An artificial intelligence-based food processing and quality control system

Abstract

The application discloses a food processing and quality control system based on artificial intelligence, comprising: a processing technology configuration center, configured to call a corresponding standardized process flow template according to food category information, and generate a target process flow structure body; a loss prediction engine, configured to output a loss rate prediction result set by using an improved quantile regression forest model; a rapid detection system docking module, configured to perform rapid detection on all food batches to be delivered out of a warehouse, generate rapid detection results, compare the rapid detection results with a food safety standard library threshold value, and obtain item-by-item qualification determination; a delivery interception controller, configured to jointly verify the loss rate prediction result set and the rapid detection results, generate a delivery permission result or a delivery interception result, and generate an alarm log; and a quality log recorder, configured to record quality log data. The improved quantile regression forest model is used to realize intelligent quality control of food processing.

Description

Technical Field

[0001] This invention relates to the field of food processing and quality control technology, and in particular to a food processing and quality control system based on artificial intelligence. Background Technology

[0002] In existing food processing and quality management processes, manual experience and static rules are typically relied upon for raw material warehousing management, process selection, processing control, and outgoing inspection. These methods have limitations when dealing with the diverse characteristics of food raw materials and dynamically changing processing environments, making it difficult to achieve refined management of raw material quality, processing losses, and outgoing safety. For example, in the raw material classification stage, fixed classification standards and manual identification are commonly used, which struggles to accurately handle raw material types with ambiguous boundaries or heterogeneous characteristics, affecting the rationality of subsequent process matching.

[0003] Especially with the increasing popularity of the "smart central kitchen" model, central kitchens, as the core of intensive, standardized, and industrialized production, face unprecedented challenges in terms of production efficiency, food safety, and cost control. Traditional management models are proving inadequate in terms of refined management, real-time response, and intelligent decision-making when dealing with the large-scale, multi-category, and high-frequency production demands of central kitchens.

[0004] Furthermore, the assessment of loss rates during processing largely relies on historical experience or linear models for prediction, lacking the ability to characterize the nonlinear relationships between multiple factors, resulting in insufficient precision in loss control. Regarding pre-shipment food inspection, although rapid testing methods are used for safety standard comparison, the lack of correlation analysis with dynamic data from the processing stage makes it difficult to assess the overall compliance of food shipments from a holistic perspective. Current systems also suffer from information silos in quality log recording and traceability, lacking unified recording and centralized management of key parameters at each stage of warehousing, processing, testing, and shipment, thus limiting the ability to conduct post-processing quality audits.

[0005] Therefore, how to provide an artificial intelligence-based food processing and quality control system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an artificial intelligence-based food processing and quality control system. This invention integrates image recognition, standardized process modeling, processing observation structure construction, improved quantile regression forest prediction model, and intelligent outbound verification mechanism. It systematically describes the intelligent control logic of the entire process from raw material warehousing judgment, process selection, loss rate prediction to food outbound permit generation. It has the advantages of high processing flow matching accuracy, strong loss prediction capability, high outbound judgment accuracy, and good traceability throughout the entire process.

[0007] According to an embodiment of the present invention, a food processing and quality control system based on artificial intelligence includes:

[0008] The processing technology configuration center is used to call the corresponding standardized process flow template based on food category information and generate the target process flow structure.

[0009] The loss prediction engine is used to perform predictions based on food category information and processing observation data. It adopts an improved quantile regression forest model and outputs a set of loss rate prediction results.

[0010] The rapid testing system interface module is used to connect with the rapid testing sampling platform to conduct rapid testing on all batches of food to be released from the warehouse, generate rapid testing results, and compare them with the threshold of the food safety standard library to obtain a pass / fail judgment for each item.

[0011] The outbound interception controller is used to jointly verify the loss rate prediction result set and the rapid detection result, generate outbound permission result or outbound interception result, and generate alarm logs.

[0012] A quality log recorder is used to record quality log data.

[0013] Optionally, modules can be integrated using the following methods:

[0014] The purchased raw materials are registered for warehousing, and the categories of the raw materials are determined by the classification and identification module to obtain food category information;

[0015] Based on the food category information, the corresponding standardized processing flow template is retrieved from the processing technology configuration center to obtain the target process flow structure;

[0016] Food processing operations are performed according to the target process flow structure. The food is weighed at the beginning and end of processing, and the operation time and equipment status information are recorded to form a processing observation data structure.

[0017] Based on food category information and processing observation data structure, an improved quantile regression forest model is used to obtain a set of loss rate prediction results;

[0018] Rapid testing samples were taken from all batches of food awaiting shipment to obtain rapid test results, which were then compared with the food safety standard database.

[0019] Before food leaves the warehouse, the loss rate prediction results and rapid test results are verified to generate outbound permission results and outbound interception results, and alarm logs are generated in case of non-compliance or risk.

[0020] Quality log data is generated during system operation and stored in the quality log recorder.

[0021] Optionally, the process of registering the purchased raw materials for warehousing and determining the category of the raw materials through a classification and identification module to obtain food category information specifically includes:

[0022] The purchased raw materials are registered for warehousing to obtain warehousing information, which includes the raw material number, supplier number, warehousing time and initial weight.

[0023] Multi-angle image acquisition is performed on the raw material to obtain image data including the top, side and overall outline, and environmental parameter data of the location of the raw material is acquired simultaneously. The environmental parameter data includes temperature, humidity, odor intensity and color reflectance.

[0024] The image data is processed by pixel matrix, edge detection and region segmentation are performed, and image feature parameters including average color vector, shape contour coefficient and surface texture gradient are extracted.

[0025] Image feature parameters and environmental parameter data are input into the classification and recognition module. Food categories are classified based on judgment rules to obtain category labels. These judgment rules include:

[0026] When the average color value of an image falls within the range of combined characteristics such as a high proportion of green channel, a relatively thin and long shape, a complex surface texture, and low color reflectance, it is classified as a leafy vegetable.

[0027] When the average color value of an image falls within the range of combined characteristics such as a high proportion of red and yellow channels, a rounded shape and outline, a smooth surface texture, and a medium to high color reflectance, it is classified as a fruit.

[0028] When the image's average color value is predominantly dark red, its shape and outline are relatively irregular, its surface texture is fibrous, and its color reflectance is low, it is identified as meat.

[0029] When the image's average color value is pale yellow, its shape and outline are relatively regular, its surface texture is fine and smooth, and its color reflectance is high, it is classified as a dairy product.

[0030] By linking the category label with the raw material number, supplier number, warehousing time, and initial weight, food category information can be obtained.

[0031] Optionally, the step of retrieving the corresponding standardized processing flow template from the processing technology configuration center based on food category information to obtain the target process flow structure specifically includes:

[0032] Extract category tags corresponding to the current food category information from the central database, and access the preset process parameter table in the processing technology configuration center using the category tags as search keywords;

[0033] Extract a set of standardized process parameters corresponding to the product category label from the process parameter table. The set of standardized process parameters includes cleaning time parameters, cutting specification parameters, and packaging method parameters. The cleaning time parameter is used to indicate the processing time of the target food in the cleaning process. The cutting specification parameter is used to indicate the knife specifications or cutting size used in the cutting operation. The packaging method parameter is used to indicate the selected packaging style, packaging material, or packaging template number.

[0034] Based on the category label and the set of standardized process parameters, a target process flow structure is constructed. The target process flow structure is used to describe the standardized processing flow corresponding to the current food category, and includes category label field, washing time parameter field, cutting specification parameter field, and packaging method parameter field.

[0035] Optionally, the step of performing food processing operations according to the target process flow structure, weighing the food at the start and end of processing, recording the operation time and equipment status information, and forming a processing observation data structure specifically includes:

[0036] Food processing operations are performed under the control of the target process flow structure, and the initial weight of raw materials is recorded using electronic weighing equipment before processing begins. The initial weight is then linked to the raw material number in the food category information.

[0037] After the processing operation is completed, the same batch of food is weighed again, the weight after processing is recorded, and a binding relationship is established with the raw material number.

[0038] Based on the recorded initial weight and processed weight, the actual loss rate of the current food batch is calculated. The actual loss rate is defined as the ratio of the weight difference obtained by subtracting the processed weight from the initial weight to the initial weight.

[0039] During the execution of the processing task, auxiliary parameter information related to the current food batch processing operation is collected. The auxiliary parameter information includes the operation time of the processing task, the set of operating status parameters of the processing equipment during processing, the set of environmental parameters of the processing site, and the experience level of the operator performing the operation. The set of environmental parameters includes the temperature, humidity, odor intensity and air color reflectance of the processing area. The operator experience level is a level label determined based on historical operation records and is used to reflect the level of operation proficiency.

[0040] The raw material number, initial weight, post-processing weight, actual loss rate, operation time, set of operating status parameters of processing equipment, set of environmental parameters, and operator experience level are combined in a structured way to generate a processing observation data structure that uniquely corresponds to the raw material number.

[0041] Optionally, the set of loss rate prediction results obtained by using an improved quantile regression forest model based on food category information and processing observation data structure specifically includes:

[0042] The food category information, environmental parameters contained in the processing observation data, operator experience level recorded in the processing observation data, equipment operating status parameters collected in the processing observation data, and operation time are all used as input features to construct an input feature vector, and the input feature vector is combined with the corresponding actual loss rate to form a training sample set.

[0043] An improved quantile regression forest model is trained based on a training sample set. The improved quantile regression forest model consists of multiple decision trees. The leaf nodes of each decision tree store the set of target values, and a distribution reconstruction mechanism is introduced at the leaf nodes.

[0044] In the prediction phase, the input feature vector corresponding to the current food batch is input into the improved quantile regression forest model. The leaf nodes corresponding to the input feature vector in each decision tree are retrieved, and the target value distribution stored after being corrected by the distribution reconstruction mechanism during the training phase is called. Then, the target value distributions of each leaf node are aggregated to form the target value distribution set of the current food batch.

[0045] The distribution set of target values ​​is calculated by quantiles to obtain a set of loss rate prediction results, which includes the loss rate prediction value at the 10th percentile, the loss rate prediction value at the 50th percentile, and the loss rate prediction value at the 90th percentile.

[0046] Optionally, the improved quantile regression forest model trained based on the training sample set, wherein the improved quantile regression forest model consists of multiple decision trees, each decision tree's leaf nodes storing the target value set, and the distribution reconstruction mechanism introduced at the leaf nodes specifically includes:

[0047] During the training phase, the training sample set is used to generate several training subsets using a bootstrap sampling method. The bootstrap sampling method involves randomly selecting samples from the training sample set using a sampling with replacement method to form a subset of the same size as the original training sample set, and constructing a decision tree for each training subset.

[0048] During the growth of each decision tree, the input feature vector is used as the basis for splitting, and the mean square error minimization criterion is used to determine the splitting node until the number of samples is less than the preset threshold or the tree depth reaches the upper limit.

[0049] At each leaf node, a set of target values ​​corresponding to all samples within the leaf node is stored. The set of target values ​​consists of the actual loss rate corresponding to the training samples falling into the leaf node. Each target value is associated with a corresponding input feature vector to characterize the distribution of observed loss rates of all samples within the leaf node. A distribution reconstruction mechanism is performed on the set of target values. The distribution reconstruction mechanism includes outlier removal, small sample smoothing, and quantile consistency correction.

[0050] An outlier removal operation is performed in the leaf node. The difference between each element in the target value set and the mean of the target value set is calculated. When the difference between an element and the mean is greater than three times the standard deviation, the element is determined to be an outlier and removed from the target value set to ensure that the target value set can reflect the normal distribution of samples in the leaf node.

[0051] Perform small sample smoothing operation in the leaf nodes. When the number of samples in the leaf node is less than the set threshold, use the kernel density estimation function to smooth the distribution of the target value.

[0052] When calculating quantiles for the smoothed target value distribution, if the obtained quantile results are non-monotonic, consistency correction is performed. This is achieved by constraining the predicted loss rate of the 10th percentile to be no greater than the predicted loss rate of the 50th percentile, and the predicted loss rate of the 50th percentile to be no greater than the predicted loss rate of the 90th percentile, thereby ensuring that the quantile results maintain a monotonically increasing relationship in numerical terms.

[0053] The target values, after being corrected by the distribution reconstruction mechanism, are distributed and stored in the corresponding leaf nodes to complete the training of the improved quantile regression forest model.

[0054] Optionally, the step of conducting rapid testing on all batches of food awaiting shipment, obtaining rapid testing results, and comparing the rapid testing results with the food safety standard library specifically includes:

[0055] Rapid testing and sampling are conducted on each batch of food in the food category information to collect key testing indicators, including pesticide residues, microbial content, heavy metal concentration and additive content. The measured values ​​of the key testing indicators are combined to form rapid test results.

[0056] The set of compliance thresholds corresponding to the food category label is retrieved from the food safety standard library. The set of compliance thresholds includes the upper limit value or allowable range standard set for key testing indicators, which is used to compare the rapid test results for compliance.

[0057] Each test indicator in the rapid test results is compared with the compliance threshold in the compliance threshold set item by item;

[0058] When the test indicator is of the upper limit type, if the test value is less than or equal to the corresponding upper limit threshold, the test indicator is deemed to be qualified; otherwise, it is deemed to be unqualified.

[0059] When the test indicator is of the range type, if the test value is between the corresponding lower and upper limits of the pass threshold, the test indicator is deemed to be qualified; otherwise, it is deemed to be unqualified, and the judgment result of each test indicator is recorded as the qualified judgment result.

[0060] Optionally, the step of verifying the loss rate prediction result set and rapid test results before food leaves the warehouse, generating outbound permission results and outbound interception results, and generating alarm logs in case of non-compliance or risk specifically includes:

[0061] Before food leaves the warehouse, the loss rate prediction result set is called, and the predicted values ​​of each quantile in the loss rate prediction result set are compared with the preset loss rate threshold to obtain the loss rate verification result.

[0062] When the 90th percentile predicted value in the loss rate prediction result set is less than or equal to the preset loss rate threshold, the loss rate verification result of the food batch is determined to be qualified.

[0063] When the 10th percentile predicted value in the loss rate prediction result set is greater than the preset loss rate threshold, the loss rate verification result of the food batch is determined to be unqualified.

[0064] When the 10th percentile predicted value in the loss rate prediction result set is less than or equal to the preset loss rate threshold and the 90th percentile predicted value is greater than the preset loss rate threshold, the loss rate verification result of the food batch is determined to be a risk batch.

[0065] Obtain the pass / fail judgment result corresponding to the rapid test result, construct the pass / fail judgment vector. When all test indicators in the pass / fail judgment vector are marked as pass, the rapid test result is judged to be overall pass; otherwise, the rapid test result is judged to be overall fail.

[0066] The loss rate verification result and the rapid test result are jointly judged. When the loss rate verification result is qualified and the rapid test result is qualified overall, an outbound permission result is generated and the food batch is allowed to carry out the outbound operation.

[0067] When the loss rate verification result is unqualified or the rapid test result is unqualified overall, an outbound interception result is generated and the outbound operation of the food batch is interrupted.

[0068] When the loss rate verification result is a risky batch and the rapid test result is qualified, a risk-marked outbound result is generated, allowing the food batch to carry out outbound operation, but at the same time it is marked to enter key monitoring.

[0069] When generating outbound interception results or risk-marked outbound results, an alarm log is also generated. The alarm log includes the raw material number, the set of loss rate prediction results, the rapid detection results, and the joint judgment results.

[0070] Optionally, the step of generating quality log data during system operation and storing the quality log data in the quality log recorder specifically includes:

[0071] During the raw material receiving process, the raw material number, supplier number, receiving time, and initial weight are recorded as receiving log entries.

[0072] During the process execution phase, the target process flow structure and the issued execution status are recorded as process log entries;

[0073] During the weighing process, the processed weight, actual loss rate, operation time, and equipment operating status parameters are recorded as weighing log entries.

[0074] In the loss prediction stage, the set of loss rate prediction results and the improved quantile regression forest model are recorded as prediction log entries.

[0075] In the rapid testing process, the rapid testing results and the pass / fail judgment vector are recorded as test log entries;

[0076] During the outbound verification process, the outbound permission result, outbound interception result, or risk-marked outbound result, along with the corresponding alarm log entries, are recorded as verification log entries.

[0077] The inbound log entries, process log entries, weighing log entries, prediction log entries, inspection log entries, and verification log entries are combined into quality log data and stored in the quality log recorder.

[0078] The beneficial effects of this invention are:

[0079] This invention, by introducing image recognition and multi-dimensional environmental perception technologies, achieves accurate identification of food categories during the raw material warehousing stage. The identification results are then linked to supplier numbers and warehousing information, enabling systematic management of warehousing data and category information. This method avoids subjective errors caused by manual classification and provides a stable data foundation for automatic matching in subsequent processing flows.

[0080] In terms of process configuration, this invention, based on food category labels, automatically retrieves corresponding templates from a standardized process configuration center to generate a target process flow structure, which guides actual processing operations. By collecting weighing records at the start and end of processing, the status of operating equipment, and operator information, structured processing observation data is formed, effectively improving the completeness and traceability of process execution data.

[0081] Regarding loss rate control, this invention constructs an improved quantile regression forest model that integrates environmental parameters, operational data, and food categories, and introduces a leaf node distribution reconstruction mechanism for prediction enhancement. This model possesses sensitivity to the loss rate distribution boundary and high prediction accuracy, providing early warnings of potential quality risks during production and enhancing the system's quality prediction capabilities.

[0082] In the outbound process, this invention jointly verifies the loss rate prediction results with the rapid inspection results, determines whether to release the goods based on a preset threshold, and automatically generates interception signals and alarm logs in abnormal situations. Combined with the full-process recording mechanism of quality logs, it enables traceability and audit support for key nodes, improving the level of intelligence in food processing inbound and outbound management. Attached Figure Description

[0083] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0084] Figure 1 This is a flowchart of a food processing and quality control system based on artificial intelligence proposed in this invention.

[0085] Figure 2 This is a schematic diagram of an artificial intelligence-based food processing and quality control system proposed in this invention.

[0086] Figure 3 This is a structural diagram of an improved quantile regression forest model in an artificial intelligence-based food processing and quality control system proposed in this invention. Detailed Implementation

[0087] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0088] refer to Figure 1-3 An artificial intelligence-based food processing and quality control system includes:

[0089] The processing technology configuration center is used to call the corresponding standardized process flow template based on food category information and generate the target process flow structure.

[0090] The loss prediction engine is used to perform predictions based on food category information and processing observation data. It adopts an improved quantile regression forest model and outputs a set of loss rate prediction results.

[0091] The rapid testing system interface module is used to connect with the rapid testing sampling platform to conduct rapid testing on all batches of food to be released from the warehouse, generate rapid testing results, and compare them with the threshold of the food safety standard library to obtain a pass / fail judgment for each item.

[0092] The outbound interception controller is used to jointly verify the loss rate prediction result set and the rapid detection result, generate outbound permission result or outbound interception result, and generate alarm logs.

[0093] A quality log recorder is used to record quality log data.

[0094] This invention achieves intelligent management of process flow, processing losses, safety inspections, and outbound control in food processing by constructing a processing technology configuration center, a loss prediction engine, a rapid inspection system integration module, an outbound interception controller, and a quality log recorder. An improved quantile regression forest model enhances the accuracy and stability of loss prediction, and combined with rapid inspection results for joint verification, effectively reducing the risk of non-conforming products leaving the warehouse. The system possesses full-process recording and traceability capabilities, enhancing the transparency and controllability of quality management.

[0095] In this embodiment, the modules are interconnected using the following method:

[0096] The purchased raw materials are registered for warehousing, and the categories of the raw materials are determined by the classification and identification module to obtain food category information;

[0097] Based on the food category information, the corresponding standardized processing flow template is retrieved from the processing technology configuration center to obtain the target process flow structure;

[0098] Food processing operations are performed according to the target process flow structure. The food is weighed at the beginning and end of processing, and the operation time and equipment status information are recorded to form a processing observation data structure.

[0099] Based on food category information and processing observation data structure, an improved quantile regression forest model is used to obtain a set of loss rate prediction results;

[0100] Rapid testing samples were taken from all batches of food awaiting shipment to obtain rapid test results, which were then compared with the food safety standard database.

[0101] Before food leaves the warehouse, the loss rate prediction results and rapid test results are verified to generate outbound permission results and outbound interception results, and alarm logs are generated in case of non-compliance or risk.

[0102] Quality log data is generated during system operation and stored in the quality log recorder.

[0103] In this embodiment, the process of registering the purchased raw materials for warehousing and determining the category of the raw materials through a classification and identification module to obtain food category information specifically includes:

[0104] The purchased raw materials are registered for warehousing to obtain warehousing information, which includes the raw material number, supplier number, warehousing time and initial weight.

[0105] Multi-angle image acquisition is performed on the raw material to obtain image data including the top, side and overall outline, and environmental parameter data of the location of the raw material is acquired simultaneously. The environmental parameter data includes temperature, humidity, odor intensity and color reflectance.

[0106] The image data is processed by pixel matrix, edge detection and region segmentation are performed, and image feature parameters including average color vector, shape contour coefficient and surface texture gradient are extracted.

[0107] Image feature parameters and environmental parameter data are input into the classification and recognition module. Food categories are classified based on judgment rules to obtain category labels. These judgment rules include:

[0108] When the average color value of an image falls within the range of combined characteristics such as a high proportion of green channel, a relatively thin and long shape, a complex surface texture, and low color reflectance, it is classified as a leafy vegetable.

[0109] When the average color value of an image falls within the range of combined characteristics such as a high proportion of red and yellow channels, a rounded shape and outline, a smooth surface texture, and a medium to high color reflectance, it is classified as a fruit.

[0110] When the image's average color value is predominantly dark red, its shape and outline are relatively irregular, its surface texture is fibrous, and its color reflectance is low, it is identified as meat.

[0111] When the image's average color value is pale yellow, its shape and outline are relatively regular, its surface texture is fine and smooth, and its color reflectance is high, it is classified as a dairy product.

[0112] By linking the category label with the raw material number, supplier number, warehousing time, and initial weight, food category information can be obtained.

[0113] In this embodiment, the step of calling the corresponding standardized processing flow template from the processing technology configuration center based on food category information to obtain the target process flow structure specifically includes:

[0114] Extract category tags corresponding to the current food category information from the central database, and access the preset process parameter table in the processing technology configuration center using the category tags as search keywords;

[0115] Extract a set of standardized process parameters corresponding to the product category label from the process parameter table. The set of standardized process parameters includes cleaning time parameters, cutting specification parameters, and packaging method parameters. The cleaning time parameter is used to indicate the processing time of the target food in the cleaning process. The cutting specification parameter is used to indicate the knife specifications or cutting size used in the cutting operation. The packaging method parameter is used to indicate the selected packaging style, packaging material, or packaging template number.

[0116] Based on the category label and standardized process parameter set, a target process flow structure is constructed. The target process flow structure is used to describe the standardized processing flow corresponding to the current food category, and includes category label field, washing time parameter field, cutting specification parameter field, and packaging method parameter field.

[0117] In this embodiment, the step of performing food processing operations according to the target process flow structure, weighing the food at the start and end of processing, recording the operation time and equipment status information, and forming a processing observation data structure specifically includes:

[0118] Food processing operations are performed under the control of the target process flow structure, and the initial weight of raw materials is recorded using electronic weighing equipment before processing begins. The initial weight is then linked to the raw material number in the food category information.

[0119] After the processing operation is completed, the same batch of food is weighed again, the weight after processing is recorded, and a binding relationship is established with the raw material number.

[0120] Based on the recorded initial weight and processed weight, the actual loss rate of the current food batch is calculated. The actual loss rate is defined as the ratio of the weight difference obtained by subtracting the processed weight from the initial weight to the initial weight.

[0121] During the execution of the processing task, auxiliary parameter information related to the current food batch processing operation is collected. The auxiliary parameter information includes the operation time of the processing task, the set of operating status parameters of the processing equipment during processing, the set of environmental parameters of the processing site, and the experience level of the operator performing the operation. The set of environmental parameters includes the temperature, humidity, odor intensity and air color reflectance of the processing area. The operator experience level is a level label determined based on historical operation records and is used to reflect the level of operation proficiency.

[0122] The raw material number, initial weight, post-processing weight, actual loss rate, operation time, set of processing equipment operating status parameters, set of environmental parameters, and operator experience level are combined in a structured way to generate a processing observation data structure that uniquely corresponds to the raw material number.

[0123] In this embodiment, the step of obtaining the loss rate prediction result set based on food category information and processing observation data structure, using an improved quantile regression forest model, specifically includes:

[0124] The food category information, environmental parameters contained in the processing observation data, operator experience level recorded in the processing observation data, equipment operating status parameters collected in the processing observation data, and operation time are all used as input features to construct an input feature vector, and the input feature vector is combined with the corresponding actual loss rate to form a training sample set.

[0125] An improved quantile regression forest model is trained based on a training sample set. The improved quantile regression forest model consists of multiple decision trees. The leaf nodes of each decision tree store the set of target values, and a distribution reconstruction mechanism is introduced at the leaf nodes.

[0126] In the prediction phase, the input feature vector corresponding to the current food batch is input into the improved quantile regression forest model. The leaf nodes corresponding to the input feature vector in each decision tree are retrieved, and the target value distribution stored after being corrected by the distribution reconstruction mechanism during the training phase is called. Then, the target value distributions of each leaf node are aggregated to form the target value distribution set of the current food batch.

[0127] The distribution set of target values ​​is calculated by quantiles to obtain a set of loss rate prediction results, which includes the loss rate prediction value at the 10th percentile, the loss rate prediction value at the 50th percentile, and the loss rate prediction value at the 90th percentile.

[0128] This invention constructs training samples with food category information, environmental parameters, operator experience level, equipment operating status, and operating time as input features. It employs a quantile regression forest model with a distribution reconstruction mechanism to effectively improve the accuracy and stability of loss rate prediction. The model can output prediction results at multiple quantiles, supports modeling the uncertainty of loss rates, enhances the responsiveness to abnormal processing conditions, and helps to achieve risk prediction and decision support before food leaves the warehouse, ensuring the quality control effect of the food processing process.

[0129] In this embodiment, the improved quantile regression forest model trained based on the training sample set consists of multiple decision trees. The leaf nodes of each decision tree store the target value set, and the distribution reconstruction mechanism introduced at the leaf nodes specifically includes:

[0130] During the training phase, the training sample set is used to generate several training subsets using a bootstrap sampling method. The bootstrap sampling method involves randomly selecting samples from the training sample set using a sampling with replacement method to form a subset of the same size as the original training sample set, and constructing a decision tree for each training subset.

[0131] During the growth of each decision tree, the input feature vector is used as the basis for splitting, and the mean square error minimization criterion is used to determine the splitting node until the number of samples is less than the preset threshold or the tree depth reaches the upper limit.

[0132] At each leaf node, a set of target values ​​corresponding to all samples within the leaf node is stored. The set of target values ​​consists of the actual loss rate corresponding to the training samples falling into the leaf node. Each target value is associated with a corresponding input feature vector to characterize the distribution of observed loss rates of all samples within the leaf node. A distribution reconstruction mechanism is performed on the set of target values. The distribution reconstruction mechanism includes outlier removal, small sample smoothing, and quantile consistency correction.

[0133] An outlier removal operation is performed in the leaf node. The difference between each element in the target value set and the mean of the target value set is calculated. When the difference between an element and the mean is greater than three times the standard deviation, the element is determined to be an outlier and removed from the target value set to ensure that the target value set can reflect the normal distribution of samples in the leaf node.

[0134] Perform small-sample smoothing in the leaf nodes. When the number of samples in a leaf node is less than a set threshold, use the kernel density estimation function to smooth the distribution of the target value.

[0135] ;

[0136] in, This represents the target value distribution function obtained based on kernel density estimation, used to smooth the actual loss rate distribution within the leaf nodes. This indicates the number of samples in the target value set of the current leaf node. This represents the bandwidth parameter, used to control the scale of kernel function smoothing. This represents the value of the loss rate during forecasting. Represents the target value set of the first The actual loss rate corresponding to each sample This represents the sample index in the target value set of the leaf node. This represents the kernel function, used to calculate the weights between sample points and prediction points. It is typically a Gaussian kernel function or other symmetric kernel functions.

[0137] When calculating quantiles for the smoothed target value distribution, if the obtained quantile results are non-monotonic, consistency correction is performed. This is achieved by constraining the predicted loss rate of the 10th percentile to be no greater than the predicted loss rate of the 50th percentile, and the predicted loss rate of the 50th percentile to be no greater than the predicted loss rate of the 90th percentile, thereby ensuring that the quantile results maintain a monotonically increasing relationship in numerical terms.

[0138] The target values, after being corrected by the distribution reconstruction mechanism, are distributed and stored in the corresponding leaf nodes to complete the training of the improved quantile regression forest model.

[0139] This invention introduces an improved quantile regression forest model for loss rate prediction. By reconstructing the distribution of the target value set of training samples within the leaf nodes, the accuracy and robustness of the prediction results are effectively improved. The model first introduces an outlier removal mechanism, judging and removing significantly deviating sample points based on a set multiple range to ensure the representativeness of the target values ​​within the leaf nodes. Then, a sample smoothing mechanism is employed, introducing a kernel density estimation method to smooth the distribution of target values ​​when the number of samples is insufficient, mitigating prediction bias caused by sample sparsity. Finally, a quantile calibration mechanism performs consistency correction on the 10th, 50th, and 90th percentile values ​​of the prediction results, ensuring that the predicted distribution matches the historical sample distribution.

[0140] The synergistic effect of these mechanisms enables the model to maintain good generalization ability even with small samples, heterogeneous distributions, or extreme data, significantly improving the robustness and reliability of loss rate prediction and providing more accurate decision-making basis for food processing quality control.

[0141] In this embodiment, the step of performing rapid testing on all batches of food to be shipped, obtaining rapid testing results, and comparing the rapid testing results with the food safety standard library specifically includes:

[0142] Rapid testing and sampling are conducted on each batch of food in the food category information to collect key testing indicators, including pesticide residues, microbial content, heavy metal concentration and additive content. The measured values ​​of the key testing indicators are combined to form rapid test results.

[0143] The set of compliance thresholds corresponding to the food category label is retrieved from the food safety standard library. The set of compliance thresholds includes the upper limit value or allowable range standard set for key testing indicators, which is used to compare the rapid test results for compliance.

[0144] Each test indicator in the rapid test results is compared with the compliance threshold in the compliance threshold set item by item;

[0145] When the test indicator is of the upper limit type, if the test value is less than or equal to the corresponding upper limit threshold, the test indicator is deemed to be qualified; otherwise, it is deemed to be unqualified.

[0146] When the test indicator is of the range type, if the test value is between the corresponding lower and upper limits of the pass threshold, the test indicator is deemed to be qualified; otherwise, it is deemed to be unqualified, and the judgment result of each test indicator is recorded as the qualified judgment result.

[0147] In this embodiment, the step of verifying the loss rate prediction result set and rapid test results before food leaves the warehouse, generating outbound permission results and outbound interception results, and generating alarm logs in the event of non-compliance or risk specifically includes:

[0148] Before food leaves the warehouse, the loss rate prediction result set is called, and the predicted values ​​of each quantile in the loss rate prediction result set are compared with the preset loss rate threshold to obtain the loss rate verification result.

[0149] When the 90th percentile predicted value in the loss rate prediction result set is less than or equal to the preset loss rate threshold, the loss rate verification result of the food batch is determined to be qualified.

[0150] When the 10th percentile predicted value in the loss rate prediction result set is greater than the preset loss rate threshold, the loss rate verification result of the food batch is determined to be unqualified.

[0151] When the 10th percentile predicted value in the loss rate prediction result set is less than or equal to the preset loss rate threshold and the 90th percentile predicted value is greater than the preset loss rate threshold, the loss rate verification result of the food batch is determined to be a risk batch.

[0152] Obtain the pass / fail judgment result corresponding to the rapid test result, construct the pass / fail judgment vector. When all test indicators in the pass / fail judgment vector are marked as pass, the rapid test result is judged to be overall pass; otherwise, the rapid test result is judged to be overall fail.

[0153] The loss rate verification result and the rapid test result are jointly judged. When the loss rate verification result is qualified and the rapid test result is qualified overall, an outbound permission result is generated and the food batch is allowed to carry out the outbound operation.

[0154] When the loss rate verification result is unqualified or the rapid test result is unqualified overall, an outbound interception result is generated and the outbound operation of the food batch is interrupted.

[0155] When the loss rate verification result is a risky batch and the rapid test result is qualified, a risk-marked outbound result is generated, allowing the food batch to carry out outbound operation, but at the same time it is marked to enter key monitoring.

[0156] When generating outbound interception results or risk-marked outbound results, an alarm log is also generated. The alarm log includes the raw material number, the set of loss rate prediction results, the rapid detection results, and the joint judgment results.

[0157] This invention establishes a joint verification mechanism before food shipment, enabling collaborative judgment of loss rate predictions and rapid testing results, effectively improving the quality assurance capabilities of the food shipment process. The system performs multi-level threshold judgments based on the 10th and 90th percentile loss rate predictions, clearly classifying batches as qualified, unqualified, and risky. Combined with the overall qualification results of rapid testing, it precisely controls shipment approvals. For risky batches exhibiting fluctuations or potential anomalies, the system implements a focused monitoring strategy and generates detailed alarm logs, providing a basis for subsequent quality traceability, thereby improving the safety and controllability of the shipment process.

[0158] In this embodiment, generating quality log data during system operation and storing the quality log data in the quality log recorder specifically includes:

[0159] During the raw material receiving process, the raw material number, supplier number, receiving time, and initial weight are recorded as receiving log entries.

[0160] During the process execution phase, the target process flow structure and the issued execution status are recorded as process log entries;

[0161] During the weighing process, the processed weight, actual loss rate, operation time, and equipment operating status parameters are recorded as weighing log entries.

[0162] In the loss prediction stage, the set of loss rate prediction results and the improved quantile regression forest model are recorded as prediction log entries.

[0163] In the rapid testing process, the rapid testing results and the pass / fail judgment vector are recorded as test log entries;

[0164] During the outbound verification process, the outbound permission result, outbound interception result, or risk-marked outbound result, along with the corresponding alarm log entries, are recorded as verification log entries.

[0165] The inbound log entries, process log entries, weighing log entries, prediction log entries, inspection log entries, and verification log entries are combined into quality log data and stored in the quality log recorder.

[0166] Example 1:

[0167] To verify the feasibility of this invention in practice, it was applied to the entire process of raw material warehousing, process execution, and rapid inspection and outbound delivery in a food processing enterprise. This enterprise routinely processes multiple food categories (such as leafy vegetables, fruits, meat, and dairy products) and operates under varying temperature and humidity conditions. To address this complex scenario, this invention deploys a comprehensive quality control mechanism consisting of food category determination, process flow structure generation, processing observation data structure construction, an improved quantile regression forest model, rapid inspection, outbound verification, and a quality log recorder.

[0168] In the initial verification phase, the system collected raw material warehousing information, image data, and environmental parameters from 20 batches of different food products. Combining rapid testing sampling results with manually labeled loss data, training and validation sets were constructed, and a quantile regression forest model integrating a leaf node distribution reconstruction mechanism was trained. After being put into use, the system achieved closed-loop management of the entire process, from raw material warehousing identification, category recognition, process selection, processing observation, loss rate prediction to rapid testing qualification verification.

[0169] Table 1 shows the change in loss rate control effectiveness before and after the system went live, with a significant narrowing of the loss rate fluctuation range. Based on this, a comparison was made with the company's existing system of manual experience-based decision-making and manual weighing recording, as shown in Table 2. The results indicate that the present invention outperforms traditional solutions in terms of average accuracy, recall, loss rate control, and anomaly detection rate. Furthermore, the system can automatically record timestamps and operation trajectories, facilitating subsequent traceability.

[0170] Furthermore, Table 3 shows the accuracy of loss rate prediction and consistency of rapid test qualification judgment under four environmental conditions (normal temperature and dryness, high temperature and high humidity, low temperature and normal humidity, and high humidity and low temperature), demonstrating good robustness and environmental adaptability.

[0171] Table 1. Processing batch data after system implementation

[0172]

[0173] As can be seen from Table 1 above, the system of this invention demonstrates high stability and predictive accuracy in key indicators of quality control throughout the entire food processing process. Firstly, regarding loss rate control, after system implementation, the average deviation between the predicted and actual loss rates for 10 batches of samples remained within ±0.3%. Specifically, the prediction errors for batches B03 and B07 were only 0.1% and 0.2%, respectively, indicating that the quantile regression forest model incorporating the leaf node distribution reconstruction mechanism possesses excellent predictive capabilities. Secondly, in terms of rapid detection, the consistency rate between some system rapid detection results and manual testing results reached 100%, with no misjudgments or omissions, demonstrating the practicality and reliability of the system's rapid detection module in high-throughput outbound scenarios.

[0174] Furthermore, all batch processing parameters were successfully recorded in the quality log recorder, ensuring full traceability of the food from warehousing to outgoing. These results demonstrate that the food quality control mechanism constructed in this invention, which integrates processing observation, intelligent prediction, rapid testing and verification, and log recording, can achieve high-precision, highly adaptable, and closed-loop quality control objectives in actual business scenarios.

[0175] Table 2 Comparison of Loss Rates Before and After Implementation

[0176]

[0177] Table 3 Predictive performance under different environmental conditions

[0178]

[0179] As can be seen from Tables 2 and 3 above, this invention optimizes several key performance indicators in the food processing and warehousing stages, particularly excelling in loss rate control and quality qualification. Firstly, after the system went live, the average loss rate for different food categories decreased from 9.37% to 6.04%, with leafy vegetables showing a reduction of 4.2 percentage points and meat loss rate decreasing from 12.5% ​​to 7.9%. This indicates that the introduced improved quantile regression forest model has a significant effect on loss prediction under complex process conditions.

[0180] Secondly, in stability tests under different environmental conditions, the system maintained an accuracy rate of over 95% in rapid inspection and qualification determination under various environments, particularly maintaining an accuracy rate of 96.1% under "high temperature + high humidity" conditions, indicating that the system still has strong robustness in extreme scenarios. Furthermore, the system also demonstrated good adaptability to different food categories, with loss rate fluctuations controlled within ±1.2% across multiple food categories, and the average prediction error remaining stable at around 1.5%. These results indicate that the system of this invention can be widely applied to multi-category food production scenarios, achieving refined loss prediction and risk control, and possesses practicality and promising prospects for widespread application.

[0181] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence based food processing and quality control system characterized in that, The method comprises the following steps: a processing process configuration center is used to call a corresponding standardized process flow template according to food category information, and generate a target process flow structure; a loss prediction engine is used to perform prediction based on food category information and processing observation data, and an improved quantile regression forest model is used to output a loss rate prediction result set; a rapid detection system docking module is used to dock a rapid detection sampling platform, perform rapid detection on all food batches to be delivered, generate rapid detection results, and compare the rapid detection results with food safety standard library thresholds to obtain item-by-item qualification determinations; a delivery interception controller is used to jointly verify the loss rate prediction result set and the rapid detection results, generate a delivery permission result or a delivery interception result, and generate an alarm log; a quality log recorder is used to record quality log data. The loss prediction engine is used to perform prediction based on food category information and processing observation data, and an improved quantile regression forest model is used to output a loss rate prediction result set, specifically comprising the following steps: environmental parameters, operator experience levels, equipment operating state parameters, and operation times contained in the food category information and the processing observation data are used to construct an input feature vector, and the input feature vector and a corresponding actual loss rate are combined to form a training sample set; an improved quantile regression forest model is trained based on the training sample set, the improved quantile regression forest model is composed of multiple decision trees, a leaf node of each decision tree stores a target value set, and a distribution reconstruction mechanism is introduced at the leaf node; in the prediction stage, the input feature vector corresponding to the current food batch is input into the improved quantile regression forest model, the leaf node corresponding to the input feature vector in each decision tree is searched, and the target value distribution stored after being corrected by the distribution reconstruction mechanism in the training stage is called to form a target value distribution set of the current food batch; quantile calculation is performed on the target value distribution set to obtain a loss rate prediction result set; the improved quantile regression forest model is trained based on the training sample set, the improved quantile regression forest model is composed of multiple decision trees, a leaf node of each decision tree stores a target value set, and a distribution reconstruction mechanism is introduced at the leaf node, specifically comprising the following steps: in the training stage, the training sample set is generated into several training subsets in a bootstrap sampling manner, and a decision tree is constructed for each training subset; in the growth process of each decision tree, the input feature vector is used as the splitting basis, the least mean square error criterion is used to determine the splitting node, and the process is continued until the number of samples is less than a preset threshold or the tree depth reaches an upper limit; a target value set corresponding to all samples in each leaf node is stored, and a distribution reconstruction mechanism is performed on the target value set, the distribution reconstruction mechanism includes outlier elimination, small sample smoothing, and quantile consistency correction; in the leaf node, the outlier elimination operation is performed, each element in the target value set is difference calculated with the mean value of the target value set, and when the difference between the element and the mean value is greater than three times the standard deviation, the element is determined as an outlier and is deleted from the target value set. The small sample smoothing operation is performed in the leaf node, and when the number of samples in the leaf node is less than a set threshold, a kernel density estimation function is used to smooth the target value distribution; When calculating the quantile point of the smoothed target value distribution, if the obtained quantile point result is not monotonic, consistency correction is performed; The target value distribution corrected by the distribution reconstruction mechanism is stored in the corresponding leaf node, and the improved quantile regression forest model training is completed.

2. The artificial intelligence based food processing and quality control system as claimed in claim 1 wherein, The modules are realized through the following methods: The raw materials purchased are registered in storage, and the classification recognition module is used to determine the category of the raw materials to obtain food category information; According to the food category information, a corresponding standardized processing flow template is called from the processing technology configuration center to obtain a target process flow structure body; According to the target process flow structure body, food processing operations are performed, and the food is weighed at the beginning and end of the processing to record the operation time and equipment state information, forming a processing observation data structure body; Based on the food category information and the processing observation data structure body, an improved quantile regression forest model is used to obtain a loss rate prediction result set; Fast detection sampling is performed on all food batches to be discharged, fast detection results are obtained, and the fast detection results are compared with the food safety standard library; Before the food is discharged, the loss rate prediction result set and the fast detection result are verified to generate a discharge permission result and a discharge interception result, and an alarm log is generated in the case of unqualified or risk; Quality log data is generated during system operation, and the quality log data is stored in the quality log recorder.

3. The artificial intelligence based food processing and quality control system as claimed in claim 2, wherein, The raw materials purchased are registered in storage, and the classification recognition module is used to determine the category of the raw materials to obtain food category information specifically includes: The raw materials purchased are registered in storage, and the classification recognition module is used to determine the category of the raw materials to obtain food category information specifically includes: The raw materials are registered in storage to obtain the storage information of the raw materials, multi-angle image acquisition is performed on the raw materials to obtain image data, and environmental parameter data of the position where the raw materials are located are synchronously acquired; The image data is subjected to pixel matrix processing to perform edge detection and region segmentation operations to extract image feature parameters; The image feature parameters and the environmental parameter data are input into the classification recognition module, and food category classification is performed based on the judgment rules to obtain a category label; 4. The artificial intelligence based food processing and quality control system as claimed in claim 2, wherein, The category label is bound with the raw material number, the supplier number, the storage time and the initial weight to obtain the food category information.

5. The artificial intelligence based food processing and quality control system as claimed in claim 2, wherein, The target process flow structure body includes a category label field, a cleaning time parameter field, a cutting specification parameter field and a packaging method parameter field. The target process flow structure body includes a category label field, a cleaning time parameter field, a cutting specification parameter field and a packaging method parameter field. The target process flow structure body includes a category label field, a cleaning time parameter field, a cutting specification parameter field and a packaging method parameter field. The target process flow structure body includes a category label field, a cleaning time parameter field, a cutting specification parameter field and a packaging method parameter field. In the process of executing the processing task, auxiliary parameter information related to the current food batch processing operation is collected, including operation time, processing equipment operating state parameter set, environment parameter set, and operator experience level; The raw material number, initial weight, weight after processing, actual loss rate, operation time, processing equipment operating state parameter set, environment parameter set, and operator experience level are structurally combined to generate a processing observation data structure.

6. The artificial intelligence based food processing and quality control system as claimed in claim 2, wherein, The fast inspection sampling is performed on all food batches to be discharged, the fast inspection detection results are obtained, and the fast inspection detection results are compared with the food safety standard library. Specifically, the fast inspection detection results are obtained by: Collecting key detection indicators from each food batch in the food category information, and combining the measurement values of the key detection indicators to form the fast inspection detection results; The compliance threshold set corresponding to the category label is retrieved from the food safety standard library, and each detection indicator in the fast inspection detection results is compared with the compliance threshold in the compliance threshold set; When the detection indicator is of the upper limit type, if the detection value is less than or equal to the corresponding qualified upper limit threshold, the detection indicator is determined to be qualified, otherwise it is determined to be unqualified; When the detection indicator is of the interval type, if the detection value is between the corresponding qualified lower limit threshold and the qualified upper limit threshold, the detection indicator is determined to be qualified, otherwise it is determined to be unqualified, and the determination results of each detection indicator are recorded as the qualified determination results.

7. The artificial intelligence based food processing and quality control system as claimed in claim 2, wherein, The loss rate prediction result set and the fast inspection detection results are verified before the food is discharged, the discharge permission result and the discharge interception result are generated, and the alarm log is generated in the case of unqualified or risk. Specifically, the loss rate prediction result set is called before the food is discharged, and each quantile prediction value in the loss rate prediction result set is compared with the preset loss rate threshold to obtain the loss rate verification result; The qualified determination results corresponding to the fast inspection detection results are obtained, and a qualified determination vector is constructed. When all detection indicators in the qualified determination vector are marked as qualified, the fast inspection detection results are determined to be overall qualified, otherwise they are determined to be overall unqualified; The loss rate verification result and the fast inspection detection result are jointly determined. When the loss rate verification result is qualified and the fast inspection detection result is overall qualified, the discharge permission result is generated and the food batch is allowed to execute the discharge operation; When the loss rate verification result is unqualified or the fast inspection detection result is overall unqualified, the discharge interception result is generated and the discharge operation of the food batch is interrupted; When the loss rate verification result is a risk batch and the fast inspection detection result is qualified, a risk marked discharge result is generated, allowing the food batch to execute the discharge operation but also marking it for key monitoring; When the discharge interception result or the risk marked discharge result is generated, an alarm log is also generated. The quality log data includes warehouse entry log entries, process log entries, weighing log entries, prediction log entries, detection log entries, and verification log entries.

8. The artificial intelligence based food processing and quality control system as claimed in claim 2, wherein, ​

Images