A calibration method and system for photoelectric sorting machines

Abstract

This invention belongs to the field of photoelectric sorting machine calibration technology, specifically involving a calibration method and system for photoelectric sorting machines. First, historical sorting data is acquired, and relevant patterns are captured using a machine learning model to generate calibration frequency adjustment coefficients. Then, a multi-dimensional calibration parameter matrix is ​​established, and analog components and optical feature reference values ​​from a standard component library are called. Next, the optimal sorting calibration scheme is obtained through an industrial internet platform based on the equipment model, etc. The scheme is dynamically corrected by combining real-time environmental data. Finally, the equipment is calibrated and sorted according to the corrected scheme, while also possessing a dynamic feedback mechanism. By adaptively adjusting the calibration frequency, integrating multi-dimensional parameters, dynamically responding to environmental changes, sharing the optimal scheme, and implementing closed-loop feedback optimization, the calibration accuracy, efficiency, and environmental adaptability of the photoelectric sorting machine are improved, ensuring the stability of sorting quality, reducing reliance on manual experience, and making it suitable for sorting various materials.

Description

Technical Field

[0001] This invention belongs to the field of photoelectric sorting machine calibration technology, and particularly relates to a device calibration method and system for photoelectric sorting machines. Background Technology

[0002] Photoelectric sorting machines are key equipment for automated material sorting in modern industrial production, widely used in ore processing, food grading, and plastic recycling. Their working principle involves using optical sensors to collect optical characteristics of materials, including reflectivity, color, and texture. Combined with preset thresholds for judging qualified and defective products, a control system drives pneumatic sorting valves and other actuators to achieve rapid material classification. Sorting accuracy directly affects product quality and production efficiency.

[0003] Before a photoelectric sorting machine can be used for material sorting, it needs to be calibrated. Only after passing the calibration can it be used for material sorting in production. The calibration result determines the accuracy of subsequent material sorting. However, the existing calibration methods for photoelectric sorting machines have the following limitations:

[0004] Traditional calibration methods often employ periodic calibration, failing to consider the differences in sorting accuracy caused by wear and tear on components and changes in material properties during long-term equipment operation. This can lead to over-calibration or under-calibration.

[0005] Existing technologies mostly focus on calibrating static parameters of optical sensors, such as exposure time and gain value, while ignoring the influence of dynamic parameters such as the time difference in the sorting response of pneumatic sorting valves to qualified and defective products. This results in calibrated equipment being prone to action delay errors in high-speed sorting scenarios.

[0006] Environmental factors such as temperature, humidity, and light intensity at the sorting site can significantly affect the detection accuracy of optical sensors. For example, high temperatures can cause sensor drift and dust adhesion can affect optical path transmission. However, existing calibration methods lack a dynamic response mechanism for real-time environmental data, and calibration parameters cannot be corrected in real time as the environment changes.

[0007] Because of the differences in hardware configuration of different equipment models and the characteristics of the materials to be sorted, the optimal calibration parameters may vary significantly. Existing methods cannot quickly obtain a suitable calibration scheme through cross-device data sharing, and the calibration process relies on human experience, which is inefficient.

[0008] The lack of continuous monitoring and dynamic adjustment of the sorting effect after calibration makes it impossible to trace the root cause of the problem and optimize the calibration plan in time when the material batch changes or the equipment ages and the misjudgment rate increases. This can easily lead to non-conforming products flowing into the next process.

[0009] Therefore, there is an urgent need to improve the calibration process of existing photoelectric sorting machines. This invention aims to improve the long-term sorting stability and intelligence level of photoelectric sorting machines by using a calibration method that adaptively adjusts the calibration frequency, integrates multi-dimensional parameters, dynamically responds to environmental changes, and has a closed-loop feedback mechanism. Summary of the Invention

[0010] The purpose of this invention is to provide a calibration method and system for photoelectric sorting machines, which improves the long-term sorting stability and intelligence level of photoelectric sorting machines by adaptively adjusting the calibration frequency, integrating multi-dimensional parameters, dynamically responding to environmental changes, and having a closed-loop feedback mechanism.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0012] A device calibration method for a photoelectric sorting machine includes the following steps:

[0013] S1: Obtain historical sorting data of the photoelectric sorting machine to be calibrated, input the preset machine learning model, capture the classification threshold drift pattern of qualified products and defective products, the defect rate fluctuation pattern, the misjudgment rate, and the material defect feature recognition accuracy decay trend in the historical sorting data, adaptively generate calibration frequency adjustment coefficient, and then obtain the calibration time node.

[0014] S2: Based on the calibration time node, establish a multi-dimensional calibration parameter matrix that includes the identification parameters of optical sensors for qualified products, the identification parameters of defective products, and the time difference of the sorting response of pneumatic sorting valves to the two types of materials.

[0015] S3: Based on the multi-dimensional calibration parameter matrix, call the standard component library for defective simulated parts and standard qualified simulated parts, including preset defect parameters and their optical characteristic reference values;

[0016] S4: Based on the equipment model, the qualified / defective product sorting threshold and optical feature reference value of the material to be sorted, the optimal sorting calibration scheme of similar equipment is called through the industrial Internet platform;

[0017] S5: Real-time acquisition of environmental data at the sorting site, calling the preset environmental data and calibration correction mapping table to obtain the correction values ​​of each parameter under the current environmental data, and obtain the optimal sorting calibration scheme after dynamic correction;

[0018] S6: The photoelectric sorter is calibrated based on the optimal sorting calibration scheme after dynamic correction, and the material to be sorted is sorted based on the calibrated photoelectric sorter.

[0019] Preferably, the specific process of step S1 is as follows:

[0020] S11: Through the control system interface of the photoelectric sorting machine, extract historical sorting data from the past 12 months in batches, and input the preprocessed historical sorting data into the preset machine learning model;

[0021] S12: The machine learning model captures the drift pattern of the classification threshold between qualified and defective products, calculates the mean and standard deviation of the classification threshold offset within different time periods; uses time series decomposition to separate the defect rate fluctuation pattern in the defect rate data within a specified time period, identifies the time node and duration of the fluctuation peak, and calculates the historical misjudgment rate; plots the material defect feature recognition accuracy curve, calculates the curve decay slope to quantify the decay trend of material defect feature recognition accuracy, and establishes the relationship between the decay rate and the cumulative operating time of the equipment;

[0022] S13: Assign specific weights to the four types of patterns—classification threshold drift, defect rate fluctuation, historical misjudgment rate, and recognition accuracy decay—and calculate the comprehensive impact factor.

[0023] S14: Set the reference calibration frequency, substitute the comprehensive influence factor into the preset nonlinear transformation function, and generate the real-time calibration frequency adjustment coefficient.

[0024] The preferred formula for calculating the comprehensive impact factor is as follows:

[0025] Comprehensive impact factor = (classification threshold drift quantification value × 30%) + (defect rate fluctuation quantification value × 25%) + (false judgment rate quantification value × 25%) + (identification accuracy decay quantification value × 20%).

[0026] Among them, the classification threshold drift quantization value is the ratio of the standard deviation of the classification threshold offset in different time periods to the preset maximum allowable standard deviation of the offset; the defect rate fluctuation quantization value is the ratio of the peak value of the defect rate fluctuation to the benchmark defect rate; the misjudgment rate quantization value is the ratio of the comprehensive misjudgment rate to the preset maximum allowable total misjudgment rate; and the recognition accuracy decay quantization value is the ratio of the decay slope of the feature recognition accuracy to the preset maximum allowable decay slope.

[0027] Preferably, the specific process of step S2 is as follows:

[0028] S21: Control the photoelectric sorting machine to continuously sort standard qualified samples and standard defective samples, and simultaneously collect the raw data of the optical sensor during the identification process. Filter out the parameters directly related to the accuracy of qualified product identification and defective product identification from the collected raw data, and extract the start-up delay time and action completion time of the pneumatic sorting valve when sorting qualified and defective products.

[0029] S22: Quantify the optical sensor identification parameters selected, convert the optical reflectivity threshold of qualified products into a standardized value between 0 and 1, and similarly process the parameters including the defect area ratio threshold of defective products; calculate the time difference of the pneumatic sorting valve for sorting the two types of materials, obtain the time difference quantification value, and standardize the time difference to the range of -1 to 1 according to the equipment design standard.

[0030] S23: Construct a three-dimensional calibration parameter matrix using the optical sensor's identification parameters for qualified products, identification parameters for defective products, and the sorting response time difference of the pneumatic sorting valve as three dimensions. In the matrix, each element corresponds to a specific combination of parameters, and the accuracy of qualified product identification, defective product identification, and overall sorting efficiency under that combination are marked, forming a complete multi-dimensional calibration parameter system.

[0031] Preferably, the specific process of step S3 is as follows:

[0032] S31: Extract the optical sensor's identification parameters for qualified products and identification parameters for defective products from the three-dimensional calibration parameter matrix to form two sets of feature parameter vectors. Hash-encode the two sets of feature parameter vectors to generate unique parameter index values.

[0033] S32: The standard component library pre-stores defective simulation parts, standard qualified simulation parts containing preset defect parameters, and a database of optical feature reference values ​​for both types of simulation parts; compares the parameter index values ​​with the feature labels of the simulation parts in the standard component library, calculates the matching degree, and selects simulation part groups with a matching degree ≥95%;

[0034] S33: For successfully matched simulated parts, retrieve the corresponding data from the optical characteristic reference value database of the standard component library, including the optical reflectivity reference curve and shape standard template of the standard qualified product, as well as the spectral response reference value of the defect area and the reference ratio of the defect area of ​​the defective simulated part.

[0035] Preferably, the specific process of step S4 is as follows:

[0036] S41: Extract the manufacturer, model, and hardware configuration version of the equipment to be calibrated, and convert the qualified / defective product sorting threshold and optical feature reference value of the material to be sorted into JSON format data compatible with the industrial internet platform; standardize the parameters;

[0037] S42: Upload standardized parameter data to the calibration scheme sharing database of the industrial internet platform for scheme retrieval. Construct primary search conditions based on equipment model to filter out historical calibration schemes for equipment of the same or compatible models. Then, use the similarity of qualified / defective product sorting thresholds and the similarity of optical feature reference values ​​as secondary search conditions to narrow down the scheme screening range.

[0038] S43: The retrieved candidate solutions are evaluated from three dimensions: the first dimension is the sorting accuracy after the solution is implemented, the second dimension is the calibration efficiency of the solution, and the third dimension is the number of times the solution is adapted to the material to be sorted. The analytic hierarchy process is used to assign weights to the three evaluation dimensions, calculate the comprehensive score of each candidate solution, and select the solution with the highest score as the optimal sorting and calibration solution.

[0039] Preferably, the specific process of step S5 is as follows:

[0040] S51: Real-time acquisition of environmental data at the sorting site, including temperature, humidity, light intensity, dust concentration, and airflow velocity. Convert temperature parameters into deviation values ​​from the reference temperature, humidity parameters into the percentage of relative humidity deviating from the reference value, light intensity parameters into the peak wavelength shift of the spectral distribution curve, dust concentration parameters into the number of particles per unit volume, and airflow velocity parameters into the ratio of average flow velocity to the reference flow velocity, and construct a set of environmental characteristic parameters.

[0041] S52: Call the preset environmental data and calibration correction mapping table, match the real-time extracted environmental feature parameter set with the mapping table, and use the interpolation algorithm to calculate the correction value when the environmental parameter is between adjacent records in the table, so as to obtain the correction value of each calibration parameter under the current environment;

[0042] S53: Based on the optimal sorting calibration scheme, the correction values ​​of each calibration parameter are applied to the corresponding parameters in the scheme to form a preliminary correction scheme; the preliminary correction scheme is verified by parameter coordination to check whether there is a conflict between the corrected parameters. If there is a conflict, a second adjustment is made based on the preset parameter priority rules to obtain the optimal sorting calibration scheme after dynamic correction.

[0043] Preferably, the specific process of step S6 is as follows:

[0044] S61: The dynamically corrected optimal sorting calibration scheme is parsed into an executable instruction sequence; the instruction sequence is written into the PLC module and sensor control unit of the photoelectric sorting machine through the API interface of the equipment control system.

[0045] S62: Optical System Calibration: Standard qualified and defective simulated parts are passed through the sorting channel along a preset trajectory. The optical sensor collects their characteristic images and compares the deviation between the measured characteristics and the reference value to determine whether the optical system calibration is qualified.

[0046] Sorting actuator calibration: Control the pneumatic sorting valve to sort the simulated parts, record the response time and execution accuracy of the sorting action, and determine whether the pneumatic system calibration is qualified;

[0047] Comprehensive calibration verification: 100 sets of mixed simulation parts are continuously fed in, and the sorting results of the equipment after calibration are statistically analyzed. If the false positive rate of qualified products is ≤1% and the false negative rate of defective products is ≤0.5%, then the overall calibration is confirmed to be complete.

[0048] S63: Based on the photoelectric sorting machine after the overall calibration is completed, the material to be sorted is sorted.

[0049] Preferably, it also includes a dynamic feedback process, the specific process of which is as follows:

[0050] The qualified and unqualified materials after sorting are sampled and inspected at specified time intervals. The sampled materials are then tested to obtain the qualified and unqualified false judgment rates. The qualified and unqualified false judgment rates are then added together to obtain the comprehensive false judgment rate.

[0051] Based on the comprehensive error rate, determine whether the dynamically corrected optimal sorting calibration scheme meets the requirements. If yes, continue to calibrate the photoelectric sorter according to the dynamically corrected optimal sorting calibration scheme. If no, repeat steps S1-S6 to obtain the dynamically corrected optimal sorting calibration scheme again.

[0052] Secondly, a device calibration system for a photoelectric sorting machine is provided, for implementing the aforementioned device calibration method for a photoelectric sorting machine, comprising:

[0053] The data acquisition module is used to acquire historical sorting data of the photoelectric sorter to be calibrated;

[0054] A machine learning model is used to capture the drift pattern of classification thresholds for qualified and defective products, the fluctuation pattern of defect rate, the false judgment rate, and the decay trend of material defect feature recognition accuracy in historical sorting data, and adaptively generate calibration frequency adjustment coefficients.

[0055] The calibration parameter matrix generation module is used to establish a multi-dimensional calibration parameter matrix that includes the identification parameters of optical sensors for qualified products, the identification parameters of defective products, and the time difference of the sorting response of pneumatic sorting valves to the two types of materials.

[0056] The data retrieval module is used to retrieve, based on the multi-dimensional calibration parameter matrix, defective simulation parts and standard qualified simulation parts from the standard component library, including preset defect parameters and their optical characteristic reference values;

[0057] The calibration scheme acquisition module is used to call the best sorting calibration scheme for similar equipment through the industrial internet platform based on the equipment model, the qualified / defective sorting threshold of the material to be sorted, and the optical feature reference value.

[0058] The calibration scheme correction module is used to obtain the correction values ​​of each parameter under the current environmental data by calling the preset environmental data and calibration correction mapping table based on environmental data, and then obtain the optimal sorting calibration scheme after dynamic correction.

[0059] The beneficial effects of this invention include:

[0060] ① By analyzing various patterns in historical sorting data, a calibration frequency adjustment coefficient is calculated, changing the traditional fixed-period calibration model. This avoids excessive downtime caused by frequent calibration and prevents sorting quality problems caused by untimely calibration, ensuring that the calibration frequency matches the actual wear and tear of the equipment, thus improving the cost-effectiveness and efficiency of calibration.

[0061] ② The calibration parameter matrix incorporates the optical sensor's identification parameters for qualified and defective products, as well as the response time difference of the pneumatic sorting valve for sorting these two types of materials. This overcomes the shortcomings of traditional methods that only consider static optical parameters. Through quantification and standardization, the coordination between the equipment in the identification and execution stages is comprehensively considered, effectively reducing the action delay error during high-speed sorting and making the equipment more consistent in sorting complex materials.

[0062] ③ Collect environmental data from the sorting site at any time. Based on the preset correspondence between environmental data and calibration correction, derive the correction values ​​for each parameter and adjust the optimal calibration scheme. In response to changes in environmental factors such as temperature, humidity, and light, verify the coordination between parameters and adjust them according to priority to ensure that the calibration parameters always meet the environmental conditions, avoid the decline in recognition accuracy caused by environmental influences, and enable the equipment to sort stably under complex working conditions.

[0063] ④ By leveraging the Industrial Internet platform to obtain the optimal calibration scheme for similar equipment, and adjusting it according to equipment model and material characteristics, the problem of poor universality of traditional calibration schemes is solved. The Analytic Hierarchy Process (AHP) is used to evaluate candidate schemes from multiple dimensions, selecting the appropriate calibration scheme, which greatly shortens preparation time, reduces operational difficulty, and makes the calibration process more intelligent.

[0064] ⑤ Regularly sample and inspect the sorting results, calculate the overall misjudgment rate, and use it to optimize the calibration process. When the scheme does not meet the requirements, a dynamically revised scheme is formulated. After forming such a virtuous cycle, the calibration scheme can be continuously optimized with changes in material batches, equipment aging, etc., ensuring the sorting accuracy of the equipment in the long term and reducing the number of unqualified products entering the next process. Attached Figure Description

[0065] Figure 1 This is a schematic flowchart of the equipment calibration method for a photoelectric sorting machine according to the present invention.

[0066] Figure 2 This is a schematic diagram of the adaptive generation of calibration frequency adjustment coefficients according to the present invention.

[0067] Figure 3 This is a schematic diagram of the recognition accuracy attenuation curve of the present invention. Detailed Implementation

[0068] The following is in conjunction with the appendix Figures 1-3 The present invention will be further described in detail below:

[0069] Example 1

[0070] See appendix Figure 1 As shown, a device calibration method for a photoelectric sorting machine includes the following steps:

[0071] S1: Obtain historical sorting data of the photoelectric sorter to be calibrated, input the preset machine learning model, capture the classification threshold drift pattern of qualified products and defective products, the defect rate fluctuation pattern, the misjudgment rate, and the attenuation trend of material defect feature recognition accuracy in the historical sorting data, adaptively generate the calibration frequency adjustment coefficient, obtain the calibration time node based on the calibration frequency adjustment coefficient, and perform the corresponding calibration process at the calibration time node.

[0072] See the recognition accuracy decay curve. Figure 3 As shown, based on Figure 3 It can be seen that at 0 months, after initial calibration, the recognition accuracy was 99%, indicating the device was in optimal condition. At 3 months, the accuracy slightly decreased to 98.5%, with no additional calibration performed during this period, resulting in a relatively slow decline. At 6 months, due to dust accumulation and slight aging of the optical lens, the accuracy dropped to 97.2%, and the decline rate accelerated. At 9 months, the accuracy further decreased to 96.0%. By this time, calibration had been performed three times, once every 5 days, at the adjusted frequency, which mitigated the decline to some extent. At 12 months, the accuracy stabilized at 95.0%, still higher than the preset minimum acceptable threshold of 90%. The correlation with calibration frequency: the high-frequency calibration initiated after 6 months (once every 5 days) resulted in a lower decline rate from 9-12 months (0.33% / month) compared to 3-6 months (0.43% / month), verifying the effectiveness of dynamically adjusting the calibration frequency.

[0073] The calibration frequency adjustment coefficient is a quantitative parameter used to dynamically correct the reference calibration cycle of the photoelectric sorting machine. Its essential attribute is a dimensionless value that comprehensively reflects the attenuation state of the equipment's sorting accuracy. It is calculated by integrating the influence weights of four key laws: classification threshold drift, defect rate fluctuation, misjudgment rate, and defect identification accuracy attenuation. The value range is usually between 0.8 and 1.5.

[0074] The calculation process is based on the analytic hierarchy process. The four types of patterns are assigned weights of 30%, 25%, 25%, and 20% respectively. The quantitative indicators of each pattern are weighted and summed to obtain the comprehensive influence factor, which is then substituted into the nonlinear transformation function (calibration frequency adjustment coefficient = 1 + 0.5 × comprehensive influence factor) to generate the comprehensive influence factor.

[0075] If the baseline calibration cycle is 200 hours, the calibration frequency adjustment coefficient serves as a dynamic correction factor for the baseline calibration cycle. When the coefficient > 1.5, it indicates that the equipment accuracy is decaying rapidly, triggering a high-frequency calibration warning, and the calibration cycle needs to be shortened, which can be reduced to within 133 hours. When the coefficient < 0.8, it indicates that the equipment is in a stable state, and the calibration cycle can be extended, but not exceeding 1.5 times the baseline cycle (i.e., 300 hours). This achieves adaptive control of the calibration frequency, optimizing calibration costs while ensuring sorting accuracy.

[0076] The calibration frequency adjustment coefficient is strongly correlated with the equipment's historical sorting data. Its value is updated in real time according to various patterns during equipment operation, directly determining the dynamic adjustment range of the calibration cycle, forming a closed-loop linkage mechanism of data characteristics, calibration frequency adjustment coefficient, and calibration frequency. Forced calibration can also be triggered when the defect rate exceeds the preset tolerance zone by 20% or the false acceptance rate of qualified products exceeds 5%.

[0077] S2: At the calibration time point, establish a multi-dimensional calibration parameter matrix that includes the optical sensor's identification parameters for qualified products, the identification parameters for defective products, and the time difference of the pneumatic sorting valve's sorting response for the two types of materials. Traditional calibration methods often adjust optical or mechanical parameters in isolation, ignoring the correlation between parameters. This matrix combines optical identification parameters with pneumatic execution parameters, intuitively demonstrating the impact of different parameter combinations on the sorting results. This ensures coordinated optimization of all parameters during calibration, avoiding system imbalance caused by adjusting a single parameter. Furthermore, since each element in the matrix is ​​labeled with the qualified product identification accuracy, defective product identification accuracy, and overall sorting efficiency under the corresponding parameter combination, the correlation between different parameter configurations and sorting effects can be clearly quantified. This provides a data foundation for subsequent use of standard components and matching of the optimal calibration scheme, reducing the blindness of calibration.

[0078] S3: Based on the multi-dimensional calibration parameter matrix, call up the standard component library containing defective product simulation parts with preset defect parameters and standard qualified product simulation parts, along with their optical feature reference values. The standard qualified product simulation parts and the defective product simulation parts with preset defect parameters are standard samples in the calibration process; their optical features have been rigorously calibrated and can serve as reference benchmarks for the identification accuracy of the photoelectric sorting machine. By comparing the features with these simulation parts, the identification deviation of the equipment's optical sensor can be accurately determined, avoiding calibration errors caused by non-standard reference materials. The multi-dimensional calibration parameter matrix already covers the optical sensor's identification parameters for two types of materials and the pneumatic valve response time difference. Calling up the matching simulation parts ensures that the calibration process closely matches the actual characteristics of the materials to be sorted, guaranteeing that the calibrated equipment can accurately identify qualified and defective products in actual production. The optical feature reference values ​​serve as a benchmark for judging whether the equipment's detection data meets the standards. In optical system calibration, by comparing the deviation between the measured features of the simulation parts and the reference values, the equipment's identification accuracy can be quantified, providing clear data for adjusting optical sensor parameters and avoiding errors from subjective experience-based judgments.

[0079] S4: Based on the equipment model, the qualified / defective product sorting threshold and optical feature reference value of the material to be sorted, the optimal sorting calibration scheme of similar equipment is called through the industrial Internet platform.

[0080] Different models of photoelectric sorting machines vary in hardware configuration, optical sensor performance, and pneumatic actuator response characteristics. Furthermore, the characteristics of the materials to be sorted, such as the sorting thresholds and optical properties of qualified and defective products, also differ. By using an industrial internet platform and searching based on equipment model and material characteristics, the optimal calibration solution for similar equipment in similar scenarios can be accurately located. This avoids poor calibration results due to mismatches between the calibration solution and the equipment or materials, thereby improving the efficiency of solution adaptation.

[0081] S5: Collect environmental data from the sorting site in real time, call the preset environmental data and calibration correction mapping table to obtain the correction values ​​of each parameter under the current environmental data, and obtain the optimal sorting calibration scheme after dynamic correction.

[0082] Environmental factors such as temperature, humidity, light intensity, and dust concentration at the sorting site directly affect the working status of the photoelectric sorting machine. For example, temperature changes may cause sensitivity drift in optical sensors, dust adhesion may alter the optical reflection characteristics of the material surface, and light fluctuations may interfere with the quality of feature image acquisition. By collecting these environmental data in real time and adjusting calibration parameters accordingly, environmental interference can be effectively counteracted, ensuring that the equipment maintains stable recognition accuracy under different environmental conditions.

[0083] S6: The photoelectric sorter is calibrated based on the optimal sorting calibration scheme after dynamic correction, and the material to be sorted is sorted based on the calibrated photoelectric sorter.

[0084] See Figure 2 As shown, the specific process of step S1 is as follows:

[0085] S11: Through the control system interface of the photoelectric sorting machine, extract historical sorting data from the past 12 months in batches, including the number of qualified products, the number of defective products, the classification timestamp, the original detection data of the optical sensor and the action record of the pneumatic sorting valve for each batch of materials, and input the historical sorting data into the preset machine learning model after preprocessing.

[0086] S12: The machine learning model captures the drift pattern of classification thresholds between qualified and defective products through a sliding window, and calculates the mean and standard deviation of the classification threshold offset within different time periods. It uses time series decomposition to separate the fluctuation pattern of the defect rate in a specified time period, identifies the time nodes and duration of fluctuation peaks, and calculates the historical misclassification rate. A material defect feature recognition accuracy curve is plotted, and the curve's decay slope is calculated to quantify the decay trend of material defect feature (crack, color difference, deformation) recognition accuracy, establishing the relationship between the decay rate and the cumulative operating time of the equipment.

[0087] S13: Assign specific weights to the four types of patterns—classification threshold drift, defect rate fluctuation, historical misjudgment rate, and recognition accuracy decay—and calculate the comprehensive impact factor.

[0088] S14: Set the benchmark calibration frequency, which can be once every 200 hours. Substitute the comprehensive influence factor into the preset nonlinear transformation function (adjustment coefficient = 1 + 0.5 × comprehensive influence factor) to generate the real-time calibration frequency adjustment coefficient. When the adjustment coefficient > 1.5, trigger a high-frequency calibration warning. When the adjustment coefficient < 0.8, extend the calibration cycle, but the maximum extension shall not exceed 1.5 times the benchmark cycle to ensure that the equipment accuracy is within a controllable range.

[0089] The formula for calculating the comprehensive impact factor is as follows:

[0090] The comprehensive impact factor is calculated as follows: (Classification threshold drift quantification value × 30%) + (Defect rate fluctuation quantification value × 25%) + (False positive rate quantification value × 25%) + (Recognition accuracy attenuation quantification value × 20%). Specifically, the classification threshold drift quantification value is the ratio of the standard deviation of the classification threshold offset over different time periods to the preset maximum allowable standard deviation, ranging from 0 to 1; the defect rate fluctuation quantification value is the ratio of the peak defect rate fluctuation to the baseline defect rate, ranging from 0 to 1; the false positive rate quantification value is the ratio of (pass false positive rate + defect false positive rate) to the preset maximum allowable total false positive rate, ranging from 0 to 1; and the recognition accuracy attenuation quantification value is the ratio of the attenuation slope of the feature recognition accuracy to the preset maximum allowable attenuation slope, ranging from 0 to 1.

[0091] The aforementioned machine learning model is a stacked ensemble model. The bottom layer consists of random forest and gradient boosting tree (XGBoost), and the top layer uses logistic regression to fuse the results. The specific selection criteria are as follows: to meet the needs of capturing multi-dimensional patterns: random forest is good at handling non-linear data, can effectively capture the fluctuation characteristics of the mean and standard deviation of the offset in the classification threshold drift pattern, and is not sensitive to outliers, making it suitable for processing historical sorting data that includes sudden fluctuations in equipment.

[0092] XGBoost, a gradient boosting tree, can accurately fit the fluctuation pattern of defect rate and identify the trend of accuracy decay in time series through the gradient boosting algorithm. In particular, when dealing with quantitative tasks such as separating the periodic term of defect rate and calculating the decay slope, its prediction accuracy is better than that of a single model.

[0093] Top-level logistic regression can weight and integrate the output of the underlying model, taking into account the influence of multiple dimensions such as classification threshold drift (30% weight) and defect rate fluctuation (25% weight), and finally generate a comprehensive influence factor that is highly consistent with the calculation logic of the calibration frequency adjustment coefficient.

[0094] To meet the data feature processing requirements: Historical sorting data contains mixed types of features such as classification threshold (continuous value), defect rate (continuous value), and misclassification rate (classification performance index). The tree model in the stacked model can automatically handle the interaction relationship between features, such as the correlation between misclassification rate and defect identification accuracy, without the need for manual feature engineering.

[0095] For data with time-series attributes, such as the decay trend of material defect feature identification accuracy, tree models can capture long-term trends through sliding window features (such as the decay slope of the last 3 months), while ensemble strategies can reduce the overfitting of a single model to local data fluctuations.

[0096] Example 2

[0097] Based on Example 1, the specific process of step S2 is as follows:

[0098] S21: Control the photoelectric sorting machine to continuously sort standard qualified samples and standard defective samples. Simultaneously collect the raw data of the optical sensor during the identification process, including spectral response value, image acquisition frame rate, feature extraction threshold, etc., as well as the action trigger time, valve core displacement data, and air pressure change curve of the pneumatic sorting valve when sorting the two types of materials. From the collected raw data, filter out the parameters directly related to the recognition accuracy of qualified products and defective products, including the optical reflectivity threshold and shape matching degree threshold of qualified products, and the defect area ratio threshold and gray value deviation threshold of defective products. At the same time, extract the start-up delay time and action completion time of the pneumatic sorting valve when sorting qualified products and defective products.

[0099] S22: Quantify the selected optical sensor recognition parameters, converting the optical reflectivity threshold of qualified products into a standardized value between 0 and 1, where 1 corresponds to the optimal recognition state. Similarly, process parameters such as the defect area ratio threshold of defective products. Calculate the sorting response time difference of the pneumatic sorting valve for the two types of materials, i.e., the total response time when sorting defective products minus the total response time when sorting qualified products, to obtain the quantified time difference value. Standardize this time difference to the range of -1 to 1 according to the equipment design standards, where 0 indicates no difference, a positive value indicates a slower sorting response for defective products, and a negative value indicates a faster sorting response for defective products than for qualified products.

[0100] S23: Construct a three-dimensional parameter matrix with the optical sensor’s identification parameters for qualified products, identification parameters for defective products, and the sorting response time difference of the pneumatic sorting valve as three dimensions; in the matrix, each element corresponds to a specific combination of parameters, and the accuracy of qualified product identification, defective product identification, and overall sorting efficiency under that combination are marked, forming a complete multi-dimensional calibration parameter system.

[0101] By using a correlation analysis algorithm, the correlation between the optical sensor identification parameters and the response time difference of the pneumatic sorting valve is verified, and redundant parameters with extremely low correlation are eliminated. Based on the best sorting performance samples in historical sorting data, the parameter weights in the parameter matrix are adjusted in reverse to make the matrix more reflective of actual sorting needs and ensure the effectiveness and practicality of multi-dimensional calibration parameters.

[0102] The specific process of step S3 is as follows:

[0103] S31: Extract the core identification parameters of the optical sensor for qualified products and the core identification parameters for defective products from the multi-dimensional calibration parameter matrix, forming two sets of feature parameter vectors. The core identification parameters for qualified products include optical reflectivity threshold and shape matching threshold, while the core identification parameters for defective products include defect area ratio threshold and grayscale deviation threshold. Hash-encode the two sets of feature parameter vectors to generate a unique parameter index value. This index value contains the optical feature difference identifier between qualified and defective products and the pneumatic sorting valve response time difference association marker.

[0104] S32: The standard component library pre-stores defective simulation parts and standard qualified simulation parts containing preset defect parameters, as well as a database of optical feature reference values ​​for both types of simulation parts. The preset defect parameters include features such as defect type, size, and location corresponding to a multi-dimensional parameter matrix. The optical feature reference value database includes standard spectral response curves, reference reflectance values, and feature size parameters. The parameter index values ​​are compared with the feature labels of the simulation parts in the standard component library to calculate the matching degree. The calculation formula is: Matching degree = Feature parameter overlap rate × Sum of weight coefficients. The weight coefficients are set according to the degree of influence of the parameters on the recognition accuracy. Simulation part groups with a matching degree ≥ 95% are selected.

[0105] S33: For successfully matched simulated parts, retrieve corresponding data from the optical feature reference value database of the standard component library, including the optical reflectivity reference curve and shape standard template of the standard qualified product, and the spectral response reference value of the defect area and the reference ratio of the defect area of ​​the defective simulated part. Verify the timeliness of the retrieved optical feature reference values. If the reference value has been stored for more than 3 months or the corresponding simulated part has been calibrated more than 50 times, automatically trigger the reference value update mechanism to obtain the latest calibrated reference data.

[0106] The compatibility of the retrieved simulation parts and optical feature reference values ​​with the pneumatic sorting valve response time difference parameters in the multi-dimensional calibration parameter matrix is ​​verified. The deviation rate between the theoretical response time difference during simulation part sorting and the preset time difference in the matrix is ​​calculated. If the deviation rate is ≤3%, the call result is confirmed to be valid. If the deviation rate is >3%, the optical feature reference values ​​of the simulation parts are fine-tuned according to the deviation value. For example, the defective reflective intensity of the defective simulation parts can be adjusted until the deviation rate meets the requirements, thus forming the final call scheme.

[0107] Example 3

[0108] Based on Example 1 or Example 2, the specific process of step S4 is as follows:

[0109] S41: Extract the manufacturer, series model, and hardware configuration version of the equipment to be calibrated, and convert the qualified / defective product sorting thresholds (optical reflectivity threshold range, defect size judgment threshold) and optical characteristic reference values ​​(standard spectral curve, characteristic wavelength reflectivity) of the materials to be sorted into JSON format data compatible with the industrial internet platform; standardize the parameters, unify the units and precision, generate a parameter data package containing equipment identification, material characteristics, and sorting standards, and attach a timestamp and a unique equipment identifier.

[0110] S42: Through the device's built-in industrial 5G gateway, standardized parameter data packets are uploaded to the calibration scheme sharing database of the industrial internet platform, triggering a scheme retrieval request. After receiving the request, the platform constructs primary search conditions based on the device model to filter out historical calibration schemes for devices of the same or compatible models. Then, using the similarity of the qualified / defective product sorting threshold and the similarity of the optical feature reference values ​​as secondary search conditions, calibration schemes with a similarity of ≥90% are filtered to narrow down the scheme selection range. The similarity is calculated using the cosine similarity algorithm.

[0111] S43: The retrieved candidate solutions are evaluated from three dimensions: The first dimension is the sorting accuracy after implementation, with a qualified product false positive rate ≤1% and a defective product missed detection rate ≤0.5% being optimal. The second dimension is the calibration efficiency, with a total calibration time ≤30 minutes being optimal. The third dimension is the number of times the solution is adapted to the material to be sorted, counting the number of successful applications of the solution in similar material sorting scenarios. The analytic hierarchy process (AHP) is used to assign weights to the three evaluation dimensions: sorting accuracy accounts for 60%, calibration efficiency for 25%, and the number of adaptations for 15%. The comprehensive score of each candidate solution is calculated, and the solutions are ranked from highest to lowest score. The solution with the highest score is selected as the optimal sorting and calibration solution.

[0112] In another embodiment of this example, the complete parameter set of the optimal candidate solution can be retrieved, including optical sensor calibration steps, pneumatic valve response adjustment parameters, environmental compensation coefficients, etc. Virtual verification is then performed, simulating the sorting effect after the solution implementation using digital twin technology. If the accuracy rate of identifying qualified products is ≥99% and the rejection rate of defective products is ≥99.5% in the simulation results, the solution is confirmed to be effective; otherwise, the candidate solution list is returned to select a suboptimal solution for re-adaptation until a solution that meets the requirements is obtained.

[0113] The specific process of step S5 is as follows:

[0114] S51: Real-time acquisition of environmental data at the sorting site, including temperature, humidity, light intensity, dust concentration, and airflow velocity. The temperature parameter is converted into a deviation value from a reference temperature of 25℃, the humidity parameter is converted into a percentage deviation from the reference value relative to 50% humidity, the light intensity parameter is converted into the peak wavelength shift of the spectral distribution curve, the dust concentration parameter is converted into the number of particles per unit volume, and the airflow velocity parameter is converted into the ratio of the average flow velocity to the reference flow velocity of 1m / s, thus constructing a set of environmental characteristic parameters.

[0115] S52: Call the preset environmental data and calibration correction mapping table. This table is indexed by a set of environmental characteristic parameters and includes correction coefficients for optical sensor parameters (exposure time, gain value), pneumatic sorting valve parameters (response delay, action force), and sorting threshold parameters (reflectivity threshold, defect judgment threshold) under different environmental conditions. Match the real-time extracted set of environmental characteristic parameters with the mapping table, and use an interpolation algorithm to calculate the correction value when the environmental parameter is between adjacent records in the table, thus obtaining the correction value of each calibration parameter under the current environment. For example, for every 1°C increase in temperature, the optical sensor exposure time increases by 0.02ms.

[0116] S53: Based on the optimal sorting calibration scheme obtained through the industrial internet platform, the correction values ​​of each calibration parameter are applied to the corresponding parameters in the scheme to form a preliminary correction scheme. Specifically, the exposure time correction value of the optical sensor is directly added to the original scheme's exposure parameters; the response delay correction value of the pneumatic sorting valve is accumulated with the original scheme's response parameters; and the sorting threshold correction value is adjusted proportionally to the original threshold range. The preliminary correction scheme undergoes parameter co-validation to check for conflicts between the corrected parameters, including mismatches between the imaging frame rate change caused by extended exposure time and the pneumatic valve response rhythm. If conflicts exist, a secondary adjustment is performed based on preset parameter priority rules to obtain the dynamically corrected optimal sorting calibration scheme. The parameter priority rules are: optical parameters take precedence over mechanical parameters, and sorting threshold parameters take precedence over auxiliary parameters.

[0117] The specific process of step S6 is as follows:

[0118] S61: The dynamically corrected optimal sorting calibration scheme is parsed into an executable instruction sequence, including the exposure time, gain value, and spectral recognition threshold of the optical sensor; the response delay and operating pressure of the pneumatic sorting valve; and the logic parameters for classifying qualified / defective products. This instruction sequence is written into the PLC module and sensor control unit of the photoelectric sorting machine via the equipment control system's API interface. After parameter configuration, the equipment self-test program is triggered to verify whether the parameters of each module are loaded correctly, such as whether the optical sensor operates according to the set exposure time and whether the pneumatic valve responds to the operating commands.

[0119] S62: First, perform optical system calibration: Standard qualified and defective simulated parts are passed through the sorting channel along a preset trajectory. Optical sensors acquire their characteristic images, and the system automatically compares the deviation between the measured features and the reference value. If the deviation is ≤0.5%, the optical calibration is considered qualified; otherwise, the sensor focal length or light source intensity is finely adjusted based on the deviation value before recalibration. Next, perform sorting actuator calibration: Control the pneumatic sorting valve to sort the simulated parts, recording the response time and execution accuracy of the sorting action, such as the rejection accuracy of defective simulated parts. When the response time deviation is ≤2ms and the rejection accuracy is ≥99.8%, the pneumatic system calibration is considered qualified. Finally, perform comprehensive calibration verification: 100 sets of mixed simulated parts are continuously fed in, containing 80% qualified parts and 20% defective parts. The sorting results of the equipment after calibration are statistically analyzed. If the false positive rate of qualified parts is ≤1% and the missed detection rate of defective parts is ≤0.5%, the overall calibration is confirmed to be complete.

[0120] S63: Based on the photoelectric sorting machine after the overall calibration is completed, the material to be sorted is sorted.

[0121] In another embodiment of this invention, a dynamic feedback process for device calibration is also provided, the specific process of which is as follows:

[0122] The qualified and unqualified materials after sorting are sampled and inspected at specified time intervals. The sampled materials are then tested to obtain the qualified and unqualified false judgment rates. The qualified and unqualified false judgment rates are then added together to obtain the comprehensive false judgment rate.

[0123] Based on the comprehensive error rate, determine whether the dynamically corrected optimal sorting calibration scheme meets the requirements. If yes, continue to calibrate the photoelectric sorter according to the dynamically corrected optimal sorting calibration scheme. If no, repeat steps S1-S6 to obtain the dynamically corrected optimal sorting calibration scheme again.

[0124] A calibration system for a photoelectric sorting machine, implementing a calibration method for the aforementioned photoelectric sorting machine, includes a data acquisition module for acquiring historical sorting data of the photoelectric sorting machine to be calibrated; a machine learning model for capturing the drift patterns of classification thresholds for qualified and defective products, the fluctuation patterns of defective product rates, the false judgment rate, and the attenuation trend of material defect feature recognition accuracy in the historical sorting data, adaptively generating calibration frequency adjustment coefficients; a calibration parameter matrix generation module for establishing a multi-dimensional calibration parameter matrix including the recognition parameters of optical sensors for qualified products, the recognition parameters of defective products, and the time difference of the sorting response of pneumatic sorting valves for the two types of materials; a data retrieval module for retrieving defective product simulation parts and standard qualified product simulation parts, including preset defect parameters and their optical feature reference values, from a standard component library based on the multi-dimensional calibration parameter matrix; and a calibration scheme acquisition module for retrieving the optimal sorting calibration scheme for similar equipment through an industrial internet platform based on the equipment model, the qualified / defective product sorting thresholds of the material to be sorted, and the optical feature reference values. The calibration scheme correction module is used to obtain the correction values ​​of each parameter under the current environmental data by calling the preset environmental data and calibration correction mapping table based on environmental data, and then obtain the optimal sorting calibration scheme after dynamic correction.

[0125] Example 4

[0126] In this embodiment, the calibration scenario of an FS-800 photoelectric separator in an iron ore beneficiation plant is used as an example. This equipment is used to separate qualified iron concentrate (qualified product) with a grade ≥65% and unqualified ore (substandard product) with a grade <65%. The specific calibration process is as follows:

[0127] Generate calibration frequency adjustment coefficients:

[0128] Data acquisition and preprocessing: Historical sorting data from the past 12 months was extracted through the equipment control system interface, including the daily number of qualified / defective products, classification threshold records, defect identification logs, etc. Abnormal data caused by equipment downtime was removed, resulting in 360 valid samples.

[0129] Pattern detection:

[0130] Classification threshold drift: The monthly deviation of the quality judgment threshold is calculated, and the standard deviation is 0.8%. The preset maximum allowable deviation standard deviation is 2.0%, so the quantified value of classification threshold drift = 0.8 / 2.0 = 0.4. Defect rate fluctuation: Using time series decomposition to separate weekly fluctuations, the weekly peak defect rate is found to be 8%, and the baseline defect rate is 5%. Therefore, the quantified value of defect rate fluctuation = 8% / 5% = 1.6. After standardization, it is taken as 1.0, which exceeds the preset upper limit. False positive rate: The historical comprehensive false positive rate is 3%, and the preset maximum allowable total false positive rate is 6%. Therefore, the quantified value of false positive rate = 3% / 6% = 0.5. Accuracy decay: The defect identification accuracy drops from the initial 99% to 95%, with a decay slope of 0.01% / hour. The preset maximum slope is 0.05% / hour. Therefore, the quantified value of decay = 0.01 / 0.05 = 0.2. Calculation of comprehensive impact factor:

[0131] The overall impact factor is calculated as follows: (0.4×30%) + (1.0×25%) + (0.5×25%) + (0.2×20%) = 0.12 + 0.25 + 0.125 + 0.04 = 0.535.

[0132] Calibration frequency adjustment: The reference calibration frequency is once a week. Substituting into the nonlinear transformation function, the adjustment coefficient = 1 + 0.5 × comprehensive influence factor, we get the adjustment coefficient = 1 + 0.5 × 0.535 = 1.267. Therefore, the actual calibration frequency is adjusted to once every 5 days, or 1.4 times a week.

[0133] Establish a multi-dimensional calibration parameter matrix: Parameter acquisition: The control equipment sorts standard iron concentrate samples with a grade of 68% and standard tailings samples with a grade of 60%, and collects the reflectivity threshold of optical sensors (qualified products ≥85%, defective products ≤70%), the proportion of defective area (defective products ≥5%), and the response time of pneumatic valves for sorting qualified products (0.3s) and defective products (0.35s).

[0134] Parameter quantification: Standardized threshold for reflectivity of qualified products: 85% / 100%=0.85; Standardized percentage of defective area of ​​substandard products: 5% / 10%=0.5 (preset maximum defect percentage 10%). Sorting response time difference = 0.35s - 0.3s = 0.05s, the maximum design time difference of the equipment is 0.1s, therefore the standardized time difference = 0.05 / 0.1 = 0.5, which is within the range of -1 to 1.

[0135] Matrix Construction: Using reflectivity threshold (0.85), defect rate (0.5), and response time difference (0.5) as three-dimensional coordinates, the accuracy rates for identifying qualified products (98.5%), defective products (97.2%), and overall sorting efficiency (96.8%) under this combination are labeled to form matrix elements. Standard component library is called:

[0136] Parameter index generation: Extract the reflectivity threshold (0.85) and defect ratio (0.5) from the matrix, and generate the index value Fe-0.85-0.5-0.5 after hash encoding.

[0137] Simulation component matching:

[0138] The standard component library retrieved a set of simulated parts with a 96% matching degree: Standard qualified simulated part: surface smoothness 90%, reflectivity baseline value 85%. Defective simulated part: contains a 3mm×5mm dent defect (preset defect parameters), defect area reflectivity baseline value 65%. Baseline value retrieval: obtain the qualified product reflectivity baseline curve (reflectivity at 550nm wavelength 85%) and the defective product defect area baseline percentage 5%.

[0139] Call the optimal calibration scheme:

[0140] Parameter encapsulation: Extract the device model FS-800 and hardware version V2.3, convert the qualified product sorting threshold (reflectivity ≥85%) and the defective product threshold (reflectivity ≤70%) into JSON format, standardize them, and upload them to the industrial internet platform.

[0141] Solution retrieval: The platform screened 10 historical solutions for the same model of equipment, and narrowed them down to 3 candidate solutions based on threshold similarity (≥90%) and optical feature similarity (≥92%).

[0142] Solution evaluation:

[0143] Sorting accuracy (60% weight): Solution A has a qualified product false positive rate of 0.8% and a defective product false negative rate of 0.3%, with a score of 0.95. Calibration efficiency (25% weight): Solution A takes 25 minutes, with a score of 0.9. Adaptation times (15% weight): Solution A has been applied 12 times in the iron ore sorting scenario, with a score of 0.8. Overall score = 0.95 × 0.6 + 0.9 × 0.25 + 0.8 × 0.15 = 0.57 + 0.225 + 0.12 = 0.915, which is determined to be the optimal solution.

[0144] Dynamic correction calibration scheme:

[0145] Environmental data acquisition: Real-time monitoring of on-site temperature 30℃ (baseline 25℃, deviation +5℃), humidity 60% (baseline 50%, deviation +20%), and dust concentration 0.02mg / m³ (baseline 0.01mg / m³, ratio 2.0).

[0146] Correction value calculation: Referring to the mapping table, we get: Temperature deviation +5℃ → Reflectivity threshold correction -2% (normalized value -0.02). Humidity deviation +20% → Pneumatic valve response time correction +0.02s (normalized time difference +0.2). Scheme correction: Based on the optimal scheme A, the reflectivity threshold is adjusted from 0.85 to 0.83, and the response time difference is adjusted from 0.5 to 0.7. Verification shows no parameter conflicts, thus forming a dynamic correction scheme.

[0147] Equipment calibration and sorting:

[0148] Parameter Configuration: The correction scheme was parsed into an instruction sequence and written to the device PLC via API. The optical sensor exposure time was set to 30ms and the pneumatic valve pressure to 0.6MPa. Calibration Verification: Optical System: The measured reflectivity of the simulated component was 83.2%. The deviation from the benchmark value of 85% was judged to be 1.8% ≤ 0.5%. Obviously, it was not valid. After readjusting the focus, the deviation was 0.3%, which is acceptable. Pneumatic System: The defective product sorting response time was 0.37s. The deviation from the standard value was judged to be 0.02s ≤ 2ms. It was not valid. After adjusting the solenoid valve delay, it met the standard. Comprehensive Verification: After sorting 100 sets of mixed simulated components, the false positive rate of qualified products was 0.7% and the missed detection rate of defective products was 0.2%, which met the requirements. Material Sorting: After calibration, the equipment sorted actual iron ore, processing 50 tons per hour, and ran continuously for 8 hours without any abnormalities. Dynamic feedback: 200 materials are randomly sampled every 2 hours, with a pass / fail error rate of 0.6% and a fail / fail error rate of 0.4%. The overall error rate is 1.0% ≤ the preset threshold of 1.5%, indicating the scheme is effective. If the overall error rate subsequently rises to 2.0%, the complete calibration process is re-executed.

Claims

1. A device calibration method for a photoelectric sorting machine, characterized in that, Includes the following steps: S1: Obtain historical sorting data of the photoelectric sorting machine to be calibrated, input the preset machine learning model, capture the classification threshold drift pattern of qualified products and defective products, the defect rate fluctuation pattern, the misjudgment rate, and the material defect feature recognition accuracy decay trend in the historical sorting data, adaptively generate calibration frequency adjustment coefficient, and then obtain the calibration time node. S2: Based on the calibration time node, establish a multi-dimensional calibration parameter matrix that includes the identification parameters of optical sensors for qualified products, the identification parameters of defective products, and the time difference of the sorting response of pneumatic sorting valves to the two types of materials. S3: Based on the multi-dimensional calibration parameter matrix, call the standard component library for defective simulated parts and standard qualified simulated parts, including preset defect parameters and their optical characteristic reference values; S4: Based on the equipment model, the qualified / defective product sorting threshold and optical feature reference value of the material to be sorted, the optimal sorting calibration scheme of similar equipment is called through the industrial Internet platform; S5: Real-time acquisition of environmental data at the sorting site, calling the preset environmental data and calibration correction mapping table to obtain the correction values ​​of each parameter under the current environmental data, and obtain the optimal sorting calibration scheme after dynamic correction; S6: The photoelectric sorter is calibrated based on the optimal sorting calibration scheme after dynamic correction, and the material to be sorted is sorted based on the calibrated photoelectric sorter.

2. The equipment calibration method for a photoelectric sorting machine according to claim 1, characterized in that, The specific process of step S1 is as follows: S11: Through the control system interface of the photoelectric sorting machine, extract historical sorting data from the past 12 months in batches, and input the preprocessed historical sorting data into the preset machine learning model; S12: The machine learning model captures the drift pattern of the classification threshold between qualified and defective products, calculates the mean and standard deviation of the classification threshold offset within different time periods; uses time series decomposition to separate the defect rate fluctuation pattern in the defect rate data within a specified time period, identifies the time node and duration of the fluctuation peak, and calculates the historical misjudgment rate; plots the material defect feature recognition accuracy curve, calculates the curve decay slope to quantify the decay trend of material defect feature recognition accuracy, and establishes the relationship between the decay rate and the cumulative operating time of the equipment; S13: Assign specific weights to the four types of patterns—classification threshold drift, defect rate fluctuation, historical misjudgment rate, and recognition accuracy decay—and calculate the comprehensive impact factor. S14: Set the reference calibration frequency, substitute the comprehensive influence factor into the preset nonlinear transformation function, and generate the real-time calibration frequency adjustment coefficient.

3. The equipment calibration method for a photoelectric sorting machine according to claim 2, characterized in that, The formula for calculating the comprehensive impact factor is as follows: Comprehensive impact factor = (classification threshold drift quantification value × 30%) + (defect rate fluctuation quantification value × 25%) + (false judgment rate quantification value × 25%) + (identification accuracy decay quantification value × 20%). Among them, the classification threshold drift quantization value is the ratio of the standard deviation of the classification threshold offset in different time periods to the preset maximum allowable standard deviation of the offset; the defect rate fluctuation quantization value is the ratio of the peak value of the defect rate fluctuation to the benchmark defect rate; the misjudgment rate quantization value is the ratio of the comprehensive misjudgment rate to the preset maximum allowable total misjudgment rate; and the recognition accuracy decay quantization value is the ratio of the decay slope of the feature recognition accuracy to the preset maximum allowable decay slope.

4. The equipment calibration method for a photoelectric sorting machine according to claim 1, characterized in that, The specific process of step S2 is as follows: S21: Control the photoelectric sorting machine to continuously sort standard qualified samples and standard defective samples, and simultaneously collect the raw data of the optical sensor during the identification process. Filter out the parameters directly related to the accuracy of qualified product identification and defective product identification from the collected raw data, and extract the start-up delay time and action completion time of the pneumatic sorting valve when sorting qualified and defective products. S22: Quantify the optical sensor identification parameters selected, convert the optical reflectivity threshold of qualified products into a standardized value between 0 and 1, and similarly process the parameters including the defect area ratio threshold of defective products; calculate the time difference of the pneumatic sorting valve for sorting the two types of materials, obtain the time difference quantification value, and standardize the time difference to the range of -1 to 1 according to the equipment design standard. S23: Construct a three-dimensional calibration parameter matrix using the optical sensor's identification parameters for qualified products, identification parameters for defective products, and the sorting response time difference of the pneumatic sorting valve as three dimensions. In the matrix, each element corresponds to a specific combination of parameters, and the accuracy of qualified product identification, defective product identification, and overall sorting efficiency under that combination are marked, forming a complete multi-dimensional calibration parameter system.

5. The equipment calibration method for a photoelectric sorting machine according to claim 4, characterized in that, The specific process of step S3 is as follows: S31: Extract the optical sensor's identification parameters for qualified products and identification parameters for defective products from the three-dimensional calibration parameter matrix to form two sets of feature parameter vectors. Hash-encode the two sets of feature parameter vectors to generate unique parameter index values. S32: The standard component library pre-stores a database of defective simulated parts, standard qualified simulated parts containing preset defect parameters, and optical characteristic reference values ​​of the two types of simulated parts. The parameter index value is compared with the feature labels of the simulated parts in the standard component library, the matching degree is calculated, and the simulated part groups with a matching degree ≥95% are selected. S33: For successfully matched simulated parts, retrieve the corresponding data from the optical characteristic reference value database of the standard component library, including the optical reflectivity reference curve and shape standard template of the standard qualified product, as well as the spectral response reference value of the defect area and the reference ratio of the defect area of ​​the defective simulated part.

6. The equipment calibration method for a photoelectric sorting machine according to claim 1, characterized in that, The specific process of step S4 is as follows: S41: Extract the manufacturer, model, and hardware configuration version of the equipment to be calibrated, and convert the qualified / defective product sorting threshold and optical feature reference value of the material to be sorted into JSON format data compatible with the industrial internet platform; standardize the parameters; S42: Upload standardized parameter data to the calibration scheme sharing database of the industrial internet platform for scheme retrieval. Construct primary search conditions based on equipment model to filter out historical calibration schemes for equipment of the same or compatible models. Then, use the similarity of qualified / defective product sorting thresholds and the similarity of optical feature reference values ​​as secondary search conditions to narrow down the scheme screening range. S43: The retrieved candidate solutions are evaluated from three dimensions: the first dimension is the sorting accuracy after the solution is implemented, the second dimension is the calibration efficiency of the solution, and the third dimension is the number of times the solution is adapted to the material to be sorted. The analytic hierarchy process is used to assign weights to the three evaluation dimensions, calculate the comprehensive score of each candidate solution, and select the solution with the highest score as the optimal sorting and calibration solution.

7. The equipment calibration method for a photoelectric sorting machine according to claim 6, characterized in that, The specific process of step S5 is as follows: S51: Real-time acquisition of environmental data at the sorting site, including temperature, humidity, light intensity, dust concentration, and airflow velocity. Convert temperature parameters into deviation values ​​from the reference temperature, humidity parameters into the percentage of relative humidity deviating from the reference value, light intensity parameters into the peak wavelength shift of the spectral distribution curve, dust concentration parameters into the number of particles per unit volume, and airflow velocity parameters into the ratio of average flow velocity to the reference flow velocity, and construct a set of environmental characteristic parameters. S52: Call the preset environmental data and calibration correction mapping table, match the real-time extracted environmental feature parameter set with the mapping table, and use the interpolation algorithm to calculate the correction value when the environmental parameter is between adjacent records in the table, so as to obtain the correction value of each calibration parameter under the current environment; S53: Based on the optimal sorting and calibration scheme, apply the correction values ​​of each calibration parameter to the corresponding parameters in the scheme to form a preliminary correction scheme; The initial correction scheme is verified by parameter coordination to check whether there are any conflicts between the corrected parameters. If there are any conflicts, a second adjustment is made based on the preset parameter priority rules to obtain the optimal sorting and calibration scheme after dynamic correction.

8. The equipment calibration method for a photoelectric sorting machine according to claim 7, characterized in that, The specific process of step S6 is as follows: S61: The dynamically corrected optimal sorting calibration scheme is parsed into an executable instruction sequence; the instruction sequence is written into the PLC module and sensor control unit of the photoelectric sorting machine through the API interface of the equipment control system. S62: Optical System Calibration: Standard qualified and defective simulated parts are passed through the sorting channel along a preset trajectory. The optical sensor collects their characteristic images and compares the deviation between the measured characteristics and the reference value to determine whether the optical system calibration is qualified. Sorting actuator calibration: Control the pneumatic sorting valve to sort the simulated parts, record the response time and execution accuracy of the sorting action, and determine whether the pneumatic system calibration is qualified; Comprehensive calibration verification: 100 sets of mixed simulation parts are continuously fed in, and the sorting results of the equipment after calibration are statistically analyzed. If the false positive rate of qualified products is ≤1% and the false negative rate of defective products is ≤0.5%, then the overall calibration is confirmed to be complete. S63: Based on the photoelectric sorting machine after the overall calibration is completed, the material to be sorted is sorted.

9. A device calibration method for a photoelectric sorting machine according to claim 1, characterized in that, It also includes a dynamic feedback process, the specific process of which is as follows: The qualified and unqualified materials after sorting are sampled and inspected at specified time intervals. The sampled materials are then tested to obtain the qualified and unqualified false judgment rates. The qualified and unqualified false judgment rates are then added together to obtain the comprehensive false judgment rate. Based on the comprehensive error rate, determine whether the dynamically corrected optimal sorting calibration scheme meets the requirements. If yes, continue to calibrate the photoelectric sorter according to the dynamically corrected optimal sorting calibration scheme. If no, repeat steps S1-S6 to obtain the dynamically corrected optimal sorting calibration scheme again.

10. A device calibration system for a photoelectric sorting machine, used to implement the device calibration method for a photoelectric sorting machine as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire historical sorting data of the photoelectric sorter to be calibrated; A machine learning model is used to capture the drift pattern of classification thresholds for qualified and defective products, the fluctuation pattern of defect rate, the false judgment rate, and the decay trend of material defect feature recognition accuracy in historical sorting data, and adaptively generate calibration frequency adjustment coefficients. The calibration parameter matrix generation module is used to establish a multi-dimensional calibration parameter matrix that includes the identification parameters of optical sensors for qualified products, the identification parameters of defective products, and the time difference of the sorting response of pneumatic sorting valves to the two types of materials. The data retrieval module is used to retrieve, based on the multi-dimensional calibration parameter matrix, defective simulation parts and standard qualified simulation parts from the standard component library, including preset defect parameters and their optical characteristic reference values; The calibration scheme acquisition module is used to call the best sorting calibration scheme for similar equipment through the industrial internet platform based on the equipment model, the qualified / defective sorting threshold of the material to be sorted, and the optical feature reference value. The calibration scheme correction module is used to obtain the correction values ​​of each parameter under the current environmental data by calling the preset environmental data and calibration correction mapping table based on environmental data, and then obtain the optimal sorting calibration scheme after dynamic correction.

Images