Aquatic product quality safety risk early warning method based on HACCP internal control data

Abstract

The invention discloses an aquatic product quality safety risk early-warning method based on HACCP internal control data, and the method comprises the steps: combing risk early-warning index details based on an HACCP system in combination with the actual business process of a to-be-detected aquatic product, obtaining corresponding time sequence data, and constructing a data set; the method comprises the following steps: forming an expert optimization scheme by combining an analytic hierarchy process and an entropy weight method, introducing a judgment matrix for index importance comparison by using the analytic hierarchy process, completing the judgment matrix by a plurality of experts, completing consistency verification, and obtaining a multi-expert comprehensive utility value; processing utility values of a plurality of experts by using an entropy weight method so as to select a representative expert to give rating data of the quality safety of the aquatic product to be detected, and adding the rating data into a data set; and training the LSTM-RBF early warning model by using the data set, and performing quality safety risk early warning on the monitored aquatic products by using the trained LSTM-RBF early warning model.

Description

Technical Field

[0001] The present invention relates to the technical field of aquatic product quality assessment, and particularly relates to a method for warning of aquatic product quality and safety risks based on HACCP internal control data. Background Art

[0002] HACCP represents the Hazard Analysis and Critical Control Points, which ensures the safety of food during the processes of production, processing, manufacturing, preparation and consumption, and is a scientific, reasonable and systematic method in terms of hazard identification, evaluation and control. The HACCP system is an important preventive control measure for food production enterprises to ensure food safety, and is also an important part of the technical trade measures adopted by various countries and regions. It is a globally recognized and accepted food safety assurance system. Many scholars have also carried out multi-angle research on the implementation of the HACCP system, demonstrating the importance of the HACCP system in monitoring and preventing hazards during the food processing process, and also indicating that the implementation of the HACCP plan can better ensure food quality and safety.

[0003] As an important safety standard for the production process of aquatic products, the HACCP plan has become the basic norm for internal safe production in enterprises. For example, Giorgio Smaldone et al. evaluated the exposure risks of parasites in different fish species through a risk classification scheme, constructed a risk-based early warning model, and enhanced the monitoring and management capabilities of parasite risks in fish products; Li Jiali et al. integrated the HACCP plan and blockchain technology, and improved the data uploading mechanism of the blockchain traceability system for the fruit supply chain; Sun Kangting et al. used white shrimp as raw materials to establish a radial basis function neural prediction model for relevant protein indicators to predict the changes in the muscle protein quality of white shrimp during micro-freezing storage; Lixing Wang et al. combined radio frequency identification with the Internet to implement a full supply chain tracking system for the HACCP plan, and so on.

[0004] However, most of the research is carried out separately for the HACCP plan or risk warning, and rarely uses the internal control data of the HACCP plan for risk warning. Since the use of the HACCP plan in the field of aquatic products is an international standard, it is of practical significance to combine the HACCP plan with risk warning. Summary of the Invention

[0005] The present invention provides a method for warning of aquatic product quality and safety risks based on HACCP internal control data. Based on the HACCP plan and combined with the existing internal data of the enterprise regarding quality and safety risk supervision, risk warning is carried out to comprehensively analyze the inherent risks of aquatic products and achieve warning, so as to make up for the deficiencies of existing research.

[0006] The present invention can be realized through the following technical solutions:

[0007] A method for early warning of aquatic product quality and safety risks based on HACCP internal control data, comprising the following steps:

[0008] Step 1: Based on the HACCP system and combined with the actual business process of the aquatic products to be inspected, sort out the details of risk early warning indicators, and obtain the time series data of each risk early warning indicator to construct a data set;

[0009] Step 2: Use the analytic hierarchy process, introduce a judgment matrix for comparing the importance of indicators, have multiple experts complete the judgment matrix respectively, and complete the consistency check to obtain the comprehensive utility value of multiple experts; then use the entropy weight method to process the utility values of multiple experts to obtain the member entropy weight corresponding to each expert, select the patent corresponding to the highest member entropy weight as the representative expert, give the rating data of the quality and safety of the aquatic products to be inspected, and add it to the data set;

[0010] Step 3: Construct an LSTM-RBF early warning model based on the LSTM neural network model and the RBF neural network model, and use the data set to train the LSTM-RBF early warning model;

[0011] Step 4: Use the trained LSTM-RBF early warning model to conduct risk early warning on the quality and safety of the aquatic products to be inspected.

[0012] Furthermore, the LSTM-RBF early warning model includes an input layer, a hidden layer and an output layer. Among them, the input layer is provided with 6 neurons to receive risk early warning indicator data. The first layer of the hidden layer is an LSTM layer containing 64 LSTM units. The second layer of the hidden layer is an LSTM layer containing 32 LSTM units. The third layer of the hidden layer is a fully connected layer containing 64 neurons, using the ReLU activation function. The fourth layer of the hidden layer is an RBF layer containing 32 RBF units. The output layer uses the softmax function to output the rating probability distribution.

[0013] Furthermore, when constructing the judgment matrix, the N risk early warning indicators in the risk early warning indicator details are respectively denoted as A1, A2,..., A N , arrange these risk early warning indicators in order horizontally and vertically. Experts compare the importance of risk early warning indicator A i and risk early warning indicator A j pairwise, and use an integer with a scale value of 1-9 as the comparison result. When A i and A j are equally important, the matrix element a ij = 1; when A i is slightly more important than A j , a ij = 3; when A i is significantly more important than A j , aij = 5; When A i is more important than A j is extremely important, a ij = 7; When A i is more strongly important than A j is strongly important, a ij = 9; The scale values 2, 4, 6, 8 represent the intermediate values of the above judgments; and the matrix element a ji = 1 / a ij , thus obtaining the judgment matrix corresponding to this expert;

[0014] Normalize the matrix elements of the judgment matrix by column, then sum by row and normalize again to obtain the index weights of the risk warning indicators of the judgment matrix corresponding to each expert. After passing the consistency check, this index weight can be used as the utility value of the corresponding expert to form the utility matrix V,

[0015]

[0016] where V i represents the utility value provided by expert i, v i1 , v i2 , …, v iN represents the utility values calculated by expert i for all risk warning indicators;

[0017] Next, use the entropy weight method to process the utility matrix V constructed based on the multi-expert utility values to obtain the member entropy weights corresponding to each expert, and select the expert corresponding to the highest member entropy weight as the representative expert.

[0018] Furthermore, the rating data is set as the quality and safety risk rating of aquatic products by the preferred expert for the provided monitoring data set. The risk levels are defined as level 1 to level 4 from low to high, with level 1 having the lowest risk and level 4 being unqualified.

[0019] The beneficial technical effects of the present invention are as follows:

[0020] (1) Based on the national HACCP standard for aquatic products, taking the HACCP plan for raw oysters as an example and combining with the deep learning algorithm, the present invention constructs an LSTM-RBF early warning model based on HACCP internal control data, exploring a new way to make good use of the large amount of internal control data accumulated by enterprises and realizing the early warning of aquatic product quality and safety risks.

[0021] (2) Taking the HACCP plan for raw oysters as an example, combining with the production and processing process, extract the monitoring points in the business process, and at the same time sort out the key internal control points to construct a detailed list of risk warning indicators to assist the refined monitoring, early warning and feedback of the internal quality of the enterprise.

[0022] (3) After testing, using the LSTM-RBF early warning model for risk early warning, the accuracy rates are 96% and 90% respectively on the qualified dataset and the complete dataset, and the test effect on the qualified dataset is significantly better.

[0023] (4) Through the quality and safety risk early warning model architecture, the internal control data that has been deposited in enterprises for many years can be activated, and the expert experience knowledge in the field and the entropy weight method are used to fully optimize group decision-making, forming the basic data source for deep learning. Then, with the help of various deep learning models, potential risk points are deeply explored, and risk points are identified earlier from the source, strangling quality and safety problems in the bud, and reducing the quality and safety risks of enterprises. Therefore, this is a method for discovering and reducing the quality and safety risks of enterprises at low cost. Description of the Drawings

[0024] Figure 1 It is a schematic diagram of the quality and safety risk early warning model architecture corresponding to the risk early warning method of the present invention;

[0025] Figure 2 It is a key internal control point diagram of the HACCP plan business process of the present invention;

[0026] Figure 3 It is a flowchart corresponding to the risk early warning method of the present invention. Detailed Embodiment

[0027] The following details the specific embodiments of the present invention in conjunction with the accompanying drawings and preferred embodiments.

[0028] Given the wide variety of aquatic products, each type of aquatic product also has its particularity in quality and safety risk prevention. For example Figure 1 、 3 As shown, the present invention provides a quality and safety risk early warning method for aquatic products based on HACCP internal control data. First, based on the HACCP system and combined with the actual business process of the aquatic products to be inspected, the detailed list of risk early warning indicators is sorted out, and the time series data of each risk early warning indicator is obtained to construct a dataset. Then, using the analytic hierarchy process, the judgment matrix for comparing the importance of indicators is introduced, and multiple experts each complete the judgment matrix and perform consistency verification to obtain the comprehensive utility value of multiple experts. Then, the entropy weight method is used to process the utility values of multiple experts, so as to select a representative expert to give the rating data of the quality and safety of the aquatic products to be inspected and add it to the dataset. Finally, an LSTM-RBF early warning model is constructed based on the LSTM neural network model and the RBF neural network model, the dataset is used to train the LSTM-RBF early warning model, and the trained LSTM-RBF early warning model is used to conduct risk early warning on the quality and safety of the aquatic products to be inspected.

[0029] Taking raw oysters as an example, based on the established quality safety risk monitoring and traceability model, on the one hand, borrowing the internal control data of daily monitoring that can be aggregated by the quality traceability model, and on the other hand, considering the dynamics of the HACCP plan, the coverage scope of the internal control data required for early warning is expanded based on the HACCP plan, and an LSTM-RBF early warning model is constructed, so as to strengthen the implicit quality safety risk management of internal control data by aquatic product cold chain processing and logistics enterprises, and achieve the goals of cost reduction and efficiency improvement.

[0030] The method specifically includes the following steps:

[0031] Step S1: Combine the HACCP plan for raw oysters with the processing and transportation business process of raw oysters to construct a detailed list of risk warning indicators.

[0032] In the HACCP plan, relevant factors affecting the quality of raw oysters and generating quality risks can be understood. The risk links in the business process are shown in Table 1. In the live oyster receiving link, for the hazards of possible pathogenic bacteria contamination, environmental chemical pollutants, and natural toxins, rejecting raw materials from closed waters, non-permitted waters, without fishing identification signs or with improper identification can effectively control. The possible hazard is the growth of pathogenic bacteria.

[0033] Table 1 Risk links in the business process

[0034]

[0035] In the dry refrigeration link, the possible hazard is the growth of pathogenic bacteria. In the meat stripping, cleaning / draining, and packaging links, there may be hazards of pathogenic bacteria growth. Among them, after analysis, the internal defect of oyster shell breaking is unlikely to cause food insecurity, and metal fragments are also unlikely to occur. In the oyster meat storage link, the growth of pathogenic bacteria is a possible hazard. Combining the enterprise research experience, through the above HACCP plan analysis, the key internal control points of the HACCP plan business process are summarized as Figure 2 shown below.

[0036] Since the meat stripping time, cleaning / draining time, and packaging time are actually accumulated in the storage time, for the sake of representativeness of the research, currently only limited to the key control points of the HACCP plan, the following detailed list of risk warning indicators is constructed:

[0037] Time from fishing to receiving, temperature in the dry refrigeration warehouse, storage time, temperature in the cold storage warehouse, time exceeding the specified temperature, and time from the dry refrigeration warehouse to the oyster storage room.

[0038] Step 2: According to the detailed list of risk warning indicators, collect, sort out and perform corresponding preprocessing on the internal control data in the enterprise production process, and construct an early warning data set.

[0039] Step 21: Obtain the internal control data of the HACCP plan based on the risk early warning index system. The specific method is as follows:

[0040] The temperature data of the specified internal control points is collected by sensors, and the completion time points of some processes are manually recorded. The time interval is calculated through the time points of the data recorded by sensors and the manually recorded data, so as to obtain the time-related data, and thus form the required internal control data, that is, the time series data corresponding to each key internal control point.

[0041] Step 22: Preprocess the internal control data. The specific method is as follows:

[0042] Adopt the standardization method, that is, scale the eigenvalue to the standard normal distribution with a mean of 0 and a standard deviation of 1. The formula is as follows:

[0043]

[0044] Among them, x i is the original monitoring data corresponding to the current risk early warning index, x i ′ is the standardized monitoring data, μ is the mean of the monitoring data of the current risk early warning index σ is the standard deviation of the monitoring data of the current risk early warning index N is the total number of monitoring data samples.

[0045] First, unify the units of the monitoring data covered in the model, and then use the above standardization formula to complete the preprocessing of the unified dimension of the monitoring data.

[0046] Step 3: Use the expert optimization method to select a representative expert to give the rating data of the quality and safety of the aquatic products to be inspected and add it to the data set.

[0047] Step 31: Provide the judgment matrix in the analytic hierarchy process to multiple experts and obtain the utility value of each expert. The specific method is as follows:

[0048] In order to obtain the judgment of the importance of the monitoring indicators by experts relatively accurately, a judgment matrix for comparing the importance of indicators is introduced. Multiple experts complete the judgment matrix respectively, calculate each judgment matrix and complete the consistency check, and then a multi-expert comprehensive utility value scheme can be obtained.

[0049] Provide the judgment matrix of the analytic hierarchy process as shown in Table 2. In this matrix, experts make pairwise comparisons of the importance of the risk early warning indicators Ai (i = 1, 2,..., 6) and Aj (j = 1, 2,..., 6), and the scale value is an integer from 1 to 9. When the two indicators Ai and Aj are equally important, the matrix element a ij = 1; when Ai is slightly more important than Aj, a ij = 3; when Ai is significantly more important than Aj, aij = 5; When Ai is extremely more important than Aj, a ij = 7; When Ai is strongly more important than Aj, a ij = 9, and the scale values 2, 4, 6, 8 represent the intermediate values of the above judgments; and the matrix element a ji = 1 / a ij . Thus, after obtaining the upper triangular or lower triangular matrix according to the expert opinions, the complete judgment matrix A can be obtained.

[0050] After obtaining the judgment matrix A, using the summation method for calculating relative importance in the Analytic Hierarchy Process (AHP), the weights of each risk warning index are obtained. After consistency check, if C.R. < 0.1, the consistency of the judgment matrix can be accepted; otherwise, it indicates that there is an inconsistency in the judgment matrix given by the experts, and it is necessary to re - judge the judgment matrix, that is, re - verify the importance level determination of each risk warning index. And the weights of the risk warning indexes that pass the consistency check can be used as the utility values of each expert.

[0051] Table 2 Expert Judgment Matrix

[0052]

[0053] The utility value schemes of multiple experts are shown in Table 3. Each row of the table represents the utility value scheme of an expert. Each value is a weight value between 0 and 1, reflecting the expert's determination of the importance of the index. The larger the weight value, the more important the expert considers the item to be for the rating.

[0054] Table 3 Comprehensive Utility Value Scheme of Multiple Experts

[0055]

[0056] Step 32: Use the entropy weight method to process the expert utility value scheme, so as to select a representative expert to give the rating data and add it to the dataset. The specific method is as follows:

[0057] Using the entropy weight model, according to the expert utility value scheme, calculate the expert weights. The expert utility value scheme shown in Table 2 constitutes a utility matrix as shown in Formula 9:

[0058]

[0059] Among them, V i is the utility value scheme provided by expert i, v i1 , v i2 , …, v iN is the utility value calculated by expert i for the j - th risk warning index. In the current risk warning index system, N = 6, that is, the six - column indexes in Table 3; M is the number of experts in Table 3.

[0060] Normalize V to obtain R = {r ij}, where

[0061]

[0062] Calculate the membership entropy value E i , and there is

[0063]

[0064] Calculate the membership entropy weight θ i , and there is

[0065]

[0066] where, 0 ≤ θi ≤ 1,

[0067] Use the entropy weight method to perform calculations based on the data in the multi-expert utility value scheme, obtain the membership entropy weights corresponding to each expert, and require the expert utility scheme corresponding to the maximum membership entropy weight as the representative expert utility value scheme. Furthermore, determine the rating data given by this expert for the quality and safety of the aquatic products to be inspected, and add it to the dataset.

[0068] Among them, the rating data is set as the risk level given by the representative expert for the quality and safety of the aquatic products to be inspected provided in the dataset. This risk level includes levels 1 to 4, where level 1 is the best, indicating that the storage and processing of oysters fully comply with the HACCP standard, the quality is in the best state, and there is no quality and safety risk; level 2 indicates that the quality of oysters is qualified, but some key internal control points are close to or slightly exceed the operating limits, or the expert believes that the appearance is a bit imperfect, belonging to low risk; level 3 indicates that the quality of oysters is qualified, but some key internal control points far exceed the operating limits, are close to the critical limits, or the expert believes that there are some problems with the appearance, there is a certain potential risk, high data risk, and measures need to be taken to reduce the risk; level 4 indicates that the quality does not meet the standard, some key internal control points exceed the critical limits, or the expert observes from the appearance and believes that it is unqualified. This is a state that is not expected in quality and safety risk control. It is also found in communication with domain experts that if the monitoring data exceeds the critical limit, without expert judgment, they will also directly process it as an unqualified product, that is, mark it as level 4. It is possible to select to store the aquatic products to be inspected in the warehouse for a period of time and then conduct quality rating, or select other time periods. Therefore, the dataset constructed including the internal control data corresponding to unqualified aquatic products is the complete dataset, and the dataset constructed without including the internal control data corresponding to unqualified aquatic products is the qualified dataset.

[0069] Step S4: Construct an LSTM-RBF early warning model and conduct early warning.

[0070] LSTM relies on the non - linear transformation of activation functions, and its non - linear expression ability is limited. While RBF maps the input data to a new feature space through the Gaussian kernel function, thus increasing the non - linear expression ability of the model, but it cannot perform deep - level feature extraction on the original data.

[0071] The LSTM neural network is a special type of recurrent neural network (RNN), specifically designed to address the vanishing gradient problem that ordinary RNNs face when dealing with long - sequence data. By introducing memory cells to capture long - term dependencies, although LSTM is typically used for processing time - series data, its gating mechanism allows it to selectively retain or discard information. Even without temporality, these mechanisms can still effectively capture complex relationships in multi - dimensional data. LSTM has a strong ability to screen input features and can effectively handle complex inter - relationships between features, which helps to more accurately extract the information required for early warning. The RBF neural network is a type of feed - forward neural network. It maps the input space of the network to a hidden - layer space, and the hidden layer usually uses the radial basis function as the activation function, while the output layer performs a linear combination to produce the final result.

[0072] Fusing LSTM and RBF to form the LSTM - RBF early - warning model. By using LSTM to extract deep - level features of the data and form meaningful feature representations, and then using RBF to perform multi - center non - linear mapping on these features, it can better handle complex classification boundaries. This enables the LSTM - RBF early - warning model to not only capture complex patterns in the data through the LSTM layer but also enhance the non - linear separation ability through the RBF layer, thereby improving the accuracy of classification.

[0073] The input layer of the early warning model constructed by the combination of LSTM and RBF sets 6 neurons to receive risk early warning index data. The first hidden layer is an LSTM layer, the second hidden layer is an LSTM layer, and the third hidden layer is a fully connected layer, containing 64 neurons, using the ReLU activation function. The fourth hidden layer is an RBF layer, and the output layer uses the softmax function to output the rating probability distribution. To avoid the problem of overfitting, a Dropout layer is added after each LSTM layer to prevent overfitting. In this way, LSTM extracts the deep features of the data to form meaningful feature representations, and the RBF layer performs multi-centered non-linear mapping on these features, which can better handle complex classification boundaries, thus effectively making up for the deficiencies of their respective single models and making the model more advantageous in processing complex and non-linear data. The first layer of LSTM selects 64 units, which is suitable for medium and small tasks and can balance the computational complexity and feature capture ability. The second layer of LSTM selects 32 units to gradually reduce the dimension of the feature representation, extract higher-order abstract features, and reducing the number of units can reduce the number of model parameters and the risk of overfitting. Since the first two LSTM layers extract features and reduce the dimension of the input data and output high-level abstract features, the RBF layer requires fewer nodes to further refine the feature space, so the number of units in the RBF layer is 32.

[0074] The LSTM-RBF model can effectively make up for the deficiencies of their respective single models and make the model more advantageous in processing complex and non-linear data. By using LSTM to extract the deep features of the data to form meaningful feature representations, the RBF layer performs multi-centered non-linear mapping on these features, which can better handle complex classification boundaries. This enables the LSTM-RBF model not only to capture complex patterns in the data through the LSTM layer but also to enhance the non-linear separation ability through the RBF layer, thereby improving the accuracy of classification.

[0075] The LSTM-RBF early warning model has an accuracy rate of 96% and 90% on the qualified dataset and the complete dataset respectively. The comparison results with the LSTM neural network model and the RBF neural network model used alone are shown in Table 4.

[0076] Table 4 Comparison of accuracy rates of different models on different datasets

[0077]

[0078] After completing the training and verification of the LSTM-RBF early warning model, early warning is carried out. Eight example data are input into the model for risk rating prediction, and the model predicts the risk level results.

[0079] Although the specific embodiments of the present invention have been described above, those skilled in the art should understand that these are only examples. Without departing from the principle and essence of the present invention, various changes or modifications can be made to these embodiments. Therefore, the protection scope of the present invention is defined by the appended claims.

Claims

1. A method for early warning of aquatic product quality and safety risks based on HACCP internal control data, characterized in that The following steps are involved: Step 1: Based on the HACCP system and the actual business process of the aquatic products to be inspected, sort out the details of the risk warning indicators, and obtain the time series data of each risk warning indicator to construct a data set; Step 2: Using the analytic hierarchy process, a judgment matrix for comparing the importance of indicators is introduced. Multiple experts complete the judgment matrix and complete the consistency check to obtain the comprehensive utility value of multiple experts. Then, the entropy weight method is used to process the utility values ​​of multiple experts to obtain the member entropy weight corresponding to each expert. The patent corresponding to the highest member entropy weight is selected as the representative expert, and the rating data of the quality and safety of the aquatic products to be inspected is given and added to the data set. Step 3: Build an LSTM-RBF early warning model based on the LSTM neural network model and the RBF neural network model, and use the data set to train the LSTM-RBF early warning model; Step 4: Use the trained LSTM-RBF early warning model to issue risk warnings for the quality and safety of the water products to be inspected.

2. The aquatic product quality safety risk early warning method based on HACCP internal control data according to claim 1, characterized in that: The LSTM-RBF early warning model includes an input layer, a hidden layer and an output layer, wherein the input layer is provided with 6 neurons to receive risk early warning indicator data, the first layer of the hidden layer includes an LSTM layer of 64 LSTM units, the second layer of the hidden layer includes an LSTM layer of 32 LSTM units, the third layer of the hidden layer is a fully connected layer, including 64 neurons, using a ReLU activation function, the fourth layer of the hidden layer includes an RBF layer of 32 RBF units, and the output layer uses a softmax function to output a rating probability distribution.

3. The aquatic product quality safety risk early warning method based on HACCP internal control data according to claim 1 is characterized in that: When constructing the judgment matrix, the N risk warning indicators in the risk warning indicator details are recorded as A1, A2, ..., A N , these risk warning indicators are arranged in order horizontally and vertically, and experts have made a i and risk warning indicator A j The importance of A is compared with A, and the integers of 1-9 are used as the comparison results. i and A j When the matrix elements a are of equal importance, ij =1; when A i Ratio A j When it is slightly more important, a ij =3; when A i Ratio A j When it is obviously important, a ij =5; when A i Ratio A j When it is extremely important, a ij =7; when A i Ratio A j When it is strongly important, a ij =9; scale values ​​2, 4, 6, 8 represent the intermediate values ​​of the above judgment; and matrix element a ji =1 / a ij , so as to obtain the judgment matrix corresponding to the expert; The matrix elements of the judgment matrix are normalized by column, and then normalized by row summation to obtain the indicator weight of the risk warning indicator of the judgment matrix corresponding to each expert. After consistency verification, the indicator weight can be used as the utility value of the corresponding expert to form the utility matrix V. Among them, V i represents the utility value provided by expert i, v i1 ,v i2 ,…,v iN represents the utility value calculated by expert i for all risk warning indicators; Next, the entropy weight method is used to process the utility matrix V constructed based on the utility values ​​of multiple experts to obtain the member entropy weight corresponding to each expert, and the expert corresponding to the highest member entropy weight is selected as the representative expert.

4. The aquatic product quality safety risk early warning method based on HACCP internal control data according to claim 1 is characterized in that: The rating data is set as the quality and safety risk rating of aquatic products of the provided monitoring data set by the preferred experts, and the risk levels are defined from low to high as one to four, with one being the lowest risk and four being unqualified.

Images