probability and moisture content variations.

[0075] To ensure real-time performance, the method includes information on how to obtain the current moisture content. By combining absorbance values ​​measured by a near-infrared spectrometer within a specific wavelength band associated with moisture with a pre-calibrated absorbance-to-moisture conversion model, the moisture content of the grain can be calculated in real time. This process leverages the efficiency and accuracy of spectral analysis to ensure that the measured data effectively represents the current moisture status of the grain.

[0076] The dynamic threshold generation process requires inputting real-time moisture content values ​​into a pre-built mapping library. If the current moisture content falls between two calibrated sample data points, the real-time threshold is calculated using linear interpolation to ensure smooth transitions and reasonableness of the calculated results. If the moisture content falls outside the calibrated range of the mapping library, the method uses the threshold corresponding to the data point closest to the current moisture content to mitigate uncertainty caused by extrapolation.

[0077] This dynamic threshold generation mechanism effectively addresses variations in grain mold severity at varying moisture contents, enabling more accurate grain quality grading in practical applications. This dynamic adjustment not only increases process flexibility but also ensures stability in the face of changing conditions, contributing to more accurate de-stoning and grading control processes and improving overall grain quality management.

[0078] In one possible implementation, a detailed feature library is first established based on the origin of the stone and the type of mineral. When this library is used, the corresponding sub-library is selected and called based on the origin of the grain. This sub-library management method effectively utilizes existing mineral feature information and ensures accurate feature matching under the same origin conditions.

[0079] In data processing, the rising edge of the current pressure waveform is normalized, with amplitude used as the standard for adjustment. Amplitude normalization reduces fluctuations caused by impact velocity differences, making waveform data more consistent and reliable when compared. This process focuses on stabilizing the quality of the basic waveform analysis data through timeline standardization.

[0080] Next, the calculation of the dynamic time warping distance is crucial for matching. After discretizing the normalized waveform and the standard sample waveform into a sequence of equally spaced points, a series of Euclidean distance matrices is formed. To identify the best match, this matrix is ​​scanned using a dynamic programming algorithm, with the minimum cumulative distance along the path from the top left corner to the bottom right corner representing the warping distance. This warping distance flexibly accommodates waveform deformation along the time axis and is a crucial indicator of waveform matching accuracy.

[0081] Finally, the warping distance and matching degree are mapped using a negative correlation. By inputting a sample set with known classification results, the matching probabilities at different warping distances are calculated to form a confusion matrix, which is then used to fit a distance-probability conversion curve. This mapping relationship provides an intuitive quantitative method from warping distance to matching degree, allowing the accuracy of stone removal to be evaluated based on the actual matching degree.

[0082] This waveform matching method achieves precise stone removal based on multi-dimensional features, improves overall recognition efficiency and reliability, and ensures robust performance under complex conditions by using dynamic time warping technology and optimized probability conversion. It is of significant significance for improving the intelligence and precision of grain processing.

[0083] In one possible implementation, during a no-load test of the equipment, the vibration amplitude is measured under different eccentric counterweight conditions to ensure that the maximum amplitude does not exceed the equipment's safety limit. These measured amplitude values ​​are used to classify the amplitude range. The amplitude range is divided into multiple equal levels, which provide the basis for subsequent vibrating screen adjustments.

[0084] After the amplitude level is determined, the next step is to adaptively adjust the amplitude based on the actual stone mixing ratio. This adjustment is based on an analysis of historical data. Specifically, the level spans for different mixing ratio ranges are specified based on the distribution density of the stone mixing ratio. In high-density ranges, the level spans are narrower to enable more precise amplitude control; in low-density ranges, the level spans are widened to reduce unnecessary frequent adjustments and improve overall control efficiency.

[0085] The level switching rules ensure stable screening results in dynamic environments. Amplitude level switching is only executed when the real-time contamination rate remains within the new density range for multiple consecutive detection cycles, thus avoiding unnecessary frequent fluctuations. During the switching process, a preset acceleration curve is used to gradually adjust the mass distribution of the eccentric block, effectively avoiding instantaneous impact loads and protecting the safety of the equipment and the objects being processed.

[0086] This control method effectively improves stone removal efficiency while reducing equipment wear and energy consumption. This method not only enhances the precision of the stone removal process but also improves the stability and reliability of the production process, resulting in more accurate grain quality grading. This not only provides a guarantee for the production process but also significantly contributes to improving overall production efficiency and product quality.

[0087] In one possible implementation, the optimal adjustment of the screen surface inclination is achieved through reverse feedback of the grain moisture content. This process can be divided into three key steps: critical value calibration, compensation amount calculation, and dynamic compensation execution.

[0088] The first step is to calibrate the critical moisture content. To do this, flowability tests are conducted on different types of grain. At a fixed screen angle, the grain moisture content is gradually increased. When the grain flow rate drops below a safe threshold, the moisture content at that point is recorded as the critical value. This critical value allows for quick identification of the critical point where grain flowability is compromised during subsequent operations.

[0089] The next step was to develop a model for the relationship between moisture content and inclination compensation. This was achieved through multiple retention time calibration tests. In these tests, the required reduction in inclination angle to achieve the target retention time was measured at various moisture content levels exceeding the critical value. This data was fitted to produce a curve comparing moisture content and compensation. This curve served as the core basis for adjusting the screen deck inclination angle.

[0090] The final step involves implementing dynamic compensation. When real-time monitoring indicates that the grain moisture content exceeds a calibrated critical value, the required inclination compensation is calculated based on a fitted curve. Based on this, the electric actuator reduces the screen surface inclination, improving grain flowability. Furthermore, when the moisture content falls below the critical value, the inclination returns to the initial base angle at a slower rate. This adjustment ensures stable flowability across varying moisture contents, preventing flow rate reduction or material accumulation due to moisture content fluctuations.

[0091] This method significantly improves grain grading and stone removal efficiency, especially when the grain moisture content varies. Proper tilt adjustment not only reduces the risk associated with moisture content fluctuations, but also improves the adaptability and efficiency of the entire process, bringing greater stability and product quality to grain processing.

[0092] In one possible implementation, the impurity residual deviation is calculated by first using the output of the cascade decision model as the expected value and comparing it with the actual impurity ratio measured by the X-ray densitometer. The absolute difference between the two is calculated and multiplied by the weighting factor of the current flow rate to obtain a weighted deviation. This process accounts for the impact of flow rate changes on impurity identification, ensuring that the deviation reflects a more realistic deviation situation.

[0093] Next, a moving average of the weighted deviations within a defined time window is calculated. If this average value continuously exceeds a pre-set dynamic tolerance threshold, retraining is triggered. This tolerance threshold is based on the impurity fluctuation range of the device under historically optimal operating conditions to ensure moderate sensitivity and avoid frequent retraining triggered by minor fluctuations.

[0094] The incremental learning mechanism operates after the triggering conditions are met. During this phase, all misclassified sample data within the triggering period is collected and their feature vectors are extracted to form a new training dataset. Gradient descent is used to iteratively update the weights of the cascaded decision model, aiming to reduce weighted bias. After each update, the change in overall recognition rate is evaluated on a validation set to ensure the effectiveness of the training improvements. When the recognition rate stops improving after multiple iterations, reaching saturation, training is terminated.

[0095] Through automated and intelligent learning, the model's accuracy in identifying impurities is continuously optimized. This approach not only improves the equipment's adaptability but also reduces the need for human intervention and adjustments by continuously updating the model's learning capabilities, thereby improving overall process efficiency and product quality. Furthermore, the introduction of incremental learning enables the equipment to quickly respond to operational anomalies, ensuring the accuracy and stability of grading and stone removal.

[0096] In one possible implementation, the first stage involves sample collection and preprocessing. Stones representing representative mineral types are collected from a specific target production area. After collection, the samples are sorted and screened based on their geological composition, with a focus on selecting stones with a particle size similar to that of grains. This helps improve the accuracy of the simulated production environment during feature extraction.

[0097] Next comes the impact waveform recording phase. During this process, a simulated free-fall device is constructed, allowing the stone to fall freely from the actual drop height on the production line. The impact waveforms generated throughout the entire process are recorded on an impact pressure station. This device is designed to accurately reproduce the actual impact scenario on the production line, thereby obtaining representative waveform data.

[0098] Next, the feature extraction and optimization phase begins. The recorded waveform data undergoes noise reduction filtering to remove noise interference caused by the external environment or equipment errors. Three core features are then extracted from the waveform: rising edge slope, peak acceleration, and oscillation decay period. These features are key indicators selected to distinguish between stones and grains. Cluster analysis is performed on the feature set of stones of the same type, identifying and eliminating anomalous samples that deviate from the main cluster. Finally, the mean and variance of the features are calculated, ensuring the accuracy and stability of the feature library.

[0099] Finally, a feature library update mechanism is key to ensuring long-term, effective operation. When a new production area is commissioned, a learning model is initially run to collect new local stone samples. After manually verifying the accuracy and representativeness of these samples, they are incorporated into the feature library for timely updates. This mechanism ensures the timeliness and regional adaptability of the stone feature library.

[0100] This scientific approach to building a stone feature library significantly improves identification accuracy during the stone removal process. The meticulous processing and analysis steps at each stage ensure the accuracy and practicality of the entire process—from sampling, recording, feature extraction, to updating the feature library—thus improving equipment reliability and reducing misjudgment rates. This provides solid data support and technical assurance for grain grading and stone removal.

[0101] In one possible implementation, the loss function is constructed. The loss function is designed to use the sum of the weighted biases of misclassified samples as the primary loss term. This setting fully considers the impact of different samples on the final model decision, ensuring that the model remains focused on improving classification. Furthermore, a regularization term is introduced to the model weights to prevent overfitting during training. This regularization term limits the model's complexity, thereby enhancing generalization and ensuring that the model performs well even on unseen samples.

[0102] Next, we need to set the adaptive learning rate. The initial learning rate is not fixed, but rather is appropriately allocated based on the estimated curvature of the loss function surface. During the iteration process, the learning rate is dynamically adjusted by monitoring the rate of change in the gradient direction. Specifically, when the gradient direction remains consistent over multiple iterations, indicating a rapid decrease in error, the learning rate can be appropriately increased to accelerate convergence. Conversely, when the gradient direction oscillates, the learning rate should be reduced to avoid missing the optimal solution and maintain stable convergence performance.

[0103] Finally, there's the early stopping mechanism. During training, if the harmonic mean of the moldy grain recognition rate and the stone misidentification rate on the validation set stops improving, the early stopping mechanism kicks in. At this point, training halts further model optimization and rolls back to the optimal weight state. This mechanism effectively prevents overfitting on the validation set, conserves computing resources, and ensures more efficient training.

[0104] Through the above optimization strategies, gradient descent not only improves the model's ability to handle complex and variable input data, but also significantly reduces the false positive rate. Furthermore, the introduction of adaptive learning rate adjustment and early stopping further enhances the flexibility and efficiency of the training process. Together, these implementation steps and strategies provide an efficient and robust scientific approach for grain grading and stone removal, ensuring the accuracy and reliability of quality control within the business process.

[0105] In one possible implementation, zoned control of the nozzle array is a crucial foundation for airflow regulation. High-pressure airflow nozzles are evenly spaced across the width of the conveyor belt, with each nozzle responsible for controlling its own coverage area. This arrangement facilitates efficient and precise zoned airflow management, ensuring independent adjustment of nozzle action within different zones.

[0106] Calculating the concentration distribution is key to dynamic adjustments. By counting the percentage of moldy particles within each nozzle zone in real time and calculating the mean and standard deviation of these data, we can determine the local concentration. This calculation provides an important basis for developing subsequent spraying strategies. As an indicator of local concentration, the standard deviation reflects the uneven distribution of moldy particles in different areas.

[0107] Spray strategy generation relies on analysis of concentration data. After identifying a zone where local concentration exceeds a certain threshold, the corresponding nozzle is activated. The spray angle is adjusted dynamically based on the conveyor belt speed: at higher speeds, the spray angle is reduced to ensure that moldy particles are exposed to the airflow for a longer period of time. At lower speeds, the angle is increased to expand the spray coverage and better disperse the moldy particles.

[0108] Finally, by optimizing the spray parameters and achieving uniform distribution of moldy particles, feedback adjustments are performed to determine the optimal spray duration and pressure combination. This step is designed to achieve the best results during the spraying process, further ensuring complete removal of moldy particles and achieving uniform distribution.

[0109] This approach dynamically adjusts airflow velocity and spray parameters, not only improving moldy kernel removal efficiency but also ensuring consistent grain treatment across the entire conveyor belt. Precise nozzle control and real-time concentration calculation enable the system to rapidly respond to changing moldy kernel distribution, maintaining efficient stone removal and grading capabilities. This automated and adaptive control significantly reduces manual intervention, increases production line automation, and ultimately ensures high-quality grain output.

[0110] In order to make the principle of the present invention clearer, the following content is used to explain it again:

[0111] The calculation principles for key features include:

[0112] Major axis / minor axis ratio calculation:

[0113] Based on a series of grain surface morphology images captured by a line scan camera, an ellipse fitting algorithm is used to extract contour boundaries. In this calculation, the major axis is defined as the maximum diagonal length of the bounding rectangle multiplied by a pixel calibration factor (0.05 mm per pixel), and the minor axis is defined as the minimum width of the bounding rectangle multiplied by the same calibration factor. The ratio of the major axis to the minor axis is the morphological characteristic parameter R. During the experimental validation phase, measurement data from 1,000 grain samples of five types of grain, including wheat and corn, showed that healthy grains had R values ​​ranging from 1.2 to 2.1, while moldy grains had abnormal R values ​​(greater than 2.3 or less than 1.1) due to structural deformation.

[0114] Surface depression depth analysis:

[0115] Using 3D point cloud reconstruction technology, a topological model of the grain surface was generated. The depth of the depression was defined as the difference in elevation between the highest and lowest points in the point cloud mesh. Measured data showed that healthy grain, due to its intact surface, had a depression depth of no more than 0.15 mm, while moldy grain, due to localized decay, typically had depressions exceeding 0.35 mm.

[0116] Calculation of mildew indication coefficient:

[0117] Based on the absorbance data of the near-infrared spectrometer at 1640 nm (starch characteristic absorption peak) and 2100 nm (mold metabolite characteristic peak), the ratio of the two peak heights K (i.e. The mold indication coefficient, M, is further calibrated by introducing a moisture content compensation factor: M = K × (1.05 - 0.005 × moisture content). In calibration tests, when the moisture content of grain was 14%, an M value exceeding 0.78 indicated moldy grains. This threshold was determined using histogram valley statistics from 2,000 samples.

[0118] Derivation of density characterization factor:

[0119] Extract the rising phase data (0 to 5 milliseconds window) from the impact pressure waveform collected by the piezoelectric weighing platform and calculate the maximum slope S of the pressure change during this period. max (i.e. the maximum value of ΔP / Δt, time resolution 0.1 milliseconds). The density characterization factor ρ is corrected by the mass compensation formula, the expression is ρ=S max ×(0.98×m -0.5 ), where m is the mass of a single grain. Mechanical verification experiments confirmed that the ρ value of quartz and other stones is greater than 1.8 g / cm³ (a steeply rising waveform), while moldy grains, due to their loose interior, have a ρ value below 1.5 g / cm³ (a flat waveform). The construction and performance verification of the cascade decision model include:

[0120] The model structure and training process include:

[0121] Primary decision-making level:

[0122] Input features include morphological parameters (R, D) and composition parameters (M). A support vector machine (SVM) classifier performs binary classification. The radial basis function (RBF) is used as the kernel function, with the penalty coefficient C set to 1.0 and the kernel parameter γ set to 0.01. The decision rule is: when the mold indication coefficient M exceeds the dynamic threshold corresponding to the current moisture content, the kernel is marked as suspected moldy.

[0123] Secondary decision-making level:

[0124] The density characterization factor ρ of the suspected particle is matched against a standard waveform in the stone signature library using dynamic time warping (DTW). The degree of match is quantified by cumulative distance, with values ​​no greater than 0.1 considered stone. The stone signature library is organized by mineral type and covers common impurities such as granite and quartz.

[0125] Training data and validation:

[0126] The model was trained using 100,000 sets of labeled samples (covering five types of grain), with 70% used for training and 30% for validation. Validation results showed that the primary decision-making process achieved a 98.5% recall rate for moldy grains, while the secondary decision-making process kept the misclassification rate for stones below 1.2%, resulting in an overall classification accuracy of 98.8%.

[0127] The collaborative control parameter generation strategy includes:

[0128] Calculation of stone mixing rate:

[0129] The number of stones marked by the cascade model per unit time is counted in real time and divided by the total amount processed to obtain the mixing rate R (in percentage form).

[0130] Amplitude control:

[0131] When the mixing rate is less than 5%, amplitude level 1 (corresponding to a 50g eccentric weight) is used; between 5% and 10%, it is increased to level 2 (80g eccentric weight); above 10%, level 3 (120g eccentric weight) is used. Amplitude switching requires the mixing rate to remain consistently above the limit for three consecutive 60-second testing cycles. The switching process gradually adjusts the eccentric weight distribution using an acceleration curve of 5 degrees per square second.

[0132] Quantification of moldy grain aggregation:

[0133] The conveyor belt is divided into eight equally spaced zones (each 15 cm wide) along its width. The standard deviation of the proportion of moldy grains in each zone is calculated in real time and defined as the aggregation degree A.

[0134] Dynamic wind speed adjustment:

[0135] The base wind speed is set at 8 m / s. When the concentration exceeds 0.3, the wind speed is increased in real time according to the formula V = 8 + (A - 0.3) × 20. At the same time, the spray angle is dynamically adjusted based on the conveyor belt speed: when the speed exceeds 1 m / s, the angle is reduced from 45 degrees to 30 degrees to extend the duration of action.

[0136] Moisture content critical calibration:

[0137] Flowability tests determined that the critical moisture content of wheat is 14%. When the actual moisture content exceeds this value, compensation is made by reducing the inclination angle by 0.3 degrees for every 1% increase in moisture content, with a maximum reduction of 5 degrees. The compensation formula is: θ = θ0 - min(5, 0.3 × (W - 14)).

[0138] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0139] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion, characterized in that: The following steps are involved: Step 1: Synchronous acquisition by multimodal sensors: A high-speed photoelectric trigger generates synchronous pulses as the grain falls, triggering a linear array camera to capture sequential images of the grain surface morphology, a near-infrared spectrometer to collect absorbance curves and moisture content data, and a piezoelectric weighing platform to collect impact pressure waveforms. Step 2: Dynamic classification decision-making, extracting morphological characteristics, composition characteristics and mechanical characteristics from the collected data, where The morphological features are based on the calculation of the major axis / minor axis ratio and the depth of the surface depressions based on the morphological sequence images. The compositional features are based on the peak height ratio of the absorbance curve at the preset mold characteristic peak to calculate the mold indication coefficient. The mechanical features are based on the maximum slope of the impact pressure waveform during the rising phase to calculate the density characterization factor. The morphological, compositional, and mechanical features are input into the cascade decision model to execute cascade decisions: the primary decision is to mark the suspected moldy grain when the mold indication coefficient exceeds the dynamic threshold. The secondary decision is to match the density characterization factor of the suspected moldy grain with the stone feature library for waveform matching and determine whether it is a stone or moldy grain based on the matching degree, and output the classification label. Step 3: Collaborative parameter generation: Calculate the stone incorporation rate and moldy grain aggregation in real time based on the classification labels. The stone incorporation rate is the percentage of stones in the total processed volume, and the moldy grain aggregation is the standard deviation of the moldy grain distribution across the width of the conveyor belt. Based on the calculated results and the moisture content data from Step 1, generate de-stoner equipment parameters: the vibration amplitude of the vibrating screen increases in a stepwise manner with increasing stone incorporation rate, the airflow velocity increases with increasing moldy grain aggregation, and the screen surface inclination is adjusted inversely based on the moisture content. Step 4: Execute and feedback optimization, and send the generated parameters to the vibrating screen, fan, and inclination adjustment mechanism for execution and control; detect the residual impurity ratio at the stone removal outlet, and trigger retraining of the cascade decision model when it exceeds the limit continuously.

2. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: The implementation of multimodal sensor synchronous acquisition in step 1 includes: The high-speed photoelectric trigger is composed of an infrared transmitter and a receiver, and generates a square wave pulse with adjustable pulse width when grains fall and block the light beam. The linear array camera starts exposure after a first fixed time window is delayed after the rising edge of the pulse. The first fixed time window is determined through a calibration test: after grains of different particle sizes are allowed to freely fall from the drop port, the average time it takes for them to reach the center of the detection area is calculated, and this average time is set as the delay window; The near-infrared spectrometer triggers high-speed scanning on the rising edge of the pulse. The scanning band covers the spectral range where the characteristic absorption peak of mold is located. The scanning resolution is adaptively adjusted according to the curvature of the grain surface. The greater the curvature, the higher the resolution is to capture local mold characteristics. The piezoelectric weighing platform starts waveform recording at the rising edge of the pulse, and the recording duration is determined by the impact dynamics test: multiple groups of grain impact waveforms are recorded, and the maximum duration of the waveform from the beginning to the decay to the baseline noise level is calculated, and this duration is set as the recording duration.

3. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: The method for generating the dynamic threshold in step 2 includes: Construct a nonlinear mapping library between moisture content and mold indication coefficient thresholds: During the equipment commissioning phase, collect representative grain samples of different moisture contents. For each sample, compile a distribution histogram of the mold indication coefficient for moldy grains. The coefficient value corresponding to the bottom of the histogram is taken as the optimal threshold for that moisture content. Real-time acquisition of current moisture content: The absorbance value of the near-infrared spectrometer in the moisture characteristic band is substituted into the pre-calibrated absorbance-moisture content conversion model for calculation; Dynamic threshold query: Input the current moisture content into the mapping library. If the moisture content is between two calibration points, linear interpolation is used to calculate the real-time threshold; if it exceeds the calibration range, the nearest calibration threshold is used.

4. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: The specific process of waveform matching in step 2 is: The stone feature database establishes a multidimensional index based on origin-mineral type, and calls the standard waveform of the specific origin-mineral type in the corresponding sub-database according to the current grain origin information; Waveform normalization processing: extract the rising edge segment of the current pressure waveform, normalize the amplitude based on the time axis, and eliminate the amplitude fluctuation caused by the impact velocity difference; Dynamic time warping distance calculation: The normalized pressure waveform and the standard waveform are discretized into a sequence of points with equal time intervals. The Euclidean distance matrix between the two waveform point pairs is constructed. A dynamic programming algorithm is used to search for the path with the minimum cumulative distance from the upper left corner to the lower right corner of the matrix. The cumulative distance of this path is the warping distance. Matching degree conversion: Establish a negative correlation mapping relationship between regularized distance and matching degree. This relationship is calibrated through the confusion matrix: input a sample set with known classification, calculate the correct matching probability under different regularized distances, and fit the distance-probability conversion curve.

5. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: In step 3, the control logic for the step-by-step increase in the vibration amplitude of the vibrating screen as the stone mixing rate increases includes: The higher the mixing rate, the higher the amplitude level, and the interval division is dynamically optimized according to the historical distribution density. The amplitude switching needs to be triggered after stable exceeding the limit for multiple consecutive cycles.

6. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: In step 3, the screen surface inclination angle is reversely adjusted according to the moisture content as follows: When the moisture content exceeds the preset critical value, the inclination angle is reduced according to the calibration curve; when the moisture content falls below the critical value, the slope returns to the basic inclination angle.

7. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: The triggering and execution of retraining in step 4 include: When the X-ray measured impurity ratio continuously exceeds the model's expected value and the weighted moving average exceeds the dynamic tolerance threshold, the incremental learning mechanism is activated to optimize the cascade decision model.

8. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 4, characterized in that: The method for constructing the stone feature library includes: Sample collection and pre-processing: Collect typical mineral types of stones from the target production area, classify them according to geological composition, and select samples with particle sizes within the grain size range; Impact waveform recording: Build a simulated free-fall device to make the stone fall from the actual production line drop height to impact the pressure station, and record the waveform of the entire impact process; Feature extraction and optimization: After performing noise reduction filtering on the waveform, three core features, namely rising edge slope, peak acceleration, and oscillation decay period, are extracted as the standard waveform for the sub-library. Cluster analysis is performed on the feature sets of stones of the same type, and samples that deviate from the main cluster are removed before calculating the feature mean and variance. Feature library update mechanism: When the system is put into use in a new production area, the learning mode is first run to collect local stone samples, which are then added to the feature library after manual verification.

9. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 7, characterized in that: The optimization process of the gradient descent method includes: Loss function construction: The sum of weighted deviations of misjudged samples is used as the main loss term, and a model weight regularization term is added to prevent overfitting; Adaptive learning rate setting: The initial learning rate is set according to the estimated value of the loss function surface curvature, and is dynamically adjusted according to the rate of change of the gradient direction during the iteration process: when the gradient direction of consecutive iterations is consistent, the learning rate is increased, and when the direction oscillates, the learning rate is reduced; Early stopping mechanism: When the harmonic mean of the moldy grain recognition rate and the stone misidentification rate on the validation set no longer increases, the system rolls back to the optimal weight state.

10. The method for dynamic grain quality grading and stone removal coordinated control based on multimodal sensor fusion according to claim 1, characterized in that: Specific methods for increasing air velocity as the concentration of moldy particles increases include: Nozzle array zoning control: high-pressure airflow nozzles are arranged equidistantly along the width of the conveyor belt, with each nozzle covering a fixed area; Aggregation distribution calculation: Real-time statistics are collected on the percentage of moldy particles in the nozzle partitions, and the standard deviation of the percentage of each partition and the average value is calculated as the local aggregation degree; Spray strategy generation: For the zones where the local concentration exceeds the limit, the corresponding nozzles are opened, and the spray angle is dynamically adjusted according to the conveyor belt speed: when the speed is high, the spray angle is reduced to extend the action time, and when the speed is low, the angle is increased to expand the coverage area; Spray parameter optimization: With the goal of uniform distribution of moldy particles, the optimal spray duration and pressure combination is determined through feedback adjustment.

Images