A training method for predicting the operating frequency of production motors

Abstract

This invention discloses a method for training a production motor operating frequency prediction model. The method involves collecting tobacco image groups and extracting tobacco image features using a three-dimensional point cloud model. This feature is converted into a series of tobacco cross-sectional area changes, which is then divided into production stages and preselected stage information with a fluctuation rate that meets requirements. Motor operating information is simultaneously collected and matched to obtain preselected motor information and its operating frequency. A Gaussian process regression algorithm is used to construct a prediction model. Trend features and local fluctuation features are extracted through multi-scale feature extraction, and preselected operating frequencies are used as output labels for training. Hyperparameters are optimized to obtain the final prediction model. A dynamic confidence interval evaluation mechanism is established to enable incremental model training. This method dynamically adjusts the motor operating frequency based on the actual state of the tobacco leaves, significantly improving drying quality and batch stability.

Description

Technical Field

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

[0002] In the tobacco processing industry, the primary processing of tobacco leaves typically includes baking, airing, and sorting. The baking process has a decisive influence on the color, aroma, and combustibility of the final product. In traditional processes, tobacco leaves are moved at a constant speed within the drying equipment on a conveyor belt, and the drying temperature and time are controlled by manual experience or fixed procedures. However, due to the natural differences in tobacco leaves during transportation, such as uneven thickness and fluctuating moisture content, existing control methods based on fixed parameters cannot achieve precise drying, which can easily lead to unstable batch quality. Summary of the Invention

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

[0004] In order to achieve the above technical objectives, the technical solution adopted in this application is: a production motor operating frequency prediction model training method, comprising:

[0005] Collecting a tobacco image group at a preset frequency, where the tobacco image group includes multiple frames of continuous tobacco image information;

[0006] Inputting the tobacco image group into the three-dimensional point cloud model for image processing to obtain tobacco image features of the preset area;

[0007] The tobacco image features are converted into a tobacco cross-sectional area change sequence, the tobacco production stages are divided according to the tobacco cross-sectional area change sequence, and the tobacco image features are segmented according to the tobacco production stages to obtain multiple production stage information;

[0008] Extracting the cross-sectional volatility of the production stage information, extracting the production stage information whose cross-sectional volatility falls within a preset volatility threshold, and recording it as pre-selected stage information;

[0009] When collecting the tobacco image group, the motor operation information is collected synchronously, and the motor operation information in the time segment to which the preselected stage information belongs is recorded as the preselected motor information;

[0010] Obtaining the motor operating frequency from the preselected motor information, recorded as the preselected operating frequency, and obtaining the tobacco cross-sectional area change sequence from the preselected stage information, recorded as the preselected area change sequence;

[0011] Fit the preselected operating frequency with the preselected area change sequence to build a basic prediction model, and use the Gaussian process regression algorithm to train the basic prediction model, including:

[0012] Set the radial basis function as the kernel function;

[0013] Perform multi-scale feature extraction on the pre-selected area change sequence to obtain the trend characteristics and local fluctuation characteristics of area change;

[0014] The trend features and local fluctuation features are used as input features, and the pre-selected running frequency is used as the output label to train the basic prediction model;

[0015] By maximizing the marginal likelihood function, the hyperparameters in the Gaussian process are optimized and the parameter information of the basic prediction model is updated to obtain the final prediction model.

[0016] A dynamic confidence interval evaluation mechanism is established to trigger incremental training of the current final prediction model when the prediction result of new input data exceeds the confidence interval.

[0017] In some embodiments, the tobacco image information includes depth image information and color image information, and collecting the tobacco image group according to a preset frequency includes:

[0018] The depth camera is used to collect the depth frame and color frame of the current tobacco, and the depth frame is converted into a depth matrix and the color frame is converted into a color matrix;

[0019] Obtain the hardware timestamp of the depth camera and time-align the depth matrix with the color matrix, including:

[0020] Adopting an adaptive time window algorithm to dynamically adjust the frame synchronization threshold, the time alignment accuracy is optimized in real time according to the tobacco movement speed;

[0021] Construct the coordinate mapping relationship between the depth matrix and the color matrix, which is expressed by formula (1). Formula (1) is as follows:

[0022] ;

[0023] In formula (1), is the pixel coordinate in the depth matrix, are the pixel coordinates in the color matrix, is the intrinsic parameter matrix of the depth camera, including the focal length and main point parameter, is the intrinsic parameter matrix of the color camera, is the color camera intrinsic parameter matrix The inverse matrix of

[0024] Furthermore, the process of constructing the coordinate mapping relationship also includes:

[0025] The confidence of the main area of tobacco is calculated by the depth matrix and expressed by formula (2). Formula (2) is as follows:

[0026] ;

[0027] In formula (2), Represents the pixel coordinate position in the depth matrix, Representing coordinates The confidence of the main area at is the natural logarithm, is the regulating factor, Represents the depth matrix at coordinates The depth measurement at is the depth threshold;

[0028] According to the confidence of the main area, an edge-sensitive weight function is established and expressed by formula (3). Formula (3) is as follows:

[0029] ;

[0030] In formula (3), is the balance coefficient, is the depth gradient amplitude, indicating the depth matrix at the coordinate The depth gradient amplitude at is the edge-sensitive weight;

[0031] Obtaining edge-sensitive weights of different mapping areas in the coordinate mapping process according to the edge-sensitive weight function, performing interpolation compensation on areas with different edge-sensitive weights using different interpolation algorithms, and obtaining an optimized coordinate mapping relationship, recorded as a final mapping relationship, wherein the interpolation algorithm includes at least two of cubic interpolation, linear interpolation, and neighbor interpolation;

[0032] The final mapping relationship is expressed by formula (4), which is as follows:

[0033] ;

[0034] In formula (4), is the number of pre-calibrated nonlinear correction basis functions, Represents the nonlinear correction basis function constructed based on the residual matching of feature points between the depth camera and the color camera. is the dynamic weight coefficient.

[0035] In some embodiments, the tobacco image group is input into a three-dimensional point cloud model for image processing to obtain tobacco image features of a preset area, including:

[0036] Convert the depth image information in the tobacco image group into 3D point cloud data, including:

[0037] The three-dimensional space coordinates are obtained by calculating the intrinsic parameter matrix of the depth camera;

[0038] Perform coordinate transformation on each pixel in the depth image information to generate raw point cloud data containing spatial position information;

[0039] Perform dynamic region segmentation on raw point cloud data, including:

[0040] Obtain the real-time speed of the conveyor belt from the motor operation information and generate the sliding window range of the current processing period;

[0041] Extract the point cloud subset within the sliding window range from the original point cloud data as the processing object;

[0042] Perform feature enhancement on a subset of the point cloud, including:

[0043] Calculate 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.

[0044] Based on the point cloud features and the color image information in the tobacco image group, the point cloud subset is segmented in color space to identify the point cloud data corresponding to the tobacco main area, which is recorded as tobacco point cloud data;

[0045] Multimodal feature fusion of tobacco point cloud data, including:

[0046] Perform feature-level fusion of the curvature descriptor and the color space segmentation results;

[0047] Perform principal component analysis on the fused multi-dimensional feature data and reduce the dimension to obtain the key feature vectors;

[0048] Perform spatiotemporal consistency checks on key feature vectors, including:

[0049] Establishing a correspondence between tobacco point cloud data of a current frame and tobacco point cloud data of a previous frame;

[0050] Optimize the correspondence relationship through the iterative closest point algorithm to ensure the accuracy of point cloud alignment in time sequence;

[0051] The verified key feature vector is compressed and encoded, including:

[0052] Octree structure is used to spatially divide tobacco point cloud data;

[0053] The statistical features of the point cloud within each octree node are extracted as the final output tobacco image features.

[0054] In some embodiments, tobacco image features are converted into a tobacco cross-sectional area change sequence, tobacco production stages are divided according to the tobacco cross-sectional area change sequence, and tobacco image features are segmented according to tobacco production stages to obtain multiple production stage information including:

[0055] Converting the three-dimensional spatial coordinate data in the tobacco image features into cross-sectional profile data includes:

[0056] Intercepting three-dimensional space coordinate data through a preset detection plane to obtain a cross-section point set;

[0057] Perform height-direction filtering on the cross-section point set to remove data points that exceed the set height range;

[0058] The contour of the filtered cross-section point set is reconstructed, including:

[0059] Projecting the cross-section point set onto a two-dimensional plane to obtain multiple projection points;

[0060] Connect the projected points to form the initial outline polygon;

[0061] Smoothing the initial contour polygon to obtain the final contour line;

[0062] Calculate the cross-sectional area at each time point, including:

[0063] Apply the area calculation formula to the final contour line to generate an area sequence;

[0064] The area values at adjacent time points are differentially calculated to obtain the area change rate series;

[0065] The production stages are divided according to the area change rate sequence, including:

[0066] Set the judgment thresholds for each stage, including the starting threshold, rising threshold, stable threshold, and declining threshold;

[0067] When the area change rate continuously exceeds the threshold of the corresponding stage, the stage transition time point is marked;

[0068] The tobacco image features are segmented according to the production stage, including:

[0069] Divide the data time period according to the stage transition time point;

[0070] Extract tobacco image features in each time period and add stage labels;

[0071] Data overlapping windows in the phase transition region are established to maintain feature continuity.

[0072] In some embodiments, extracting the cross-sectional fluctuation rate of the production stage information, and extracting the production stage information whose cross-sectional fluctuation rate falls within a preset fluctuation threshold, and recording it as pre-selected stage information, includes:

[0073] Calculate cross-sectional volatility within each production stage, including:

[0074] Obtaining tobacco cross-sectional area change sequence in production stage information;

[0075] The area change sequence is processed by sliding window, and the area fluctuation rate in each window is calculated and expressed by formula (5). The formula (5) is as follows:

[0076] ;

[0077] In formula (5), For the moment The cross-sectional area fluctuation rate at is the length of the sliding window, represents the degree of freedom adjustment term, For the The cross-sectional area at each time point, is the average area within the window;

[0078] Determine the preset fluctuation threshold range, including:

[0079] Calculate the volatility distribution characteristics during the historical normal production phase and set the upper and lower volatility thresholds;

[0080] Screening pre-selection stage information, including:

[0081] Compare the area fluctuation rate in each production stage with the preset fluctuation threshold range, and record the data segment corresponding to the area fluctuation rate within the preset fluctuation threshold range as a valid data segment;

[0082] Merge consecutive valid data segments to generate pre-selection stage information.

[0083] In some embodiments, synchronously collecting motor operation information when collecting the tobacco image group, and recording the motor operation information within the time period of the preselected stage information as preselected motor information includes:

[0084] The operating parameters output by the motor controller are read in real time through the industrial bus interface to generate motor operating information. The operating parameters include operating frequency, current value and speed;

[0085] Time-aligning the motor operation information with the tobacco image group, and interpolating and compensating the motor operation information whose time deviation exceeds a preset threshold;

[0086] Filter motor operation information according to the time period of the pre-selected stage information, including:

[0087] Determine the start and end time points of the pre-selection stage information;

[0088] Extracting data within a time segment corresponding to the preselected stage information from the time-aligned motor operation information;

[0089] Eliminate abnormal data that does not meet the operating parameter threshold within the time period;

[0090] The filtered motor operation information is marked as pre-selected motor information, including:

[0091] Add corresponding pre-selection stage information identifier for pre-selected motor information;

[0092] Establish an associated mapping relationship between the pre-selected motor information and the pre-selected stage information.

[0093] In some embodiments, obtaining a tobacco cross-sectional area change sequence of preselected stage information, denoted as a preselected area change sequence, includes:

[0094] Obtaining cross-sectional area data at each time point in the preselected stage information;

[0095] Perform smoothing and filtering on the cross-sectional area data;

[0096] Verify the continuity and integrity of the cross-sectional area data to obtain the tobacco cross-sectional area change sequence;

[0097] A preselected stage identifier is added to the tobacco cross-sectional area change sequence to obtain a preselected area change sequence.

[0098] In some embodiments, setting the radial basis function as the kernel function is expressed by formula (6), which is as follows:

[0099] ;

[0100] in, is the signal variance parameter, is the length scale parameter, is the noise variance parameter, is the Kroneckerdelta function, For the input feature vectors, For the input feature vectors;

[0101] Multi-scale feature extraction is performed on the pre-selected area change sequence to obtain the trend characteristics and local fluctuation characteristics of the area change, which are expressed by formula (7). Formula (7) is as follows:

[0102] ;

[0103] in, For the moment The trend characteristics of For the moment The fluctuation characteristics of For the The cross-sectional area at each time point, For dynamic window size.

[0104] In some embodiments, using trend features and local fluctuation features as input features and preselecting running frequencies as output labels to train a basic prediction model includes:

[0105] Construct a multi-scale feature fusion input vector, including:

[0106] Standardize the trend features to obtain a standardized trend feature vector;

[0107] Normalize the local fluctuation characteristics to obtain the normalized fluctuation characteristic vector;

[0108] The standardized trend feature vector and the normalized fluctuation feature vector are combined using the feature cross method to generate a combined feature item to form a fusion feature matrix;

[0109] Create a training sample set, including:

[0110] Divide the fused feature matrix into overlapping sample blocks according to time series;

[0111] Label each overlapping sample block with the corresponding pre-selected running frequency median;

[0112] Construct a time series cross validation set and a training sample set;

[0113] 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:

[0114] Calculate the mean square error between the predicted frequency and the actual frequency;

[0115] Verify the prediction consistency of the basic prediction model in the stable and unstable periods.

[0116] In some embodiments, optimizing hyperparameters in the Gaussian process by maximizing the marginal likelihood function and updating parameter information of the basic prediction model to obtain the final prediction model includes:

[0117] Establish the Gaussian process marginal likelihood function, including:

[0118] Calculate the kernel function matrix based on the current hyperparameter combination, construct the marginal likelihood function expression including the noise term, initialize the hyperparameter search space and optimize the constraints;

[0119] Perform hyperparameter optimization, including:

[0120] The conjugate gradient method is used to iteratively optimize the marginal likelihood function;

[0121] The length scale parameter and signal variance parameter are updated in each iteration;

[0122] Monitor the changing trend and convergence status of the likelihood function value;

[0123] Verify the hyperparameter optimization results, including:

[0124] Check the physical plausibility of optimized hyperparameters;

[0125] Verify the positive definiteness of the kernel function matrix;

[0126] Evaluate the stability of the gradient descent process;

[0127] Update basic forecast model parameters, including:

[0128] Inject the optimized hyperparameters into the Gaussian process model;

[0129] Recalculate the posterior distribution of the training data;

[0130] Update the model's memory representation and prediction interface;

[0131] Generate the final prediction model, including:

[0132] Solidify the optimized model parameters and structure;

[0133] Save metadata required for model deployment;

[0134] Establish model version management and rollback mechanism.

[0135] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0136] The above-mentioned technical solution provides a method for training a production motor operating frequency prediction model. This method collects tobacco image groups and extracts tobacco image features using a three-dimensional point cloud model. This feature is converted into a series of tobacco cross-sectional area changes, which is then divided into production stages. Preselected stage information with a fluctuation rate that meets the required level is then selected. Motor operating information is simultaneously collected and matched to obtain preselected motor information and its operating frequency. A Gaussian process regression algorithm is used to construct a prediction model. Trend features and local fluctuation features are extracted through multi-scale features, and preselected operating frequencies are used as output labels for training. Hyperparameters are optimized to obtain the final prediction model. A dynamic confidence interval evaluation mechanism is established to enable incremental model training. This technical solution dynamically adjusts the motor operating frequency based on the actual state of the tobacco leaves, significantly improving drying quality and batch stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0137] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0138] Figure 1 is a method step diagram of steps S101 to S107 of the training method described in the specific embodiment;

[0139] Figure 2 It is a method step diagram of steps S201 to S206 of the training method described in the specific implementation method. DETAILED DESCRIPTION

[0140] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It is particularly noted that the following examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. Similarly, the following examples are only some embodiments of the present invention and are not intended to be exhaustive. All other embodiments obtained by those of ordinary skill in the art without creative effort are intended to fall within the scope of protection of the present invention.

[0141] See also Figure 1 This embodiment provides a method for training a production motor operating frequency prediction model, comprising:

[0142] S101, collecting a tobacco image group according to a preset frequency, where the tobacco image group includes multiple frames of continuous tobacco image information;

[0143] S102, inputting the tobacco image group into the three-dimensional point cloud model for image processing to obtain tobacco image features of a preset area;

[0144] S103, converting tobacco image features into a tobacco cross-sectional area change sequence, dividing tobacco production stages according to the tobacco cross-sectional area change sequence, and segmenting tobacco image features according to tobacco production stages to obtain multiple production stage information;

[0145] S104, extracting cross-sectional fluctuations of production stage information, extracting production stage information whose cross-sectional fluctuations fall within a preset fluctuation threshold, and recording it as pre-selected stage information;

[0146] S105, synchronously collecting motor operation information when collecting the tobacco image group, and recording the motor operation information within the time segment to which the preselected stage information belongs as preselected motor information;

[0147] S106, obtaining the motor operating frequency from the preselected motor information, recorded as the preselected operating frequency, and obtaining the tobacco cross-sectional area change sequence from the preselected stage information, recorded as the preselected area change sequence;

[0148] S107, fitting the preselected operating frequency with the preselected area change sequence to construct a basic prediction model, and using a Gaussian process regression algorithm to perform model training on the basic prediction model, including:

[0149] Set the radial basis function as the kernel function;

[0150] Perform multi-scale feature extraction on the pre-selected area change sequence to obtain the trend characteristics and local fluctuation characteristics of area change;

[0151] The trend features and local fluctuation features are used as input features, and the pre-selected running frequency is used as the output label to train the basic prediction model;

[0152] By maximizing the marginal likelihood function, the hyperparameters in the Gaussian process are optimized and the parameter information of the basic prediction model is updated to obtain the final prediction model.

[0153] A dynamic confidence interval evaluation mechanism is established to trigger incremental training of the current final prediction model when the prediction result of new input data exceeds the confidence interval.

[0154] In step S101, the preset frequency is determined based on a comprehensive consideration of the conveyor belt's operating speed and tobacco processing requirements, ensuring that the captured tobacco image set fully records the continuous state changes of the tobacco leaves during transportation. Preferably, when the conveyor belt runs at a speed of 0.3 m / s, a 1-second acquisition interval ensures that adjacent images maintain a 30% overlap. The resulting continuous tobacco image set provides the necessary data foundation for subsequent 3D reconstruction.

[0155] In step S102, the tobacco image group is processed using a three-dimensional point cloud model. Preferably, the three-dimensional coordinates of the preset monitoring area are calculated using a stereo vision algorithm to reconstruct the three-dimensional morphology of the tobacco surface. The extracted tobacco image features include key parameters such as spatial curvature, which directly serve the subsequent cross-sectional area calculation.

[0156] In step S103, a tobacco cross-sectional area change sequence is calculated using three-dimensional feature data, preferably using an equally spaced slicing algorithm. The slope change obtained by differentially processing the tobacco cross-sectional area change sequence data serves as the basis for dividing production stages. When the slope change exceeds a preset slope threshold, a tobacco production stage transition point is automatically marked, resulting in multiple production stage information. The preset slope threshold is determined by analyzing the area change rate characteristics of high-quality batches in historical production data, and a critical change rate that can significantly distinguish different process stages is selected as the threshold.

[0157] In step S104, the degree of dispersion of the cross-sectional area is calculated within the divided tobacco production stages, and the stable stages whose fluctuation rate remains within the preset fluctuation threshold are screened as pre-selected stage information. Preferably, the preset fluctuation threshold is determined based on the statistical analysis results of historical high-quality batch data.

[0158] In step S105 , the motor operation information acquisition and image acquisition are synchronized through a hardware timing device to ensure high precision of time alignment and obtain pre-selected motor information.

[0159] In step S106, the preselected motor information is precisely matched with the corresponding preselected stage information. Preferably, the motor operating frequency in the preselected motor information is filtered 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 the standardized processing provides reliable input for subsequent model training.

[0160] In step S107, the basic prediction model is constructed using a Gaussian process regression algorithm. The initial parameters of the radial basis function kernel are set based on the typical fluctuation range of the area series. The multi-scale feature extraction process simultaneously captures trend features and local fluctuation features. Trend features reflect the overall trend of change within a stage, while local fluctuation features capture transient anomalies. A dynamic confidence interval assessment mechanism continuously monitors forecast deviations and automatically triggers model updates when thresholds are consistently exceeded, ensuring the continuous optimization capabilities of the forecast system.

[0161] The implementation principles of this embodiment can be understood as establishing a nonlinear mapping relationship between the dynamic changes in tobacco's physical form and motor control parameters. Multi-scale feature extraction is used to capture key control features at different process stages. The probabilistic nature of Gaussian processes is leveraged to assess the credibility of prediction results. This enables closed-loop predictive control, from raw material state perception to control parameter optimization, ensuring that process parameters dynamically adapt to raw material state changes. For example, when a sudden change in tobacco leaf moisture content is detected, resulting in increased fluctuations in cross-sectional area, the model automatically adjusts the motor frequency to extend the drying time for that batch. Simultaneously, incremental learning is used to update the correspondence between fluctuation features and frequency.

[0162] This embodiment establishes a mapping relationship between tobacco cross-sectional area change sequences and motor operating frequency, and employs multi-scale feature extraction and Gaussian process regression to build a prediction model. This accurately correlates the dynamic changes in tobacco physical form with motor control parameters. A dynamic confidence interval evaluation mechanism ensures continuous model optimization, enabling the motor operating frequency to adapt to state changes throughout the tobacco production phase, effectively improving control accuracy and stability during tobacco processing.

[0163] In some embodiments, the tobacco image information includes depth image information and color image information, and collecting the tobacco image group according to a preset frequency includes:

[0164] The depth camera is used to collect the depth frame and color frame of the current tobacco, and the depth frame is converted into a depth matrix and the color frame is converted into a color matrix;

[0165] Obtain the hardware timestamp of the depth camera and time-align the depth matrix with the color matrix, including:

[0166] Adopting an adaptive time window algorithm to dynamically adjust the frame synchronization threshold, the time alignment accuracy is optimized in real time according to the tobacco movement speed;

[0167] Construct the coordinate mapping relationship between the depth matrix and the color matrix, which is expressed by formula (1). Formula (1) is as follows:

[0168] ;

[0169] In formula (1), is the pixel coordinate in the depth matrix, are the pixel coordinates in the color matrix, is the intrinsic parameter matrix of the depth camera, including the focal length and main point parameter, is the intrinsic parameter matrix of the color camera, is the color camera intrinsic parameter matrix The inverse matrix of

[0170] Furthermore, the process of constructing the coordinate mapping relationship also includes:

[0171] The confidence of the main area of tobacco is calculated by the depth matrix and expressed by formula (2). Formula (2) is as follows:

[0172] ;

[0173] In formula (2), Represents the pixel coordinate position in the depth matrix, where Represents the index of the pixel in the width direction of the image, Represents the index of the pixel in the height direction; Representing coordinates The confidence level of the main area at , the output value is between 0 and 1, indicating the possibility that the pixel belongs to the tobacco main area; is the natural logarithm; is the regulating factor; Represents the depth matrix at coordinates The depth measurement value at represents the vertical distance between the actual spatial position corresponding to the pixel and the camera; is the depth threshold, when Much greater than hour, Approaching 0, Approaching 1; when Much smaller than hour, Approach , Approaching 0; at the depth threshold A smooth transition zone is formed near the slope, and the excessive steepness is determined by the adjustment factor control;

[0174] According to the confidence of the main area, an edge-sensitive weight function is established and expressed by formula (3). Formula (3) is as follows:

[0175] ;

[0176] In formula (3), is the balance coefficient, which is used to adjust the weight ratio of the subject confidence and edge features; is the depth gradient amplitude, indicating the depth matrix at the coordinate The depth gradient amplitude at the point represents the drastic degree of depth change in the area around the point; The edge-sensitive weight comprehensively considers the main attributes and edge features of the region to guide the selection of interpolation strategies for subsequent coordinate mapping;

[0177] Obtaining edge-sensitive weights of different mapping areas in the coordinate mapping process according to the edge-sensitive weight function, performing interpolation compensation on areas with different edge-sensitive weights using different interpolation algorithms, and obtaining an optimized coordinate mapping relationship, recorded as a final mapping relationship, wherein the interpolation algorithm includes at least two of cubic interpolation, linear interpolation, and neighbor interpolation;

[0178] The final mapping relationship is expressed by formula (4), which is as follows:

[0179] ;

[0180] In formula (4), is the number of pre-calibrated nonlinear correction basis functions, It represents the nonlinear correction basis function constructed based on the residual matching of feature points between the depth camera and the color camera (i.e., the deviation between the actual coordinates and the linear mapping coordinates), which is used to compensate for lens distortion and assembly errors, thereby ensuring the accuracy of 3D point cloud reconstruction. is the dynamic weight coefficient.

[0181] In this embodiment, the depth image information uses the time-of-flight principle to obtain scene depth data, while the color image information uses a standard RGB three-channel format to record tobacco surface features. Preferably, the depth frame is converted into a regular depth matrix after correction using camera calibration parameters, and the color frame is gamma-corrected and white-balanced to generate a standardized color matrix, ensuring the basic consistency of multimodal data.

[0182] Dynamically adjusting the frame synchronization threshold using an adaptive time window algorithm includes the following steps:

[0183] The displacement of tobacco surface feature points in continuous depth frames is obtained through feature point tracking algorithms. The average displacement per unit time is calculated as the estimated value of tobacco movement speed. Feature points are selected from prominent areas of tobacco texture to ensure tracking stability.

[0184] A mapping relationship between movement speed and frame synchronization threshold is established. When the tobacco movement speed increases, the frame synchronization threshold is proportionally reduced; when the movement speed decreases, the frame synchronization threshold is correspondingly relaxed. The proportional relationship is determined through pre-experimental calibration to ensure that the optimal alignment accuracy can be maintained when the speed changes;

[0185] A sliding time window mechanism is used to maintain the current motion state. The time window length is adaptively adjusted according to the speed change rate: for uniform motion, a fixed-length time window is maintained; when acceleration or deceleration is detected, the time window length is automatically shortened to improve response speed.

[0186] Based on the frame synchronization threshold calculated in real time, the timestamp difference between the depth frame and the color frame is judged: when the timestamp difference exceeds the current threshold, a linear interpolation algorithm is used to compensate for the earlier acquired frame to ensure the time alignment accuracy of the depth matrix and the color matrix.

[0187] The coordinate mapping relationship between the depth matrix and the color matrix is constructed, and a bidirectional nearest neighbor search algorithm is used to handle pixel matching in occluded areas. Forward matching prioritizes depth continuity, while reverse verification focuses on maintaining texture consistency. Areas with differences in bidirectional matching are marked as special processing areas.

[0188] In formula (1), the depth camera intrinsic parameter matrix and the color camera intrinsic parameter matrix It can be pre-calibrated using the Zhang Zhengyou calibration method.

[0189] In formula (2), the depth threshold Take the median of the effective working distance above the conveyor belt plane and adjust the factor Controls the steepness of the confidence changes.

[0190] In formula (3), the balance coefficient According to the depth gradient amplitude The statistical distribution of the image is determined to ensure that the detail features are not lost while retaining the main structure.

[0191] The edge-sensitive weighting function dynamically adjusts weight distribution using pre-trained balance coefficients, enhancing detail preservation in tobacco edge regions and optimizing computational efficiency in flat areas. The final mapping relationship is spatially aligned using a thin plate spline interpolation algorithm, ensuring accurate 3D feature extraction.

[0192] This embodiment ensures the consistency of multimodal image data in 3D reconstruction by establishing a precise spatiotemporal alignment mechanism and optimized coordinate mapping. For example, when a sudden depth change occurs at the edge of a tobacco leaf, the edge-sensitive weighting function automatically enhances the mapping accuracy of that area, avoiding feature distortion caused by depth discontinuities.

[0193] This embodiment uses an adaptive time window algorithm to achieve precise temporal alignment of the depth matrix and color matrix. Combined with an optimized coordinate mapping relationship and edge-sensitive weighting function, it effectively solves the problem of synchronous matching of multimodal image data in tobacco motion. Furthermore, a bidirectional nearest neighbor search algorithm is used to handle occluded region matching, ensuring feature consistency between depth information and color textures. This provides an accurate and reliable data foundation for subsequent 3D reconstruction, significantly improving the integrity and accuracy of tobacco physical morphology feature extraction.

[0194] See also Figure 2In some embodiments, the tobacco image group is input into the three-dimensional point cloud model for image processing to obtain tobacco image features of the preset area, including:

[0195] S201, converting depth image information in the tobacco image group into three-dimensional point cloud data, including:

[0196] The three-dimensional space coordinates are obtained by calculating the intrinsic parameter matrix of the depth camera;

[0197] Perform coordinate transformation on each pixel in the depth image information to generate raw point cloud data containing spatial position information;

[0198] S202, performing dynamic region segmentation on the original point cloud data, including:

[0199] Obtain the real-time speed of the conveyor belt from the motor operation information and generate the sliding window range of the current processing period;

[0200] Extract the point cloud subset within the sliding window range from the original point cloud data as the processing object;

[0201] S203, performing feature enhancement processing on the point cloud subset, including:

[0202] Calculate 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.

[0203] Based on the point cloud features and the color image information in the tobacco image group, the point cloud subset is segmented in color space to identify the point cloud data corresponding to the tobacco main area, which is recorded as tobacco point cloud data;

[0204] S204, performing multimodal feature fusion on the tobacco point cloud data, including:

[0205] Perform feature-level fusion of the curvature descriptor and the color space segmentation results;

[0206] Perform principal component analysis on the fused multi-dimensional feature data and reduce the dimension to obtain the key feature vectors;

[0207] S205: Performing spatiotemporal consistency check on the key feature vectors, including:

[0208] Establishing a correspondence between tobacco point cloud data of a current frame and tobacco point cloud data of a previous frame;

[0209] Optimize the correspondence relationship through the iterative closest point algorithm to ensure the accuracy of point cloud alignment in time sequence;

[0210] S206: compress and encode the verified key feature vector, including:

[0211] The octree structure is used to spatially divide tobacco point cloud data;

[0212] The statistical features of the point cloud in each octree node are extracted as the final output tobacco image features.

[0213] In step S201, the process of converting the depth image information into three-dimensional point cloud data is realized by perspective projection transformation, and the intrinsic parameter matrix of the depth camera is Used to convert the pixel coordinate system to the camera coordinate system. When calculating the three-dimensional space coordinates of each pixel, the depth value and the focal length parameter in the intrinsic parameter matrix are combined. After completing the coordinate transformation, the generated raw point cloud data contains spatial position information, ensuring that each depth pixel can be accurately mapped to the three-dimensional space.

[0214] In step S202, dynamic region segmentation is implemented using a sliding window mechanism, with the real-time speed of the conveyor belt used to determine the spatiotemporal extent of the window. The sliding window range is calculated by taking into account the conveyor belt speed and point cloud processing latency, ensuring that the extracted point cloud subset fully covers the tobacco area during the current processing period. Discrete points in the original point cloud data that fall outside the window range are automatically filtered out, retaining valid processing objects.

[0215] In step S203, preferably, point cloud feature calculation utilizes a local surface fitting method based on k-nearest neighbors, with the normal vector obtained by least squares plane fitting of neighborhood points, and the curvature descriptor characterizing the concave-convex characteristics of the local surface. Preferably, color space segmentation combines the hue component of the HSV color model and point cloud curvature features, using a region growing algorithm to identify the main tobacco area and exclude interfering point clouds such as conveyor belt background.

[0216] In step S204, multimodal feature fusion jointly optimizes geometric features and color features in the feature space. The curvature descriptor and color segmentation results are concatenated to form a high-dimensional feature vector. Preferably, principal component analysis retains the top three principal components with the highest contribution rate. The key feature vector after dimensionality reduction contains both shape and color information.

[0217] In step S205, the spatiotemporal consistency check is achieved by establishing the correspondence between the point clouds of consecutive frames. The point-to-surface distance metric is used during the iterative nearest point algorithm optimization to ensure that the tobacco point clouds of adjacent frames maintain shape continuity during the movement, thereby effectively compensating for the point cloud position deviation caused by the vibration of the conveyor belt.

[0218] In step S206, the octree encoding adaptively divides the space according to the spatial distribution characteristics of the point cloud. Preferably, the point cloud statistical features of each node include descriptors such as density, normal vector distribution and curvature mean, which not only retains the global structural features of the point cloud but also extracts local detail information.

[0219] This embodiment achieves accurate extraction of tobacco's three-dimensional features through the system's point cloud processing process. For example, when processing tobacco leaves with high moisture content, color space segmentation can effectively distinguish moist areas, while curvature features can identify the curling shape of the leaves. The feature vector formed after multimodal fusion comprehensively represents the physical state of the tobacco.

[0220] This embodiment achieves precise extraction of tobacco point cloud data through dynamic region segmentation and feature enhancement. It then combines multimodal feature fusion and spatiotemporal consistency verification to construct a comprehensive 3D feature representation of the tobacco's physical state. Octree-encoded compressed tobacco image features retain both global structural information and local details, providing an accurate and reliable 3D morphological basis for subsequent motor control parameter prediction, significantly improving the accuracy of state perception during tobacco processing.

[0221] In some embodiments, tobacco image features are converted into a tobacco cross-sectional area change sequence, tobacco production stages are divided according to the tobacco cross-sectional area change sequence, and tobacco image features are segmented according to tobacco production stages to obtain multiple production stage information including:

[0222] Converting the three-dimensional spatial coordinate data in the tobacco image features into cross-sectional profile data includes:

[0223] Intercepting three-dimensional space coordinate data through a preset detection plane to obtain a cross-section point set;

[0224] Perform height-direction filtering on the cross-section point set to remove data points that exceed the set height range;

[0225] The contour of the filtered cross-section point set is reconstructed, including:

[0226] Projecting the cross-section point set onto a two-dimensional plane to obtain multiple projection points;

[0227] Connect the projected points to form the initial outline polygon;

[0228] Smoothing the initial contour polygon to obtain the final contour line;

[0229] Calculate the cross-sectional area at each time point, including:

[0230] Apply the area calculation formula to the final contour line to generate an area sequence;

[0231] The area values at adjacent time points are differentially calculated to obtain the area change rate series;

[0232] The production stages are divided according to the area change rate sequence, including:

[0233] Set the judgment thresholds for each stage, including the starting threshold, rising threshold, stable threshold, and declining threshold;

[0234] When the area change rate continuously exceeds the threshold of the corresponding stage, the stage transition time point is marked;

[0235] The tobacco image features are segmented according to the production stage, including:

[0236] Divide the data time period according to the stage transition time point;

[0237] Extract tobacco image features in each time period and add stage labels;

[0238] Data overlapping windows in the phase transition region are established to maintain feature continuity.

[0239] In this embodiment, the process of converting three-dimensional spatial coordinate data into cross-sectional profile data is achieved through spatial geometric transformation.

[0240] A pre-set detection plane is set parallel to the conveyor belt's direction of motion, and its height position is dynamically adjusted based on the thickness of the tobacco deposit. Preferably, the cross-sectional point set is acquired using a spatial index acceleration algorithm, retaining only the three-dimensional point cloud data that intersects the detection plane. Height-direction filtering eliminates abnormal outliers by setting upper and lower thresholds to ensure the accuracy of contour reconstruction.

[0241] During contour reconstruction, the cross-sectional point set is first projected onto a two-dimensional plane perpendicular to the detection plane. The projected points are then used to construct an initial topological relationship using the Delaunay triangulation algorithm. Preferably, the initial contour polygon is extracted using the α-shape algorithm, and then smoothed using a B-spline-based curve fitting method. This method eliminates measurement noise while preserving characteristic inflection points, ultimately resulting in a smooth and continuous final contour line.

[0242] When calculating the cross-sectional area at each time point, the area calculation formula is applied to the final contour line to generate an area sequence. The area calculation formula uses the improved contour integral method, including:

[0243] Uniformly resample the point set on the final contour line;

[0244] Calculate the signed area of the resampled polygonal region;

[0245] The edge compensation coefficient is introduced to correct the measurement error.

[0246] Preferably, the improved contour integral method ensures the consistency of the distribution density of contour points through uniform resampling; the signed area calculation can use Green's formula to achieve 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 of adjacent time points uses the central difference method to effectively suppress the influence of measurement noise on the change rate calculation.

[0247] When dividing the production stages according to the area change rate sequence, set the judgment thresholds for each stage, including the starting stage threshold, rising stage threshold, stable stage threshold, and declining stage threshold. The determination of the stage judgment thresholds includes:

[0248] Calculate the typical area change rate range in each stage of historical production data;

[0249] Sliding window variance analysis was used to dynamically adjust the threshold boundaries;

[0250] When the environmental parameters change beyond the set range, the threshold is triggered to recalibrate.

[0251] The stage determination thresholds are based on statistical analysis of historical production data. The start-up, ramp-up, stabilization, and decline thresholds correspond to typical variations in different process stages. Sliding window variance analysis monitors production status fluctuations in real time. When significant changes in ambient parameters such as temperature and humidity are detected, threshold recalibration is automatically triggered to ensure accurate stage delineation.

[0252] When segmenting tobacco image features by production stage, the data overlap window in the stage transition region is preferably set to a 50% overlap between adjacent stages, with linear weighted fusion achieving a smooth feature transition. In addition to adding stage labels, tobacco image features for each time period also record the corresponding process parameters and environmental conditions, providing complete information for subsequent analysis.

[0253] The steps in this embodiment can be understood as achieving reliable classification of tobacco production stages through precise geometric calculations and adaptive threshold adjustment. For example, during the tobacco leaf rehumidification process, when the area change rate continuously exceeds the rising threshold, the system accurately identifies the rapid moisture absorption stage and automatically adjusts subsequent processing parameters.

[0254] This embodiment uses an improved contour integral method and a dynamic threshold adjustment mechanism to accurately calculate tobacco cross-sectional area variation sequences and reliably classify production stages. Cross-sectional profile data, based on 3D coordinate data conversion and contour reconstruction, combined with adaptive stage determination thresholds and transition zone processing, ensures accurate segmentation of tobacco image features by production stage, providing a reliable state identification basis for process parameter optimization.

[0255] In some embodiments, extracting the cross-sectional fluctuation rate of the production stage information, and extracting the production stage information whose cross-sectional fluctuation rate falls within a preset fluctuation threshold, and recording it as pre-selected stage information, includes:

[0256] Calculate cross-sectional volatility within each production stage, including:

[0257] Obtaining tobacco cross-sectional area change sequence in production stage information;

[0258] Perform sliding window processing on the area change sequence and calculate the area fluctuation rate within each window, which is expressed by formula (5). Formula (5) is as follows:

[0259] ;

[0260] In formula (5), For the moment The cross-sectional area fluctuation rate at is the length of the sliding window (i.e., the total number of time points), Represents the degree of freedom adjustment term, which is used to estimate volatility unbiasedly and avoid the number of samples in the window The statistical bias caused by For the The cross-sectional area at each time point, is the average area within the window;

[0261] Determine the preset fluctuation threshold range, including:

[0262] Calculate the volatility distribution characteristics during the historical normal production phase and set the upper and lower volatility thresholds;

[0263] Screening pre-selection stage information, including:

[0264] Compare the area fluctuation rate in each production stage with the preset fluctuation threshold range, and record the data segment corresponding to the area fluctuation rate within the preset fluctuation threshold range as a valid data segment;

[0265] Merge consecutive valid data segments to generate pre-selection stage information.

[0266] In this embodiment, the calculation of cross-sectional fluctuation rate is achieved through sliding window statistical analysis. The tobacco cross-sectional area change sequence reflects the dynamic change characteristics of tobacco physical morphology. The sliding window processing of the area change sequence includes:

[0267] Adaptive window size adjustment strategy is adopted, the window size Conveyor belt speed Satisfies the following formula:

[0268] ;

[0269] in, is the base window size, is the adjustment coefficient;

[0270] Apply Gaussian weight decay to the window edge data:

[0271] ;

[0272] in, For the Gaussian weights of data points, is the attenuation coefficient. A sliding window is used to process the area change sequence, ensuring that the window size automatically adjusts with the conveyor belt speed. Weight attenuation is also used to smooth the window edge effect, improving the accuracy of volatility calculation.

[0273] The preset fluctuation threshold range is determined based on statistical analysis of historical production data. The upper and lower fluctuation thresholds correspond to the reasonable range of area fluctuation under normal production conditions. Valid data segments are selected by comparing the current fluctuation rate against the threshold ranges on a frame-by-frame basis, retaining stable production stages that meet process requirements.

[0274] Merge consecutive valid data segments to generate pre-selection stage information. Furthermore, establish a pre-selection stage information quality assessment mechanism, including:

[0275] Calculate the stability index of the area sequence in each preselection stage ;

[0276] Eliminate data segments whose stability index is lower than the set standard;

[0277] Perform smooth transition processing on the boundary area.

[0278] The stability index calculation includes:

[0279] Calculate the autocorrelation coefficient of the area series within the preselected period ;

[0280] Mean binding area and variance Construct a stability scoring function:

[0281] ;

[0282] in, is the first weight coefficient, is the second weight coefficient, The above-mentioned pre-selection stage information quality assessment mechanism ensures the reliability of pre-selection stage information through multi-dimensional quantitative analysis. The autocorrelation coefficient reflects the time series correlation, the variance index represents the fluctuation amplitude, and the comprehensive scoring function comprehensively evaluates the stage stability.

[0283] The steps in this example can be understood as accurately filtering production stage information through adaptive window processing and rigorous quality assessment. For example, during the tobacco drying process, the system can automatically identify and retain process stages where moisture content steadily changes, providing a reliable data foundation for subsequent motor control parameter optimization.

[0284] This embodiment achieves precise screening and stability assessment of production stage information through adaptive sliding window processing and quality assessment mechanisms. A preselected stage information extraction method based on cross-sectional fluctuations, combined with adaptive window adjustment and stability index calculation, ensures reliable process stability of the preselected stage information, providing high-quality input data for subsequent motor control parameter optimization.

[0285] In some embodiments, synchronously collecting motor operation information when collecting the tobacco image group, and recording the motor operation information within the time period of the preselected stage information as preselected motor information includes:

[0286] The operating parameters output by the motor controller are read in real time through the industrial bus interface to generate motor operating information. The operating parameters include operating frequency, current value and speed;

[0287] Time-aligning the motor operation information with the tobacco image group, and interpolating and compensating the motor operation information whose time deviation exceeds a preset threshold;

[0288] Filter the motor operation information according to the time period of the pre-selected stage information, including:

[0289] Determine the start and end time points of the pre-selection stage information;

[0290] Extracting data within a time segment corresponding to the preselected stage information from the time-aligned motor operation information;

[0291] Eliminate abnormal data that does not meet the operating parameter threshold within the time period;

[0292] The filtered motor operation information is marked as pre-selected motor information, including:

[0293] Add corresponding pre-selection stage information identifier for pre-selected motor information;

[0294] Establish an associated mapping relationship between the pre-selected motor information and the pre-selected stage information.

[0295] In this embodiment, the motor operation information is collected through the industrial bus interface. The operation parameters include the operating frequency that characterizes the working state of the motor, the current value that reflects the load condition, and the speed that directly reflects the movement state of the conveyor belt.

[0296] Preferably, the time alignment process uses the hardware timing signal as a benchmark, and compensates for data with time deviation through a linear interpolation algorithm to ensure the time synchronization accuracy of the motor operation information and the tobacco image group.

[0297] The determination of the time segment corresponding to the pre-selected stage information is based on accurate timestamp matching, and the positioning of the start and end time points takes into account the overlapping window range of the stage transition area.

[0298] The setting of the operating parameter threshold refers to the rated operating parameters of the motor. Optionally, the elimination of abnormal data is achieved through the principle of three times the standard deviation, and the main working range that conforms to the normal distribution is retained.

[0299] The marking process of pre-selected motor information establishes a complete data traceability chain, and the pre-selected stage information identification includes the stage type and process feature description; preferably, the associated mapping relationship is quickly retrieved through a hash table structure, supporting a two-way query function, which can not only find the corresponding process stage through the motor operation information, but also obtain relevant motor parameters according to the process stage.

[0300] The steps in this example can be understood as ensuring that motor operating information accurately corresponds to the physical state of the tobacco through strict time alignment and data screening. For example, during the tobacco leaf rehumidification stage, the system can accurately correlate motor frequency changes with tobacco moisture content variations during this process, providing reliable data support for process optimization.

[0301] This embodiment accurately collects motor operating information through an industrial bus interface and employs hardware timing signals to ensure time synchronization with the tobacco image set, establishing a reliable correlation and mapping relationship between preselected motor information and preselected stage information. By filtering operating parameter thresholds and eliminating abnormal data, the validity of the preselected motor information is ensured, providing accurately matched motor operating status data for subsequent process parameter optimization, achieving a precise correspondence between the physical state of the tobacco processing process and the motor control parameters.

[0302] In some embodiments, obtaining a tobacco cross-sectional area change sequence of preselected stage information, denoted as a preselected area change sequence, includes:

[0303] Obtaining cross-sectional area data at each time point in the preselected stage information;

[0304] Perform smoothing and filtering on the cross-sectional area data;

[0305] Verify the continuity and integrity of the cross-sectional area data to obtain the tobacco cross-sectional area change sequence;

[0306] A preselected stage identifier is added to the tobacco cross-sectional area change sequence to obtain a preselected area change sequence.

[0307] In this embodiment, the tobacco cross-sectional area data of the pre-selected stage information can be obtained through a time series database, and the cross-sectional area data at each time point is derived from the aforementioned three-dimensional point cloud model processing results.

[0308] Preferably, the smoothing filter processing is implemented using a Savitzky-Golay filter, which effectively suppresses measurement noise while retaining the area change trend characteristics, and the window length is dynamically adjusted according to the duration of the preselected stage.

[0309] Preferably, the continuity check of the cross-sectional area data is achieved by detecting the area jump value at adjacent time points, and triggering a data integrity check when the jump value exceeds the process allowable range; the integrity check includes timestamp continuity verification and numerical validity check to ensure that there is no data missing or abnormal value in the area change sequence.

[0310] The pre-selected stage identification contains metadata such as stage type, process parameter range and environmental conditions, providing complete information for subsequent analysis.

[0311] The steps in this embodiment can be understood as ensuring that the preselected area change sequence accurately reflects the evolution of the tobacco's physical state through rigorous filtering and verification. For example, during the tobacco leaf drying stage, the processed area change sequence can clearly demonstrate the correspondence between tobacco shrinkage characteristics and drying process parameters, providing a reliable basis for process optimization.

[0312] This implementation uses a Savitzky-Golay filter and a rigorous data validation mechanism to ensure that the preselected area change sequence accurately reflects the characteristics of tobacco's physical state changes. Complete metadata annotation based on preselected stage identifiers enables precise correspondence between cross-sectional area changes and process parameters during tobacco processing, providing a reliable data foundation for subsequent process optimization and quality control.

[0313] In some embodiments, setting the radial basis function as the kernel function is expressed by formula (6), which is as follows:

[0314] ;

[0315] in, is the signal variance parameter, is the length scale parameter, is the noise variance parameter, is the Kroneckerdelta function, For the input feature vectors, For the input feature vectors;

[0316] Multi-scale feature extraction is performed on the pre-selected area change sequence to obtain the trend characteristics and local fluctuation characteristics of the area change, which are expressed by formula (7). Formula (7) is as follows:

[0317] ;

[0318] in, For the moment The trend characteristics of For the moment The fluctuation characteristics of For the The cross-sectional area at each time point, For dynamic window size.

[0319] In this embodiment, the setting of the radial basis function kernel function is realized by formula (6), where the signal variance parameter Control function fluctuation amplitude, length scale parameter Determines the speed at which feature correlation decays with distance, noise variance parameter Used to adjust the observation noise level, Kronecker delta function In the input Input feature vector Hedi Input feature vector If they are the same, the value is 1, otherwise it is 0. This kernel function can effectively characterize the similarity relationship between input feature vectors and provide a suitable covariance structure for subsequent Gaussian process regression modeling.

[0320] 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 within 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, it ensures that feature extraction adapts to tobacco state changes at different process speeds.

[0321] The solution in this embodiment can be understood as achieving a comprehensive characterization of tobacco area changes through the use of radial basis function kernels and multi-scale feature extraction. For example, during the tobacco leaf drying process, trend features can reflect the overall degree of shrinkage, while fluctuation features can capture localized uneven drying. The combination of these two provides a multi-dimensional reference for process optimization.

[0322] This implementation uses a radial basis function kernel to precisely characterize the similarity relationships between feature vectors and employs a dynamic window multi-scale feature extraction method to comprehensively characterize the trend and local fluctuation characteristics of tobacco cross-sectional area changes. Kernel function configuration based on signal variance parameters, length scale parameters, and noise variance parameters, combined with trend and fluctuation analysis using an adaptive window, provides a multi-dimensional, high-precision feature description foundation for tobacco processing state monitoring and process optimization.

[0323] In some embodiments, using trend features and local fluctuation features as input features and preselecting running frequencies as output labels to train a basic prediction model includes:

[0324] Construct a multi-scale feature fusion input vector, including:

[0325] Standardize the trend features to obtain a standardized trend feature vector;

[0326] Normalize the local fluctuation characteristics to obtain the normalized fluctuation characteristic vector;

[0327] The standardized trend feature vector and the normalized fluctuation feature vector are combined using the feature cross method to generate a combined feature item to form a fusion feature matrix;

[0328] Create a training sample set, including:

[0329] Divide the fused feature matrix into overlapping sample blocks according to time series;

[0330] Label each overlapping sample block with the corresponding pre-selected running frequency median;

[0331] Construct a time series cross validation set and a training sample set;

[0332] 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:

[0333] Calculate the mean square error between the predicted frequency and the actual frequency;

[0334] Verify the prediction consistency of the basic prediction model in the stable and unstable periods.

[0335] In this example, when constructing a multi-scale feature fusion input vector, trend features are calculated using a sliding window mean, representing the macroscopic trend of tobacco cross-sectional area changes. Z-scores are used to normalize the feature vector to eliminate dimensionality effects, and the normalized trend feature vector preserves the relative change relationship at each time point. Local fluctuation features reflect the instantaneous fluctuation intensity of the area series. Normalization uses a maximum-minimum scaling method to map the fluctuation amplitude to the [0, 1] interval. Normalized fluctuation feature vectors ensure comparability across different batches of data. A feature crossover method uses an outer product operation to generate combined feature terms. The fused feature matrix contains both trend-fluctuation interaction features and original features, enhancing the model's representational capabilities by increasing feature dimensionality.

[0336] Furthermore, the construction of the multi-scale feature fusion input vector also includes:

[0337] Perform principal component analysis and dimensionality reduction on trend characteristics and fluctuation characteristics respectively;

[0338] Use attention mechanism to dynamically adjust feature weights;

[0339] Adding time delay features enhances timing correlation.

[0340] 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.

[0341] When establishing the training sample set, overlapping sample blocks were segmented based on the duration of the process phases, and adjacent sample blocks were set to overlap by 30% to ensure temporal continuity. The median of the preselected running frequency was labeled based on the statistical characteristics of the frequency data within the window to eliminate interference from transient fluctuations. The temporal cross-validation set was partitioned using a forward chaining method, strictly maintaining chronological order to verify the model's generalization capability.

[0342] The steps in this embodiment can be understood as: through multi-scale feature fusion and time series verification mechanisms, a nonlinear mapping relationship between tobacco physical state and motor control parameters is constructed. For example, during the drying stage, the model can identify the overall shrinkage rate through trend features, while simultaneously using fluctuation features to detect local over-dry areas. Combined with time delay features, it predicts the subsequent required motor speed adjustment amplitude, achieving precise control of process parameters. The application of feature cross-talk and attention mechanisms effectively improves the model's adaptability to complex working conditions, and time series verification ensures the reliability of the prediction results during continuous operation of the production line.

[0343] In some embodiments, optimizing hyperparameters in the Gaussian process by maximizing the marginal likelihood function and updating parameter information of the basic prediction model to obtain the final prediction model includes:

[0344] Establish the Gaussian process marginal likelihood function, including:

[0345] Calculate the kernel function matrix based on the current hyperparameter combination, construct the marginal likelihood function expression including the noise term, initialize the hyperparameter search space and optimize the constraints;

[0346] Perform hyperparameter optimization, including:

[0347] The conjugate gradient method is used to iteratively optimize the marginal likelihood function;

[0348] The length scale parameter and signal variance parameter are updated in each iteration;

[0349] Monitor the changing trend and convergence status of the likelihood function value;

[0350] Verify the hyperparameter optimization results, including:

[0351] Check the physical plausibility of optimized hyperparameters;

[0352] Verify the positive definiteness of the kernel function matrix;

[0353] Evaluate the stability of the gradient descent process;

[0354] Update basic forecast model parameters, including:

[0355] Inject the optimized hyperparameters into the Gaussian process model;

[0356] Recalculate the posterior distribution of the training data;

[0357] Update the model's memory representation and prediction interface;

[0358] Generate the final prediction model, including:

[0359] Solidify the optimized model parameters and structure;

[0360] Save metadata required for model deployment;

[0361] Establish model version management and rollback mechanism.

[0362] In this embodiment, when establishing the Gaussian process marginal likelihood function, the kernel function matrix is calculated using a radial basis function kernel. Its construction process includes a noise variance parameter to characterize the influence of observation noise. The hyperparameter search space set during the initialization phase explicitly constrains the range of values for the length scale parameter and the signal variance parameter. These boundary conditions are determined based on the statistical properties of a preselected area change sequence, ensuring that the parameter optimization direction conforms to the laws of process physics.

[0363] During hyperparameter optimization, the conjugate gradient method iteratively updates the results by calculating the partial derivatives of the marginal likelihood function with respect to the length scale parameter and the signal variance parameter. These two key hyperparameters are adjusted simultaneously with each iteration to maintain consistent process response characteristics. The optimization process monitors the rate of change of the likelihood function value in real time, and the search is automatically terminated when the improvement in successive iterations falls below a preset convergence threshold, which is adaptively adjusted based on the size of the training sample.

[0364] During the verification phase, the optimized length scale parameters are first checked to ensure they fall within the typical process response time range and that the signal variance parameters match the actual fluctuation level at the preselected operating frequency. Eigenvalue analysis is used to verify the positive definiteness of the kernel matrix, ensuring that the covariance calculations meet mathematical requirements. Gradient descent stability is assessed by tracking the smoothness of the parameter update trajectory; abnormal fluctuations trigger an optimization restart.

[0365] When updating model parameters, the optimized length scale and signal variance parameters are injected into the Gaussian process covariance function. Preferably, the posterior distribution of the training data is updated via Cholesky decomposition. The model's memory representation is restructured to take into account the latency requirements of real-time predictions, and the prediction interface retains the full variance output functionality to support confidence assessment.

[0366] During the generation of the final prediction model, the solidification operation synchronizes the kernel function parameters and feature scaling factors. Metadata management includes records of the hyperparameter optimization path and validation set evaluation results. The versioning mechanism enables fast retrieval through parameter fingerprints, and the rollback function relies on checkpoint snapshots during the optimization process. For example, if a new production batch is detected to exhibit unusual fluctuations, the system can automatically roll back to a model version adapted for highly volatile conditions to ensure prediction stability.

[0367] This example utilizes a Gaussian process hyperparameter optimization method based on marginal likelihood functions to precisely match length scale parameters and signal variance parameters with tobacco processing characteristics. Combined with rigorous mathematical verification and engineering management mechanisms, a final prediction model with physical rationality and numerical stability is constructed. Iterative optimization using the conjugate gradient method and positive definiteness verification of the kernel function matrix ensures the optimality of the model parameters. Version management and rollback mechanisms ensure the adaptability and reliability of the prediction system under different operating conditions.

[0368] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0369] The present invention accurately extracts the change sequence of tobacco cross-sectional area through a three-dimensional point cloud model, combines multi-scale feature extraction methods to comprehensively characterize trend characteristics and local fluctuation characteristics, and constructs a nonlinear mapping relationship between tobacco physical form and motor operating frequency. The prediction model constructed using the Gaussian process regression algorithm optimizes the length scale parameter and signal variance parameter through the marginal likelihood function, achieving accurate prediction of tobacco leaf state changes. The dynamic confidence interval evaluation mechanism and model version management are combined to achieve continuous optimization and working condition adaptability of the prediction system, 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 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.

[0370] In addition, the functional units in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0371] 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 this understanding, the technical solution of the present invention, 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 enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0372] The above descriptions are only some embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for training a production motor operating frequency prediction model, characterized in that: include: Collecting a tobacco image group at a preset frequency, where the tobacco image group includes multiple frames of continuous tobacco image information; Inputting the tobacco image group into the three-dimensional point cloud model for image processing to obtain tobacco image features of the preset area; The tobacco image features are converted into a tobacco cross-sectional area change sequence, the tobacco production stages are divided according to the tobacco cross-sectional area change sequence, and the tobacco image features are segmented according to the tobacco production stages to obtain multiple production stage information; Extracting the cross-sectional volatility of the production stage information, extracting the production stage information whose cross-sectional volatility falls within a preset volatility threshold, and recording it as pre-selected stage information; When collecting the tobacco image group, the motor operation information is collected synchronously, and the motor operation information in the time segment to which the preselected stage information belongs is recorded as the preselected motor information; Obtaining the motor operating frequency from the preselected motor information, recorded as the preselected operating frequency, and obtaining the tobacco cross-sectional area change sequence from the preselected stage information, recorded as the preselected area change sequence; Fit the preselected operating frequency with the preselected area change sequence to build a basic prediction model, and use the Gaussian process regression algorithm to train the basic prediction model, including: Set the radial basis function as the kernel function; Perform multi-scale feature extraction on the pre-selected area change sequence to obtain the trend characteristics and local fluctuation characteristics of area change; The trend features and local fluctuation features are used as input features, and the pre-selected running frequency is used as the output label to train the basic prediction model; By maximizing the marginal likelihood function, the hyperparameters in the Gaussian process are optimized and the parameter information of the basic prediction model is updated to obtain the final prediction model. A dynamic confidence interval evaluation mechanism is established to trigger incremental training of the current final prediction model when the prediction result of new input data exceeds the confidence interval.

2. The method for training a production motor operating frequency prediction model according to claim 1, wherein: The tobacco image information includes depth image information and color image information. Collecting tobacco image groups according to a preset frequency includes: Collecting a depth frame and a 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 a hardware timestamp of a 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, the time alignment accuracy is optimized in real time according to the tobacco movement speed; The coordinate mapping relationship between the depth matrix and the color matrix is constructed and expressed by formula (1). The formula (1) is as follows: ； In formula (1), is the pixel coordinate in the depth matrix, are the pixel coordinates in the color matrix, is the intrinsic parameter matrix of the depth camera, including the focal length and main point parameter, is the intrinsic parameter matrix of the color camera, is the color camera intrinsic parameter matrix The inverse matrix of Furthermore, the process of constructing the coordinate mapping relationship also includes: The confidence of the main area of tobacco is calculated by the depth matrix and expressed by formula (2), which is as follows: ； In formula (2), Represents the pixel coordinate position in the depth matrix, Representing coordinates The confidence of the main area at is the natural logarithm, is the regulating factor, Represents the depth matrix at coordinates The depth measurement at is the depth threshold; According to the confidence of the subject area, an edge-sensitive weight function is established and expressed by formula (3). The formula (3) is as follows: ； In formula (3), is the balance coefficient, is the depth matrix at coordinates The depth gradient amplitude at is the edge-sensitive weight; Obtaining edge-sensitive weights of different mapping areas in a coordinate mapping process according to the edge-sensitive weight function, performing interpolation compensation on areas with different edge-sensitive weights using different interpolation algorithms to obtain an optimized coordinate mapping relationship, recorded as a final mapping relationship, wherein the interpolation algorithm includes at least two of cubic interpolation, linear interpolation, and neighbor interpolation; The final mapping relationship is expressed by formula (4), which is as follows: ； In formula (4), is the number of pre-calibrated nonlinear correction basis functions, Represents the nonlinear correction basis function constructed based on the residual matching of feature points between the depth camera and the color camera. is the dynamic weight coefficient.

3. The production motor operating frequency prediction model training method according to claim 2, characterized in that: The tobacco image group is input into the 3D point cloud model for image processing, and the tobacco image features of the preset area are obtained, including: Converting the depth image information in the tobacco image group into three-dimensional point cloud data includes: Obtaining three-dimensional space coordinates by calculating the intrinsic parameter matrix of the depth camera; Performing coordinate transformation on each pixel in the depth image information to generate original point cloud data containing spatial position information; Performing dynamic region segmentation on the original point cloud data, including: Obtain the real-time speed of the conveyor belt from the motor operation information and generate the sliding window range of the current processing period; Extracting a point cloud subset within the sliding window from the original point cloud data as a processing object; Performing feature enhancement processing on the point cloud subset includes: Calculating point cloud features of each point in the point cloud subset, wherein the point cloud features include a normal vector of each point and a curvature descriptor of local surface characteristics; performing color space segmentation on the point cloud subset according to the point cloud features and the color image information in the tobacco image group, identifying point cloud data corresponding to the tobacco main body area, and recording the data as tobacco point cloud data; Multimodal feature fusion of tobacco point cloud data, including: Performing feature-level fusion of the curvature descriptor and the color space segmentation result; Perform principal component analysis on the fused multi-dimensional feature data and reduce the dimension to obtain the key feature vectors; Perform spatiotemporal consistency checks on key feature vectors, including: Establishing a correspondence between tobacco point cloud data of a current frame and tobacco point cloud data of a previous frame; Optimizing the correspondence relationship through an iterative closest point algorithm to ensure the accuracy of point cloud alignment in time sequence; The verified key feature vector is compressed and encoded, including: Using an octree structure to spatially divide the tobacco point cloud data; The point cloud statistical features within each octree node are extracted as the tobacco image features that are finally output.

4. The method for training a production motor operating frequency prediction model according to claim 1, wherein: The tobacco image features are converted into a tobacco cross-sectional area change sequence. The tobacco production stages are divided according to the tobacco cross-sectional area change sequence. The tobacco image features are then segmented according to the tobacco production stages to obtain multiple production stage information including: Converting the three-dimensional spatial coordinate data in the tobacco image features into cross-sectional profile data comprises: intercepting the three-dimensional space coordinate data through a preset detection plane to obtain a cross-section point set; Performing height-direction filtering on the cross-section point set to remove data points that exceed a set height range; The contour of the filtered cross-section point set is reconstructed, including: Projecting the cross-sectional point set onto a two-dimensional plane to obtain a plurality of projection points; Connect the projected points to form the initial outline polygon; Smoothing the initial contour polygon to obtain a final contour line; Calculate the cross-sectional area at each time point, including: Applying an area calculation formula to the final contour line to generate an area sequence; The area values at adjacent time points are differentially calculated to obtain the area change rate series; The production stages are divided according to the area change rate sequence, including: Set the judgment thresholds for each stage, including the starting threshold, rising threshold, stable threshold, and declining threshold; When the area change rate continuously exceeds the threshold of the corresponding stage, a stage transition time point is marked; Segmenting the tobacco image features according to production stages includes: Divide the data time period according to the phase transition time points; Extracting the tobacco image features in each time period and adding stage labels; Data overlapping windows in the phase transition region are established to maintain feature continuity.

5. The method for training a production motor operating frequency prediction model according to claim 1, wherein: Extracting the cross-sectional volatility of the production stage information, extracting the production stage information whose cross-sectional volatility falls within the preset volatility threshold, recorded as pre-selected stage information, includes: Calculate cross-sectional volatility within each production stage, including: Obtaining a tobacco cross-sectional area change sequence in the production stage information; The area change sequence is subjected to sliding window processing, and the area fluctuation rate in each window is calculated and expressed by formula (5). The formula (5) is as follows: ； In formula (5), For the moment The cross-sectional area fluctuation rate at is the length of the sliding window, represents the degree of freedom adjustment term, For the The cross-sectional area at each time point, is the average area within the window; Determine the preset fluctuation threshold range, including: Calculate the volatility distribution characteristics during the historical normal production phase and set the upper and lower volatility thresholds; Screening pre-selection stage information, including: Compare the area fluctuation rate in each production stage with the preset fluctuation threshold range, and record the data segment corresponding to the area fluctuation rate within the preset fluctuation threshold range as a valid data segment; The consecutive valid data segments are merged to generate pre-selection stage information.

6. The production motor operating frequency prediction model training method according to claim 1, characterized in that: When collecting the tobacco image group, the motor operation information is collected synchronously, and the motor operation information in the time segment to which the preselected stage information belongs is recorded as preselected motor information, including: The operating parameters output by the motor controller are read in real time through the industrial bus interface to generate motor operating information, wherein the operating parameters include operating frequency, current value and speed; Time-aligning the motor operation information with the tobacco image group, and performing interpolation compensation on the motor operation information whose time deviation exceeds a preset threshold; Filter the motor operation information according to the time period of the pre-selected stage information, including: Determining the starting time point and the ending time point of the pre-selection stage information; Extracting data within a time segment corresponding to the preselected stage information from the time-aligned motor operation information; Eliminate abnormal data that does not meet the operating parameter threshold within the time period; The filtered motor operation information is marked as pre-selected motor information, including: Adding a corresponding preselected stage information identifier to the preselected motor information; An associated mapping relationship between the pre-selected motor information and the pre-selected stage information is established.

7. The method for training a production motor operating frequency prediction model according to claim 1, wherein: The tobacco cross-sectional area change sequence for obtaining pre-selection stage information, denoted as the pre-selection area change sequence, includes: Obtaining cross-sectional area data at each time point in the preselected stage information; Perform smoothing and filtering on the cross-sectional area data; Verifying the continuity and integrity of the cross-sectional area data to obtain the tobacco cross-sectional area change sequence; A preselected stage identifier is added to the tobacco cross-sectional area change sequence to obtain a preselected area change sequence.

8. The method for training a production motor operating frequency prediction model according to claim 1, wherein: The radial basis function is set as the kernel function and is expressed by formula (6), which is as follows: ； in, is the signal variance parameter, is the length scale parameter, is the noise variance parameter, is the Kroneckerdelta function, For the input feature vectors, For the input feature vectors; Multi-scale feature extraction is performed on the pre-selected area change sequence to obtain the trend characteristics and local fluctuation characteristics of the area change, which are expressed by formula (7). Formula (7) is as follows: ； in, For the moment The trend characteristics of For the moment The fluctuation characteristics of For the The cross-sectional area at each time point, For dynamic window size.

9. The method for training a production motor operating frequency prediction model according to claim 1, wherein: The trend features and local fluctuation features are used as input features, and the pre-selected running frequency is used as the output label to train the basic prediction model. The training includes: Construct a multi-scale feature fusion input vector, including: performing normalization processing on the trend feature to obtain a normalized trend feature vector; Normalizing the local fluctuation characteristics to obtain a normalized fluctuation characteristic vector; The standardized trend feature vector and the normalized fluctuation feature vector are combined using the feature cross method to generate a combined feature item to form a fusion feature matrix; Create a training sample set, including: Dividing 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.

10. The production motor operating frequency prediction model training method according to claim 1, characterized in that: By maximizing the marginal likelihood function to optimize the hyperparameters in the Gaussian process and updating the parameter information of the basic prediction model, the final prediction model is obtained, 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, initialize the hyperparameter search space and optimize the constraints; Perform hyperparameter optimization, including: The conjugate gradient method is used to iteratively optimize the marginal likelihood function; The length scale parameter and signal variance parameter are updated in each iteration; Monitor the changing trend and convergence status of the likelihood function value; Verify the hyperparameter optimization results, including: Check the physical plausibility of optimized hyperparameters; Verify the positive definiteness of the kernel function matrix; Evaluate the stability of the gradient descent process; Update basic forecast model parameters, including: Inject the optimized hyperparameters into the Gaussian process model; Recalculate the posterior distribution of the training data; Update the model's memory representation and prediction interface; Generate the final prediction model, including: Solidify the optimized model parameters and structure; Save metadata required for model deployment; Establish model version management and rollback mechanism.

Images