A big data-based pear denuding and clamping intelligent control method and system

Abstract

The application relates to the technical field of agricultural product processing control, and discloses a pear denuding and clamping intelligent control method and system based on big data, which realizes continuous iteration of system performance through surface characteristic self-adaptive compensation; through double-waveband diffuse reflection detection and identification of pear fruit surface states, after matching corresponding friction coefficient benchmarks, the clamping force benchmark value and the clamping surface pressure distribution are corrected by a friction compensation branch, the friction coefficient fluctuation in low friction scenarios such as waxing and water carrying can be effectively offset, and the clamping positioning consistency and processing reliability are improved; meanwhile, after each processing is completed, the system automatically feeds back the single fruit characteristics, the actual clamping parameters and the processing results to a pear clamping big database, and the prediction model precision is continuously optimized through regular incremental training. The data closed-loop mechanism can enable the system to independently improve the adaptation capability to special working conditions and new variety fruits as the production scale expands, and realize continuous independent evolution of system performance.

Description

Technical Field

[0001] This invention relates to the technical field of agricultural product processing control, specifically to a big data-based intelligent control method and system for pear pitting and clamping. Background Technology

[0002] With the large-scale development of the pear deep processing industry, pitting has become a core process in the processing of pear canned goods, dried pears, pear juice, and other products. The clamping process, as a crucial pre-pitting step, directly determines the pitting coaxiality, residual pit rate, and finished product yield based on its positioning accuracy and clamping stability. Traditional pear pitting clamping methods mostly use fixed tooling fixtures, relying on uniform mechanical parameters to clamp batches of fruit. This results in poor overall adaptability and adjustment flexibility, making it difficult to meet the demands of high-quality, highly automated deep processing production. As a typical differentiated agricultural product, pears exhibit significant multi-dimensional differences between individuals, posing three core technical challenges to clamping control: First, the large variation in size and shape makes universal adaptation difficult. Mainstream varieties such as Ya pear, Xue pear, Xiang pear, and Dangshan pear have large diameter ranges and generally exhibit eccentricity, not being standard spheres. A single clamp and fixed clamping parameters cannot achieve full-size compatibility, easily leading to clamping errors. First, misalignment leads to poor coaxiality of pitting and a high rate of residual pits. Second, inconsistent maturity and firmness make it difficult to uniformly match the clamping force: due to the influence of maturity and storage time, the firmness of the flesh of the same batch of fruit may vary by a factor of two. There is an irreconcilable contradiction in uniform clamping force parameters. If the clamping force is too strong on the soft fruit with high maturity, it will be crushed and trapped. If the clamping force is insufficient on the hard fruit with low maturity, it will be unstable and slip. Third, the surface characteristics are variable, resulting in insufficient clamping reliability: pears can be in various states before processing, such as with the peel, with wax for preservation, or with water after washing. The coefficient of friction of the peel fluctuates significantly under different states, which further aggravates the problems of clamping slippage and misalignment.

[0003] Chinese invention patent application CN120848220A, published on October 28, 2025, discloses a control method and system for an intelligent pitting device. The method involves acquiring image information of fruit through a vision acquisition device, identifying the external features of the fruit in the image information, and obtaining a prediction of the fruit type and hardness. Based on the prediction of the fruit type and hardness, the method analyzes the constraint forces for clamping and pitting, and performs target balancing configuration based on the clamping constraint force and the pitting constraint force to determine the cooperative control parameters. Control is executed according to the cooperative control parameters, and execution feedback data is tracked. Adaptive analysis is performed based on the execution feedback data to identify execution compensation adjustment parameters. The cooperative control parameters are dynamically adjusted using the execution compensation adjustment parameters to complete the pitting operation control. This method solves the technical problem of existing pitting device control failing to adapt to different fruit characteristics, resulting in poor pitting effects. It achieves precise pitting of fruits with different characteristics, thereby improving the quality of juice. However, the above technical solutions can only mechanically control the pitting clamping force based on fruit type. This method aims to intelligently and precisely control the pitting clamping force considering the external size, hardness, and surface friction characteristics of the fruit. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the three core technical challenges in clamping control mentioned above: First, the large variation in fruit size and shape makes universal adaptation difficult: mainstream varieties such as Ya pear, Xue pear, Xiang pear, and Dangshan pear have a wide range in single fruit diameter and generally exhibit eccentricity, not being standard spheres. A single clamp and fixed clamping parameters cannot achieve full-size compatibility, easily leading to clamping misalignment, resulting in poor coaxiality of pitting and a high rate of residual pits. Second, inconsistent maturity and firmness make it difficult to uniformly match clamping force: affected by maturity and storage time, the firmness of the flesh of the same batch of fruit may vary by as much as 200%. There is an irreconcilable contradiction in uniformly setting clamping force parameters. Excessive clamping force on highly mature, soft fruit will damage and sink the fruit, while insufficient clamping force on less mature, firm fruit will result in instability and slippage. Thirdly, the surface characteristics are highly variable, leading to insufficient clamping reliability: pears exist in various states before processing, such as with peel, after waxing for preservation, or with water after washing. The coefficient of friction of the peel fluctuates significantly under different states, further exacerbating the problems of clamping slippage and positioning deviation. The above-mentioned solution is to achieve precise matching of single-fruit clamping parameters based on multi-source sensing and quantification of the geometric, mechanical, and surface characteristics of pears, and relying on a big data multi-task model. Combined with closed-loop correction with safety constraints and anti-impact smooth control, it simultaneously solves the three major industry pain points of poor fruit shape adaptation, clamping force mismatch, and slippage and damage, which can significantly reduce the scrap rate in the pitting process. The data feedback iteration mechanism supports the system's autonomous evolution, achieving the goals of high adaptability, low operating loss, and continuous optimization capabilities.

[0006] (II) Technical Solution

[0007] This invention is achieved through the following technical solution: a pear pitting and clamping intelligent control method based on big data, the method comprising the following steps:

[0008] S1. At the pre-inspection station before the pear fruit is depitted and clamped, the position of the pear fruit is synchronously triggered by the positioning encoder. The visual geometric features, near-infrared spectral features and surface diffuse reflectance features of the pear fruit to be processed are collected synchronously. All data are aligned with a unified timestamp to construct a single fruit multi-source feature dataset.

[0009] S2. Based on the pre-built pear clamping database, geometric features, mechanical features, and friction features are extracted from the single fruit multi-source feature dataset, and standardization is performed to obtain a core feature vector of the pear to be clamped with uniform dimensions.

[0010] S3. Input the core feature vector into the pre-trained multi-task clamping parameter prediction model, and output the initial optimal clamping parameter set adapted to the characteristics of a single fruit through shared feature extraction and three-branch parallel inference.

[0011] S4. During the clamping process, the displacement feedback data and force feedback data of the clamping mechanism are collected in real time, compared with the initial optimal clamping parameter set item by item, the deviation is calculated and the clamping risk state is identified, and an 8-dimensional clamping state deviation matrix is ​​constructed.

[0012] S5. Input the clamping state deviation matrix into the adaptive closed-loop control algorithm with safety constraints, output real-time adjustment instructions and convert them into hardware drive signals to dynamically adjust the clamping stroke, clamping force and clamping posture to complete stable clamping before de-kerneling; at the same time, the final clamping result is correlated with the single fruit feature and sent back to the pear clamping big data database to realize database expansion and model iterative optimization.

[0013] Preferably, the pre-detection station in S1 that collects multi-source individual feature data integrates a binocular vision unit, a near-infrared spectroscopy detection unit, a surface diffuse reflection detection unit, and a positioning encoder;

[0014] The binocular vision unit is set directly above and to the side of the conveying station to collect panoramic stereoscopic images of the pear from the front and side views, and to analyze the geometric dimensions and eccentricity characteristics of the pear.

[0015] The near-infrared spectroscopy detection unit is located on the side of the conveying station. It uses a diffuse reflection probe to collect near-infrared spectral data of the pear peel and flesh to predict the flesh firmness and maturity characteristics.

[0016] The surface diffuse reflection detection unit is set on both sides of the conveying station. It identifies the surface condition of pears by emitting dual-band detection light and collecting reflectivity data, including three scenarios: native with peel, waxed, and washed with water.

[0017] The positioning encoder is installed on the drive shaft of the conveyor line and is used to detect the conveying position of the pear in real time. When the pear reaches the center of the detection station, the binocular vision unit, near-infrared spectroscopy detection unit and surface diffuse reflection detection unit are triggered to start data acquisition, ensuring that the multi-source data corresponds to the same spatial position of the same pear.

[0018] Preferably, step S2 specifically includes the following steps:

[0019] S21. Geometric feature extraction: Perform image segmentation and 3D reconstruction on the binocular vision image, extract the maximum and minimum fruit diameters of the pear fruit, and the two-dimensional centroid coordinates in the transport plane. Calculate the fruit shape eccentricity by dividing the difference between the maximum and minimum fruit diameters by the average fruit diameter, and construct a 5-dimensional geometric feature subset.

[0020] S22. Mechanical feature extraction: After performing multivariate scattering correction and first derivative preprocessing on near-infrared spectral data, the data is input into a pre-calibrated partial least squares hardness prediction model to calculate the average hardness of the pear flesh and the hardness of the peel, and to construct a 2D mechanical feature subset.

[0021] S23. Friction feature extraction: Calculate the reflectance ratio of dual-band light in the surface diffuse reflection detection data, compare the ratio with the preset surface state threshold range, and match the surface state type corresponding to the pear fruit; then query the rubber-pear skin friction coefficient benchmark value corresponding to the surface state from the pear clamping database to construct a 1-dimensional surface friction feature subset.

[0022] S24. Feature standardization processing: Based on the historical maximum and minimum values ​​of each feature dimension in the pear clamping database, range normalization processing is performed on the geometric feature subset, mechanical feature subset, and surface friction feature subset respectively, mapping all feature values ​​to the [0,1] interval, and splicing to generate an 8-dimensional core feature vector of the pear to be clamped.

[0023] Preferably, the construction method of the pear-clamping big data includes:

[0024] Pear samples covering the mainstream varieties of Ya pear, Xue pear, Xiang pear and Dangshan pear were collected, and the samples were labeled with four basic attributes: variety, fruit diameter range, maturity level and surface condition.

[0025] Through standardized clamping tests, the optimal clamping stroke, optimal initial clamping force, critical force for crushing, critical force for slippage, and coaxiality data for different samples were recorded.

[0026] The basic attributes, core feature vectors, corresponding clamping parameters, and processing results of the samples are associated and stored to form a basic database with tens of thousands of samples. After each processing is completed during the production process, the newly added single fruit data and processing results are automatically stored in the database to complete regular expansion and updates.

[0027] Preferably, the multi-task clamping parameter prediction model training process in S3 specifically includes the following steps:

[0028] S311. Extract labeled historical samples from the Pear Clamping Big Data Database and extract the core feature vectors of the samples as the training input set;

[0029] S312. Set the geometric adaptation parameters corresponding to the sample as the first training label, including the total clamping stroke, the three-dimensional coordinates of the clamping center point, and the clamping claw deflection angle, for a total of 5 dimensions.

[0030] S313. Set the mechanical control parameters corresponding to the sample as the second training label, including the initial clamping force, the critical force threshold for crushing injury, and the force threshold for slippage warning, for a total of 3 dimensions;

[0031] S314. Set the friction compensation parameters corresponding to the sample as the third training label, including clamping force friction compensation coefficient and pressure distribution balance coefficient, for a total of 2 dimensions;

[0032] S315. Construct a multi-branch joint loss function, calculate the mean squared error loss of the three branch labels respectively, and sum them according to the geometric weight of 0.4, the mechanical weight of 0.4, and the friction weight of 0.2 to obtain the total loss;

[0033] S316. Using the Adam optimizer and backpropagation algorithm, iteratively update the network weights of the shared feature extraction layer and the three fully connected branches until the total loss converges to below the preset threshold, thus completing the model training.

[0034] Preferably, the initial optimal clamping parameter set output by S3 includes three categories: geometric adaptation parameters, mechanical control parameters, and friction compensation parameters.

[0035] The geometric adaptation parameters are used to match the size and shape eccentricity of the pear fruit, determine the initial clamping position and opening and closing stroke, and adapt to the discrete differences in the shape of multiple varieties of pear fruit.

[0036] The mechanical control parameters are used to match the hardness of the pear flesh, set the initial value and safety boundary of the clamping force, and balance the needs of preventing soft fruit from being crushed and hard fruit from slipping.

[0037] The friction compensation parameters are used to match the surface condition of the pear, correct the clamping force reference value and the pressure distribution of the clamping surface, and offset the influence of friction coefficient fluctuations on clamping stability.

[0038] Preferably, step S4 specifically includes the following steps:

[0039] S41. Geometric deviation calculation: Real-time clamping stroke data is collected by the displacement encoder of the clamping mechanism, and the difference between the actual clamping stroke and the target clamping stroke is calculated to obtain the stroke deviation; The X and Y position offsets of the clamping center point and the fruit centroid, as well as the clamping posture angle deviation, are calculated by collecting data through clamping positioning vision, and combined to form a 4-dimensional geometric deviation subset.

[0040] S42. Mechanical Deviation Calculation: Real-time clamping force data is collected by multi-dimensional force sensors on the clamping jaws. The difference between the actual clamping force and the initial set clamping force is calculated to obtain the clamping force deviation. The standard deviation of the clamping force of all clamping jaws is calculated to obtain the pressure distribution uniformity deviation. These are combined to form a 2D mechanical deviation subset.

[0041] S43. Clamping Risk Status Judgment: Calculate the clamping force change rate and displacement change rate within a unit control cycle. If the clamping force change rate is lower than the first set threshold and the displacement change rate is higher than the second set threshold, it is determined that there is a slippage trend. If the displacement change rate is lower than the third set threshold and the clamping force change rate is higher than the fourth set threshold, it is determined that there is a crushing trend. Generate corresponding slippage warning value and crushing warning value.

[0042] S44. Deviation Matrix Construction: The geometric deviation subset, mechanical deviation subset, slippage warning value, and indentation warning value are normalized by range and combined to generate an 8-dimensional clamping state deviation matrix.

[0043] Preferably, step S5 specifically includes the following steps:

[0044] S51. Construct an adaptive fuzzy PID control algorithm with safety constraints. Define the algorithm input as the clamping state deviation matrix and define the algorithm output as three types of continuous control quantities: stroke adjustment, clamping force adjustment, and clamping jaw angle adjustment.

[0045] S52. Set safety constraint judgment rules, and the safety constraint priority is higher than the PID adjustment logic: Read the current clamping force value in real time. If the current clamping force reaches the critical force threshold for crushing the corresponding pear, the first-level safety constraint is triggered, the clamping action is forcibly stopped and the position is locked. If a slippage trend is detected, the friction compensation logic is triggered, and the clamping force is increased stepwise according to the friction compensation coefficient, and the clamping force does not exceed the slippage warning force threshold.

[0046] S53. Input the clamping state deviation matrix at the current moment into the adaptive fuzzy PID control algorithm, and calculate the real-time adjustment command that minimizes the deviation under the premise of satisfying the safety constraint rules.

[0047] S54. The adjustment command is converted into a pulse control signal for the stepper motor and a torque control signal for the servo motor, respectively, to drive the clamping mechanism to complete the dynamic adjustment of stroke, clamping force and posture.

[0048] S55. After the clamping and pitting process is completed, the core characteristics of the single fruit processed this time, the actual clamping parameters, and the clamping result data are correlated and sent back to the pear clamping big data database.

[0049] Preferably, before performing dynamic adjustment in step S54, an anti-impact smoothing process is further included, specifically:

[0050] S541. Read the current values ​​of the clamping parameters from the previous control cycle, including the current stroke, current clamping force, and current clamping angle.

[0051] S542. Calculate the difference between the target parameter value corresponding to the adjustment command and the current value of the corresponding parameter respectively to obtain three types of adjustment differences;

[0052] S543. Determine whether the three types of adjustment differences exceed the preset maximum single adjustment step size. If a certain type of adjustment difference exceeds the corresponding maximum single adjustment step size, then the actual adjustment amount of that type issued this time will be truncated to the maximum single adjustment step size; if it does not exceed the limit, then the original adjustment instruction will be issued.

[0053] A big data-based intelligent control system for pear pitting and clamping, used to implement the big data-based intelligent control method for pear pitting and clamping, the system comprising:

[0054] The multi-source fruit sensing module is installed at the pre-inspection conveying station of the pitting equipment. It integrates a binocular vision unit, a near-infrared spectroscopy detection unit, a surface diffuse reflection detection unit, and a positioning encoder. It is responsible for simultaneously collecting geometric, mechanical, and surface-related multi-source individual characteristic data of the pears to be processed.

[0055] The big data edge computing module has a built-in storage unit and a computing unit. It stores a large database of pear clamping and a pre-trained multi-task clamping parameter prediction model. It is responsible for feature extraction, standardization processing, initial clamping parameter prediction and clamping state deviation matrix calculation.

[0056] The adaptive clamping control module has a built-in adaptive fuzzy PID control algorithm with safety constraints. It is responsible for receiving the clamping state deviation matrix, generating real-time adjustment instructions that meet the safety rules, and converting them into hardware drive signals.

[0057] The core-free clamping execution module includes an adaptive multi-jaw clamping mechanism, a multi-dimensional force sensor, a displacement encoder, and a drive unit. It is responsible for receiving drive signals to complete the clamping action and providing real-time feedback of displacement and force data.

[0058] (III) Beneficial Effects

[0059] This invention provides a pear pitting and clamping intelligent control method and system based on big data. It has the following beneficial effects:

[0060] I. Significantly Improved Equipment Versatility Through Full-Dimensional Fruit Shape Adaptive Compatibility: By acquiring 3D images of pears using a binocular vision unit and extracting geometric features such as fruit diameter, eccentricity, and 2D centroid, and relying on a large database of pear clamping samples covering multiple varieties and specifications, combined with the geometric parameter branch of a multi-task clamping parameter prediction model, customized output of clamping stroke, clamping center, and clamping posture parameters can be generated for each individual fruit. Alignment deviations are then corrected in real-time through a displacement feedback closed loop. This technology can adapt to various pear varieties with a wide range of fruit diameters, accurately matching the eccentricity characteristics of different fruit shapes. It solves the adaptation problem of traditional fixed clamps that rely on a single clamp for all fruit types, eliminating the need for frequent tooling adjustments due to variety changes. It supports continuous processing in mixed batches, significantly improving equipment versatility and pitting alignment accuracy.

[0061] Second, by precisely and dynamically matching the clamping force, the processing loss rate is effectively reduced. Near-infrared spectroscopy is used for non-destructive testing of fruit flesh hardness and maturity. The mechanical parameter branches of the multi-task prediction model output the initial clamping force, critical crushing force, and slippage warning force threshold. During clamping, a real-time force feedback closed loop is constructed based on multi-dimensional force sensors. Combined with an adaptive control algorithm with safety constraints, the clamping force is dynamically adjusted. Simultaneously, an anti-impact smoothing mechanism is added to avoid instantaneous force impacts. This technical solution can match differentiated clamping force ranges for pears of different maturity levels, achieving stable clamping within the critical crushing threshold. This avoids crushing damage to highly mature soft fruits and prevents slippage in low-maturity hard fruits, resolving the contradiction of soft fruits being easily damaged and hard fruits being easily slipped under uniform parameters. This significantly reduces the scrap rate caused by poor clamping.

[0062] Third, through adaptive compensation of surface characteristics, the system achieves continuous iteration of performance. By identifying the surface state of the pear fruit through dual-band diffuse reflection detection and matching it with the corresponding friction coefficient benchmark, the friction compensation branch corrects the clamping force benchmark value and the pressure distribution on the clamping surface. This effectively offsets friction coefficient fluctuations in low-friction scenarios such as waxing and water application, improving clamping positioning consistency and processing reliability. Simultaneously, after each processing cycle, the system automatically correlates and transmits the individual fruit characteristics, actual clamping parameters, and processing results back to the pear clamping database. Regular incremental training continuously optimizes the accuracy of the prediction model. This data closed-loop mechanism allows the system to autonomously improve its adaptability to special working conditions and new fruit varieties as production scales up, achieving continuous autonomous evolution of system performance. Attached Figure Description

[0063] Figure 1 A flowchart of a pear pitting and clamping intelligent control method based on big data provided by the present invention;

[0064] Figure 2 This is a schematic diagram of a pear pitting and clamping intelligent control system based on big data, provided by the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] An embodiment of a pear pitting and clamping intelligent control method and system based on big data is as follows:

[0067] Example 1:

[0068] Please see Figure 1 A big data-based intelligent control method for pear pitting and clamping, comprising the following steps:

[0069] S1. At the pre-inspection station before the pear fruit is depitted and clamped, the position of the pear fruit is synchronously triggered by the positioning encoder. The visual geometric features, near-infrared spectral features and surface diffuse reflectance features of the pear fruit to be processed are collected synchronously. All data are aligned with a unified timestamp to construct a single fruit multi-source feature dataset.

[0070] The pre-detection station in S1, which collects multi-source individual feature data, integrates a binocular vision unit, a near-infrared spectroscopy detection unit, a surface diffuse reflection detection unit, and a positioning encoder.

[0071] The binocular vision unit is set above and to the side of the conveyor station to collect panoramic stereo images of the pears from the front and side views, and to analyze the geometric size and eccentricity characteristics of the pears.

[0072] Specifically, the binocular vision unit consists of two 5-megapixel global shutter industrial cameras, which are vertically arranged above the conveyor station and horizontally arranged to the side of the conveyor station, respectively, with their optical axes perpendicular to each other. A ring-shaped diffused white light source is installed to eliminate highlights and shadows on the surface of the pear. When the pear enters the detection area, the two cameras are exposed simultaneously to acquire RGB images with a resolution of 2448×2048, which are used for subsequent 3D contour reconstruction and geometric parameter calculation.

[0073] The near-infrared spectroscopy detection unit is set on the side of the conveyor station. It uses a diffuse reflection probe to collect near-infrared spectral data of the pear peel and flesh to predict the flesh firmness and maturity characteristics.

[0074] Specifically, the near-infrared spectroscopy detection unit uses a miniature fiber optic spectral probe in the 900nm-1700nm band. The probe is positioned directly above the equator of the pear fruit, with a fixed detection distance of 20mm and an integration time of 20ms. During data acquisition, the spectral probe is in close contact with the pear fruit skin to emit near-infrared light, receives diffuse reflection light signals from the pulp tissue, and outputs continuous spectral data for inverting the internal firmness and maturity indicators of the pulp.

[0075] The surface diffuse reflection detection unit is set on both sides of the conveying station. It identifies the surface condition of pears by emitting dual-band detection light and collecting reflectivity data, including three scenarios: native with peel, waxed, and washed with water.

[0076] Specifically, the surface diffuse reflection detection unit includes two symmetrically arranged photoelectric detection components. Each component integrates a 650nm visible light source, a 940nm near-infrared light source, and a corresponding photoelectric receiver. Detection light is emitted vertically from both sides of the conveyor line onto the pear surface at the same time. The intensity of the reflected light is collected and the reflectivity value is calculated. The surface condition of the pear is distinguished by the ratio of the two-band reflectivity.

[0077] The positioning encoder is installed on the drive shaft of the conveyor line to detect the conveying position of the pears in real time. When the pears reach the center of the detection station, the binocular vision unit, near-infrared spectroscopy detection unit and surface diffuse reflection detection unit are triggered to start data acquisition, ensuring that the multi-source data corresponds to the same spatial position of the same pear.

[0078] Specifically, the positioning encoder is an incremental rotary encoder, coaxially mounted on the drive roller shaft of the conveyor line, with a resolution of 1000 pulses / revolution. It outputs pulse signals in real time as the roller rotates, accurately reflecting the conveying displacement of the pears. The system pre-calibrates the pulse threshold corresponding to the center of the detection station. When the pulse count reaches the threshold, it outputs a synchronous trigger signal to simultaneously send to the binocular vision, near-infrared spectroscopy, and surface diffuse reflectance units, ensuring that all three start collecting data at the same time. The system uses the trigger signal time as a reference to stamp all collected data with a unified millisecond-level timestamp, binding image, spectral, and reflectance data with the fruit position coordinates to generate a multi-source feature dataset corresponding to a single pear, eliminating the temporal misalignment of multi-sensor data.

[0079] S2. Based on the pre-built pear clamping database, geometric features, mechanical features, and friction features are extracted from the single fruit multi-source feature dataset, and standardization is performed to obtain a core feature vector of the pear to be clamped with uniform dimensions.

[0080] S2 specifically includes the following steps:

[0081] S21. Geometric feature extraction: Perform image segmentation and 3D reconstruction on the binocular vision image, extract the maximum and minimum fruit diameters of the pear fruit, and the two-dimensional centroid coordinates in the transport plane. Calculate the fruit shape eccentricity by dividing the difference between the maximum and minimum fruit diameters by the average fruit diameter, and construct a 5-dimensional geometric feature subset.

[0082] Specifically, in embodiment S21, the system first performs threshold segmentation and edge extraction on the binocular image, and completes the three-dimensional contour reconstruction of the pear fruit through a binocular stereo matching algorithm; extracts the maximum transverse diameter, minimum transverse diameter and centroid coordinates in the X and Y directions in the transport plane from the three-dimensional contour, and then calculates the fruit shape eccentricity using the formula: fruit shape eccentricity = (maximum fruit diameter - minimum fruit diameter) / average fruit diameter, where the average fruit diameter is the arithmetic mean of the maximum fruit diameter and the minimum fruit diameter; finally, the maximum fruit diameter, minimum fruit diameter, centroid X coordinate, centroid Y coordinate, and fruit shape eccentricity constitute a 5-dimensional geometric feature subset to quantify the size specifications and irregularity of the pear fruit.

[0083] S22. Mechanical feature extraction: After performing multivariate scattering correction and first derivative preprocessing on near-infrared spectral data, the data is input into a pre-calibrated partial least squares hardness prediction model to calculate the average hardness of the pear flesh and the hardness of the peel, and to construct a 2D mechanical feature subset.

[0084] Specifically, in Example S22, multivariate scattering correction is first performed on the original near-infrared spectral data to eliminate the difference in scattered light intensity caused by the unevenness of the pear surface; then, first-order derivative processing is performed to eliminate the interference of spectral baseline drift and highlight the characteristic absorption peaks of the pulp tissue; the preprocessed spectral data is input into a pre-calibrated partial least squares (PLS) hardness prediction model, which is pre-calibrated through hundreds of standard hardness test samples and can output two values: average pulp hardness and peel hardness, in kg / cm². The two values ​​constitute a 2D mechanical feature subset to quantify the mechanical load-bearing characteristics of the pear.

[0085] S23. Friction feature extraction: Calculate the reflectance ratio of dual-band light in the surface diffuse reflection detection data, compare the ratio with the preset surface state threshold range, and match the surface state type corresponding to the pear fruit; then query the rubber-pear skin friction coefficient benchmark value corresponding to the surface state from the pear clamping database to construct a 1-dimensional surface friction feature subset.

[0086] Specifically, in embodiment S23, the reflectance ratio of 940nm and 650nm wavelength light is calculated, and three threshold ranges are preset: a reflectance ratio > 1.2 is judged as a waxed state, 0.8 ≤ ratio ≤ 1.2 is judged as a raw state with peel, and a ratio < 0.8 is judged as a washed and wet state. After matching the surface state, the friction coefficient benchmark value between the rubber clamping pad and the pear peel under the corresponding state is retrieved from the pear clamping database to form a 1-dimensional surface friction feature subset, thereby quantifying the friction characteristics of the pear surface.

[0087] S24. Feature standardization processing: Based on the historical maximum and minimum values ​​of each feature dimension in the pear clamping database, range normalization processing is performed on the geometric feature subset, mechanical feature subset, and surface friction feature subset respectively, mapping all feature values ​​to the [0,1] interval, and splicing to generate an 8-dimensional core feature vector of the pear to be clamped.

[0088] Specifically, in embodiment S24, the historical maximum and minimum values ​​of each of the eight feature dimensions in the pear clamp database are retrieved. The range normalization formula is used: standardized value = (actual value - minimum value of dimension) / (maximum value of dimension - minimum value of dimension). All features are calculated one by one, and features of different dimensions are uniformly mapped to the [0,1] interval. They are then concatenated in sequence to generate an 8-dimensional core feature vector, eliminating the interference of dimension differences on subsequent model predictions.

[0089] The construction methods of Pear Clip's big data include:

[0090] Pear samples covering the mainstream varieties of Ya pear, Xue pear, Xiang pear and Dangshan pear were collected, and the samples were labeled with four basic attributes: variety, fruit diameter range, maturity level and surface condition.

[0091] Through standardized clamping tests, the optimal clamping stroke, optimal initial clamping force, critical force for crushing, critical force for slippage, and coaxiality data for different samples were recorded.

[0092] The basic attributes, core feature vectors, corresponding clamping parameters, and processing results of the samples are associated and stored to form a basic database with tens of thousands of samples. After each processing is completed during the production process, the newly added single fruit data and processing results are automatically stored in the database to complete regular expansion and updates.

[0093] Specifically, the pear clamping database is constructed as follows: samples of four mainstream pear varieties—Ya pear, Xue pear, Xiang pear, and Dangshan pear—are collected in advance, covering the entire fruit diameter range, three maturity levels, and three surface conditions, with a total sample size of ≥10,000 groups; the optimal clamping stroke, optimal initial clamping force, critical force for crushing, critical force for slippage, and coaxiality for core removal are determined for each sample through standardized clamping tests; the basic attributes, core feature vectors, corresponding clamping parameters, and processing results of the samples are associated and stored to form a basic database; during the production process, after each pear is processed, the system automatically adds new data to the database, and incremental model training is conducted once a month to achieve dynamic expansion and accuracy iteration of the database.

[0094] By achieving full-dimensional fruit shape adaptive adaptation, the equipment's versatility is significantly improved. It acquires 3D images of pears using a binocular vision unit and extracts geometric features such as fruit diameter, eccentricity, and 2D centroid. Leveraging a large database of pear clamping samples covering multiple varieties and specifications, and combining the geometric parameter branches of a multi-task clamping parameter prediction model, it can customize clamping stroke, clamping center, and clamping posture parameters for each individual fruit. Furthermore, it uses a displacement feedback closed-loop system to correct alignment deviations in real time. This technology is adaptable to various pear varieties with a wide fruit diameter range, accurately matching the eccentricity characteristics of different fruit shapes. It solves the adaptation problem of traditional fixed clamps that rely on a single clamp for all fruit types, eliminating the need for frequent tooling adjustments due to variety changes. It supports continuous processing in mixed batches, significantly improving equipment versatility and pitting alignment accuracy.

[0095] S3. Input the core feature vector into the pre-trained multi-task clamping parameter prediction model, and output the initial optimal clamping parameter set that adapts to the characteristics of a single fruit through shared feature extraction and three-branch parallel inference.

[0096] The training process of the multi-task clamping parameter prediction model in S3 specifically includes the following steps:

[0097] S311. Extract labeled historical samples from the Pear Clamping Big Data Database and extract the core feature vectors of the samples as the training input set;

[0098] S312. Set the geometric adaptation parameters corresponding to the sample as the first training label, including the total clamping stroke, the three-dimensional coordinates of the clamping center point, and the clamping claw deflection angle, for a total of 5 dimensions.

[0099] S313. Set the mechanical control parameters corresponding to the sample as the second training label, including the initial clamping force, the critical force threshold for crushing injury, and the force threshold for slippage warning, for a total of 3 dimensions;

[0100] S314. Set the friction compensation parameters corresponding to the sample as the third training label, including clamping force friction compensation coefficient and pressure distribution balance coefficient, for a total of 2 dimensions;

[0101] S315. Construct a multi-branch joint loss function, calculate the mean squared error loss of the three branch labels respectively, and sum them according to the geometric weight of 0.4, the mechanical weight of 0.4, and the friction weight of 0.2 to obtain the total loss;

[0102] S316. Using the Adam optimizer and backpropagation algorithm, iteratively update the network weights of the shared feature extraction layer and the three fully connected branches until the total loss converges to below the preset threshold, thus completing the model training.

[0103] Specifically, the constructed multi-task network includes an input layer (receiving an 8-dimensional feature vector), a shared feature extraction layer (two fully connected layers with 64 and 32 neurons respectively, using the ReLU activation function), and three parallel output branches: the geometric parameter branch is a 5-neuron fully connected layer that outputs clamping stroke and position / pose parameters; the mechanical parameter branch is a 3-neuron fully connected layer that outputs clamping force and critical threshold; and the friction compensation branch is a 2-neuron fully connected layer that outputs compensation coefficients. During training, 10,000 labeled samples are extracted from the database and divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer is used, with an initial learning rate set to 0.001. The joint loss function is composed of the weighted average squared error loss of the three branches. When the total loss of the validation set does not decrease for 10 consecutive iterations, an early stopping mechanism is triggered to solidify the model weights. Finally, the model's single inference time is less than 10ms, meeting the production line cycle time requirements.

[0104] The initial optimal clamping parameter set output by S3 includes three categories: geometric adaptation parameters, mechanical control parameters, and friction compensation parameters.

[0105] Geometric adaptation parameters are used to match the size and shape eccentricity of pears, determine the initial clamping position and opening / closing stroke, and adapt to the shape dispersion differences of multiple pear varieties.

[0106] Mechanical control parameters are used to match the hardness of pear flesh, set the initial value and safety boundary of clamping force, and balance the needs of preventing soft fruit from being crushed and hard fruit from slipping.

[0107] Friction compensation parameters are used to match the surface condition of the pear, correct the clamping force reference value and the pressure distribution of the clamping surface, and offset the influence of friction coefficient fluctuations on clamping stability.

[0108] Specifically, in actual operation, the 8-dimensional core feature vector is first input into the shared feature extraction layer. After completing the deep feature mapping, it is divided into three paths and enters the corresponding branches. The geometric branch outputs the total clamping stroke, the three-dimensional coordinates of the clamping center point, and the deflection angle of the clamping claw to ensure that the clamping mechanism is accurately aligned with the fruit's centroid. The mechanical branch outputs three parameters: initial clamping force, critical crushing force, and slippage warning force, to define the safe clamping range. The friction branch outputs the clamping force friction compensation coefficient and pressure distribution balance coefficient to correct the clamping reliability under different surface conditions. The three types of parameters are output simultaneously to form the initial optimal clamping parameter set adapted to the single fruit.

[0109] By precisely and dynamically matching the clamping force, the processing loss rate is effectively reduced. Near-infrared spectroscopy is used for non-destructive testing of fruit flesh hardness and maturity. The mechanical parameter branches of the multi-task prediction model output the initial clamping force, critical crushing force, and slippage warning force threshold. During clamping, a real-time force feedback closed loop is constructed based on multi-dimensional force sensors. Combined with an adaptive control algorithm with safety constraints, the clamping force is dynamically adjusted. An anti-impact smoothing mechanism is also added to avoid instantaneous force impacts. This technical solution can match differentiated clamping force ranges for pears of different maturity levels, achieving stable clamping within the critical crushing threshold. This avoids crushing of highly mature soft fruits and prevents slippage of low-maturity hard fruits, resolving the contradiction of soft fruits being easily damaged and hard fruits being easily slipped under uniform parameters. This significantly reduces the scrap rate caused by poor clamping.

[0110] S4. During the clamping process, the displacement feedback data and force feedback data of the clamping mechanism are collected in real time, compared with the initial optimal clamping parameter set item by item, the deviation is calculated and the clamping risk state is identified, and an 8-dimensional clamping state deviation matrix is ​​constructed.

[0111] S4 specifically includes the following steps:

[0112] S41. Geometric deviation calculation: Real-time clamping stroke data is collected by the displacement encoder of the clamping mechanism, and the difference between the actual clamping stroke and the target clamping stroke is calculated to obtain the stroke deviation; The X and Y position offsets of the clamping center point and the fruit centroid, as well as the clamping posture angle deviation, are calculated by collecting data through clamping positioning vision, and combined to form a 4-dimensional geometric deviation subset.

[0113] Specifically, in embodiment S41, a displacement encoder is installed on the opening and closing transmission shaft of the clamping mechanism, with a resolution of 0.01mm, to collect the opening and closing stroke of the clamping claw in real time; an auxiliary positioning camera is arranged directly above the clamping station to collect the position image of the pear after clamping, calculate the two-dimensional offset of the clamping center point and the centroid of the fruit in the X and Y directions, and simultaneously collect the deflection angle deviation of the clamping claw; the stroke deviation, X-direction position deviation, Y-direction position deviation, and angle deviation together constitute a subset of four-dimensional geometric deviations, reflecting the degree of deviation between the clamping position and the target position.

[0114] S42. Mechanical Deviation Calculation: Real-time clamping force data is collected by multi-dimensional force sensors on the clamping jaws. The difference between the actual clamping force and the initial set clamping force is calculated to obtain the clamping force deviation. The standard deviation of the clamping force of all clamping jaws is calculated to obtain the pressure distribution uniformity deviation. These are combined to form a 2D mechanical deviation subset.

[0115] Specifically, in embodiment S42, miniature multidimensional force sensors are installed at the ends of the three gripping claws, with a sampling frequency of 50Hz, to collect the clamping force values ​​of each claw in real time; the average value of the clamping force of the three claws is taken as the actual total clamping force, and the difference between it and the initial set clamping force is used to obtain the clamping force deviation; the standard deviation of the clamping force of the three claws is calculated to obtain the pressure distribution uniformity deviation, which reflects the degree of balance of the clamping force; the two together constitute a 2D mechanical deviation subset.

[0116] S43. Clamping Risk Status Judgment: Calculate the clamping force change rate and displacement change rate within a unit control cycle. If the clamping force change rate is lower than the first set threshold and the displacement change rate is higher than the second set threshold, it is determined that there is a slippage trend. If the displacement change rate is lower than the third set threshold and the clamping force change rate is higher than the fourth set threshold, it is determined that there is a crushing trend. Generate corresponding slippage warning value and crushing warning value.

[0117] Specifically, in embodiment S43, the system uses 20ms as a control cycle to calculate the rate of change of clamping force and the rate of change of displacement within a single cycle. The first threshold (clamping force change rate threshold) is preset to 0.1kg / cm² / cycle, and the second threshold (displacement change rate threshold) is preset to 0.5mm / cycle. If the clamping force change rate is lower than the first threshold and the displacement change rate is higher than the second threshold, it indicates that the clamping force has not increased effectively but the pear fruit has shifted, and it is determined that there is a slippage trend. The third threshold (displacement change rate threshold) is preset to 0.1mm / cycle, and the fourth threshold (clamping force change rate threshold) is preset to 0.3kg / cm² / cycle. If the displacement change rate is lower than the third threshold and the clamping force change rate is higher than the fourth threshold, it indicates that the stroke has not changed significantly but the clamping force continues to increase, and the pear fruit is at risk of being crushed, and it is determined that there is a crushing trend. Finally, a slippage warning value and a crushing warning value quantified from 0 to 1 are generated. Under normal conditions, the warning value is 0, and when the risk is triggered, the warning value is 1.

[0118] S44. Deviation Matrix Construction: The geometric deviation subset, mechanical deviation subset, slippage warning value, and indentation warning value are normalized by range and combined to generate an 8-dimensional clamping state deviation matrix.

[0119] Specifically, in embodiment S44, the eight data items, including 4-dimensional geometric deviation, 2-dimensional mechanical deviation, slippage warning value, and indentation warning value, are normalized according to their ranges and mapped to the [0,1] interval. They are then combined in a fixed order to generate an 8-dimensional clamping state deviation matrix, which serves as the standardized input for subsequent closed-loop control.

[0120] S5. Input the clamping state deviation matrix into the adaptive closed-loop control algorithm with safety constraints, output real-time adjustment commands and convert them into hardware drive signals to dynamically adjust the clamping stroke, clamping force and clamping posture to complete stable clamping before de-kerneling; at the same time, the final clamping result is correlated with the single fruit features and sent back to the pear clamping big data database to realize database expansion and model iterative optimization.

[0121] S5 specifically includes the following steps:

[0122] S51. Construct an adaptive fuzzy PID control algorithm with safety constraints. Define the algorithm input as the clamping state deviation matrix and define the algorithm output as three types of continuous control quantities: stroke adjustment, clamping force adjustment, and clamping jaw angle adjustment.

[0123] Specifically, in embodiment S51, the adaptive fuzzy PID control algorithm takes an 8-dimensional deviation matrix as input and automatically adjusts the proportional, integral, and derivative parameters of the PID according to the magnitude and trend of the deviation to achieve rapid and stable convergence of the deviation; the output three types of control quantities correspond to the adjustment of clamping stroke, clamping force, and clamping posture, respectively, to achieve multi-dimensional synchronous correction.

[0124] S52. Set safety constraint judgment rules, and the safety constraint priority is higher than the PID adjustment logic: Read the current clamping force value in real time. If the current clamping force reaches the critical force threshold for crushing the corresponding pear, the first-level safety constraint is triggered, the clamping action is forcibly stopped and the position is locked. If a slippage trend is detected, the friction compensation logic is triggered, and the clamping force is increased stepwise according to the friction compensation coefficient, and the clamping force does not exceed the slippage warning force threshold.

[0125] Specifically, in embodiment S52, the safety constraint rules have a higher priority than the PID control logic; the first-level safety constraint is pressure damage protection: the current clamping force is compared with the critical pressure damage threshold in real time, and once the threshold is reached, the drive signal of the clamping servo is immediately cut off, the clamping position is locked, and clamping is prohibited from continuing, thus preventing pressure damage to the fruit flesh; the second-level constraint is slippage compensation: when a slippage trend warning is detected, the clamping force is increased stepwise in an increment of 0.2 kg / cm² per cycle according to the friction compensation coefficient, and the upper limit of the clamping force does not exceed the slippage warning force threshold, thus eliminating slippage while avoiding exceeding the clamping force limit.

[0126] S53. Input the clamping state deviation matrix at the current moment into the adaptive fuzzy PID control algorithm, and calculate the real-time adjustment command that minimizes the deviation under the premise of satisfying the safety constraint rules.

[0127] S54. The adjustment command is converted into a pulse control signal for the stepper motor and a torque control signal for the servo motor, respectively, to drive the clamping mechanism to complete the dynamic adjustment of stroke, clamping force and posture.

[0128] Specifically, in embodiment S54, the stroke and angle adjustment amounts are converted into pulse control signals for the stepper motor, and the clamping force adjustment amount is converted into torque control signals for the servo motor, which are then sent to the drive unit of the clamping actuator. Before the command is issued, anti-shock smoothing processing is performed, specifically: S541, reading the current values ​​of the clamping parameters from the previous control cycle, including the current stroke, current clamping force, and current clamping angle; S542, calculating the difference between the target parameter value corresponding to the adjustment command and the corresponding current parameter value, obtaining three types of adjustments. Difference; S543, determine whether the three types of adjustment differences exceed the preset single maximum adjustment step size. If a certain type of adjustment difference exceeds the corresponding single maximum adjustment step size, the actual adjustment amount issued this time will be cut off to the single maximum adjustment step size; if it does not exceed, the original adjustment instruction will be issued. The preset single maximum adjustment step size is: stroke 1mm, clamping force 0.5kg / cm², angle 2°. By cutting off the step size, the clamping force is suddenly increased or the fast movement of the mechanism is avoided to prevent impact damage to the pear fruit, and mechanical wear is reduced.

[0129] S55. After the clamping and pitting process is completed, the core characteristics of the single fruit processed this time, the actual clamping parameters, and the clamping result data are correlated and sent back to the pear clamping big data database.

[0130] Specifically, in embodiment S55, when all dimensions of the clamping state deviation matrix are below the stable threshold and remain stable for three consecutive control cycles, the clamping is determined to be complete and the de-coreing process begins. After de-coreing, the system associates the core features of the single fruit processed this time, the actual clamping parameters, and the clamping results (whether it is damaged, whether it slips, and whether the de-coreing is qualified) and sends them back to the pear clamping database for subsequent database expansion and incremental model training.

[0131] Through adaptive compensation of surface characteristics, the system achieves continuous iteration of performance. By identifying the pear fruit surface condition through dual-band diffuse reflection detection and matching it to the corresponding friction coefficient benchmark, the friction compensation branch corrects the clamping force benchmark value and the pressure distribution on the clamping surface. This effectively offsets friction coefficient fluctuations in low-friction scenarios such as waxing and water application, improving clamping positioning consistency and processing reliability. Simultaneously, after each processing cycle, the system automatically correlates and transmits the individual fruit characteristics, actual clamping parameters, and processing results back to the pear clamping database. Regular incremental training continuously optimizes the accuracy of the prediction model. This data closed-loop mechanism allows the system to autonomously improve its adaptability to special working conditions and new fruit varieties as production scales up, achieving continuous autonomous evolution of system performance.

[0132] For example, consider a typical scenario of a mixed-batch pitting production line in a pear canning plant;

[0133] The production line is running smoothly, the clamping mechanism is in standby mode with general reference parameters, and the conveyor line transports the pears through the workstations at a uniform speed. The fruits to be processed include Ya pears and Xiang pears, which are of uneven ripeness and have water on their surface after washing.

[0134] At T=0s, a pear enters the pre-inspection station along the conveyor line. The positioning encoder pulse count reaches the trigger threshold, simultaneously triggering all sensors to collect data: binocular vision measures the maximum fruit diameter to be 75mm, the minimum fruit diameter to be 68mm, and the eccentricity to be 0.12; near-infrared spectroscopy predicts the average flesh hardness to be 4.2kg / cm²; the surface diffuse reflectance detection shows a reflectance ratio of 0.72, indicating that the pear is in a washed and wet state; the system adds a unified timestamp to all data, constructing a single-fruit multi-source feature dataset for this pear.

[0135] At T=0.1s, the system completes the extraction and normalization of three types of features, generating an 8-dimensional core feature vector, which is then input into the multi-task clamping parameter prediction model. After one forward inference, the model outputs the initial optimal clamping parameter set: clamping stroke 68mm, clamping center point corresponding to the fruit's centroid; initial clamping force 3.0kg / cm², critical crushing force 4.0kg / cm², slippage warning force 3.8kg / cm²; friction compensation coefficient 1.2.

[0136] At T=0.15s, the clamping mechanism performs the clamping action according to the initial parameters. During the clamping process, the force sensor reports an actual clamping force of 2.8kg / cm². The system detects a clamping force change rate of 0.05kg / cm² and a displacement change rate of 0.8mm within a single cycle, which meets the slippage trend judgment conditions. The slippage warning value is set to 1, and the system constructs an 8-dimensional clamping state deviation matrix.

[0137] At T=0.2s, the adaptive fuzzy PID control algorithm receives the deviation matrix and simultaneously triggers the friction compensation logic, outputting a clamping force increase command. After shock-resistant smoothing, the system gradually increases the clamping force in two control cycles: 0.4 kg / cm² in the first cycle and 0.2 kg / cm² in the second cycle, finally stabilizing the clamping force at 3.4 kg / cm².

[0138] At T=0.28s, the slippage trend is eliminated, all dimensions of the clamping state deviation matrix are below the stable threshold, and remain stable for 3 consecutive cycles. The system determines that the clamping is complete, and the denuclearization station starts the denuclearization process.

[0139] T=2s, the pitting process is completed. The system determines that there is no pressure damage, no slippage, and the pitting coaxiality is qualified. All characteristics of the pear, the clamping parameters and the processing results are associated and transmitted back to the pear clamping database.

[0140] Subsequently, after the small-diameter fragrant pears (55mm in diameter, 3.0kg / cm² in hardness, and with original skin) are conveyed into the workstation, the system automatically outputs smaller clamping strokes and lower clamping force parameters. No manual adjustment of the tooling is required throughout the process, enabling continuous adaptive processing of mixed batches.

[0141] Example 2:

[0142] Please see Figure 2 A pear pitting and clamping intelligent control system based on big data, used to implement a pear pitting and clamping intelligent control method based on big data, the system includes:

[0143] The multi-source fruit sensing module is installed at the pre-inspection conveying station of the pitting equipment. It integrates a binocular vision unit, a near-infrared spectroscopy detection unit, a surface diffuse reflection detection unit, and a positioning encoder. It is responsible for simultaneously collecting geometric, mechanical, and surface-related multi-source individual characteristic data of the pears to be processed.

[0144] The big data edge computing module has a built-in storage unit and a computing unit. It stores a large database of pear clamping and a pre-trained multi-task clamping parameter prediction model. It is responsible for feature extraction, standardization processing, initial clamping parameter prediction and clamping state deviation matrix calculation.

[0145] The adaptive clamping control module has a built-in adaptive fuzzy PID control algorithm with safety constraints. It is responsible for receiving the clamping state deviation matrix, generating real-time adjustment instructions that meet the safety rules, and converting them into hardware drive signals.

[0146] The core-free clamping execution module includes an adaptive multi-jaw clamping mechanism, a multi-dimensional force sensor, a displacement encoder, and a drive unit. It is responsible for receiving drive signals to complete the clamping action and providing real-time feedback of displacement and force data.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pear pitting and clamping intelligent control method based on big data, characterized in that, The method includes the following steps: S1. At the pre-inspection station before the pear fruit is depitted and clamped, the position of the pear fruit is synchronously triggered by the positioning encoder. The visual geometric features, near-infrared spectral features and surface diffuse reflectance features of the pear fruit to be processed are collected synchronously. All data are aligned with a unified timestamp to construct a single fruit multi-source feature dataset. S2. Based on the pre-built pear clamping database, geometric features, mechanical features, and friction features are extracted from the single fruit multi-source feature dataset, and standardization is performed to obtain a core feature vector of the pear to be clamped with uniform dimensions. S3. Input the core feature vector into the pre-trained multi-task clamping parameter prediction model, and output the initial optimal clamping parameter set adapted to the characteristics of a single fruit through shared feature extraction and three-branch parallel inference. S4. During the clamping process, the displacement feedback data and force feedback data of the clamping mechanism are collected in real time, compared with the initial optimal clamping parameter set item by item, the deviation is calculated and the clamping risk state is identified, and an 8-dimensional clamping state deviation matrix is ​​constructed. S5. Input the clamping state deviation matrix into the adaptive closed-loop control algorithm with safety constraints, output real-time adjustment instructions and convert them into hardware drive signals to dynamically adjust the clamping stroke, clamping force and clamping posture to complete stable clamping before de-kerneling; at the same time, the final clamping result is correlated with the single fruit feature and sent back to the pear clamping big data database to realize database expansion and model iterative optimization.

2. The intelligent control method for pear pitting and clamping based on big data according to claim 1, characterized in that, The pre-detection station in S1, which collects multi-source individual feature data, integrates a binocular vision unit, a near-infrared spectroscopy detection unit, a surface diffuse reflection detection unit, and a positioning encoder. The binocular vision unit is set directly above and to the side of the conveying station to collect panoramic stereoscopic images of the pear from the front and side views, and to analyze the geometric dimensions and eccentricity characteristics of the pear. The near-infrared spectroscopy detection unit is located on the side of the conveying station. It uses a diffuse reflection probe to collect near-infrared spectral data of the pear peel and flesh to predict the flesh firmness and maturity characteristics. The surface diffuse reflection detection unit is set on both sides of the conveying station. It identifies the surface condition of pears by emitting dual-band detection light and collecting reflectivity data, including three scenarios: native with peel, waxed, and washed with water. The positioning encoder is installed on the drive shaft of the conveyor line and is used to detect the conveying position of the pear in real time. When the pear reaches the center of the detection station, the binocular vision unit, near-infrared spectroscopy detection unit and surface diffuse reflection detection unit are triggered to start data acquisition, ensuring that the multi-source data corresponds to the same spatial position of the same pear.

3. The intelligent control method for pear pitting and clamping based on big data according to claim 1, characterized in that, S2 specifically includes the following steps: S21. Geometric feature extraction: Perform image segmentation and 3D reconstruction on the binocular vision image, extract the maximum and minimum fruit diameters of the pear fruit, and the two-dimensional centroid coordinates in the transport plane. Calculate the fruit shape eccentricity by dividing the difference between the maximum and minimum fruit diameters by the average fruit diameter, and construct a 5-dimensional geometric feature subset. S22. Mechanical feature extraction: After performing multivariate scattering correction and first derivative preprocessing on near-infrared spectral data, the data is input into a pre-calibrated partial least squares hardness prediction model to calculate the average hardness of the pear flesh and the hardness of the peel, and to construct a 2D mechanical feature subset. S23. Friction feature extraction: Calculate the reflectance ratio of dual-band light in the surface diffuse reflection detection data, compare the ratio with the preset surface state threshold range, and match the surface state type corresponding to the pear fruit; then query the rubber-pear skin friction coefficient benchmark value corresponding to the surface state from the pear clamping database to construct a 1-dimensional surface friction feature subset. S24. Feature standardization processing: Based on the historical maximum and minimum values ​​of each feature dimension in the pear clamping database, range normalization processing is performed on the geometric feature subset, mechanical feature subset, and surface friction feature subset respectively, mapping all feature values ​​to the [0,1] interval, and splicing to generate an 8-dimensional core feature vector of the pear to be clamped.

4. The intelligent control method for pear pitting and clamping based on big data according to claim 3, characterized in that, The construction methods of the pear-shaped big data include: Pear samples covering the mainstream varieties of Ya pear, Xue pear, Xiang pear and Dangshan pear were collected, and the samples were labeled with four basic attributes: variety, fruit diameter range, maturity level and surface condition. Through standardized clamping tests, the optimal clamping stroke, optimal initial clamping force, critical force for crushing, critical force for slippage, and coaxiality data for different samples were recorded. The basic attributes, core feature vectors, corresponding clamping parameters, and processing results of the samples are associated and stored to form a basic database with tens of thousands of samples. After each processing is completed during the production process, the newly added single fruit data and processing results are automatically stored in the database to complete regular expansion and updates.

5. The intelligent control method for pear pitting and clamping based on big data according to claim 1, characterized in that, The training process of the multi-task clamping parameter prediction model in S3 specifically includes the following steps: S311. Extract labeled historical samples from the Pear Clamping Big Data Database and extract the core feature vectors of the samples as the training input set; S312. Set the geometric adaptation parameters corresponding to the sample as the first training label, including the total clamping stroke, the three-dimensional coordinates of the clamping center point, and the clamping claw deflection angle, for a total of 5 dimensions. S313. Set the mechanical control parameters corresponding to the sample as the second training label, including the initial clamping force, the critical force threshold for crushing injury, and the force threshold for slippage warning, for a total of 3 dimensions; S314. Set the friction compensation parameters corresponding to the sample as the third training label, including clamping force friction compensation coefficient and pressure distribution balance coefficient, for a total of 2 dimensions; S315. Construct a multi-branch joint loss function, calculate the mean squared error loss of the three branch labels respectively, and sum them according to the geometric weight of 0.4, the mechanical weight of 0.4, and the friction weight of 0.2 to obtain the total loss; S316. Using the Adam optimizer and backpropagation algorithm, iteratively update the network weights of the shared feature extraction layer and the three fully connected branches until the total loss converges to below the preset threshold, thus completing the model training.

6. The intelligent control method for pear pitting and clamping based on big data according to claim 1, characterized in that, The initial optimal clamping parameter set output by S3 includes three categories: geometric adaptation parameters, mechanical control parameters, and friction compensation parameters. The geometric adaptation parameters are used to match the size and shape eccentricity of the pear fruit, determine the initial clamping position and opening and closing stroke, and adapt to the discrete differences in the shape of multiple varieties of pear fruit. The mechanical control parameters are used to match the hardness of the pear flesh, set the initial value and safety boundary of the clamping force, and balance the needs of preventing soft fruit from being crushed and hard fruit from slipping. The friction compensation parameters are used to match the surface condition of the pear, correct the clamping force reference value and the pressure distribution of the clamping surface, and offset the influence of friction coefficient fluctuations on clamping stability.

7. The intelligent control method for pear pitting and clamping based on big data according to claim 1, characterized in that, S4 specifically includes the following steps: S41. Geometric deviation calculation: Real-time clamping stroke data is collected by the displacement encoder of the clamping mechanism, and the difference between the actual clamping stroke and the target clamping stroke is calculated to obtain the stroke deviation; The X and Y position offsets of the clamping center point and the fruit centroid, as well as the clamping posture angle deviation, are calculated by collecting data through clamping positioning vision, and combined to form a 4-dimensional geometric deviation subset. S42. Mechanical Deviation Calculation: Real-time clamping force data is collected by multi-dimensional force sensors on the clamping jaws. The difference between the actual clamping force and the initial set clamping force is calculated to obtain the clamping force deviation. The standard deviation of the clamping force of all clamping jaws is calculated to obtain the pressure distribution uniformity deviation. These are combined to form a 2D mechanical deviation subset. S43. Clamping Risk Status Judgment: Calculate the clamping force change rate and displacement change rate within a unit control cycle. If the clamping force change rate is lower than the first set threshold and the displacement change rate is higher than the second set threshold, it is determined that there is a slippage trend. If the displacement change rate is lower than the third set threshold and the clamping force change rate is higher than the fourth set threshold, it is determined that there is a crushing trend. Generate corresponding slippage warning value and crushing warning value. S44. Deviation Matrix Construction: The geometric deviation subset, mechanical deviation subset, slippage warning value, and indentation warning value are normalized by range and combined to generate an 8-dimensional clamping state deviation matrix.

8. The intelligent control method for pear pitting and clamping based on big data according to claim 1, characterized in that, S5 specifically includes the following steps: S51. Construct an adaptive fuzzy PID control algorithm with safety constraints. Define the algorithm input as the clamping state deviation matrix and define the algorithm output as three types of continuous control quantities: stroke adjustment, clamping force adjustment, and clamping jaw angle adjustment. S52. Set safety constraint judgment rules, and the safety constraint priority is higher than the PID adjustment logic: Read the current clamping force value in real time. If the current clamping force reaches the critical force threshold for crushing the corresponding pear, the first-level safety constraint is triggered, the clamping action is forcibly stopped and the position is locked. If a slippage trend is detected, the friction compensation logic is triggered, and the clamping force is increased stepwise according to the friction compensation coefficient, and the clamping force does not exceed the slippage warning force threshold. S53. Input the clamping state deviation matrix at the current moment into the adaptive fuzzy PID control algorithm, and calculate the real-time adjustment command that minimizes the deviation under the premise of satisfying the safety constraint rules. S54. The adjustment command is converted into a pulse control signal for the stepper motor and a torque control signal for the servo motor, respectively, to drive the clamping mechanism to complete the dynamic adjustment of stroke, clamping force and posture. S55. After the clamping and pitting process is completed, the core characteristics of the single fruit processed this time, the actual clamping parameters, and the clamping result data are correlated and sent back to the pear clamping big data database.

9. The intelligent control method for pear pitting and clamping based on big data according to claim 8, characterized in that, Before performing dynamic adjustment in S54, an anti-impact smoothing process is also included, specifically: S541. Read the current values ​​of the clamping parameters from the previous control cycle, including the current stroke, current clamping force, and current clamping angle. S542. Calculate the difference between the target parameter value corresponding to the adjustment command and the current value of the corresponding parameter respectively to obtain three types of adjustment differences; S543. Determine whether the three types of adjustment differences exceed the preset single maximum adjustment step size. If a certain type of adjustment difference exceeds the corresponding single maximum adjustment step size, then truncate the actual adjustment amount of that type issued this time to the single maximum adjustment step size. If the limit is not exceeded, the original adjustment instruction will be issued.

10. A pear pitting and clamping intelligent control system based on big data, characterized in that, The system is used to implement the big data-based intelligent control method for pear pitting and clamping as described in any one of claims 1-9, the system comprising: The multi-source fruit sensing module is installed at the pre-inspection conveying station of the pitting equipment. It integrates a binocular vision unit, a near-infrared spectroscopy detection unit, a surface diffuse reflection detection unit, and a positioning encoder. It is responsible for simultaneously collecting geometric, mechanical, and surface-related multi-source individual characteristic data of the pears to be processed. The big data edge computing module has a built-in storage unit and a computing unit. It stores a large database of pear clamping and a pre-trained multi-task clamping parameter prediction model. It is responsible for feature extraction, standardization processing, initial clamping parameter prediction and clamping state deviation matrix calculation. The adaptive clamping control module has a built-in adaptive fuzzy PID control algorithm with safety constraints. It is responsible for receiving the clamping state deviation matrix, generating real-time adjustment instructions that meet the safety rules, and converting them into hardware drive signals. The core-free clamping execution module includes an adaptive multi-jaw clamping mechanism, a multi-dimensional force sensor, a displacement encoder, and a drive unit. It is responsible for receiving drive signals to complete the clamping action and providing real-time feedback of displacement and force data.

Images