Production motor operation frequency prediction model training method

Abstract

The invention discloses a production motor operation frequency prediction model training method, which comprises the following steps: acquiring a tobacco image group, extracting tobacco image features by using a three-dimensional point cloud model, converting the tobacco image features into a tobacco cross section area change sequence, dividing production stages, and screening out pre-selected stage information of which the fluctuation ratio meets requirements; synchronously acquiring motor operation information and matching to obtain pre-selected motor information and operation frequency thereof; a Gaussian process regression algorithm is adopted to construct a prediction model, trend features and local fluctuation features are obtained through multi-scale feature extraction to serve as input, the preselected operation frequency serves as an output label for training, and hyper-parameters are optimized to obtain a final prediction model; the established dynamic confidence interval evaluation mechanism can realize model increment training. The operation frequency of the motor is dynamically adjusted according to the actual state of the tobacco leaves, and the drying quality and the batch stability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a method for training a prediction model for the operating frequency of a motor used in production. Background Art

[0002] In the field of tobacco processing, the primary processing of tobacco leaves usually includes processes such as baking, sun-drying, and sorting. Among them, the baking process has a decisive impact on the color, aroma, and combustibility of the final product. In traditional processes, tobacco leaves are moved uniformly in a drying device through a conveyor belt, and the drying temperature and time are controlled by manual experience or fixed programs. However, due to natural characteristic differences such as uneven thickness and fluctuating moisture content during the conveying process of tobacco leaves, the existing regulation methods based on fixed parameters are difficult to achieve precise drying, easily leading to unstable batch quality. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to propose a method for training a prediction model for the operating frequency of a motor used in production, so as to solve the problem that the adjustment of the conveying motor in existing tobacco processing cannot dynamically adapt to the morphological changes of tobacco leaves, resulting in unstable drying quality.

[0004] To achieve the above technical purpose, the technical solution adopted in the present application is: A method for training a prediction model for the operating frequency of a motor used in production, including: Collect a group of tobacco images at a preset frequency. The group of tobacco images includes multiple frames of continuous tobacco image information; Input the group of tobacco images into a three-dimensional point cloud model for image processing to obtain the tobacco image features of a preset area; Convert the tobacco image features into a sequence of changes in the cross-sectional area of tobacco, divide the tobacco production stages according to the sequence of changes in the cross-sectional area of tobacco, and segment the tobacco image features according to the tobacco production stages to obtain multiple pieces of production stage information; Extract the cross-sectional volatility of the production stage information, and extract the production stage information with the cross-sectional volatility within a preset volatility threshold range, denoted as preselected stage information; Synchronously collect the motor operation information when collecting the group of tobacco images, and record the motor operation information within the time period to which the preselected stage information belongs as preselected motor information; Obtain the motor operating frequency in the preselected motor information, denoted as the preselected operating frequency, and obtain the sequence of changes in the cross-sectional area of tobacco in the preselected stage information, denoted as the preselected area change sequence; Fit the preselected operating frequency with the preselected area change sequence to construct a basic prediction model, and use the Gaussian process regression algorithm to train the model of the basic prediction model, including: Set the radial basis function as the kernel function; Extract multi-scale features from the preselected area change sequence to obtain the trend features and local fluctuation features of the area change; Use the trend features and local fluctuation features as input features together, and the preselected operating frequency as the output label to train the basic prediction model; Optimize the hyperparameters in the Gaussian process by maximizing the marginal likelihood function, and update the parameter information of the basic prediction model to obtain the final prediction model; Establish a dynamic confidence interval evaluation mechanism, and trigger the incremental training of the current final prediction model when the prediction result of the new input data exceeds the confidence interval.

[0005] In some embodiments, the tobacco image information includes depth image information and color image information. Collecting the tobacco image group according to the preset frequency includes: Collect the depth frame and color frame of the current tobacco through a depth camera, convert the depth frame into a depth matrix, and convert the color frame into a color matrix; Obtain the hardware timestamp of the depth camera, and perform time alignment on the depth matrix and the color matrix, including: Adopt an adaptive time window algorithm to dynamically adjust the frame synchronization threshold, and optimize the time alignment accuracy in real time according to the tobacco movement speed; Construct the coordinate mapping relationship between the depth matrix and the color matrix, which is represented by formula (1). Formula (1) is as follows: ; In formula (1), is the pixel coordinate in the depth matrix, is the pixel coordinate in the color matrix, is the internal parameter matrix of the depth camera, including the focal length and the principal point parameters, is the internal parameter matrix of the color camera, is the inverse matrix of the internal parameter matrix of the color camera; And, during the construction of the coordinate mapping relationship, it also includes: Calculate the confidence degree of the main body area of the tobacco through the depth matrix, which is represented by formula (2). Formula (2) is as follows: ; In formula (2), represents the pixel coordinate position in the depth matrix, represents the coordinate at the confidence degree of the main body area, is the natural logarithm, is the adjustment factor, represents the depth measurement value of the depth matrix at the coordinate ; is the depth threshold; Based on the confidence of the main body region, an edge-sensitive weight function is established, which is represented by formula (3). Formula (3) is as follows: ; In formula (3), is the balance coefficient, is the amplitude of the depth gradient, representing the amplitude of the depth gradient of the depth matrix at the coordinate ; is the edge-sensitive weight; According to the edge-sensitive weight function, the edge-sensitive weights of different mapping regions in the coordinate mapping process are obtained. Different interpolation algorithms are used for interpolation compensation in regions with different edge-sensitive weights to obtain an optimized coordinate mapping relationship, denoted as the final mapping relationship. The interpolation algorithms include at least two of cubic interpolation, linear interpolation, and nearest neighbor interpolation; The final mapping relationship is represented by formula (4). Formula (4) is as follows: ; In formula (4), is the number of pre-calibrated non-linear correction basis functions, represents the non-linear correction basis function constructed based on the matching residuals of feature points between the depth camera and the color camera, is the dynamic weight coefficient.

[0006] In some embodiments, the tobacco image group is input into the 3D point cloud model for image processing. The obtained tobacco image features in the preset region include: Converting the depth image information in the tobacco image group into 3D point cloud data includes: Calculating the 3D space coordinates through the internal parameter matrix of the depth camera; Performing coordinate transformation on each pixel point in the depth image information to generate the original point cloud data containing spatial position information; Performing dynamic region segmentation on the original point cloud data includes: Obtaining the real-time movement speed of the conveyor belt in the motor operation information to generate the sliding window range of the current processing period; Extracting the point cloud subset located within the sliding window range from the original point cloud data as the processing object; Performing feature enhancement processing on the point cloud subset includes: Calculating the point cloud features of each point in the point cloud subset. The point cloud features include the normal vector of each point and the curvature descriptor of the local surface characteristics; According to the point cloud features and the color image information in the tobacco image group, performing color space segmentation on the point cloud subset to identify the point cloud data corresponding to the tobacco main body region, denoted as tobacco point cloud data; Perform multi-modal feature fusion on tobacco point cloud data, including: Perform feature-level fusion of curvature descriptors and color space segmentation results; Perform principal component analysis on the fused multi-dimensional feature data and reduce the dimension to obtain key feature vectors; Perform spatio-temporal consistency verification on the key feature vectors, including: Establish the correspondence between the tobacco point cloud data of the current frame and the tobacco point cloud data of the previous frame; Optimize the correspondence through the iterative closest point algorithm to ensure the alignment accuracy of point clouds in time series; Perform compression encoding on the verified key feature vectors, including: Use an octree structure to partition the tobacco point cloud data in space; Extract the point cloud statistical features within each octree node as the tobacco image features of the final output.

[0007] In some embodiments, convert the tobacco image features into a sequence of changes in the cross-sectional area of tobacco, divide the tobacco production stages according to the sequence of changes in the cross-sectional area of tobacco, and perform data segmentation on the tobacco image features according to the tobacco production stages to obtain multiple production stage information, including: Convert the three-dimensional space coordinate data in the tobacco image features into cross-sectional contour data, including: Intercept the three-dimensional space coordinate data through a preset detection plane to obtain a cross-sectional point set; Perform height-direction filtering on the cross-sectional point set to remove data points outside the set height range; Perform contour reconstruction on the filtered cross-sectional point set, including: Project the cross-sectional point set onto a two-dimensional plane to obtain multiple projection points; Connect the projection points to form an initial contour polygon; Smooth the initial contour polygon to obtain the final contour line; Calculate the cross-sectional area at each time point, including: Apply the area calculation formula to the final contour line to generate an area sequence; Perform differential calculation on the area values at adjacent time points to obtain an area change rate sequence; Divide the production stages according to the area change rate sequence, including: Set the determination thresholds for each stage, including the starting stage threshold, the rising stage threshold, the stable stage threshold, and the falling stage threshold; When the area change rate continuously exceeds the threshold of the corresponding stage, mark the stage conversion time point; Perform segmentation on the tobacco image features according to the production stages, including: Divide the data time period according to the stage conversion time point; Extract the tobacco image features in each time period and add phase labels; Establish a data overlapping window for the phase transition region to maintain feature continuity.

[0008] In some embodiments, extract the cross-sectional volatility of the production stage information, and extract the production stage information whose cross-sectional volatility is within a preset volatility threshold range, which is recorded as the preselected stage information, including: Calculate the cross-sectional volatility within each production stage, including: Obtain the sequence of changes in the cross-sectional area of tobacco in the production stage information; Perform a sliding window process on the area change sequence, calculate the area volatility within each window, which is represented by formula (5), and the formula (5) is as follows: ; In formula (5), is the cross-sectional area volatility at time , is the length of the sliding window, represents the degrees of freedom adjustment term, is the cross-sectional area at the th time point, is the average area within the window; Determine the preset volatility threshold range, including: Statistically analyze the volatility distribution characteristics of the historical normal production stage, and set the upper and lower volatility threshold values; Screen the preselected stage information, including: Compare the area volatility within each production stage with the preset volatility threshold range, and record the data segment corresponding to the area volatility within the preset volatility threshold range as the valid data segment; Merge the continuous valid data segments to generate the preselected stage information.

[0009] In some embodiments, synchronously collect the motor operation information when collecting the tobacco image group, and record the motor operation information within the time period to which the preselected stage information belongs as the preselected motor information, including: Read the operation parameters output by the motor controller in real time through the industrial bus interface to generate the motor operation information, and the operation parameters include the operation frequency, current value, and rotation speed; Align the motor operation information with the tobacco image group in time, and perform interpolation compensation on the motor operation information whose time deviation exceeds the preset threshold; Screen the motor operation information according to the time period of the preselected stage information, including: Determine the start time point and end time point of the preselected stage information; Extract data within the time period corresponding to the preselected stage information from the motor operation information after time alignment; Eliminate abnormal data that does not meet the operation parameter thresholds within the time period; Mark the filtered motor operation information as preselected motor information, including: Add the corresponding preselected stage information identifier to the preselected motor information; Establish an association mapping relationship between the preselected motor information and the preselected stage information.

[0010] In some embodiments, obtain the tobacco cross-sectional area change sequence of the preselected stage information, denoted as the preselected area change sequence, including: Obtain the cross-sectional area data at each time point in the preselected stage information; Perform a smoothing filter process on the cross-sectional area data; Verify the continuity and integrity of the cross-sectional area data to obtain the tobacco cross-sectional area change sequence; Add the preselected stage identifier to the tobacco cross-sectional area change sequence to obtain the preselected area change sequence.

[0011] In some embodiments, set the radial basis function as the kernel function, which is represented by formula (6). Formula (6) is as follows: ; Where, is the signal variance parameter, is the length scale parameter, is the noise variance parameter, is the Kronecker delta function, is the th input feature vector, is the th input feature vector; Perform multi-scale feature extraction on the preselected area change sequence to obtain the trend feature and local fluctuation feature of the area change, which is represented by formula (7). Formula (7) is as follows: ; Where, is the trend feature at time , is the fluctuation feature at time , is the cross-sectional area at the th time point, is the dynamic window size.

[0012] In some embodiments, use the trend feature and the local fluctuation feature as input features together, and use the preselected operating frequency as the output label to train the basic prediction model, including: Construct a multi-scale feature fusion input vector, including: Standardize the trend features to obtain a standardized trend feature vector; Normalize the local fluctuation features to obtain a normalized fluctuation feature vector; Generate combined feature terms from the standardized trend feature vector and the normalized fluctuation feature vector using the feature cross method to form a fusion feature matrix; Establish a training sample set, including: Slice the fusion feature matrix into overlapping sample blocks according to the time series; Label each overlapping sample block with the corresponding preselected median operating frequency; Construct a time series cross-validation set and a training sample set; Use the training sample set to execute the model training process and monitor the model training effect on the time series cross-validation set, including: Calculate the mean squared error between the predicted frequency and the actual frequency; Verify the prediction consistency of the basic prediction model in the stable period and the non-stable period.

[0013] In some embodiments, optimize the hyperparameters in the Gaussian process by maximizing the marginal likelihood function and update the parameter information of the basic prediction model to obtain the final prediction model, including: Establish a Gaussian process marginal likelihood function, including: Calculate the kernel function matrix based on the current hyperparameter combination, construct an expression of the marginal likelihood function including the noise term, and initialize the hyperparameter search space and optimization constraints; Execute the hyperparameter optimization process, including: Iteratively optimize the marginal likelihood function using the conjugate gradient method; Update the length scale parameter and the signal variance parameter in each iteration; Monitor the change trend and convergence state of the likelihood function value; Verify the hyperparameter optimization result, including: Check the physical rationality of the optimized hyperparameters; Verify the positive definiteness of the kernel function matrix; Evaluate the stability of the gradient descent process; Update the basic prediction model parameters, including: Inject the optimized hyperparameters into the Gaussian process model; Recalculate the posterior distribution of the training data; Update the memory representation and prediction interface of the model; Generate the final prediction model, including: Solidify the optimized model parameters and structure; Save the metadata required for model deployment; Establish a model version management and rollback mechanism.

[0014] Adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are as follows: The above technical solution provides a method for training a prediction model for the operating frequency of a production motor. By collecting a group of tobacco images and using a three-dimensional point cloud model to extract the features of the tobacco images, converting them into a sequence of changes in the cross-sectional area of the tobacco, dividing the production stage, and screening out the preselected stage information with a volatility meeting the requirements; synchronously collecting the motor operating information and matching to obtain the preselected motor information and its operating frequency; using the Gaussian process regression algorithm to construct a prediction model, obtaining trend features and local fluctuation features as inputs through multi-scale feature extraction, using the preselected operating frequency as the output label for training, and optimizing the hyperparameters to obtain the final prediction model; the established dynamic confidence interval evaluation mechanism can realize incremental training of the model. The above technical solution dynamically adjusts the operating frequency of the motor according to the actual state of the tobacco leaves, significantly improving the drying quality and batch stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0016] Figure 1 It is a method step diagram of steps S101 to S107 of the training method described in the specific implementation manner; Figure 2 It is a method step diagram of steps S201 to S206 of the training method described in the specific implementation manner. SPECIFIC IMPLEMENTATION MANNER

[0017] The following will further describe the present invention in detail in conjunction with the drawings and embodiments. It should be specifically noted that the following embodiments are only used to illustrate the present invention, but do not limit the scope of the present invention. Similarly, the following embodiments are only some embodiments of the present invention rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0018] Please refer to Figure 1 , this embodiment provides a method for training a prediction model for the operating frequency of a production motor, including: S101. Collect a group of tobacco images according to a preset frequency. The group of tobacco images includes information of multiple consecutive tobacco images; S102. Input the tobacco image group into the three-dimensional point cloud model for image processing to obtain the tobacco image features in the preset area; S103. Convert the tobacco image features into a sequence of changes in the cross-sectional area of tobacco, divide the tobacco production stages according to the sequence of changes in the cross-sectional area of tobacco, and perform data segmentation on the tobacco image features according to the tobacco production stages to obtain multiple production stage information; S104. Extract the cross-sectional volatility of the production stage information, and extract the production stage information whose cross-sectional volatility is within the preset volatility threshold range, denoted as the preselected stage information; S105. Synchronously collect the motor operation information when collecting the tobacco image group, and record the motor operation information within the time period to which the preselected stage information belongs as the preselected motor information; S106. Obtain the motor operation frequency in the preselected motor information, denoted as the preselected operation frequency, and obtain the sequence of changes in the cross-sectional area of tobacco for the preselected stage information, denoted as the preselected area change sequence; S107. Fit the preselected operation frequency with the preselected area change sequence to construct a basic prediction model, and use the Gaussian process regression algorithm to train the model of the basic prediction model, including: Set the radial basis function as the kernel function; Perform multi-scale feature extraction on the preselected area change sequence to obtain the trend feature and local fluctuation feature of the area change; Use the trend feature and local fluctuation feature as input features and the preselected operation frequency as the output label to train the basic prediction model; Optimize the hyperparameters in the Gaussian process by maximizing the marginal likelihood function, and update the parameter information of the basic prediction model to obtain the final prediction model; Establish a dynamic confidence interval evaluation mechanism, and trigger the incremental training of the current final prediction model when the prediction result of the newly input data exceeds the confidence interval.

[0019] In step S101, the determination of the preset frequency is based on the comprehensive consideration of the conveyor belt running speed and the tobacco processing technology requirements to ensure that the collected tobacco image group can completely record the continuous state changes of the tobacco leaves during the conveying process. Preferably, when the conveyor belt runs at a speed of 0.3 m / s, a 1-second acquisition interval can ensure that adjacent images have a 30% overlapping area, and the obtained continuous tobacco image group provides a necessary data basis for subsequent three-dimensional reconstruction.

[0020] In step S102, use the three-dimensional point cloud model to process the tobacco image group. Preferably, calculate the three-dimensional coordinates of the preset monitoring area through the stereo vision algorithm, reconstruct the three-dimensional topography of the tobacco surface, and the extracted tobacco image features include key parameters such as spatial curvature, which directly serve the subsequent cross-sectional area calculation.

[0021] In step S103, the tobacco cross-sectional area change sequence is obtained by calculating three-dimensional feature data. Preferably, it is implemented using an equidistant slicing algorithm. The slope change obtained by differentiating the tobacco cross-sectional area change sequence data is used as the basis for dividing the production stage. When it is detected that the slope change exceeds a preset slope threshold, the tobacco production stage conversion point is automatically marked, and multiple production stage information is obtained. Among them, the preset slope threshold is determined by analyzing the area change rate characteristics of high-quality batches in historical production data, and the critical change rate that can significantly distinguish different process stages is selected as the threshold.

[0022] In step S104, the cross-sectional area dispersion is calculated within the divided tobacco production stage, and the stable stage with the volatility maintained within the preset fluctuation threshold range is selected as the preselected stage information. Preferably, the determination of the preset fluctuation threshold comes from the statistical analysis results of historical high-quality batch data.

[0023] In step S105, the acquisition of motor operation information and image acquisition are synchronized through a hardware timekeeping device to ensure high-precision time alignment, and the preselected motor information is obtained.

[0024] In step S106, the preselected motor information is accurately matched with the corresponding preselected stage information. Preferably, the motor operation frequency in the preselected motor information is passed through an IIR filter to eliminate high-frequency interference; the tobacco cross-sectional area change sequence of the preselected stage information is normalized to eliminate batch differences, and after being standardized, it provides reliable input for subsequent model training.

[0025] In step S107, the Gaussian process regression algorithm is used to construct the basic prediction model. The initial parameters of the radial basis function kernel can be set with reference to the typical fluctuation range of the area sequence. The multi-scale feature extraction process simultaneously obtains the trend feature and the local fluctuation feature. The trend feature reflects the overall change trend within the stage, and the local fluctuation feature captures instantaneous abnormal situations. The dynamic confidence interval evaluation mechanism continuously monitors the prediction deviation, and when the situation of continuously exceeding the threshold occurs, the model update is automatically triggered to ensure that the prediction system has the ability of continuous optimization.

[0026] The implementation principle of this embodiment can be understood as: establishing a non-linear mapping relationship between the dynamic changes of the tobacco physical form and the motor control parameters, capturing the key control features of different process stages through multi-scale feature extraction, using the probability characteristics of the Gaussian process to realize the credibility evaluation of the prediction results, realizing the closed-loop predictive control from raw material state perception to control parameter optimization, and ensuring that the process parameters are dynamically adapted to the changes in the raw material state. For example, when it is detected that the sudden change in the moisture content of the tobacco leaves causes an increase in the cross-sectional area fluctuation, the model automatically adjusts the motor frequency to extend the drying time of this batch, and at the same time updates the corresponding relationship between the fluctuation feature and the frequency through incremental learning.

[0027] In this embodiment, by establishing the mapping relationship between the sequence of changes in the cross-sectional area of tobacco and the operating frequency of the motor, a prediction model is constructed using multi-scale feature extraction and Gaussian process regression algorithm to achieve the precise correlation between the dynamic changes in the physical form of tobacco and the motor control parameters. Through the dynamic confidence interval evaluation mechanism, the continuous optimization of the model is ensured, enabling the operating frequency of the motor to adapt to the state changes in the tobacco production stage and effectively improving the control accuracy and stability of the tobacco processing process.

[0028] In some embodiments, the tobacco image information includes depth image information and color image information. Collecting the tobacco image group according to a preset frequency includes: Collecting the depth frame and color frame of the current tobacco through a depth camera, converting the depth frame into a depth matrix, and converting the color frame into a color matrix; Obtaining the hardware timestamp of the depth camera and aligning the time of the depth matrix and the color matrix, including: Adopting an adaptive time window algorithm to dynamically adjust the frame synchronization threshold and optimizing the time alignment accuracy in real time according to the tobacco movement speed; Constructing the coordinate mapping relationship between the depth matrix and the color matrix, which is represented by formula (1). Formula (1) is as follows: ; In formula (1), is the pixel coordinate in the depth matrix, is the pixel coordinate in the color matrix, is the internal parameter matrix of the depth camera, including the focal length and the principal point parameters, is the internal parameter matrix of the color camera, is the inverse matrix of the internal parameter matrix of the color camera; In addition, during the construction of the coordinate mapping relationship, it also includes: Calculating the confidence degree of the main body area of tobacco through the depth matrix, which is represented by formula (2). Formula (2) is as follows: ; In formula (2), represents the pixel coordinate position in the depth matrix, where represents the index of the pixel in the image width direction, represents the index of the pixel in the height direction; represents the coordinate at the main body area confidence degree at this point, and the output value is between 0 and 1, representing the possibility degree that this pixel point belongs to the main body area of tobacco; is the natural logarithm; is the adjustment factor; represents the depth matrix at the coordinate The depth measurement value at [location], characterizing the vertical distance between the actual spatial position corresponding to this pixel point and the camera; is the depth threshold. When is much greater than , approaches 0, approaches 1; when is much less than , approaches , approaches 0; a smooth transition region is formed near the depth threshold , and the oversteepness is controlled by the adjustment factor ; Based on the confidence of the main region, an edge-sensitive weight function is established, expressed by formula (3), and formula (3) is as follows: ; In formula (3), is the balance coefficient, used to adjust the weight ratio between the main confidence and the edge feature; is the amplitude of the depth gradient, representing the amplitude of the depth gradient of the depth matrix at the coordinate , characterizing the severity of the depth change in the surrounding area of this point; is the edge-sensitive weight, comprehensively considering the main body attributes and edge features of the region, and used to guide the selection of the interpolation strategy for the subsequent coordinate mapping; According to the edge-sensitive weight function, the edge-sensitive weights of different mapping regions in the coordinate mapping process are obtained, and different interpolation algorithms are used for interpolation compensation in regions with different edge-sensitive weights to obtain an optimized coordinate mapping relationship, denoted as the final mapping relationship. The interpolation algorithms include at least two of cubic interpolation, linear interpolation, and nearest neighbor interpolation; The final mapping relationship is expressed by formula (4), and formula (4) is as follows: ; In formula (4), is the number of pre-calibrated non-linear correction basis functions, represents the non-linear correction basis function constructed based on the residual error of the feature point matching between the depth camera and the color camera (i.e., the deviation between the actual coordinate and the linear mapping coordinate), used to compensate for lens distortion and assembly errors, and thus ensure the accuracy of the 3D point cloud reconstruction, is the dynamic weight coefficient.

[0029] In this embodiment, the depth image information obtains the scene depth data through the time-of-flight principle, and the color image information records the tobacco surface features in the standard RGB three-channel format. Preferably, the depth frame is converted into a regular depth matrix after being corrected by the camera calibration parameters, and the color frame generates a standardized color matrix after gamma correction and white balance processing to ensure the basic consistency of the multi-modal data.

[0030] The dynamic adjustment of the frame synchronization threshold by using the adaptive time window algorithm includes the following steps: Obtain the displacement of the tobacco surface feature points in the continuous depth frames through the feature point tracking algorithm, and calculate the average displacement per unit time as the estimated value of the tobacco movement speed. The feature points are selected from the regions with significant tobacco textures to ensure tracking stability; Establish the mapping relationship between the movement speed and the frame synchronization threshold. When the tobacco movement speed increases, the frame synchronization threshold is reduced proportionally; when the movement speed decreases, the frame synchronization threshold is relaxed accordingly. The proportional relationship is determined through pre-experimental calibration to ensure the best alignment accuracy when the speed changes; Adopt the sliding time window mechanism to maintain the current motion state. The length of the time window is adaptively adjusted according to the speed change rate: for the uniform motion state, maintain a fixed-length time window; when acceleration or deceleration is detected, automatically shorten the time window length to improve the response speed; Judge the time stamp difference between the depth frame and the color frame according to the frame synchronization threshold calculated in real time: when the time stamp difference exceeds the current threshold, use the linear interpolation algorithm to compensate the earlier acquired frame to ensure the time alignment accuracy between the depth matrix and the color matrix.

[0031] Construct the coordinate mapping relationship between the depth matrix and the color matrix, and use the bidirectional nearest neighbor search algorithm to process the pixel matching in the occluded area. The forward matching gives priority to depth continuity, and the reverse verification focuses on maintaining texture consistency. The regions with bidirectional matching differences are marked as special processing regions.

[0032] In formula (1), the depth camera intrinsic matrix and the color camera intrinsic matrix can be pre-calibrated and obtained through the Zhang Zhengyou calibration method.

[0033] In formula (2), the depth threshold takes the median of the effective working distance above the conveyor belt plane, and the adjustment factor controls the steepness of the confidence change.

[0034] In formula (3), the balance coefficient is determined according to the statistical distribution of the depth gradient amplitude to ensure that while retaining the main structure, the detailed features are not lost.

[0035] The edge-sensitive weight function dynamically adjusts the weight distribution through the pre-trained balance coefficient, enhancing the detail retention in the tobacco edge area and optimizing the calculation efficiency in the flat area. The final mapping relationship can be achieved by the thin-plate spline interpolation algorithm for spatial alignment to ensure the accuracy of 3D feature extraction.

[0036] In this embodiment, by establishing an accurate spatio-temporal alignment mechanism and an optimized coordinate mapping relationship, the consistency of multi-modal image data in 3D reconstruction is ensured. For example, when there is a depth mutation at the tobacco edge, the edge-sensitive weight function can automatically enhance the mapping accuracy of this area, avoiding feature distortion caused by depth discontinuity.

[0037] In this embodiment, the accurate time alignment between the depth matrix and the color matrix is achieved through the adaptive time window algorithm. Combining the optimized coordinate mapping relationship and the edge-sensitive weight function, the synchronous matching problem of multi-modal image data in the tobacco motion state is effectively solved. Further, the bidirectional nearest neighbor search algorithm is used to process the occlusion area matching to ensure the feature consistency between the depth information and the color texture, providing an accurate and reliable data basis for subsequent 3D reconstruction and significantly improving the integrity and accuracy of tobacco physical shape feature extraction.

[0038] Please refer to Figure 2 , in some embodiments, inputting the tobacco image group into the 3D point cloud model for image processing, the obtained tobacco image features in the preset area include: S201. Converting the depth image information in the tobacco image group into 3D point cloud data, including: Calculating the 3D space coordinates through the internal parameter matrix of the depth camera; Performing coordinate transformation on each pixel point in the depth image information to generate the original point cloud data containing spatial position information; S202. Dynamically segmenting the original point cloud data, including: Obtaining the real-time movement speed of the conveyor belt in the motor operation information to generate the sliding window range of the current processing period; Extracting the point cloud subset located within the sliding window range from the original point cloud data as the processing object; S203. Performing feature enhancement processing on the point cloud subset, including: Calculating the point cloud features of each point in the point cloud subset, where the point cloud features include the normal vector of each point and the curvature descriptor of the local surface characteristics; According to the point cloud features and the color image information in the tobacco image group, performing color space segmentation on the point cloud subset to identify the point cloud data corresponding to the tobacco main area, denoted as tobacco point cloud data; S204. Performing multi-modal feature fusion on the tobacco point cloud data, including: Perform feature-level fusion of the curvature descriptor and the color space segmentation result; Perform principal component analysis on the fused multi-dimensional feature data to reduce the dimension and obtain the key feature vector; S205. Perform spatio-temporal consistency verification on the key feature vector, including: Establish the correspondence between the tobacco point cloud data of the current frame and the tobacco point cloud data of the previous frame; Optimize the correspondence through the iterative closest point algorithm to ensure the alignment accuracy of the point cloud in time series; S206. Compress and encode the verified key feature vector, including: Use the octree structure to perform spatial partitioning on the tobacco point cloud data; Extract the point cloud statistical features within each octree node as the tobacco image features of the final output.

[0039] In step S201, the process of converting the depth image information into three-dimensional point cloud data is realized through perspective projection transformation, and the internal parameter matrix of the depth camera is used to convert the pixel coordinate system to the camera coordinate system. When calculating the three-dimensional spatial coordinates of each pixel point, the depth value and the focal length parameter in the internal parameter matrix are combined to complete the coordinate transformation. The generated original point cloud data contains spatial position information, ensuring that each depth pixel can be accurately mapped to the three-dimensional space.

[0040] In step S202, the dynamic area segmentation is realized through the sliding window mechanism, and the real-time movement speed of the conveyor belt is used to determine the spatio-temporal range of the window. The calculation of the sliding window range takes into account the conveyor belt movement speed and the point cloud processing delay to ensure that the extracted point cloud subset completely covers the tobacco area in the current processing period. The discrete points outside the window range in the original point cloud data will be automatically filtered, and the effective processing objects will be retained.

[0041] In step S203, preferably, the point cloud feature calculation adopts the local surface fitting method based on k-nearest neighbors. The normal vector is obtained by least squares plane fitting of the neighborhood points, and the curvature descriptor characterizes the concave-convex characteristics of the local surface. Preferably, the color space segmentation combines the hue component of the HSV color model and the point cloud curvature feature, and identifies the main tobacco area through the region growing algorithm to exclude the interference point cloud such as the conveyor belt background.

[0042] In step S204, the multi-modal feature fusion jointly optimizes the geometric feature and the color feature in the feature space. The curvature descriptor and the color segmentation result form a high-dimensional feature vector through feature splicing. Preferably, the principal component analysis retains the top three principal components with the highest contribution rate, and the key feature vector after dimension reduction contains both shape and color information.

[0043] In step S205, the spatio-temporal consistency check is achieved by establishing the correspondence relationship of point clouds between consecutive frames. When optimizing the iterative closest point algorithm, the point-to-plane distance metric is adopted to ensure the shape continuity of the tobacco point clouds in adjacent frames during the movement process, thereby effectively compensating for the position deviation of the point clouds caused by the conveyor belt vibration.

[0044] In step S206, the octree encoding adaptively divides the space according to the spatial distribution characteristics of the point cloud. Preferably, the statistical features of the point cloud for each node include descriptors such as density, normal vector distribution, and mean curvature, which not only retain the global structural features of the point cloud but also extract local detail information.

[0045] In this embodiment, the accurate extraction of the three-dimensional features of tobacco is realized through the systematic point cloud processing flow. For example, when processing tobacco leaves with high moisture content, color space segmentation can effectively distinguish the wet areas, while the curvature feature can identify the curling shape of the leaves. The feature vector formed after multi-modal fusion comprehensively characterizes the physical state of the tobacco.

[0046] In this embodiment, the accurate extraction of tobacco point cloud data is realized through dynamic region segmentation and feature enhancement processing. Combining multi-modal feature fusion and spatio-temporal consistency check to construct three-dimensional features that comprehensively characterize the physical state of tobacco. The tobacco image features compressed by octree encoding not only retain the global structural information but also contain local detail features, providing an accurate and reliable three-dimensional morphological basis for the prediction of subsequent motor control parameters and significantly improving the state perception accuracy of the tobacco processing process.

[0047] In some embodiments, the tobacco image features are converted into a sequence of changes in the cross-sectional area of the tobacco. The tobacco production stages are divided according to the sequence of changes in the cross-sectional area of the tobacco, and the tobacco image features are sliced according to the tobacco production stages to obtain multiple production stage information, including: Converting the three-dimensional space coordinate data in the tobacco image features into cross-sectional contour data, including: Intercepting the three-dimensional space coordinate data through a preset detection plane to obtain a cross-sectional point set; Filtering the cross-sectional point set in the height direction to remove the data points beyond the set height range; Performing contour reconstruction on the filtered cross-sectional point set, including: Projecting the cross-sectional point set onto a two-dimensional plane to obtain multiple projection points; Connecting the projection points to form an initial contour polygon; Smoothing the initial contour polygon to obtain the final contour line; Calculating the cross-sectional area at each time point, including: Applying the area calculation formula to the final contour line to generate an area sequence; Performing a difference calculation on the area values at adjacent time points to obtain a sequence of area change rates; Dividing the production stages according to the area change rate sequence, including: Setting the determination thresholds for each stage, including the start-up threshold, the rising threshold, the stable threshold, and the declining threshold; When the area change rate continuously exceeds the threshold of the corresponding stage, marking the stage conversion time point; Segmenting the tobacco image features according to the production stages, including: Dividing the data time periods according to the stage conversion time points; Extracting the tobacco image features within each time period and adding stage labels; Establishing a data overlapping window in the stage transition area to maintain feature continuity.

[0048] In this embodiment, the process of converting the three-dimensional space coordinate data into cross-sectional profile data is realized through spatial geometric transformation.

[0049] The preset detection plane is set parallel to the movement direction of the conveyor belt, and its height position is dynamically adjusted according to the tobacco stacking thickness. Preferably, the acquisition of the cross-sectional point set adopts a spatial indexing acceleration algorithm, and only the three-dimensional point cloud data intersecting with the detection plane is retained; the height-direction filtering eliminates abnormal outlier points by setting upper and lower threshold values to ensure the accuracy of the contour reconstruction.

[0050] During the contour reconstruction process, first project the cross-sectional point set onto a two-dimensional plane perpendicular to the detection plane, and the projection points can construct the initial topological relationship through the Delaunay triangulation algorithm. Preferably, the initial contour polygon extracts the outer contour features through the α-shape algorithm, and then adopts the curve fitting method based on B-spline for smoothing processing, eliminating the measurement noise while retaining the feature inflection points, and finally obtaining a smooth and continuous final contour line.

[0051] When calculating the cross-sectional area at each time point, applying the area calculation formula to the final contour line to generate an area sequence, and the area calculation formula adopts an improved contour integral method, including: Performing uniform resampling on the point set on the final contour line; Calculating the signed area of the resampled polygon region; Introducing an edge compensation coefficient to correct the measurement error.

[0052] Preferably, the improved contour integral method ensures the consistent distribution density of contour points through uniform resampling; the signed area calculation can be realized by the Green's formula for efficient integration; the edge compensation coefficient is dynamically adjusted according to the local curvature of the contour line to correct the area calculation error caused by insufficient sampling; the area difference calculation between adjacent time points adopts the central difference method to effectively suppress the influence of measurement noise on the change rate calculation.

[0053] When dividing the production stages according to the sequence of area change rates, set the decision thresholds for each stage, including the start-up period threshold, the rising period threshold, the stable period threshold, and the declining period threshold. The determination of the stage decision thresholds includes: Statistically analyze the typical area change rate ranges for each stage in the historical production data; Use moving window variance analysis to dynamically adjust the threshold boundaries; When the environmental parameters change beyond the set range, trigger the threshold recalibration.

[0054] Among them, the setting of the stage decision thresholds is based on the statistical analysis of historical production data. The start-up period threshold, the rising period threshold, the stable period threshold, and the declining period threshold respectively correspond to the typical change characteristics of different process stages. The moving window variance analysis monitors the production status fluctuations in real time. When significant changes in parameters such as environmental temperature and humidity are detected, it automatically triggers the threshold recalibration process to ensure the accuracy of stage division.

[0055] When segmenting the tobacco image features according to the production stages, preferably, the data overlap window in the stage transition area is set as a cross area where each adjacent stage accounts for 50%, and the feature smooth transition is achieved through linear weighted fusion. In addition to adding stage labels to the tobacco image features in each time period, the corresponding process parameters and environmental status are also recorded to provide complete information for subsequent analysis.

[0056] The steps of this embodiment can be understood as: achieving a reliable division of the tobacco production stages through precise geometric calculations and adaptive threshold adjustments. For example, during the tobacco leaf rewetting process, when the area change rate continuously exceeds the rising period threshold, the system accurately identifies the entry into the rapid moisture absorption stage and automatically adjusts the subsequent processing parameters.

[0057] This embodiment realizes the accurate calculation of the tobacco cross-sectional area change sequence and the reliable division of the production stages through an improved contour integration method and a dynamic threshold adjustment mechanism. Based on the cross-sectional contour data obtained from the three-dimensional space coordinate data conversion and contour reconstruction, combined with the adaptive stage decision threshold and transition area processing, it ensures the accuracy of segmenting the tobacco image features according to the production stages and provides a reliable basis for state recognition for process parameter optimization.

[0058] In some embodiments, extract the cross-sectional volatility of the production stage information, and extract the production stage information where the cross-sectional volatility is within the preset volatility threshold range, denoted as the preselected stage information, including: Calculate the cross-sectional volatility within each production stage, including: Obtain the tobacco cross-sectional area change sequence in the production stage information; Perform a moving window process on the area change sequence, calculate the area volatility within each window, which is represented by formula (5), and formula (5) is as follows: ; In formula (5), is the cross-sectional area volatility at time , is the length of the sliding window (i.e., the total number of time points), represents the degrees of freedom adjustment term, which is used for unbiased estimation of volatility to avoid the statistical bias caused by the number of samples within the window, is the th cross-sectional area at the th time point, and is the average area within the window; Determine the preset volatility threshold range, including: Statistically analyze the volatility distribution characteristics of the historical normal production stage, and set the upper and lower volatility threshold values; Screen the preselected stage information, including: Compare the area volatility within each production stage with the preset volatility threshold range, and record the data segment corresponding to the area volatility within the preset volatility threshold range as the valid data segment;

[0059] In this embodiment, the calculation of the cross-sectional volatility is achieved through sliding window statistical analysis, and the tobacco cross-sectional area change sequence reflects the dynamic change characteristics of the physical form of tobacco. Performing a sliding window process on the area change sequence includes: Adopt an adaptive window size adjustment strategy, and the window size is related to the conveyor belt speed and satisfies the following formula: ; where is the reference window size, is the adjustment coefficient; Apply Gaussian weight attenuation to the window edge data: ; where is the Gaussian weight value of the th data point, is the attenuation coefficient. By performing a sliding window process on the area change sequence, it is ensured that the window size is automatically adjusted according to the conveyor belt speed, and at the same time, the window edge effect is smoothed through weight attenuation to improve the accuracy of volatility calculation.

[0060] The determination of the preset volatility threshold range is based on the statistical analysis of historical production data. The upper and lower volatility threshold values respectively correspond to the reasonable range of area fluctuations under normal production conditions. The screening of the valid data segment is achieved by comparing the current volatility with the threshold range frame by frame, and the stable production stages that meet the process requirements are retained.

[0061] Merge consecutive valid data segments to generate pre-selection stage information. Further, establish a quality evaluation mechanism for pre-selection stage information, including: Calculate the stability index of the area sequence within each pre-selection stage ; Eliminate data segments with stability indices lower than the set standard; Perform a smooth transition process on the boundary region.

[0062] The calculation of the stability index includes: Calculate the autocorrelation coefficient of the area sequence within the pre-selection stage ; Combine the area mean and variance to construct a stability scoring function: ; wherein, is the first weight coefficient, is the second weight coefficient, is the maximum lag order. The above quality evaluation mechanism for pre-selection stage information ensures the reliability of pre-selection stage information through multi-dimensional quantitative analysis. The autocorrelation coefficient reflects the temporal correlation, the variance index characterizes the fluctuation amplitude, and the comprehensive scoring function comprehensively evaluates the stage stability.

[0063] The steps of this embodiment can be understood as: achieving precise screening of production stage information through adaptive window processing and strict quality evaluation. For example, during the tobacco leaf drying process, the system can automatically identify and retain the process stages with stable moisture content changes, providing a reliable data basis for subsequent optimization of motor control parameters.

[0064] This embodiment realizes precise screening and stability evaluation of production stage information through adaptive sliding window processing and a quality evaluation mechanism. The pre-selection stage information extraction method based on cross-sectional volatility, combined with adaptive window adjustment and stability index calculation, ensures that the selected pre-selection stage information has reliable process stability and provides high-quality input data for subsequent optimization of motor control parameters.

[0065] In some embodiments, when collecting a group of tobacco images, synchronously collect motor operation information, and record the motor operation information within the time period to which the pre-selection stage information belongs as pre-selection motor information, including: Read the operation parameters output by the motor controller in real time through the industrial bus interface to generate motor operation information, and the operation parameters include operation frequency, current value, and rotation speed; Align the motor operation information with the group of tobacco images in time, and perform interpolation compensation on the motor operation information with a time deviation exceeding the preset threshold; Filter the motor operation information according to the time period of the preselected stage information, including: Determine the start time point and end time point of the preselected stage information; Extract the data within the time period corresponding to the preselected stage information from the motor operation information after time alignment; Eliminate the abnormal data that does not meet the operation parameter threshold within the time period; Mark the filtered motor operation information as preselected motor information, including: Add the corresponding preselected stage information identifier to the preselected motor information; Establish an associated mapping relationship between the preselected motor information and the preselected stage information.

[0066] In this embodiment, the acquisition of the motor operation information is implemented through an industrial bus interface, and the operation parameters include the operation frequency representing the motor working state, the current value reflecting the load condition, and the rotation speed directly reflecting the movement state of the conveyor belt.

[0067] Preferably, the time alignment process uses the hardware time synchronization signal as a reference, and compensates the data with time deviation through the linear interpolation algorithm to ensure the time synchronization accuracy between the motor operation information and the tobacco image group.

[0068] The determination of the time period corresponding to the preselected stage information is based on accurate timestamp matching, and the positioning of the start time point and end time point considers the overlapping window range of the stage transition area.

[0069] The setting of the operation parameter threshold refers to the rated working parameters of the motor. Optionally, the elimination of abnormal data is achieved through the three - standard - deviation principle, and the main working interval that conforms to the normal distribution is retained.

[0070] The marking process of the preselected motor information establishes a complete data traceability chain. The preselected stage information identifier includes the stage type and process feature description; preferably, the associated mapping relationship is implemented through a hash table structure to achieve fast retrieval and support the two - way query function, that is, both the corresponding process stage can be found through the motor operation information, and the relevant motor parameters can be obtained according to the process stage.

[0071] The steps of this embodiment can be understood as: through strict time alignment and data screening, ensure the accurate correspondence between the motor operation information and the physical state of the tobacco. For example, in the tobacco leaf moisture - regain stage, the system can accurately associate the motor frequency change in this process stage with the tobacco moisture content change characteristics, providing reliable data support for process optimization.

[0072] In this embodiment, the precise acquisition of motor operation information is achieved through an industrial bus interface, and a hardware time synchronization signal is used to ensure time synchronization with the tobacco image group, establishing a reliable association mapping relationship between the preselected motor information and the preselected stage information. By screening the operation parameter thresholds and eliminating abnormal data, the effectiveness of the preselected motor information is ensured, providing accurately matched motor operation state data for subsequent process parameter optimization, and realizing the precise correspondence between the physical state and the motor control parameters in the tobacco processing process.

[0073] In some embodiments, obtaining the sequence of changes in the cross-sectional area of tobacco in the preselected stage information, denoted as the preselected area change sequence, includes: Obtaining the cross-sectional area data at each time point in the preselected stage information; Performing smoothing filtering on the cross-sectional area data; Verifying the continuity and integrity of the cross-sectional area data to obtain the sequence of changes in the cross-sectional area of tobacco; Adding a preselected stage identifier to the sequence of changes in the cross-sectional area of tobacco to obtain the preselected area change sequence.

[0074] In this embodiment, the acquisition of the cross-sectional area data of tobacco in the preselected stage information can be realized through a time series database, and the cross-sectional area data at each time point is derived from the processing results of the aforementioned three-dimensional point cloud model.

[0075] Preferably, the smoothing filtering is implemented using a Savitzky-Golay filter, which effectively suppresses measurement noise while retaining the characteristics of the area change trend, and the window length is dynamically adjusted according to the duration of the preselected stage.

[0076] Preferably, the continuity verification of the cross-sectional area data is achieved by detecting the jump variable of the area between adjacent time points, and when the jump variable exceeds the process allowable range, a data integrity check is triggered; the integrity verification includes the verification of the continuity of the time stamps and the verification of the numerical validity, ensuring that there are no data missing or abnormal values in the area change sequence.

[0077] The preselected stage identifier includes metadata such as the stage type, process parameter range, and environmental conditions, providing complete information for subsequent analysis.

[0078] The steps of this embodiment can be understood as: through a strict filtering process and verification mechanism, ensuring that the preselected area change sequence accurately reflects the evolution process of the tobacco physical state. For example, in the tobacco leaf drying stage, the processed area change sequence can clearly present the corresponding relationship between the tobacco shrinkage characteristics and the drying process parameters, providing a reliable basis for process optimization.

[0079] In this embodiment, through the Savitzky-Golay filter and a strict data verification mechanism, it is ensured that the preselected area change sequence accurately reflects the characteristics of the change in the physical state of tobacco. Based on the complete metadata annotation marked in the preselection stage, the accurate correspondence between the cross-sectional area change and process parameters in the tobacco processing process is realized, providing a reliable data basis for subsequent process optimization and quality control.

[0080] In some embodiments, a radial basis function is set as the kernel function, which is represented by formula (6), and formula (6) is as follows: ; where is the signal variance parameter, is the length scale parameter, is the noise variance parameter, is the Kronecker delta function, is the th input feature vector, is the th input feature vector; Multi-scale feature extraction is performed on the preselected area change sequence to obtain the trend feature and local fluctuation feature of the area change, which is represented by formula (7), and formula (7) is as follows: ; where is the trend feature at time , is the fluctuation feature at time , is the cross-sectional area at the th time point, is the dynamic window size.

[0081] In this embodiment, the setting of the radial basis function kernel function is realized through formula (6). Among them, the signal variance parameter controls the fluctuation amplitude of the function, the length scale parameter determines the speed at which the feature correlation decays with distance, the noise variance parameter is used to adjust the observation noise level, and the Kronecker delta function takes the value of 1 when the th input feature vector and the th input feature vector are the same, and 0 otherwise. Through this kernel function, the similarity relationship between input feature vectors can be effectively characterized, providing a suitable covariance structure for subsequent Gaussian process regression modeling.

[0082] The multi-scale feature extraction process is implemented by formula (7), where the trend feature Characterizes the macro-change trend and fluctuation characteristics reflected by the mean value of the area in the window The local fluctuation characteristics are described by the standard deviation of the area within the window. Dynamic window size According to the real-time speed of the conveyor belt Dynamic adjustment, expressed by the formula , preferably, the benchmark window size Set to 30 sampling points, speed adjustment coefficient Adjustable within the range of 0.5 to 1.5, ensuring that feature extraction adapts to tobacco state changes at different process speeds.

[0083] The solution of this embodiment can be understood as: through the radial basis function kernel and multi-scale feature extraction, a comprehensive characterization of tobacco area changes is achieved. For example, in the tobacco leaf drying process, the trend feature can reflect the overall shrinkage degree, while the fluctuation feature can capture the local uneven drying phenomenon. The combination of the two provides a multi-dimensional reference basis for process optimization.

[0084] This embodiment uses the radial basis function kernel to accurately characterize the similarity relationship between feature vectors, and adopts a dynamic window multi-scale feature extraction method to achieve a comprehensive characterization of the trend characteristics of tobacco cross-sectional area changes and local fluctuation characteristics. Based on the kernel function configuration of signal variance parameters, length scale parameters and noise variance parameters, combined with the trend fluctuation analysis of the adaptive window, it provides a multi-dimensional and high-precision feature description foundation for tobacco processing state monitoring and process optimization.

[0085] In some embodiments, taking the trend feature and the local fluctuation feature as input features and preselecting the running frequency as the output label to train the basic prediction model includes: Construct a multi-scale feature fusion input vector, including: Standardize the trend features to obtain a standardized trend feature vector; Normalize the local fluctuation characteristics to obtain a normalized fluctuation characteristic vector; The standardized trend feature vector and the normalized fluctuation feature vector are combined by using the feature cross method to generate a combined feature item to form a fusion feature matrix; Create a training sample set, including: Divide the fused feature matrix into overlapping sample blocks according to time series; Label each overlapping sample block with the corresponding pre-selected running frequency median; Construct a time series cross validation set and a training sample set; Use the training sample set to perform the model training process and monitor the model training effect on the time series cross validation set, including: Calculate the mean square error between the predicted frequency and the actual frequency; Verify the prediction consistency of the basic prediction model in the stable and unstable periods.

[0086] In this embodiment, when constructing a multi-scale feature fusion input vector, the trend feature is obtained by sliding window mean calculation to characterize the macro trend of tobacco cross-sectional area changes. The standardization process uses the Z-score method to eliminate the dimension effect, and the standardized trend feature vector retains the relative change relationship at each time point. The local fluctuation feature reflects the instantaneous fluctuation intensity of the area sequence. The normalization process uses the maximum and minimum value scaling method to map the fluctuation amplitude to the [0,1] interval. The normalized fluctuation feature vector ensures the comparability of data from different batches. The feature crossover method uses the outer product operation to generate combined feature items. The fused feature matrix contains both trend-fluctuation interaction features and original features, and the model representation capability is improved by increasing the feature dimension.

[0087] Furthermore, the construction of the multi-scale feature fusion input vector also includes: Perform principal component analysis and dimensionality reduction on trend characteristics and fluctuation characteristics respectively; Use attention mechanism to dynamically adjust feature weights; Adding time delay features enhances timing correlation.

[0088] Among them, principal component analysis dimensionality reduction retains the principal component with the highest contribution rate, eliminating feature redundancy while maintaining key information; the attention mechanism realizes dynamic weighting by calculating feature importance scores, focusing on feature dimensions that are strongly related to motor control; the time delay feature captures the lag effect of process parameter changes by introducing feature data of historical time steps.

[0089] When establishing the training sample set, the segmentation of overlapping sample blocks takes into account the duration of the process stage, and a 30% overlap area is set for adjacent sample blocks to ensure time series continuity. The annotation of the median of the pre-selected operating frequency is based on the statistical characteristics of the frequency data in the window to eliminate the interference of instantaneous fluctuations. The forward chain partitioning method is used for the time series cross-validation set to strictly maintain the time sequence to verify the generalization ability of the model.

[0090] The steps of this embodiment can be understood as: through multi-scale feature fusion and timing verification mechanism, a nonlinear mapping relationship between tobacco physical state and motor control parameters is constructed. For example, in the drying stage, the model can identify the overall shrinkage rate through trend characteristics, and use fluctuation characteristics to detect local over-dry areas, and combine time delay characteristics to predict the subsequent required motor speed adjustment amplitude, so as to achieve precise control of process parameters. The application of feature crossover and attention mechanism effectively improves the model's adaptability to complex working conditions, and timing verification ensures the reliability of the prediction results in the continuous operation of the production line.

[0091] In some embodiments, the hyperparameters in the Gaussian process are optimized by maximizing the marginal likelihood function, and the parameter information of the basic prediction model is updated to obtain the final prediction model, including: Establish the Gaussian process marginal likelihood function, including: Calculate the kernel function matrix based on the current hyperparameter combination, construct the marginal likelihood function expression including the noise term, and initialize the hyperparameter search space and optimization constraints; Execute the hyperparameter optimization process, including: Iteratively optimize the marginal likelihood function using the conjugate gradient method; Update the length scale parameter and the signal variance parameter in each iteration; Monitor the change trend and convergence status of the likelihood function value; Verify the hyperparameter optimization results, including: Check the physical rationality of the optimized hyperparameters; Verify the positive definiteness of the kernel function matrix; Evaluate the stability of the gradient descent process; Update the basic prediction model parameters, including: Inject the optimized hyperparameters into the Gaussian process model; Recalculate the posterior distribution of the training data; Update the memory representation and prediction interface of the model; Generate the final prediction model, including: Solidify the optimized model parameters and structure; Save the metadata required for model deployment; Establish a model version management and rollback mechanism.

[0092] In this embodiment, when establishing the Gaussian process marginal likelihood function, the kernel function matrix is obtained by calculating the radial basis function kernel, and its construction process includes the noise variance parameter to characterize the influence of the observation noise. The hyperparameter search space set in the initialization stage clearly constrains the value ranges of the length scale parameter and the signal variance parameter, and the above boundary conditions are determined based on the statistical characteristics of the preselected area change sequence to ensure that the parameter optimization direction conforms to the process physical laws.

[0093] During the execution of the hyperparameter optimization process, the conjugate gradient method realizes iterative update by calculating the partial derivatives of the marginal likelihood function with respect to the length scale parameter and the signal variance parameter, and synchronously adjusts these two key hyperparameters in each iteration to maintain the consistency of the process response characteristics. The change rate of the likelihood function value is monitored in real time during the optimization process, and the search is automatically terminated when the continuous iterative improvement amount is lower than the preset convergence threshold, and the preset convergence threshold is adaptively adjusted according to the training sample size.

[0094] In the verification stage, first, it is checked whether the optimized length scale parameter is within the range of typical process response time, and whether the signal variance parameter matches the actual fluctuation level of the preselected operating frequency. The positive definiteness verification of the kernel function matrix adopts the eigenvalue analysis method to ensure that the covariance calculation meets the mathematical requirements. The stability of gradient descent is evaluated by tracking the smoothness of the parameter update trajectory, and abnormal fluctuations will trigger the restart of optimization.

[0095] When updating the model parameters, the optimized length scale parameter and signal variance parameter are injected into the Gaussian process covariance function. Preferably, the posterior distribution of the training data is updated through Cholesky decomposition. The memory representation reconstruction of the model takes into account the latency requirements of real-time prediction, and the prediction interface retains the complete variance output function to support confidence evaluation.

[0096] In the process of generating the final prediction model, the solidification operation keeps the synchronous storage of the kernel function parameters and the feature scaling factor. Metadata management includes the recording of the hyperparameter optimization path and the evaluation results of the validation set. The version mechanism enables fast retrieval through parameter fingerprints, and the rollback function relies on the checkpoint snapshots during the optimization process. For example, when abnormal fluctuation characteristics are detected in a new production batch, the system can automatically roll back to the model version adapted to high-fluctuation working conditions to ensure prediction stability.

[0097] In this embodiment, through the Gaussian process hyperparameter optimization method based on the marginal likelihood function, the accurate matching of the length scale parameter and the signal variance parameter with the tobacco processing characteristics is achieved. Combining strict mathematical verification and engineering management mechanisms, a final prediction model with physical rationality and numerical stability is constructed. Through the iterative optimization of the conjugate gradient method and the positive definiteness verification of the kernel function matrix, the optimality of the model parameters is ensured, while the version management and rollback mechanisms guarantee the adaptability and reliability of the prediction system under different working conditions.

[0098] Adopting the above technical solutions, compared with the prior art, the beneficial effects of the present invention are as follows: The present invention accurately extracts the sequence of changes in the cross-sectional area of tobacco through a three-dimensional point cloud model, comprehensively characterizes the trend features and local fluctuation features by combining multi-scale feature extraction methods, and constructs a non-linear mapping relationship between the physical form of tobacco and the motor operating frequency; the prediction model constructed by the Gaussian process regression algorithm optimizes the length scale parameter and the signal variance parameter through the marginal likelihood function, realizing the accurate prediction of the changes in the state of tobacco leaves; combined with the dynamic confidence interval evaluation mechanism and model version management, the continuous optimization and working condition adaptation ability of the prediction system are realized, effectively improving the control accuracy and stability of the tobacco processing process. The closed-loop predictive control from raw material state perception to control parameter optimization enables the motor operating frequency to adapt to the state changes in the tobacco production stage, significantly improving the adaptability and stability of the drying process, and providing a more accurate and reliable control method for tobacco processing.

[0099] In addition, in each embodiment of the present invention, each functional unit can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0100] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods in each embodiment of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0101] The above are only partial embodiments of the present invention, and thus do not limit the protection scope of the present invention. Any equivalent device or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A training method for a motor operating frequency prediction model for production, characterized in that, Including: Collecting a group of tobacco images at a preset frequency, where the group of tobacco images includes multiple frames of continuous tobacco image information; Inputting the group of tobacco images into a three-dimensional point cloud model for image processing to obtain tobacco image features in a preset area; Converting the tobacco image features into a sequence of changes in the cross-sectional area of tobacco, dividing the tobacco production stages according to the sequence of changes in the cross-sectional area of tobacco, and splitting the tobacco image features according to the tobacco production stages to obtain multiple pieces of production stage information; Extracting the cross-sectional volatility of the production stage information, and extracting the production stage information with the cross-sectional volatility within a preset volatility threshold range, denoted as preselected stage information; Collecting motor operation information synchronously when collecting the group of tobacco images, and denoting the motor operation information within the time period to which the preselected stage information belongs as preselected motor information; Obtaining the motor operation frequency in the preselected motor information, denoted as the preselected operation frequency, and obtaining the sequence of changes in the cross-sectional area of tobacco of the preselected stage information, denoted as the preselected area change sequence; Fitting the preselected operation frequency and the preselected area change sequence to construct a basic prediction model, and training the basic prediction model using the Gaussian process regression algorithm, including: Setting a radial basis function as the kernel function; Performing multi-scale feature extraction on the preselected area change sequence to obtain the trend feature and local fluctuation feature of the area change; Using the trend feature and the local fluctuation feature together as input features and the preselected operation frequency as the output label to train the basic prediction model; Optimizing the hyperparameters in the Gaussian process by maximizing the marginal likelihood function and updating the parameter information of the basic prediction model to obtain the final prediction model; Establishing a dynamic confidence interval evaluation mechanism, and triggering incremental training of the current final prediction model when the prediction result of the newly input data exceeds the confidence interval.

2. The training method for the operating frequency prediction model of the motor for production according to claim 1, characterized in that The tobacco image information includes depth image information and color image information. Collecting the group of tobacco images at a preset frequency includes: Collecting the depth frame and color frame of the current tobacco through a depth camera, converting the depth frame into a depth matrix, and converting the color frame into a color matrix; Obtaining the hardware timestamp of the depth camera, and performing time alignment on the depth matrix and the color matrix, including: Adopting an adaptive time window algorithm to dynamically adjust the frame synchronization threshold, and optimizing the time alignment accuracy in real time according to the tobacco movement speed; Constructing a coordinate mapping relationship between the depth matrix and the color matrix, which is represented by formula (1), and the formula (1) is as follows: ； In formula (1), is the pixel coordinate in the depth matrix, is the pixel coordinate in the color matrix, is the internal parameter matrix of the depth camera, including the focal length and the principal point parameters, is the internal parameter matrix of the color camera, is the inverse matrix of the internal parameter matrix of the color camera ; And, in the process of constructing the coordinate mapping relationship, it also includes: Calculating the confidence degree of the main body area of tobacco through the depth matrix, which is represented by formula (2), and the formula (2) is as follows: ； In formula (2), represents the pixel coordinate position in the depth matrix, represents the coordinate where the confidence of the main region is located, is the natural logarithm, is the adjustment factor, represents the depth measurement value of the depth matrix at the coordinate where it is located, is the depth threshold; Establishing an edge-sensitive weight function according to the confidence degree of the main body area, which is represented by formula (3), and the formula (3) is as follows: ； In Equation (3), is the balance coefficient, is the magnitude of the depth gradient of the depth matrix at the coordinate and is the edge sensitivity weight. Obtaining the edge-sensitive weights of different mapping regions in the coordinate mapping process according to the edge-sensitive weight function, and performing interpolation compensation on different regions with edge-sensitive weights using different interpolation algorithms to obtain the optimized coordinate mapping relationship, denoted as the final mapping relationship, and the interpolation algorithms include at least two of cubic interpolation, linear interpolation, and nearest neighbor interpolation; The final mapping relationship is represented by formula (4), and the formula (4) is as follows: ； In formula (4), is the number of pre-calibrated non-linear correction basis functions, represents the non-linear correction basis function constructed based on the matching residuals of feature points between the depth camera and the color camera, is the dynamic weight coefficient.

3. The training method for the operating frequency prediction model of the motor for production according to claim 2, wherein Inputting the tobacco image group into the three-dimensional point cloud model for image processing, the obtained tobacco image features in the preset area include: Converting the depth image information in the tobacco image group into three-dimensional point cloud data, including: Calculating the three-dimensional space coordinates through the internal parameter matrix of the depth camera; Performing coordinate transformation on each pixel point in the depth image information to generate the original point cloud data containing spatial position information; Performing dynamic region segmentation on the original point cloud data, including: Obtaining the real-time movement speed of the conveyor belt in the motor operation information to generate the sliding window range of the current processing period; Extracting the point cloud subset located within the sliding window range from the original point cloud data as the processing object; Performing feature enhancement processing on the point cloud subset, including: Calculating the point cloud features of each point in the point cloud subset, where the point cloud features include the normal vector of each point and the curvature descriptor of the local surface characteristics; According to the point cloud features and the color image information in the tobacco image group, performing color space segmentation on the point cloud subset to identify the point cloud data corresponding to the tobacco main body area, denoted as tobacco point cloud data; Performing multi-modal feature fusion on the tobacco point cloud data, including: Performing feature-level fusion on the curvature descriptor and the color space segmentation result; Performing principal component analysis on the fused multi-dimensional feature data to reduce the dimension and obtain the key feature vector; Performing spatio-temporal consistency verification on the key feature vector, including: Establishing the correspondence relationship between the tobacco point cloud data of the current frame and the tobacco point cloud data of the previous frame; Optimizing the correspondence relationship through the iterative closest point algorithm to ensure the point cloud alignment accuracy in time series; Performing compression encoding on the verified key feature vector, including: Using an octree structure to perform spatial partitioning on the tobacco point cloud data; Extracting the point cloud statistical features within each octree node as the tobacco image features of the final output.

4. The training method for the operating frequency prediction model of the production motor according to claim 1, characterized in that, Converting the tobacco image features into a sequence of tobacco cross-sectional area changes, dividing the tobacco production stages according to the sequence of tobacco cross-sectional area changes, and performing data segmentation on the tobacco image features according to the tobacco production stages to obtain multiple production stage information, including: Converting the three-dimensional space coordinate data in the tobacco image features into cross-sectional contour data, including: Intercepting the three-dimensional space coordinate data through a preset detection plane to obtain a cross-sectional point set; Performing height-direction filtering on the cross-sectional point set to remove the data points exceeding the set height range; Performing contour reconstruction on the filtered cross-sectional point set, including: Projecting the cross-sectional point set onto a two-dimensional plane to obtain multiple projection points; Connecting the projection points to form an initial contour polygon; Performing smoothing processing on the initial contour polygon to obtain the final contour line; Calculating the cross-sectional area at each time point, including: Applying the area calculation formula to the final contour line to generate an area sequence; Performing differential calculation on the area values of adjacent time points to obtain an area change rate sequence; Dividing the production stages according to the area change rate sequence, including: Set the determination thresholds for each stage, including the starting stage threshold, the rising stage threshold, the stable stage threshold, and the falling stage threshold; When the area change rate continuously exceeds the threshold of the corresponding stage, mark the stage transition time point; Segment the tobacco image features according to the production stage, including: Divide the data time period according to the stage transition time point; Extract the tobacco image features within each time period and add stage labels; Establish a data overlap window for the stage transition area to maintain feature continuity.

5. The training method of the motor operating frequency prediction model for production according to claim 1, wherein Extract the cross-sectional volatility of the production stage information, and extract the production stage information with the cross-sectional volatility within the preset volatility threshold range, denoted as the preselected stage information, including: Calculate the cross-sectional volatility within each production stage, including: Obtain the tobacco cross-sectional area change sequence in the production stage information; Perform a sliding window process on the area change sequence, calculate the area volatility within each window, which is represented by formula (5), and the formula (5) is as follows: ； In formula (5), is the cross-sectional area volatility at time , is the length of the sliding window, represents the degrees of freedom adjustment term, is the cross-sectional area at the -th time point, is the average area within the window; Determine the preset volatility threshold range, including: Statistically analyze the volatility distribution characteristics of the historical normal production stage, and set the upper volatility threshold and the lower volatility threshold; Screen the preselected stage information, including: Compare the area volatility within each production stage with the preset volatility threshold range, and record the data segment corresponding to the area volatility within the preset volatility threshold range as the valid data segment; Merge the continuous valid data segments to generate the preselected stage information.

6. The training method for the motor operating frequency prediction model for production according to claim 1, characterized in that, Synchronously collect the motor operation information when collecting the tobacco image group, and record the motor operation information within the time period to which the preselected stage information belongs as the preselected motor information, including: Read the operation parameters output by the motor controller in real time through the industrial bus interface to generate the motor operation information, and the operation parameters include the operation frequency, the current value, and the rotation speed; Align the time of the motor operation information with the tobacco image group, and perform interpolation compensation on the motor operation information with a time deviation exceeding the preset threshold; Screen the motor operation information according to the time period of the preselected stage information, including: Determine the start time point and the end time point of the preselected stage information; Extract the data within the time period corresponding to the preselected stage information from the time-aligned motor operation information; Eliminate the abnormal data that does not meet the operation parameter threshold within the time period; Mark the screened motor operation information as the preselected motor information, including: Add the corresponding preselected stage information identifier to the preselected motor information; Establish the association mapping relationship between the preselected motor information and the preselected stage information.

7. The training method for the operating frequency prediction model of the motor for production according to claim 1, characterized in that Obtain the tobacco cross-sectional area change sequence of the preselected stage information, denoted as the preselected area change sequence, including: Obtain the cross-sectional area data at each time point in the preselected stage information; Perform smoothing filtering on the cross-sectional area data; Verify the continuity and integrity of the cross-sectional area data to obtain the tobacco cross-sectional area change sequence; Add the preselected stage identifier to the tobacco cross-sectional area change sequence to obtain the preselected area change sequence.

8. The training method for the motor operating frequency prediction model for production according to claim 1, wherein Set the radial basis function as the kernel function, which is represented by formula (6), and the formula (6) is as follows: ； wherein, is the signal variance parameter, is the length scale parameter, is the noise variance parameter, is the Kronecker delta function, is the th input feature vector, is the th input feature vector; Perform multi-scale feature extraction on the preselected area change sequence to obtain the trend feature and local fluctuation feature of the area change, which are represented by formula (7). The formula (7) is as follows: ； Among them, is the trend feature at time , is the fluctuation feature at time , is the cross-sectional area at the -th time point, is the dynamic window size.

9. The training method for the operating frequency prediction model of the motor for production according to claim 1, wherein Use the trend feature and local fluctuation feature as input features together, and use the preselected operating frequency as the output label to train the basic prediction model, including: Construct a multi-scale feature fusion input vector, including: Perform standardization processing on the trend feature to obtain a standardized trend feature vector; Perform normalization processing on the local fluctuation feature to obtain a normalized fluctuation feature vector; Use the feature cross method to generate combined feature terms from the standardized trend feature vector and the normalized fluctuation feature vector to form a fusion feature matrix; Establish a training sample set, including: Slice the fusion feature matrix into overlapping sample blocks according to the time series; Label the corresponding median value of the preselected operating frequency for each overlapping sample block; Construct a time series cross-validation set and a training sample set; Use the training sample set to execute the model training process and monitor the model training effect on the time series cross-validation set, including: Calculate the mean square error between the predicted frequency and the actual frequency; Verify the prediction consistency of the basic prediction model in the stable period and the non-stable period.

10. The training method of the motor operating frequency prediction model for production according to claim 1, characterized in that, Optimize the hyperparameters in the Gaussian process by maximizing the marginal likelihood function and update the parameter information of the basic prediction model to obtain the final prediction model, including: Establish a Gaussian process marginal likelihood function, including: Calculate the kernel function matrix based on the current hyperparameter combination, construct an expression of the marginal likelihood function including the noise term, and initialize the hyperparameter search space and optimization constraints; Execute the hyperparameter optimization process, including: Iteratively optimize the marginal likelihood function using the conjugate gradient method; Update the length scale parameter and the signal variance parameter in each iteration; Monitor the change trend and convergence state of the likelihood function value; Verify the hyperparameter optimization result, including: Check the physical rationality of the optimized hyperparameters; Verify the positive definiteness of the kernel function matrix; Evaluate the stability of the gradient descent process; Update the parameters of the basic prediction model, including: Inject the optimized hyperparameters into the Gaussian process model; Recalculate the posterior distribution of the training data; Update the memory representation and prediction interface of the model; Generate the final prediction model, including: Solidify the optimized model parameters and structure; Save the metadata required for model deployment; Establish a model version management and rollback mechanism.

Images