A food processing process dynamic quality monitoring and intelligent control method and system

Abstract

The application provides a food processing process dynamic quality monitoring and intelligent control method and system, the method comprises the following steps: S1, data acquisition; S2, data preprocessing; S3, establishing a quality prediction model; S4, quality monitoring; S5, intelligent control. The application adopts the closed-loop interaction of four algorithms of BP, RF, SVR and LSTM, combines with dynamic weight fusion, reduces the quality index prediction error, and improves the shelf life prediction accuracy; the data acquisition frequency and model calculation time consumption are matched, the quality index is predicted in quasi real time; the quantitative control strategy based on the prediction model replaces the experience operation, reduces the processing parameter fluctuation amplitude; the prediction results of the multi-algorithm fusion provide a reliable basis for the control, improve the product qualified rate, and reduce the raw material loss rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent regulation of food processing process, and in particular to a dynamic quality monitoring and intelligent regulation method and system for food processing process. BACKGROUND

[0002] The quality stability of food processing process directly affects product safety and market competitiveness, and the traditional quality monitoring and regulation method has the following limitations:

[0003] Detection lag: relying on manual sampling offline detection (such as laboratory detection of hardness and color), from sampling to result feedback needs several hours, which cannot reflect the processing dynamics in real time, and is easy to lead to the production of unqualified products in batches.

[0004] Single prediction model: multiple single algorithms (such as single BP neural network) are used to predict quality indicators, ignoring the time sequence characteristics (such as the cumulative effect of temperature over time) and nonlinear relationship (such as the complex mapping of ingredient content and color), and the prediction error is large.

[0005] Experience-dependent regulation: the regulation strategy is mostly based on the experience of operators (such as adjusting temperature according to hand feeling), and lacks quantitative basis, which is easy to cause over-regulation or under-regulation (such as temperature fluctuation ± 5℃ or more), resulting in poor product quality consistency.

[0006] Data island problem: the data of processing parameters (temperature, humidity, etc.), raw material ingredients, and quality indicators are stored separately, and do not form a closed loop interaction, which cannot realize the linkage optimization of "parameter-quality-regulation". SUMMARY

[0007] The present application provides a dynamic quality monitoring and intelligent regulation method and system for food processing process to solve one or more of the above problems.

[0008] To achieve the above purpose, the present application adopts the following technical scheme:

[0009] A dynamic quality monitoring and intelligent regulation method for food processing process, comprising:

[0010] S1. Data acquisition; acquiring the processing temperature, relative humidity, internal pressure, protein content and fat content signals in the food processing process, and performing synchronous storage processing to obtain a data set D1;

[0011] S2. Data preprocessing; performing outlier detection and removal processing on the data set D1 to obtain a data set D2; performing missing value filling processing on the data set D2 to obtain a data set D3; and performing standardization processing on the data set D3 to obtain a data set D4;

[0012] S3. Establish a quality prediction model; based on the data set D4, build and train four models of BP neural network, random forest, support vector regression, LSTM neural network, and optimize through the following interaction mechanism:

[0013] The hidden layer output of the BP neural network is used as the input feature of the random forest, and the feature importance of the random forest is fed back to the BP neural network to correct the input layer weight;

[0014] The support vector weight of the support vector regression selects the key samples of the random forest, and the out-of-bag error of the random forest adjusts the penalty coefficient of the support vector regression;

[0015] The time trend feature of the LSTM neural network is input to the BP neural network, and the cell state of the LSTM is used as the tree weight of the random forest;

[0016] The distance feature of the support vector regression is input to the BP neural network, and the predicted value of the BP neural network is used as the input of the LSTM neural network;

[0017] After the four models pass the above-mentioned interactive transmission of features and parameters, the quality index prediction value Y including hardness, color and shelf life is obtained through weighted fusion;

[0018] S4. Quality monitoring; compare the quality index prediction value Y with the preset threshold range, and generate a warning signal if it exceeds the threshold;

[0019] S5. Intelligent regulation; according to the warning signal and the quality index prediction value Y, formulate a processing parameter adjustment strategy, and feed back the adjusted parameters to the processing equipment to complete the regulation.

[0020] In this specification, in S1, the processing temperature is measured by a thermocouple sensor, the relative humidity is measured by a capacitive humidity sensor, the internal pressure is measured by a piezoelectric pressure sensor, and the protein and fat content is measured by a near-infrared spectrum sensor.

[0021] In this specification, in S2, the outlier detection and removal process adopts the mean plus or minus three times the standard deviation criterion, and the data outside the range is removed; the missing value filling process adopts the linear interpolation method, and the missing value is calculated according to the known data points before and after the missing point; the standardization process converts the data to the [0, 1] interval, which is calculated by (data value-minimum value) / (maximum value-minimum value).

[0022] In this specification, in S3, the input layer of the BP neural network contains the time trend feature output by the data set D4 and the LSTM, the hidden layer contains two layers, and the output layer uses the Adam optimizer and the mean square error loss function during training, and the output corresponds to the quality index of hardness, color and shelf life.

[0023] In the specification, in S3, the random forest is composed of 50 decision trees, the input features include the processing parameters of the data set D4 and the output of the hidden layer of the BP neural network, the support vector weight is trained based on the support vector regression to screen key samples, and the output feature importance and the out-of-bag error.

[0024] In the specification, in S3, the support vector regression adopts a radial basis kernel function, the penalty coefficient is adjusted based on the out-of-bag error of the random forest, the insensitive loss parameter is adjusted based on the prediction error of the BP neural network, and the support vector and the distance feature are output for optimizing the random forest and the BP neural network.

[0025] In the specification, in S3, the LSTM neural network input includes D4 and the BP neural network quality prediction value, the time trend feature is extracted through the gating mechanism, and the output hidden state mean is used to optimize the BP neural network input and the cell state is used to optimize the tree weight of the random forest.

[0026] In the specification, in S4, the threshold range is determined according to the food industry standard and the product design requirement, and the early warning signal includes the sound and light alarm of the device buzzer and warning light, and the system prompt information of the monitoring terminal

[0027] In the specification, in S5, the control strategy is specifically: if the hardness is lower than the lower limit of the threshold, the processing temperature is increased by the temperature adjustment coefficient; if the hardness is higher than the upper limit of the threshold, the processing temperature is reduced by the temperature adjustment coefficient; if the color and luster exceed the threshold range, the protein content in the raw material is adjusted by the ingredient adjustment coefficient; if the shelf life is lower than the lower limit of the threshold, the processing humidity is reduced by the humidity adjustment coefficient; the adjusted parameters are fed back through the device control system interface, and are re-collected by S1 into the next round of processing.

[0028] A food processing process dynamic quality monitoring and intelligent control system, the food processing process dynamic quality monitoring and intelligent control method of any one of the above, the food processing process dynamic quality monitoring and intelligent control system comprises:

[0029] A data acquisition module for acquiring processing temperature, relative humidity, internal pressure, protein content, and fat content signals during food processing, for synchronous storage and processing to obtain a data set D1.

[0030] A data preprocessing module for detecting and removing outliers from the data set D1 to obtain a data set D2, filling missing values in the data set D2 to obtain a data set D3, and standardizing the data set D3 to obtain a data set D4.

[0031] A quality prediction model establishment module for constructing and training four models of BP neural network, random forest, support vector regression, and LSTM neural network based on the data set D4, and optimizing through the following interaction mechanism:

[0032] The hidden layer output of the BP neural network is taken as the input feature of the random forest, and the feature importance of the random forest is fed back to the BP neural network to correct the input layer weight;

[0033] The support vector weight of the support vector regression selects the key sample of the random forest, and the out-of-bag error of the random forest adjusts the penalty coefficient of the support vector regression;

[0034] The time trend feature of the LSTM neural network is input to the BP neural network, and the cell state of the LSTM is taken as the tree weight of the random forest;

[0035] The distance feature of the support vector regression is input to the BP neural network, and the predicted value of the BP neural network is taken as the input of the LSTM neural network;

[0036] After the four models pass the above-mentioned interactive transmission of features and parameters, the quality index prediction value Y including hardness, color and shelf life is obtained through weighted fusion;

[0037] The quality monitoring module is used for comparing the quality index prediction value Y with the preset threshold range, and generating a warning signal if the threshold is exceeded;

[0038] The intelligent control module is used for formulating a processing parameter adjustment strategy according to the warning signal and the quality index prediction value Y, feeding back the adjusted parameters to the processing equipment, and completing the control.

[0039] In summary, the present application has at least the following beneficial effects:

[0040] Improved prediction accuracy: the four algorithms of BP, RF, SVR and LSTM are closed-loop interacted (feature complementation and error correction), combined with dynamic weight fusion, which reduces the quality index prediction error; the time series modeling capability of LSTM captures the time accumulation effect of processing parameters (such as the influence of continuous low temperature on shelf life), which improves the prediction accuracy of shelf life.

[0041] Enhanced real-time performance: the data acquisition frequency is matched with the model calculation time consumption, realizing the quasi-real-time prediction of quality index, and the closed-loop interaction termination condition (iteration number or error convergence) ensures the calculation efficiency.

[0042] Intelligent control: the quantitative control strategy based on the prediction model replaces the empirical operation, reduces the fluctuation amplitude of the processing parameters; the prediction results of multi-algorithm fusion provide reliable basis for control, improve the product qualification rate, and reduce the raw material loss rate.

[0043] Full-process closed-loop collaboration: build a closed-loop system of "data collection-preprocessing-prediction-monitoring-regulation", and form real-time feedback of processing parameters, quality indicators and regulation amounts (such as the adjustment parameters of S5 are re-collected by S1), so as to realize the transformation from "passive detection" to "active optimization".

[0044] Adaptive enhancement: the algorithm interaction mechanism can adapt to different food categories (such as baking and canning), dynamically adjust the model weight through feature importance, and adapt to new production lines without retraining the model. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0046] Figure 1 The flowchart of the food processing process dynamic quality monitoring and intelligent control method involved in the present application.

[0047] Figure 2 The flowchart of data collection involved in the present application.

[0048] Figure 3 The flowchart of data preprocessing involved in the present application.

[0049] Figure 4 The flowchart of quality monitoring and intelligent control involved in the present application. DETAILED DESCRIPTION

[0050] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the embodiments of the present application. Therefore, the drawings and the description are considered to be exemplary in nature rather than limiting.

[0051] The following disclosure provides many different embodiments or examples for implementing different structures of the embodiments of the present application. In order to simplify the disclosure of the embodiments of the present application, the components and settings of specific examples are described in the following. Of course, they are only examples, and the purpose is not to limit the embodiments of the present application. In addition, the embodiments of the present application can repeatedly refer to numerals and / or reference letters in different examples, and such repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0052] The embodiments of the present application will be described in detail below with reference to the drawings.

[0053] As Figure 1 shown, the embodiment provides a food processing process dynamic quality monitoring and intelligent control method, comprising:

[0054] S1. Data acquisition; collect the processing temperature, relative humidity, internal pressure, protein content, fat content signals in the food processing process, and perform synchronous storage processing to obtain a data set D1;

[0055] S2. Data preprocessing; performing outlier detection and removal processing on the data set D1 to obtain a data set D2; performing missing value filling processing on the data set D2 to obtain a data set D3; performing standardization processing on the data set D3 to obtain a data set D4;

[0056] S3. Establishing a quality prediction model; based on the data set D4, four models of BP neural network, random forest, support vector regression, and LSTM neural network are constructed and trained, and are optimized through the following interaction mechanism:

[0057] The hidden layer output of the BP neural network is used as the input feature of the random forest, and the feature importance of the random forest is fed back to the BP neural network to correct the input layer weight;

[0058] The support vector weight of the support vector regression selects the key samples of the random forest, and the out-of-bag error of the random forest adjusts the penalty coefficient of the support vector regression;

[0059] The time trend feature of the LSTM neural network is input to the BP neural network, and the cell state of the LSTM is used as the tree weight of the random forest;

[0060] The distance feature of the support vector regression is input to the BP neural network, and the predicted value of the BP neural network is used as the input of the LSTM neural network;

[0061] After the four models pass the features and parameters through the above interaction, the quality index prediction value Y including hardness, color, and shelf life is obtained through weighted fusion;

[0062] S4. Quality monitoring; comparing the quality index prediction value Y with the preset threshold range, if it exceeds the threshold, a warning signal is generated;

[0063] S5. Intelligent control; according to the warning signal and the quality index prediction value Y, a processing parameter adjustment strategy is developed, and the adjusted parameters are fed back to the processing equipment to complete the control.

[0064] In some embodiments, in S1, the processing temperature is measured by a thermocouple sensor, the relative humidity is measured by a capacitive humidity sensor, the internal pressure is measured by a piezoelectric pressure sensor, and the protein and fat content is measured by a near-infrared spectrum sensor.

[0065] In some embodiments, in S2, the outlier detection and removal process adopts the mean plus or minus three standard deviation criterion to eliminate data outside the range; the missing value filling process adopts the linear interpolation method to calculate the missing value according to the known data points before and after the missing point; the standardization process converts the data to the [0, 1] interval by calculating (data value-minimum value) / (maximum value-minimum value).

[0066] In some embodiments, in S3, the input layer of the BP neural network contains the data set D4 and the time trend features output by the LSTM, the hidden layer contains two layers, and the output layer adopts the Adam optimizer and the mean square error loss function during training, and the output corresponds to the quality indicators of hardness, color and shelf life.

[0067] In some embodiments, in S3, the random forest is composed of 50 decision trees, the input features include the processing parameters of the data set D4 and the output of the BP neural network hidden layer, the support vector weight is selected based on the support vector regression to train the key samples, and the feature importance and the out-of-bag error are output.

[0068] In some embodiments, in S3, the support vector regression adopts the radial basis kernel function, the penalty coefficient is adjusted based on the out-of-bag error of the random forest, the insensitive loss parameter is adjusted based on the prediction error of the BP neural network, and the support vector and distance features are output for optimizing the random forest and the BP neural network.

[0069] In some embodiments, in S3, the input of the LSTM neural network includes D4 and the quality prediction value of the BP neural network, the time trend features are extracted through the gating mechanism, the mean of the output hidden state is used to optimize the input of the BP neural network, and the cell state is used to optimize the tree weight of the random forest.

[0070] In some embodiments, in S4, the threshold range is determined according to the food industry standards and product design requirements, the early warning signal includes the sound and light alarm of the device buzzer and warning light, and the system prompt information of the monitoring terminal

[0071] In some embodiments, in S5, the control strategy is specifically: if the hardness is lower than the lower limit of the threshold, the processing temperature is increased by the temperature adjustment coefficient; if the hardness is higher than the upper limit of the threshold, the processing temperature is reduced by the temperature adjustment coefficient; if the color exceeds the threshold range, the protein content in the raw material is adjusted by the ingredient adjustment coefficient; if the shelf life is lower than the lower limit of the threshold, the processing humidity is reduced by the humidity adjustment coefficient; the adjusted parameters are fed back through the device control system interface, and are re-collected by S1 to enter the next round of processing.

[0072] The technical concept of the application is as follows:

[0073] A food processing process dynamic quality monitoring and intelligent control method, comprising:

[0074] S1. Data collection, the data collection process is shown in Figure 2 .

[0075] Key parameters in food processing are collected, including processing temperature T (unit: ℃), relative humidity H (unit: %), internal pressure P (unit: Pa), protein content in raw materials (unit: %, affecting the nutritional value and toughness of the food), and fat content in raw materials (unit: %, affecting the flavor and shelf life of the food); .

[0076] The sensors and methods used for collection are as follows: temperature is measured by a thermocouple sensor, relative humidity is measured by a capacitive humidity sensor, internal pressure is measured by a piezoelectric pressure sensor, and protein and fat content is analyzed by a near-infrared spectrum sensor. The corresponding parameters are collected synchronously at a frequency of every 10 seconds, and the 5 parameter values collected each time are recorded in chronological order to form a data set D1.

[0077] S2. Data preprocessing, the data preprocessing is shown in Figure 3 .

[0078] 1. Abnormal value detection and removal: 3σ criterion is used to detect abnormal values. First, calculate the mean and standard deviation of each parameter. If the value of a data point is outside the range of "mean minus 3 times standard deviation" to "mean plus 3 times standard deviation", it is determined that the data point is an abnormal value and is removed from the data set, and the processed data set D2 is obtained.

[0079] 2. Missing value filling: if there are missing data in the data set D2, linear interpolation method is used for filling. The specific operation is to find the nearest two known data points before and after the missing data point, and according to the values of the two points and their time intervals with the missing point, the value of the missing point is calculated by linear interpolation, and the filled data set D3 is obtained.

[0080] 3. Standardization processing: the data set D3 is standardized to convert all parameter values to the [0, 1] interval; the processing method is: for each parameter, subtract the minimum value of the parameter from each value of the parameter, and then divide by the difference between the maximum and minimum values of the parameter. The standardized data set D4 is obtained.

[0081] S3. Establishment of quality prediction model:

[0082] 1. BP neural network (BP):

[0083] Model construction:

[0084] Structure design:

[0085] ​Input layer: consists of 6 neurons, corresponding to 6 input features respectively. Among them: the first 5 features are respectively corresponding to the processing parameters T, H, P, , ; the 6th feature is the time trend feature output by the LSTM neural network (comprehensively reflects the parameter change rule of the previous 10 time steps, used to capture the time sequence dependence of the processing process).

[0086] Hidden layer 1: contains 10 neurons, using ReLU activation function: , , the input value of the neuron; at the same time, the dropout mechanism (randomly ignore 20% of the neuron connections each time) is introduced to prevent overfitting.

[0087] Hidden layer 2: consists of 5 neurons, the input is the splicing vector of the output of hidden layer 1 and the distance feature passed in by the SVR model (fusion of BP's own features and SVR support vector information).

[0088] Output layer: 3 neurons correspond to quality hardness Y1, color Y2, shelf life Y3 respectively, using linear activation function: .

[0089] Weight correction mechanism: the input layer weight needs to be corrected according to the feature importance provided by the RF model ( , indicating the importance of the th input feature), the formula is:

[0090] ;

[0091] Among them is the corrected input layer weight, is the initial weight (subject to [-0.05, 0.05] uniform distribution), is the maximum value of all RF feature importance, through this correction, BP pays more attention to key features.

[0092] Model training:

[0093] Data preparation: use the D4 extended set of S2 output (contains features), divide it into training set and test set according to 7:3.

[0094] Optimization algorithm: use Adam optimizer (learning rate 0.01), learning rate (control parameter update step), momentum parameter =0.9 (first-order moment estimation decay rate), (second-order moment estimation decay rate) iteration 1000 times, 32 samples are processed each time.

[0095] Loss function: Mean Squared Error with L2 regularization, formula:

[0096] ;

[0097] where is the number of training samples, is the predicted value of the BP for the th sample, the th index, is the index of the quality index =1, 2, 3 corresponding to hardness, color, shelf life respectively, is the actual value, =0.001 is the regularization coefficient, is the hidden layer index, is the set of all connection weights in the th hidden layer, is a single weight value in the weight set .

[0098] Model application:

[0099] Forward propagation process:

[0100] Hidden layer 1 calculation: signal transmission from input layer to hidden layer 1 is ;

[0101] where is the weighted input sum of the th neuron in the hidden layer 1 of the BP neural network , which is the linear combination result of input features and weights; is the weight value connecting the th feature in the input layer to the th neuron in the hidden layer 1; is the bias term of the th neuron in the hidden layer 1, and the output after ReLU activation is , which is used as the input feature of RF.

[0102] Hidden layer 2 calculation: the input is the concatenation of hidden layer 1 output and SVR distance features, and after calculation and activation, we get is the neuron index of hidden layer 2, .

[0103] Output layer calculation: the final predicted value is , where is the predicted value of the BP for the th quality index, is the​ The connection weights from each neuron to the j-th quality metric in the output layer, output .

[0104] 2. Random Forest (RF):

[0105] Model building:

[0106] Structural design: Consists of 50 CART decision trees (tree index) Each tree has a maximum depth of 8 and a minimum number of leaf node samples of 5.

[0107] Input features: Contains 15 dimensions, including processing temperature, relative humidity, internal pressure, protein content, fat content, and 10 outputs from the first hidden layer of the backpropagation layer. ( ).

[0108] Key Sample Selection: Support Vector Weights Based on SVR ; S represents the number of support vectors; the top 20% of samples by weight are selected as the key sample set. It is prioritized for training.

[0109] Model training:

[0110] Sampling strategy: Each tree is sampled from the bootstrap sample. 63.2% of the samples were extracted as training data, and the remaining 36.8% were used as out-of-bag data for evaluation.

[0111] Decision tree construction: 5 features are randomly selected for each node, and the Gini index is used to construct the tree.

[0112] ;

[0113] in For category index ( ), indicating the category into which the sample is divided in the node; This represents the total number of categories for that node (for regression problems, this refers to the total number of intervals after the sample values ​​are discretized). For the first Split nodes based on the principle of minimizing the proportion of class samples.

[0114] Feature importance calculation: The formula is as follows:

[0115] ;

[0116] in For the first The importance of each feature ( (corresponding to the 15 input features of RF) ) is the Tree feature The decrease of Gini index after split, is the Gini index of root node.

[0117] Out-of-bag error calculation:

[0118] Out-of-bag error is used to evaluate the generalization ability of RF model, and the calculation formula is:

[0119] ;

[0120] Where is the out-of-bag error, is the number of out-of-bag samples, is the index of out-of-bag samples ), is the prediction value of the RF for the th out-of-bag sample, is the actual quality index value of the th out-of-bag sample (obtained by laboratory detection).

[0121] Model application:

[0122] Prediction process: each tree outputs a prediction value , combined with the cell state of LSTM to calculate the tree weight ;

[0123] Where, is the cell state of LSTM at time step , here is one-to-one corresponding to the tree index , that is, the th tree corresponds to the cell state of the th time step; is the maximum value of all time step cell states of LSTM, and the final prediction value is .

[0124] Output delivery: the prediction result is delivered to BP and LSTM, the feature importance is delivered to BP, and the out-of-bag error is delivered to SVR.

[0125] 3. Support vector regression (SVR):

[0126] Model construction:

[0127] Kernel function selection: radial basis kernel function is adopted:

[0128] ;

[0129] Where The first and the second sample feature vectors, = 0.1, is the Euclidean distance.

[0130] Optimization objective and constraints:

[0131] The objective function is: ;

[0132] The constraints are:

[0133] ;

[0134] ;

[0135] ;

[0136] where is the weight vector of SVR in high-dimensional feature space, b is the bias term; is the slack variable, is the error corresponding to the sample predicted value lower than the true value, is the error corresponding to the sample predicted value higher than the true value; is the penalty coefficient, which controls the punishment degree of slack variable, and the constraint condition contains error range and non-negative slack variable, ( is the out-of-bag error variance of RF); is the insensitive loss parameter, (BP error correction), is the predicted value of SVR.

[0137] Model training: solve by Lagrange dual method, get support vector ( , is the total number of support vectors) and multiplier , the support vector is the sample that satisfies .

[0138] Model application:

[0139] Prediction process: the predicted value of the feature vector of the new sample is (b is the bias).

[0140] Feature transmission: the prediction result , the distance feature is transmitted to BP.

[0141] 4. LSTM neural network (LSTM):

[0142] Model construction:

[0143] Cell structure: LSTM processes time series data through gating mechanism;

[0144] Forget gate: where is the output of the forget gate at time step (range [0, 1], controls the proportion of historical information retained), is the sigmoid activation function, is the weight matrix of the forget gate, is the hidden state at time step (reflects the features of the previous time step); is the input vector at time step ; is the bias term of the forget gate;

[0145] Input gate: where is the output of the input gate at time step , is the weight matrix of the input gate, is the bias term of the input gate;

[0146] Candidate cell state: where is the candidate cell state at time step (range [-1, 1], represents new information to be added); is the hyperbolic tangent function, is the weight matrix of the candidate cell state; is the bias term of the candidate cell state;

[0147] Cell state: where is the cell state at time step (stores long-term memory information); is the cell state at time step ; is the element-wise multiplication operator (corresponding elements are multiplied);

[0148] Output gate: where is the output of the output gate at time step (range [0, 1], controls the proportion of output information); is the weight matrix of the output gate; is the bias term of the output gate;

[0149] Hidden state: where For time steps The hidden state is output as short-term memory information for subsequent prediction.

[0150] Input extension: The input vector includes processing parameters and BP prediction values. ;in For time steps The processing temperature; For time step relative humidity; For time step Internal pressure, For time step The protein content of the raw materials; For time step The fat content of the raw materials; For BP neural networks at time steps The predicted values ​​of quality indicators.

[0151] Model training:

[0152] The data consists of a parameter sequence over 10 consecutive time steps (predicting the metric at step 11), and the loss function is:

[0153] ;

[0154] in The loss value of LSTM; This represents the total length of the training sequence. For LSTM at time steps The predicted value, The target output is the first Actual values ​​of quality indicators at each time step The square of the Euclidean distance; This is a balancing coefficient used to weigh consistency with BP forecasts;

[0155] Optimization process: The Adam optimizer (learning rate 0.001) was used for 500 iterations, combined with an early stopping mechanism to prevent overfitting.

[0156] Model Application:

[0157] Prediction process: Input 10 time step sequences and calculate the hidden state. and cell state Predicting the first Step indicators ( This is the output layer weight matrix; (This refers to the output layer bias term).

[0158] Feature transfer: Prediction results Time trend characteristics The cell state is passed to the BP and used for RF tree weight calculation.

[0159] 5. Integrated Applications:

[0160] Interaction contribution calculation:

[0161] The formula is ;

[0162] in For the first The interaction contribution of each model is used to measure the consistency between the model and the true value and the predictive complementarity with other models; These correspond to BP neural network, random forest, support vector regression, and LSTM neural network, respectively; corr is the Pearson correlation coefficient. Another index for the fusion model ( =1,2,3,4 and ); For the first The predicted values ​​of each model; These are the actual measured values ​​of the quality indicators (obtained through laboratory testing).

[0163] Dynamic weight allocation: ,in For the first The dynamic weights of each model reflect the proportion of its contribution to the fusion prediction (the sum of the weights is 1). The sum of the contributions of the four model interactions.

[0164] Final prediction: , , , , The dynamic weights for BP, RF, SVR, and LSTM are respectively used to transmit Y to S4.

[0165] The interaction of the four algorithms in S3 (BP neural network, random forest, support vector regression, and LSTM neural network) forms a closed loop, and its core loop path is:

[0166] The closed-loop logic of BP→RF→SVR→LSTM→BP is as follows:

[0167] BP outputs the hidden layer Transmitted to RF;

[0168] RF will reduce bag outside error Pass to SVR;

[0169] SVR will support vector Passed to LSTM;

[0170] LSTM incorporates time trend features Passed to BP;

[0171] The above process forms a complete loop, in which each algorithm continuously exchanges feature and error information and dynamically optimizes the prediction results.

[0172] Loop termination condition: The closed-loop interaction terminates when any of the following conditions are met:

[0173] 1. Number of iterations met: The preset maximum number of iterations N (e.g., N=50) has been achieved, ensuring controllable computational efficiency;

[0174] 2. Error Convergence: Root Mean Square Error (RMSE) of each model's predictions:

[0175] ;

[0176] Where M is the sample size. For the first The model prediction value for each sample. For the first The actual measured value of each sample The error threshold is a non-negative real number, such as... =0.01; In K consecutive iterations (e.g., K=5), the RMSE is less than the threshold. This indicates that the prediction accuracy has met the requirements;

[0177] Feature stability: The feature importance of the RF output over K consecutive iterations. Distance features compared to SVR output The rate of change is less than 1%, indicating that feature transfer has become stable.

[0178] S4. Quality Control:

[0179] Set the threshold range for the quality index; the threshold range for hardness Y1 is... The threshold range for color Y2 is The threshold range for shelf life Y3 is: These thresholds are determined based on food industry standards (such as national food quality and safety standards) and product design requirements. The predicted value Y obtained in S3 is compared with the threshold range. If any quality indicator exceeds the corresponding threshold range, an early warning signal is issued. The early warning signal includes an audible and visual alarm (implemented through a buzzer and warning light on the device) and system prompts (displayed on the monitoring terminal).

[0180] The process of quality monitoring involves comparing the predicted value Y with the set threshold ranges to determine whether it exceeds the threshold. If it does, a corresponding warning signal is triggered.

[0181] S5: Intelligent Control

[0182] Based on the monitoring results of S4 and the quality prediction model of S3, formulate control strategies;

[0183] If the hardness Y1 is lower than Then increase the processing temperature and adjust the amount as follows: Where k1 is the temperature adjustment coefficient (valued at 0.5℃ / N, obtained by fitting historical control data); if the hardness Y1 is higher than Then reduce the processing temperature by the amount adjusted. ;

[0184] If the color Y2 exceeds Then adjust the protein content in the raw materials. The adjustment amount is Where k2 is the component adjustment coefficient (valued at 0.02% of the color difference value, obtained by fitting historical control data). The target value for color (take the middle value of the threshold range, i.e.) ;

[0185] If the shelf life Y3 is lower Then reduce the processing humidity, and adjust the amount as follows: Where k3 is the humidity adjustment coefficient (valued at 0.3% / day, obtained by fitting historical control data).

[0186] Adjusted parameters , , Feedback is sent to the processing equipment (via the equipment's control system interface) to control the processing. The controlled parameters are then re-acquired by the sensors in S1 and enter the next round of data processing and monitoring.

[0187] The process of intelligent control involves calculating the adjustment amount of each parameter according to the corresponding adjustment formula based on the quality indicators that exceed the threshold in the monitoring results, and then feeding the adjusted parameters back to the processing equipment to complete the control.

[0188] Quality monitoring and intelligent control processes, such as Figure 4 As shown.

[0189] In some embodiments, S3. Establish a quality prediction model (fusion and interactive optimization):

[0190] Based on BP neural networks (BP), random forests (RF), support vector regression (SVR), and LSTM neural networks, a correlation fusion model is formed by the pairwise interaction of influencing parameters to predict quality indicators. .

[0191] (1) Interaction between BP and RF:

[0192] The impact of RF on BP: Importance of RF output characteristics Correct the weights of the BP input layer.

[0193] Feature importance calculation formula: 50 represents the number of trees. ) is the first Tree characteristics The decrease in the Gini index after the split. The root node is the Gini index.

[0194] BP input layer weight correction: ,in As the initial weights, This represents the maximum importance of all RF features.

[0195] The impact of backpropagation (BP) on radio frequency response (RF): BP hidden layer output As an additional feature of RF.

[0196] RF input feature set expanded to ,in .

[0197] (2) Interaction between BP and SVR:

[0198] The impact of SVR on BP: SVR support vectors Calculate distance features to enhance the representation of the BP hidden layer;

[0199] Distance features: The input to the second hidden layer of the BP layer is .

[0200] The impact of BP on SVR: BP output layer error (The actual value) Corrects the SVR insensitivity loss parameter ;

[0201] Correction formula: ,in =0.01 is the initial value. =0.5 is the correction factor.

[0202] (3) Interaction between BP and LSTM:

[0203] The impact of LSTM on BP: The hidden state of LSTM Extract time trend features as additional input to the backpropagation (BP) algorithm.

[0204] Trend characteristics: (Taking the average of the last 10 time steps), the BP input layer is expanded to... .

[0205] The impact of BP on LSTM: BP's predicted quality index As the target correction term of LSTM, it optimizes its time series prediction;

[0206] LSTM loss function correction:

[0207] ,in 0.3 is the mean square error, and 0.3 is the correction factor.

[0208] (4) Interaction between RF and SVR:

[0209] The impact of SVR on RF: SVR support vector weights (Lagrange multipliers) are used to select key samples in RF and improve tree training efficiency;

[0210] Key sample set: RF is only based on Perform bootstrap sampling training. It refers to the first support vector in the support vector set output by SVR (Support Vector Regression). One sample.

[0211] The impact of RF on SVR: RF out-of-bag prediction error (Out-of-bag sample error) Optimize the penalty coefficient for SVR ;

[0212] Optimized formula: ,in =10 is the initial value. =2 is the attenuation coefficient. Let Variance be the variance.

[0213] (5) Interaction between RF and LSTM:

[0214] The effect of LSTM on RF: Cell state of LSTM (Cell state at time step t) Calculate time weights and adjust tree weights in the RF.

[0215] Tree weight: RF prediction value .

[0216] The impact of RF on LSTM: the importance of RF features Modify the input weights of the LSTM to enhance the time sensitivity of key parameters;

[0217] LSTM input weight correction: ( (Initial input weights).

[0218] (6) Interaction between SVR and LSTM:

[0219] The impact of LSTM on SVR: Multi-step prediction sequences using LSTM (Predicting the next 5 steps) Constructing the timing constraints for SVR, Indicates that LSTM is applied to the first... The predicted values ​​of the quality indicators for the first step (future step 1) optimize its decision boundary;

[0220] SVR constraints: , This represents the minimum value among the predictions made by the LSTM over the next 5 steps. This represents the maximum value, and this constraint ensures that the SVR prediction results conform to the time series trend.

[0221] The impact of SVR on LSTM: SVR support vectors Construct reference points and correct the hidden state updates of the LSTM;

[0222] Hidden state correction: ,in =0.2 is the correction factor. For the revised version Hide your status at all times. To compensate for the deviation between the support vectors and the hidden state, the hidden state is adjusted by averaging the deviation, thereby enhancing the sensitivity of the LSTM to key samples.

[0223] (7) Integrated and merged output:

[0224] The output of the four models based on pairwise interactions Through dynamic weight fusion: calculate the interaction contribution: ;

[0225] Dynamic weights: ;

[0226] Final prediction: .

[0227] Processing procedure: Implement the pairwise interaction of the four algorithms (feature correction, parameter optimization, and constraint enhancement) in sequence according to the above formula, calculate the interaction contribution to determine the dynamic weight, and obtain the final predicted value Y by weighting, and transmit it to S4.

[0228] A dynamic quality monitoring and intelligent control system for food processing, comprising the dynamic quality monitoring and intelligent control method for food processing described in any one of the above-mentioned methods, wherein the dynamic quality monitoring and intelligent control system for food processing comprises:

[0229] The data acquisition module is used to synchronously store and process signals such as processing temperature, relative humidity, internal pressure, protein content, and fat content during food processing to obtain dataset D1.

[0230] The data preprocessing module is used to perform outlier detection and removal on dataset D1 to obtain dataset D2; to perform missing value imputation on dataset D2 to obtain dataset D3; and to perform standardization on dataset D3 to obtain dataset D4.

[0231] A quality prediction model module is established to build and train four models—BP neural network, random forest, support vector regression, and LSTM neural network—based on dataset D4, and optimize them through the following interaction mechanism:

[0232] The hidden layer output of the BP neural network serves as the input feature of the random forest, and the feature importance of the random forest is fed back to the BP neural network to correct the input layer weights.

[0233] Support vector weights from support vector regression are used to select key samples for random forests, and the out-of-bag error of random forests is used to adjust the penalty coefficient of support vector regression.

[0234] The time trend features of the LSTM neural network are input into the BP neural network, and the cell states of the LSTM are used as the tree weights of the random forest.

[0235] The distance features from support vector regression are input into a backpropagation (BP) neural network, and the predictions from the BP neural network are used as inputs to an LSTM neural network.

[0236] After the four models transmit features and parameters through the above interaction, they are then weighted and fused to obtain the predicted value Y of the quality index, including hardness, color and shelf life.

[0237] The quality monitoring module is used to compare the predicted value Y of the quality indicator with a preset threshold range. If the value exceeds the threshold, an early warning signal is generated.

[0238] The intelligent control module is used to formulate processing parameter adjustment strategies based on early warning signals and the predicted value Y of the quality index, and then feeds the adjusted parameters back to the processing equipment to complete the control.

[0239] This invention is applicable to food processing scenarios that are sensitive to processing parameters (temperature, humidity, pressure) and raw material composition (protein, fat), and require strict control of quality indicators such as hardness, color, and shelf life. Specific examples are as follows:

[0240] 1. Baked goods (bread, cookies)

[0241] Processing characteristics: Precise control of baking temperature (150-220℃), relative humidity (30-80%) and oven pressure (slight negative pressure) is required. The protein (gluten protein in flour, which affects the dough's toughness) and fat (butter, grease, which affects crispness) content in the raw materials are crucial to the quality of the finished product.

[0242] Correlation of quality indicators:

[0243] Hardness: The softness of bread and the crispness of biscuits need to be controlled by temperature (e.g., high temperature for short time to reduce hardness, low temperature for long time to increase hardness).

[0244] Color: The golden yellow of bread crust and the brown of biscuits are determined by the Maillard reaction, which needs to be regulated by temperature and protein content (gluten protein participates in the reaction);

[0245] Shelf life: related to baking humidity (low humidity extends shelf life) and the degree of fat oxidation, and processing humidity needs to be controlled.

[0246] 2. Meat processing (sausages, ham)

[0247] Processing characteristics: It involves pickling, steaming, smoking and other steps. It is necessary to control the steaming temperature (70-90℃), the smoking humidity (50-70%) and the sterilization pressure (0.1-0.2MPa). The content of muscle protein (affecting gelation) and fat (affecting flavor and juiciness) in the raw materials is key.

[0248] Correlation of quality indicators:

[0249] Hardness: The elasticity of sausages and the slicability of ham are determined by the strength of protein thermogelation, which needs to be controlled by temperature (degree of protein denaturation);

[0250] Color: The pink color formed by the oxidation of myoglobin needs to be controlled by processing temperature (to avoid fading due to high temperature) and fat content (fat encapsulates and protects the pigment);

[0251] Shelf life: related to sterilization pressure (high-pressure sterilization extends shelf life) and moisture content (humidity control reduces water activity).

[0252] 3. Dairy products (cheese, yogurt)

[0253] Processing characteristics: During the fermentation stage, temperature (25-45℃) and humidity (80-90%) in the fermentation chamber need to be controlled. During the pressing stage, pressure (0.05-0.1MPa) needs to be controlled. The protein (casein, which affects curdling) and fat (milk fat, which affects taste) content in the raw milk directly affect the quality.

[0254] Correlation of quality indicators:

[0255] Hardness: The texture (soft / hard) of cheese is controlled by fermentation temperature (low temperature and slow fermentation result in soft cheese) and pressing pressure (high pressure results in hard cheese);

[0256] Color: The milky white color of natural cheese needs to be maintained by avoiding high-temperature browning, which is achieved by controlling the processing temperature and ensuring even distribution of fat (fat particles affect light reflection).

[0257] Shelf life: related to fermentation humidity (low humidity reduces mold contamination) and protein content (high protein forms a dense structure and delays spoilage).

[0258] 4. Puffed snacks (potato chips, rice cakes)

[0259] Processing characteristics: During the puffing stage, it is necessary to control the heating temperature (180-250℃), the pressure inside the equipment (sudden drop to form puffing), and the moisture content of the raw materials (10-15%). The starch (but protein needs to be added to enhance the taste) and fat (oil content of fried puffed foods) in the raw materials affect the characteristics of the finished product.

[0260] Correlation of quality indicators:

[0261] Hardness: The degree of expansion determines the hardness (high pressure difference forms low hardness and loose structure), which needs to be controlled by pressure;

[0262] Color: The golden color of fried potato chips is determined by oil oxidation and temperature. The oil temperature needs to be controlled (to avoid it being too high and causing it to burn).

[0263] Shelf life: related to processing humidity (low moisture extends shelf life) and fat content (addition of antioxidants reduces rancidity).

[0264] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0265] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0266] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0267] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0268] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0269] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0270] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0271] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0272] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0273] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A method for dynamic quality monitoring and intelligent control of food processing, characterized in that, include: S1. Data Acquisition; Collect signals of processing temperature, relative humidity, internal pressure, protein content, and fat content during food processing, perform synchronous storage and processing, and obtain dataset D1; S2. Data preprocessing; Outlier detection and removal are performed on dataset D1 to obtain dataset D2; Missing values ​​were imputed in dataset D2 to obtain dataset D3; Standardize dataset D3 to obtain dataset D4; S3. Establish a quality prediction model; Based on dataset D4, four models—BP neural network, random forest, support vector regression, and LSTM neural network—were built and trained, and optimized through the following interaction mechanism: The hidden layer output of the BP neural network serves as the input feature of the random forest, and the feature importance of the random forest is fed back to the BP neural network to correct the input layer weights. Support vector weights from support vector regression are used to select key samples for random forests, and the out-of-bag error of random forests is used to adjust the penalty coefficient of support vector regression. The time trend features of the LSTM neural network are input into the BP neural network, and the cell states of the LSTM are used as the tree weights of the random forest. The distance features from support vector regression are input into a backpropagation (BP) neural network, and the predictions from the BP neural network are used as inputs to an LSTM neural network. After the four models transmit features and parameters through the above interaction, they are then weighted and fused to obtain the predicted value Y of the quality index, including hardness, color and shelf life. S4. Quality monitoring; compare the predicted value of the quality indicator Y with the preset threshold range, and generate an early warning signal if it exceeds the threshold. S5. Intelligent control; Based on the early warning signal and the predicted value Y of the quality index, formulate a processing parameter adjustment strategy, and feed the adjusted parameters back to the processing equipment to complete the control; In S3, the input layer of the BP neural network contains the time trend features of the dataset D4 and the output of LSTM. The hidden layer contains two layers. The output layer is trained using the Adam optimizer and the mean square error loss function, and outputs the corresponding quality indicators of hardness, color and shelf life. In S3, the random forest consists of 50 decision trees. The input features include the processing parameters of dataset D4 and the output of the hidden layer of the BP neural network. Key samples are selected for training based on the support vector weights of support vector regression. The output features are the importance and out-of-bag error. In S3, support vector regression uses a radial basis function kernel, the penalty coefficient is adjusted based on the out-of-bag error of random forest, and the insensitive loss parameter is adjusted based on the prediction error of BP neural network. The output support vectors and distance features are used to optimize random forest and BP neural network. In S3, the input of the LSTM neural network includes the quality prediction values ​​of the D4 and BP neural networks. The time trend features are extracted through the gating mechanism. The mean of the hidden state output is used to optimize the input of the BP neural network, and the cell state is used to optimize the tree weights of the random forest. S3 is based on BP neural network, random forest, support vector regression, and LSTM neural network. It forms a correlation and fusion model by passing influencing parameters through pairwise interactions to predict quality indicators. Y1 represents the mass hardness, Y2 represents the color, and Y3 represents the shelf life. (1) Interaction between BP and RF: The impact of RF on BP: Importance of RF output characteristics Correct the weights of the BP input layer; Feature importance calculation formula: 50 represents the number of trees. ) is the first Tree characteristics The decrease in the Gini index after the split. The root node's Gini index; BP input layer weight correction: ,in As the initial weights, This represents the maximum importance of all RF features. The impact of backpropagation (BP) on radio frequency response (RF): BP hidden layer output As an additional feature of RF; RF input feature set expanded to ,in T is the processing temperature, H is the relative humidity, and P is the internal pressure. The protein content in the raw materials. Fat content in raw materials For the first hidden layer of the BP neural network The weighted sum of inputs to each neuron; (2) Interaction between BP and SVR: The impact of SVR on BP: SVR support vectors Calculate distance features to enhance the representation of the BP hidden layer; Distance features: The input to the second hidden layer of the BP layer is , The distance feature output by SVR; , This represents the total number of support vectors. It is a natural exponential function; The distance is Euclidean. for The representation of k=1,2…10; for The representation of k=1,2…10; The impact of BP on SVR: BP output layer error , The actual value is used to correct the SVR insensitive loss parameter. ; Correction formula: ,in =0.01 is the initial value. =0.5 is the correction factor; (3) Interaction between BP and LSTM: The impact of LSTM on BP: The hidden state of LSTM Extract time trend features as additional input to BP; Trend characteristics: , As a time trend feature, For time step The hidden state, the BP input layer is expanded to ; The impact of BP on LSTM: BP's predicted quality index As the target correction term of LSTM, it optimizes its time series prediction; LSTM loss function correction: ,in The mean squared error is 0.3, and the correction factor is 0.

3. This represents the total length of the training sequence. For LSTM at time steps The predicted value, The target output is the first Actual values ​​of quality indicators at each time step The square of the Euclidean distance. For BP at time step Predicted values ​​of quality indicators; (4) Interaction between RF and SVR: The impact of SVR on RF: Support vector weights of SVR Lagrange multipliers are used to select key samples in RF (Radical Randomization) to improve tree training efficiency. Key sample set: RF is only based on Perform bootstrap sampling training. It refers to the first support vector in the SVR output set. One sample; The impact of RF on SVR: RF out-of-bag prediction error Optimize the penalty coefficient of SVR , For RF to the first Predicted values ​​for samples outside the bag, For the first The actual quality index values ​​of the samples outside the bag; Optimized formula: ,in =10 is the initial value. =2 is the attenuation coefficient. For variance; (5) Interaction between RF and LSTM: The effect of LSTM on RF: Cell state of LSTM Calculate the time weights and adjust the tree weights in the RF algorithm; Tree weight: RF prediction value , For LSTM neural networks at time steps The state of cells, This represents the maximum value of cell states across all time steps of the LSTM process. The impact of RF on LSTM: the importance of RF features Modify the input weights of the LSTM to enhance the time sensitivity of key parameters; LSTM input weight correction: , These are the initial input weights; (6) Interaction between SVR and LSTM: The impact of LSTM on SVR: Multi-step prediction sequences using LSTM Constructing timing constraints for SVR, Indicates that LSTM is applied to the first... The predicted values ​​of the quality indicators are used to optimize the decision boundary; SVR constraints: , This represents the minimum value among the predictions made by the LSTM over the next 5 steps. This represents the maximum value, and this constraint ensures that the SVR prediction results conform to the time series trend; The impact of SVR on LSTM: SVR support vectors Construct reference points and correct the hidden state updates of the LSTM; Hidden state correction: ,in =0.2 is the correction factor. For the revised version Hide your status at all times. To compensate for the deviation between the support vectors and the hidden state, the hidden state is adjusted by averaging the deviation, thereby enhancing the sensitivity of the LSTM to key samples. (7) Integrated and merged output: The output of the four models based on pairwise interactions Through dynamic weight fusion: calculate the interaction contribution: ; Dynamic weights: ; Final prediction: ; For the first The interaction contribution of each model These correspond to BP neural network, random forest, support vector regression, and LSTM neural network, respectively; corr is the Pearson correlation coefficient. As another index for the fusion model, =1,2,3,4 and ; For the first The predicted values ​​of each model; These are the actual measured values ​​of the quality indicators; For the first The dynamic weights of each model The sum of the contributions of the four model interactions. , , , These are the dynamic weights of BP, RF, SVR, and LSTM, respectively. These are the predicted values ​​for BP's quality indicators; Processing procedure: Implement the pairwise interaction of the four algorithms in sequence according to the above formula, calculate the interaction contribution to determine the dynamic weight, and obtain the final predicted value Y by weighting, and transmit it to S4.

2. The method for dynamic quality monitoring and intelligent control of food processing according to claim 1, characterized in that, In S1, the processing temperature is measured using a thermocouple sensor, the relative humidity is measured using a capacitive humidity sensor, the internal pressure is measured using a piezoelectric pressure sensor, and the protein and fat content is measured using a near-infrared spectroscopy sensor.

3. The method for dynamic quality monitoring and intelligent control of food processing according to claim 1, characterized in that, In S2, outlier detection and removal use the mean plus or minus three standard deviations criterion to remove data that exceeds this range; missing value imputation uses linear interpolation to estimate missing values ​​based on known data points before and after the missing point; standardization transforms the data to the [0,1] interval and calculates it using (data value - minimum value) / (maximum value - minimum value).

4. The method for dynamic quality monitoring and intelligent control of food processing according to claim 1, characterized in that, In S4, the threshold range is determined according to food industry standards and product design requirements. The warning signals include audible and visual alarms from the equipment buzzer and warning lights, as well as system prompts from the monitoring terminal.

5. The method for dynamic quality monitoring and intelligent control of food processing according to claim 1, characterized in that, In S5, the specific control strategy is as follows: if the hardness is lower than the lower threshold, the processing temperature is increased according to the temperature adjustment coefficient; if the hardness is higher than the upper threshold, the processing temperature is decreased according to the temperature adjustment coefficient; if the color exceeds the threshold range, the protein content in the raw material is adjusted according to the composition adjustment coefficient; if the shelf life is lower than the lower threshold, the processing humidity is decreased according to the humidity adjustment coefficient; the adjusted parameters are fed back through the equipment control system interface and are collected again by S1 to enter the next round of processing.

6. A dynamic quality monitoring and intelligent control system for food processing, characterized in that, The method for dynamic quality monitoring and intelligent control of food processing as described in any one of claims 1 to 5, wherein the dynamic quality monitoring and intelligent control system for food processing comprises: The data acquisition module is used to collect signals of processing temperature, relative humidity, internal pressure, protein content, and fat content during food processing, and to synchronously store and process them to obtain dataset D1. The data preprocessing module is used to perform outlier detection and removal on dataset D1 to obtain dataset D2; to perform missing value imputation on dataset D2 to obtain dataset D3; and to perform standardization on dataset D3 to obtain dataset D4. A quality prediction model module is established to build and train four models—BP neural network, random forest, support vector regression, and LSTM neural network—based on dataset D4, and optimize them through the following interaction mechanism: The hidden layer output of the BP neural network serves as the input feature of the random forest, and the feature importance of the random forest is fed back to the BP neural network to correct the input layer weights. Support vector weights from support vector regression are used to select key samples for random forests, and the out-of-bag error of random forests is used to adjust the penalty coefficient of support vector regression. The time trend features of the LSTM neural network are input into the BP neural network, and the cell states of the LSTM are used as the tree weights of the random forest. The distance features from support vector regression are input into a backpropagation (BP) neural network, and the predictions from the BP neural network are used as inputs to an LSTM neural network. After the four models transmit features and parameters through the above interaction, they are then weighted and fused to obtain the predicted value Y of the quality index, including hardness, color and shelf life. The quality monitoring module is used to compare the predicted value Y of the quality indicator with a preset threshold range. If the value exceeds the threshold, an early warning signal is generated. The intelligent control module is used to formulate processing parameter adjustment strategies based on early warning signals and the predicted value Y of the quality index, and then feeds the adjusted parameters back to the processing equipment to complete the control.

Images