Intelligent cephalopod food nutritional quality regulation and control method

Abstract

The invention relates to the technical field of food processing, and discloses an intelligent cephalopod food nutritional quality regulation and control method, which comprises the following steps: S1, data acquisition and deployment: respectively arranging sensor combinations in a frying area, a coating area, a conveying area and an environment monitoring point of a cephalopod food coating and frying production line; the method comprises the following steps: S1, generating a processing variable feature set and a quality feature set, S3, screening key data, S4, generating regulation and control parameters, and S5, performing real-time regulation and control. According to the method, multiple types of sensors are deployed in a key area of a production line, data such as oil temperature and humidity are collected according to the frequency of 50 ms / time, feature extraction, a random forest and a BP neural network coupling algorithm are combined, the nonlinear relation between processing variables and quality and the dynamic relation between the material progress and a failure event are analyzed, and the processing accuracy is improved. The problems that in traditional production, real-time information is difficult to obtain, and the variable influence rule is unknown are solved, and comprehensive data support and scientific basis are provided for accurate regulation and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food processing, in particular to an intelligent cephalopod food nutritional quality control method. BACKGROUND

[0002] The coating and frying processing of cephalopod food materials (such as squid, octopus, etc.) is a core link in the production of high-quality processed food, and its nutritional quality (such as protein retention rate, fat oxidation degree) is closely related to the processing process. It is necessary to accurately control the processing and environmental variables such as oil temperature, humidity, pressure, powder ratio, conveying speed, combined with the control of production status such as material progress and capacity loss rate, to realize the stable improvement of product quality.

[0003] In the production of high-quality cephalopod processed food, the existing technology has carried out related research on nutritional quality control, including the development of intelligent large-scale quality control technology and supporting processing equipment, which is used to preliminarily collect part of the parameters (such as oil temperature, conveying speed) of the production line, and through basic data statistical analysis, the influence of processing variables on product quality is analyzed to provide reference for manual adjustment or simple automatic control.

[0004] However, in the existing technology, there are obvious limitations in real-time processing information acquisition of cephalopod coating and frying production line: it is difficult to build a real-time acquisition system covering the whole process through existing technology and equipment, it is difficult to deeply analyze the influence law of oil temperature, humidity and other variables on processing quality, and the correlation between material progress, capacity loss rate and processing failure event probability, resulting in that the multi-source dynamic data closely related to the quality control of the processing process cannot be clearly determined, and the real-time and effective acquisition of production line information is difficult to achieve. In view of this, we propose an intelligent cephalopod food nutritional quality control method. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an intelligent cephalopod food nutritional quality control method, which solves the problem of obvious limitations in real-time processing information acquisition of cephalopod coating and frying production line in the prior art.

[0006] To achieve the above purpose, the present application is realized by the following technical scheme:

[0007] An intelligent cephalopod food nutritional quality control method, comprising the following steps:

[0008] S1 data acquisition deployment: set sensor combination in the frying zone, coating zone, conveying zone and environmental monitoring point of the cephalopod food coating and frying production line, collect oil temperature, humidity, pressure, powder ratio, conveying speed and material progress data, and transmit them to the data processing unit through industrial Ethernet;

[0009] S2 variable correlation analysis: the data processing unit calls the algorithm suite to process the data collected in S1, extracts features from the raw data to generate a set of processing variable features and a set of quality features, and sorts the importance of the features to filter out the core variables, and simultaneously constructs a nonlinear mapping relationship between the processing and environmental variables and the processing quality by taking the core variables as input, and simultaneously calculates the dynamic correlation degree of material progress, capacity loss rate and processing failure event occurrence probability through the time series analysis submodule;

[0010] S3 key data filtering: based on the variable importance sorting results and nonlinear mapping relationship strength output in S2, set the correlation strength threshold, and determine the processing and environmental variables and material progress data with correlation strength exceeding the threshold as strongly related production line multi-source dynamic data for processing process quality control;

[0011] S4 control parameter generation: the control decision unit loads the key data filtered in S3 and calls the intelligent optimization algorithm; the algorithm takes the output of the nonlinear mapping relationship constructed in S2 as the prediction benchmark, combines the physical constraint parameters of the production line equipment, and calculates and generates the adjustment amount sequence of the processing parameters through the rolling optimization mechanism, which includes oil temperature adjustment value, humidity adjustment value, pressure adjustment value, powder ratio adjustment value and conveying speed adjustment value;

[0012] S5 real-time control execution: the execution unit receives the adjustment amount sequence generated in S4, sends the oil temperature adjustment value to the controller of the frying equipment, sends the humidity adjustment value to the controller of the ventilation equipment, sends the pressure adjustment value to the pressure control valve of the coating cavity, sends the powder ratio adjustment value to the controller of the powder conveying pump, and sends the conveying speed adjustment value to the driving device of the conveying belt. Each control component executes the adjustment instruction.

[0013] Preferably, in the S1 data collection deployment, at least three oil temperature sensors are arranged in the frying zone in a triangular distribution, the oil temperature sensor uses a K-type thermocouple, and the measurement range is 50-300℃; the humidity sensor uses a capacitive humidity sensor, and the sampling rate is consistent with the sensor acquisition frequency; the pressure sensor uses a diffused silicon pressure transmitter, which is installed at the inlet and outlet positions of the coating cavity.

[0014] Preferably, in the S2 variable correlation analysis, the feature extraction includes time domain feature and frequency domain feature conversion of the raw data; the time domain features include mean, variance, peak factor and kurtosis of the data, and the frequency domain features are obtained through Fourier transform; the dimension of the processing variable feature set is 8, and the dimension of the quality feature set is 6, which is based on the color parameters, texture parameters and nutritional ingredient parameters of cephalopod food.

[0015] Preferably, in the S2 variable correlation analysis, the random forest model is used for variable importance sorting, and the core variables with the top 30% of the influence weight are screened out, the random forest model contains 500 decision trees, and the maximum depth is set to 15; the BP neural network model is used to construct the nonlinear mapping relationship between the processing and environmental variables and the processing quality with the core variables as the input, the BP neural network model contains 3 layers of hidden layers, the number of neurons of each layer is 64, 32 and 16 respectively, the hidden layer activation function uses the ReLU function, and the output layer uses the linear activation function.

[0016] Preferably, in the S2 variable correlation analysis, the training of the BP neural network model uses the Adam optimizer, the learning rate is set to 0.001, the iteration number is set to 1000 times, the division ratio of the training set and the test set is 7:3, and the training termination condition is that the loss function value is lower than 0.001.

[0017] Preferably, in the S2 variable correlation analysis, the time series analysis submodule uses the sliding window method to calculate the correlation degree, the window length is set to 50 sampling periods, the time delay parameter is set to 5 sampling periods, and the correlation degree is quantified by mutual information entropy.

[0018] Preferably, in the S3 key data screening, the correlation strength threshold is determined by the 90th percentile of the historical 3-month production data, the correlation strength between the processing and environmental variables and the processing quality is calculated by the Pearson correlation coefficient, and when the value range is 0.6-1.0, it is determined as strong correlation.

[0019] Preferably, in the S4 control parameter generation, the intelligent optimization algorithm is a model predictive control algorithm, the prediction time domain is set to 10 sampling periods, and the control time domain is set to 5 sampling periods; the device physical constraint parameters include the oil temperature adjustment value range of-10℃ to +10℃, the humidity adjustment value range of-5% RH to +5% RH, and the pressure adjustment value range of-0.02MPa to +0.02MPa.

[0020] Preferably, in the S4 control parameter generation, the rolling optimization objective function of the model predictive control algorithm contains three items: the square sum of the processing parameter adjustment amount, the square sum of the predicted quality deviation, and the square sum of the parameter adjustment rate, and the weight proportion of the three items is 2:6:2.

[0021] Preferably, in the S5 real-time control execution, the controller of the frying equipment is a PID controller, the proportional coefficient is set to 5.0, the integral time is set to 10s, and the differential time is set to 2s; the speed control accuracy of the conveyor belt driving device is ±0.1m / min.

[0022] The application provides an intelligent cephalopod food nutrition quality control method.

[0023] 1. The present application solves the problems of difficult real-time information acquisition and unknown variable influence law in traditional production by deploying multiple types of sensors in key areas of the production line, collecting data such as oil temperature and humidity at a frequency of 50ms / time, combining feature extraction, random forest and BP neural network coupling algorithm, analyzing the nonlinear relationship between processing variables and quality and the dynamic correlation between material progress and failure events, and providing comprehensive data support and scientific basis for precise control.

[0024] 2. The present application solves the problems of difficult real-time information acquisition and unknown variable influence law in traditional production by deploying multiple types of sensors in key areas of the production line, collecting data such as oil temperature and humidity at a frequency of 50ms / time, combining feature extraction, random forest and BP neural network coupling algorithm, analyzing the nonlinear relationship between processing variables and quality and the dynamic correlation between material progress and failure events, and providing comprehensive data support and scientific basis for precise control.

[0025] 3. The present application solves the problems of difficult real-time information acquisition and unknown variable influence law in traditional production by deploying multiple types of sensors in key areas of the production line, collecting data such as oil temperature and humidity at a frequency of 50ms / time, combining feature extraction, random forest and BP neural network coupling algorithm, analyzing the nonlinear relationship between processing variables and quality and the dynamic correlation between material progress and failure events, and providing comprehensive data support and scientific basis for precise control. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The present application is an intelligent cephalopod food nutrition quality control method flowchart. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the present application specification.

[0028] Embodiment:

[0029] Please refer to the accompanying Figure 1 The present application provides an intelligent cephalopod food nutrition quality control method, comprising the following steps:

[0030] S1 data acquisition deployment: In the oil frying area, coating area, conveying area and environmental monitoring point of the cephalopod food coating and frying production line, a combination of sensors is set up, wherein at least three oil temperature sensors are set up in the oil frying area in a triangular distribution, pressure sensors and powder flow sensors are set up in the coating area, speed sensors and material position sensors are set up in the conveying area, and humidity sensors are set up in the environmental monitoring point; each sensor collects oil temperature, humidity, pressure, powder ratio, conveying speed and material progress data at a frequency of 50ms / time, and the collected data is transmitted to the data processing unit through industrial Ethernet;

[0031] In the S1 data acquisition deployment, the oil temperature sensor adopts a K-type thermocouple, the measurement range is 50-300℃; the humidity sensor adopts a capacitive humidity sensor, the sampling rate is consistent with the sensor acquisition frequency; the pressure sensor adopts a diffused silicon pressure transmitter, which is installed at the inlet and outlet positions of the coating cavity;

[0032] S2 variable correlation analysis: the data processing unit calls an algorithm suite to process the data collected in S1, the algorithm suite includes a feature extraction module and a double model coupled analysis module; the feature extraction module converts the original data into time domain features and frequency domain features to generate a processing variable feature set and a quality feature set; in the double model coupled analysis module, the first model sorts the feature set by variable importance and selects the top 30% of the core variables with the highest impact weight, and the second model constructs a nonlinear mapping relationship between the core variables, processing and environmental variables, and processing quality, and simultaneously calculates the dynamic correlation degree of material progress, capacity loss rate, and processing failure event occurrence probability through a time series analysis submodule;

[0033] In the S2 variable correlation analysis, the time domain features of the feature extraction module include the mean, variance, peak factor, and kurtosis of the data, and the frequency domain features are obtained through Fourier transform; the processing variable feature set has a dimension of 8, and the quality feature set is based on the color parameters, texture parameters, and nutritional ingredient parameters of cephalopod food, with a dimension of 6, including the following algorithms:

[0034] First, calculate the time domain features of the collected original data, first calculate the mean, this step is to obtain the central tendency representative value of the data, the specific method is to sum all the collected sample data, and then divide by the total number of samples, that is Where N represents the total number of samples participating in the calculation, x i represents the i-th specific sample value;

[0035] Then calculate the variance to reflect the dispersion degree of the data, when calculating, first find the difference between each sample value and the mean, square the difference values, and then divide by the total number of samples minus 1, the formula is Here σ 2 represents the variance, and μ is the mean calculated earlier;

[0036] Then calculate the peak factor, which is used to reflect the prominence of the signal peak, the calculation method is to divide the maximum value of the absolute value of all sample values by the effective value of the sample value, and the effective value is the square root of the mean of the square of the sample value, the formula is expressed as Where max(|x i |) is the maximum value of the absolute value of the sample value;

[0037] The kurtosis is recalculated to reflect the steepness of the signal distribution. The difference between each sample value and the mean is divided by the standard deviation to obtain a normalized value. The fourth power of the normalized value is taken, and the mean is subtracted by 3, that is, Here, σ is the standard deviation, which is the square root of the variance.

[0038] After the time domain feature calculation is completed, the frequency domain feature extraction is performed. The Fourier transform is used to convert the time domain signal to the frequency domain, so as to extract the frequency characteristics of the signal. The formula is (where k=0, 1,..., N-1), x(n) is the nth signal value in the time domain, X(k) is the kth component in the converted frequency domain, and j is the imaginary unit. Through this transformation, the periodic characteristics hidden in the data can be captured, such as periodic fluctuations in device operation.

[0039] The obtained 4 time domain features and 4 frequency domain features are combined together to form an 8-dimensional processing variable feature set; at the same time, based on the color parameters (such as brightness, color difference) of cephalopod food, the texture parameters (such as hardness, elasticity) and the nutritional ingredient parameters (such as protein content, fat content), 2 key indicators are selected respectively to construct a 6-dimensional quality feature set;

[0040] In the S2 variable correlation analysis, the first model of the double model coupling analysis module is a random forest model, containing 500 decision trees, and the maximum depth is set to 15; the second model is a BP neural network model, containing 3 hidden layers, and the number of neurons in each layer is 64, 32 and 16 respectively. The hidden layer activation function uses the ReLU function, and the output layer uses the linear activation function, including the following algorithms:

[0041] First, a random forest model is constructed, which is composed of 500 decision trees, and the maximum depth is set to 15. A single decision tree is constructed using the CART algorithm, and the Gini index is used as the criterion for node splitting during the construction process. The calculation method of the Gini index is as follows:

[0042]

[0043] Wherein:

[0044] D is the current sample set to be split;

[0045] K is the total number of classes in the sample set;

[0046] p k is the proportion of the kth class of samples in the sample set D;

[0047] The Gini index under different splitting modes is calculated, the splitting mode that can make the Gini index decrease the most is selected to split the node, and finally a plurality of decision trees are constructed. The random forest model obtains the final output by voting or taking the mean value of the results of the decision trees, so as to evaluate and sort the variable importance of the extracted feature set, and screen out the core variables with the top 30% of the influence weight.

[0048] Then a BP neural network model is constructed, which contains 3 hidden layers, and the number of neurons in each layer is 64, 32 and 16 respectively. The hidden layer adopts a ReLU activation function. The output of a neuron is calculated by receiving the output of the previous layer of neurons, multiplying it by the corresponding connection weight, summing it up, adding the bias of the neuron, and finally processing it through the ReLU activation function, that is, wherein is the output of the mth layer and the lth neuron, w l i m is the connection weight from the (m-1)th layer and the ith neuron to the mth layer and the lth neuron, is the output of the (m-1)th layer and the ith neuron, is the bias of the mth layer and the lth neuron, and ReLU(x) = max(0, x). The output layer adopts a linear activation function, and the output is the sum of the product of the output of the previous layer of neurons and the corresponding weight plus the bias. The formula is wherein y k is the output of the kth node of the output layer, w k l M is the connection weight from the (m-1)th layer and the lth neuron to the mth layer and the kth neuron, is the output of the (m-1)th layer and the lth neuron is the bias of the kth neuron of the output layer. The model takes the core variables screened out by the random forest as input, and constructs a nonlinear mapping relationship between the processing and environmental variables and the processing quality.

[0049] In the S2 variable correlation analysis, the training of the BP neural network model uses the Adam optimizer, the learning rate is set to 0.001, the number of iterations is set to 1000, the division ratio of the training set and the test set is 7:3, and the training termination condition is that the loss function value is lower than 0.001. The algorithm includes the following algorithms:

[0050] The network parameters are updated and optimized by using the Adam optimizer, the learning rate is set to 0.001, the iteration number is set to 1000 times, the data set is divided into training set and test set according to the proportion of 7:3, the loss function between the predicted value and the actual value of the model is calculated during training, the weights and biases of the network are continuously adjusted by using the Adam optimizer, and the training is stopped when the loss function value is lower than 0.001, the Adam optimizer dynamically adjusts the learning rate by combining the first-order moment estimation and the second-order moment estimation, and the parameter update formula is:

[0051]

[0052] Wherein:

[0053] w t is the parameter at the t time;

[0054] w t+1 is the updated parameter;

[0055] α is the learning rate;

[0056] and are the first-order moment and second-order moment estimates after bias correction respectively;

[0057] ∈ is a small constant to prevent the denominator from being 0;

[0058] Through such a training process, the model can better fit the relationship between input and output;

[0059] In the S2 variable correlation analysis, the time series analysis submodule uses the sliding window method to calculate the correlation degree, the window length is set to 50 sampling periods, the time delay parameter is set to 5 sampling periods, and the correlation degree is quantified by mutual information entropy, including the following algorithm:

[0060] In the implementation process of the time series analysis submodule, the sliding window method is used to process the time series data, and the continuous time series data is divided into multiple continuous subsequences according to the set window length (50 sampling periods), each subsequence is a window, and the formula is:

[0061] W i =[x i ,x i+1 ,...,x i+L-1 ](where i=1,2,...,N-L+1)W i is the sequence of the i-th window, L is the window length, N is the total length of the time series data, x i is the i-th time series data point, in this way, the local features of the data in different time periods can be captured; at the same time, the time delay parameter is set to 5 sampling periods, which is used to analyze the correlation relationship of variables at different time points.

[0062] The dynamic correlation degree between the material progress, the production capacity loss rate and the processing failure event occurrence probability is quantified by mutual information entropy, and the calculation of the mutual information entropy is based on the joint probability distribution and the marginal probability distribution of the variables, and the formula is:

[0063]

[0064] Wherein:

[0065] X is the material progress or the production capacity loss rate variable;

[0066] Y is the processing failure event probability variable;

[0067] p(x,y) is the joint probability distribution of X and Y;

[0068] p(x) and p(y) are the marginal probability distributions of X and Y, respectively, and the strength of the correlation between the variables is determined by calculating the size of the mutual information entropy value;

[0069] S3 key data screening: based on the variable importance ranking result and the nonlinear mapping relationship strength output by S2, set the correlation strength threshold, and determine the processing and environmental variables and the material progress data whose correlation strength exceeds the threshold as the production line multi-source dynamic data strongly related to the processing process quality control;

[0070] In the S3 key data screening, the correlation strength threshold is determined by the 90th percentile of the historical 3-month production data, the correlation strength of the processing and environmental variables and the processing quality is calculated by the Pearson correlation coefficient, and when the value range is 0.6-1.0, it is determined as strong correlation;

[0071] S4 control parameter generation: the control decision unit loads the key data screened out by S3, and calls the intelligent optimization algorithm; the algorithm takes the output of the second model in S2 as the prediction benchmark, combines the physical constraint parameters of the production line equipment, and calculates and generates the adjustment amount sequence of the processing parameters through the rolling optimization mechanism, the adjustment amount sequence includes oil temperature adjustment value, humidity adjustment value, pressure adjustment value, powder ratio adjustment value and conveying speed adjustment value;

[0072] In the S4 control parameter generation, the intelligent optimization algorithm is a model predictive control algorithm, the prediction time domain is set to 10 sampling periods, and the control time domain is set to 5 sampling periods; the device physical constraint parameters include the oil temperature adjustment value range of-10℃ to +10℃, the humidity adjustment value range of-5% RH to +5% RH, and the pressure adjustment value range of-0.02MPa to +0.02MPa;

[0073] In the S4 regulation parameter generation, the rolling optimization objective function of the model predictive control algorithm contains three terms: the square sum of the processing parameter adjustment amount, the square sum of the predicted quality deviation, and the square sum of the parameter adjustment rate, and the weight proportion of the three terms is 2:6:2, the established function contains three weighted square sums, and the formula is:

[0074]

[0075] The first term is the square sum of the predicted quality deviation, λ1=6 is the weight thereof, is the predicted quality value at the future i time, y ref is a preset target quality value (determined based on the quality characteristics of qualified products), and the term ensures that the regulation direction is consistent with the quality improvement target by minimizing the deviation between the predicted quality and the target quality;

[0076] The second term is the square sum of the processing parameter adjustment amount, λ2=2 is the weight thereof, Δu(k+j|k represents the processing parameter adjustment amount (such as the oil temperature adjustment value ΔT, the humidity adjustment value ΔH, etc.) at the future j time planned at the current time k, and the term is used to limit the size of the adjustment amount to avoid parameter mutation to cause impact on the equipment.

[0077] The third term is the square sum of the parameter adjustment rate, λ3=2 is the weight thereof, and Δu(k+j|k)-Δu(k+j-1|k) represents the change value of the adjustment amount at adjacent times. The term is used to smooth the adjustment process and ensure stable parameter change. N c =5 is the control time domain, representing the adjustment amount planning of the actual execution of the future 5 sampling periods, and the weight proportion of the three terms is 2:6:2, which prioritizes the minimization of the quality deviation, and echoes the importance analysis of the quality characteristics in S2;

[0078] And the device physical constraint parameters, the optimization process needs to satisfy:

[0079]

[0080] Wherein:

[0081] ΔT is the oil temperature adjustment value;

[0082] ΔH is the humidity adjustment value;

[0083] ΔP is the pressure adjustment value;

[0084] These constraint ranges are determined based on the fluctuation range of the key data screened in S3 in the history of qualified production, ensuring that the adjustment amount is within the physical range of safe operation of the equipment, and closely linked with the screening results of the key data in the foregoing;

[0085] Finally, the rolling optimization calculation is performed. At each sampling period, based on the current key data and the prediction model, the control quantity sequence that minimizes the objective function J under the constraint condition is solved, and only the first adjustment quantity in the sequence is taken as the execution instruction at the current time. The above process is repeated at the next sampling period, the latest key data is used to update the prediction and optimization, and an adjustment quantity sequence containing the oil temperature adjustment value, the humidity adjustment value, the pressure adjustment value, the powder proportion adjustment value and the conveying speed adjustment value is generated. The sequence will be sent to the execution unit of S5 to realize real-time regulation and control of the production line. The whole process ensures that the regulation and control can dynamically respond to the changes in the processing process through the rolling optimization mechanism, and is adapted to the analysis result of the dynamic correlation degree in the time sequence analysis submodule S2, forming a complete closed loop from data collection, correlation analysis to parameter regulation.

[0086] S5 real-time regulation and control execution: the execution unit receives the adjustment quantity sequence generated by S4, sends the oil temperature adjustment value to the controller of the frying equipment, sends the humidity adjustment value to the controller of the ventilation equipment, sends the pressure adjustment value to the pressure control valve of the coating cavity, sends the powder proportion adjustment value to the controller of the powder conveying pump, and sends the conveying speed adjustment value to the driving device of the conveying belt. Each control component executes the adjustment instruction at a frequency of 100 ms / time.

[0087] In the S5 real-time regulation and control execution, the controller of the frying equipment is a PID controller, the proportional coefficient is set to 5.0, the integral time is set to 10 s, and the differential time is set to 2 s. The speed control accuracy of the conveying belt driving device is ±0.1 m / min.

[0088] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application. The scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for regulating the nutritional quality of cephalopod food products, characterized in that, Comprise the following steps: S1 data acquisition deployment: In the cephalopod food coating and frying production line, at least three oil temperature sensors are arranged in the frying area, and the oil temperature sensors are arranged in a triangular distribution. The oil temperature sensor adopts K-type thermocouple, and the measurement range is 50-300℃; the humidity sensor adopts a capacitive humidity sensor, and the sampling rate is consistent with the sensor acquisition frequency; The pressure sensor adopts a diffusion silicon pressure transmitter, which is installed at the inlet and outlet positions of the coating cavity. In the S2 variable correlation analysis, the feature extraction includes time domain feature and frequency domain feature conversion of the original data; The time domain features include mean, variance, peak factor and kurtosis of the data, and the frequency domain features are obtained by Fourier transform; the processing variable feature set has a dimension of 8, and the quality feature set is constructed based on the color parameters, texture parameters and nutritional ingredient parameters of cephalopod food, and has a dimension of 6. S3 key data screening: Based on the variable importance ranking results and nonlinear mapping relationship strength output by S2, set the correlation strength threshold, and determine the processing and environmental variables and material progress data with correlation strength exceeding the threshold as the production line multi-source dynamic data strongly related to the quality control of the processing process; 2. The method according to claim 1, wherein S4 control parameter generation: The control decision unit loads the key data screened out by S3, and calls an intelligent optimization algorithm; the algorithm takes the output of the nonlinear mapping relationship constructed in S2 as the prediction benchmark, combines the physical constraint parameters of the production line equipment, and calculates and generates the adjustment amount sequence of the processing parameters through the rolling optimization mechanism. The adjustment amount sequence includes oil temperature adjustment value, humidity adjustment value, pressure adjustment value, powder ratio adjustment value and conveying speed adjustment value; S5 real-time control execution: The execution unit receives the adjustment amount sequence generated by S4, sends the oil temperature adjustment value to the controller of the frying equipment, sends the humidity adjustment value to the controller of the ventilation equipment, sends the pressure adjustment value to the pressure control valve of the coating cavity, sends the powder ratio adjustment value to the controller of the powder conveying pump, and sends the conveying speed adjustment value to the driving device of the conveying belt. Each control component executes the adjustment instruction.

3. The method according to claim 1, wherein In the S1 data acquisition deployment, at least three oil temperature sensors are arranged in the frying area, and the oil temperature sensors are arranged in a triangular distribution. The oil temperature sensor adopts K-type thermocouple, and the measurement range is 50-300℃; the humidity sensor adopts a capacitive humidity sensor, and the sampling rate is consistent with the sensor acquisition frequency; The pressure sensor adopts a diffusion silicon pressure transmitter, which is installed at the inlet and outlet positions of the coating cavity. In the S2 variable correlation analysis, the feature extraction includes time domain feature and frequency domain feature conversion of the original data; The time domain features include mean, variance, peak factor and kurtosis of the data, and the frequency domain features are obtained by Fourier transform; the processing variable feature set has a dimension of 8, and the quality feature set is constructed based on the color parameters, texture parameters and nutritional ingredient parameters of cephalopod food, and has a dimension of 6.

4. The method according to claim 1, wherein In the S2 variable correlation analysis, the random forest model is used to sort the variable importance, and the core variables with the top 30% of the influence weight are screened out. The random forest model contains 500 decision trees, and the maximum depth is set to 15. The BP neural network model is used to build the nonlinear mapping relationship between the processing and environmental variables and the processing quality with the core variables as the input. The BP neural network model contains 3 hidden layers, and the number of neurons in each layer is 64, 32, and 16 respectively. The hidden layer activation function uses the ReLU function, and the output layer uses the linear activation function.

5. The method according to claim 4, wherein In the S2 variable correlation analysis, the Adam optimizer is used for training the BP neural network model, the learning rate is set to 0.001, the number of iterations is set to 1000 times, the division ratio of the training set and the test set is 7:3, and the training termination condition is that the loss function value is less than 0.

001.

6. The method according to claim 1, wherein In the S2 variable correlation analysis, the sliding window method is used to calculate the correlation degree in the time series analysis submodule, the window length is set to 50 sampling periods, the time delay parameter is set to 5 sampling periods, and the correlation degree is quantified by mutual information entropy.

7. The method according to claim 1, wherein the method is characterized by, In the S3 key data screening, the correlation strength threshold is determined by the 90th percentile of the historical 3-month production data, and the correlation strength between the processing and environmental variables and the processing quality is calculated by the Pearson correlation coefficient. When the value range is 0.6-1.0, it is determined as strong correlation.

8. The method according to claim 1, wherein the method is characterized by, In the S4 control parameter generation, the intelligent optimization algorithm is the model predictive control algorithm, the prediction time domain is set to 10 sampling periods, and the control time domain is set to 5 sampling periods. The device physical constraint parameters include the oil temperature adjustment value range of-10℃ to +10℃, the humidity adjustment value range of-5%RH to +5%RH, and the pressure adjustment value range of-0.02MPa to +0.02MPa.

9. The method according to claim 8, wherein the method is characterized by, In the S4 control parameter generation, the rolling optimization objective function of the model predictive control algorithm contains three items: the square sum of the processing parameter adjustment amount, the square sum of the predicted quality deviation, and the square sum of the parameter adjustment rate. The weight proportion of the three items is 2:6:

2.

10. The method according to claim 1, wherein the method is characterized by, In the S5 real-time control execution, the controller of the frying equipment is a PID controller, the proportional coefficient is set to 5.0, the integral time is set to 10s, and the differential time is set to 2s. The speed control accuracy of the conveyor belt driving device is ±0.1m / min.

Images