Intelligent brine salinity prediction method and system based on multi-source data fusion

Abstract

The application discloses a kind of brine salinity intelligent prediction method and system based on multi-source data fusion, belong to brine product intelligent control field.For the problem of low salinity control precision, big batch fluctuation, the method fuses temperature, salinity, pH and other multi-source data to construct three-layer feedforward bayesian regularization neural network, and outputs water replenishment, material replenishment and salt replenishment;First, the average relative error and batch variation coefficient are used as double targets, and the genetic algorithm is used to determine the number of hidden layer nodes and initial weight threshold;Then, the model is trained by using the Bayesian regularization LM algorithm, which can be generalized to different brine product process flows;In actual application, multi-source data is collected in real time and input into the model, and the output mapping is control parameters such as feeding frequency and valve opening;When the salinity deviation exceeds the limit, adaptive feedback regulation is started, and dynamic correction is formed to form a closed loop;After the salinity meets the standard, a linkage signal is triggered to drive the production line to run continuously.The method is mainly used for precise control and continuous production of brine products.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for braised products. More specifically, this invention relates to an intelligent prediction method and system for braising broth salinity based on multi-source data fusion, applicable to real-time prediction and automatic precise control of braising broth salinity on industrial continuous braising production lines. Background Technology

[0002] Braised products are a unique traditional food in my country, and salinity is a core indicator that determines the flavor and quality of the product, improves texture, adjusts pH value, and extends shelf life. However, the precise control of salinity in braised products has long faced many industry challenges. Traditional braising processing relies entirely on manual experience: the capacity of a single workstation is only 10-15 pots / day, resulting in low production efficiency and making it difficult to meet the needs of large-scale industrial continuous production; the braising broth temperature is maintained above 85℃, and the workshop environment is hot and humid, requiring workers to frequently taste, replenish ingredients, and stir, creating harsh working conditions and a high risk of burns and other safety accidents; the accuracy of salinity control depends entirely on worker experience, easily leading to problems such as excessive salinity or uneven flavor, poor salinity uniformity between batches, and frequent issues such as product texture deterioration and shortened shelf life due to improper replenishment timing, resulting in raw material waste and economic losses. Breaking free from dependence on manual experience and achieving automated and intelligent salinity control has become an urgent need for the transformation and upgrading of the braised product industry.

[0003] In recent years, a number of salinity detection and control technologies have emerged in the industry. For example, Chinese patent CN117269413A discloses a salting quality control system based on the salinity detection of old brine. This system uses a conductivity sensor module to measure the conductivity of the solution, indirectly calculates the salinity value based on the linear relationship between ion concentration and conductivity, and combines it with a digital food quality control traceability system to achieve salinity monitoring and visualization. This solution, to a certain extent, achieves real-time salinity detection and data visualization, providing a technical means for salinity control in the old brine processing process. However, this technical solution still has significant limitations: First, the core of the system is based on conductivity-based salinity detection and simple calculation, lacking the ability to integrate and analyze multi-source operating data such as temperature, pH value, and brine loss during the braising process, making it difficult to cope with complex dynamic changes such as water evaporation and the dissolution of flavorings during the braising process; Second, the system can only achieve single-point detection and static calculation of salinity, with the detection and control links disconnected, making it impossible to form a dynamic and continuous closed-loop control of salinity; Third, it lacks predictive modeling capabilities based on intelligent algorithms such as machine learning, and cannot proactively predict salinity change trends based on real-time operating conditions and output precise replenishment control instructions.

[0004] Besides the solutions mentioned above, other salinity detection and control methods exist in the prior art. For example, Chinese patent CN117368425A discloses a salinity fluctuation detection device and its detection method for braised food processing, mainly solving the problems of manual handheld operation and probe cleaning of salinity detectors. It achieves automatic adjustment of the detection position and probe cleaning through a moving device and an automatic cleaning mechanism. This solution focuses on the convenience of automated operation of the detection device and does not address salinity prediction modeling and intelligent control. Another example is Chinese patent CN118091058A, which discloses a rapid prediction method for the amount of salt and monosodium glutamate added simultaneously to braising liquid. It uses an electronic tongue to collect taste sensor signals and combines them with a neural network model to quantitatively predict the amount of flavoring agents added. Although this solution introduces a neural network model, its prediction object is the amount added to a static sample. It does not involve the real-time fusion and dynamic prediction of multi-source operating data (temperature, pH value, braising liquid loss, etc.) during the braising process, nor does it construct an integrated prediction-control closed-loop system.

[0005] Overall, existing salinity detection and control technologies for braised products still suffer from the following core shortcomings: First, most solutions can only achieve single-point timed salinity detection, with the detection and control processes disconnected, making it difficult to form a dynamic and continuous closed-loop control of braising liquid salinity. Second, existing prediction algorithms mostly use traditional BP neural networks and other basic models, which are prone to overfitting, have poor generalization ability, and weak noise resistance, making them unsuitable for the complex working conditions of batch fluctuations in raw materials and continuous loss of braising liquid in braising production lines. Third, there is a lack of systematic prediction-control integrated design, which cannot directly convert prediction results into executable feeding control parameters, still requiring manual intervention, making it difficult to meet the requirements of precise, standardized, and intelligent salinity control in industrial production. Summary of the Invention

[0006] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0007] Another objective of this invention is to provide an intelligent prediction method for brine salinity based on multi-source data fusion. By constructing a two-layer fusion architecture of global optimization using a genetic algorithm and local fine training using a Bayesian regularized neural network, it effectively overcomes the shortcomings of traditional neural networks, such as sensitivity to initial parameters, susceptibility to local optima, and overfitting. This method achieves high-precision prediction with a salinity prediction error of ≤5% and a batch salinity fluctuation rate of ≤3%. Based on the prediction results, it automatically generates executable feeding and control instructions, providing intelligent closed-loop salinity control technology support for the entire process of industrial production of braised products.

[0008] To achieve these objectives and other advantages according to the present invention, a method for intelligent prediction of brine salinity based on multi-source data fusion is provided, comprising: S1. Define the network architecture of the salinity prediction model. The network architecture adopts a three-layer feedforward Bayesian regularized neural network framework. Its input layer nodes correspond to the real-time collected temperature, salinity, and pH value, and the output layer nodes correspond to the water replenishment, feed replenishment, and salt replenishment. S2. Based on historical operating condition datasets, with the dual objectives of minimizing the mean relative error (MRE) of salinity prediction and minimizing the coefficient of variation (CV) of salinity between batches, a fitness function F is constructed, using the network architecture with parameters to be determined as the evaluation object: Using a genetic algorithm to maximize the evolutionary objective with the fitness function F, a global search is performed to determine the number of hidden layer nodes, initial weights, and initial thresholds of the network architecture in S1, thereby obtaining the globally optimal combination of initial parameters. S3. Assign the globally optimal initial parameter combination obtained in step S2 to the network architecture described in step S1 as the starting state for network training, under the Bayesian regularization framework, to minimize the objective function G=αE. d +βE w To achieve the training objective, the Levenberg-Marquardt algorithm is used to iteratively converge the training of the network. During the training process, Bayesian regularization automatically balances the fitting accuracy and model complexity to suppress overfitting, generating a trained salinity prediction model. S4. Input the real-time collected temperature, salinity, and pH values ​​into the salinity prediction model trained in step S3. The model outputs a brine quality control instruction set including water replenishment, material replenishment, and salt replenishment. The control instruction set is combined with the real-time collected brine weight to convert it into control parameters executable by the feeding system. The control parameters include the number of feedings, the flow valve opening, the pump running time, and the feeding sequence. The feeding mechanism is driven to perform water replenishment, material replenishment, and salt replenishment operations according to the control parameters. During the feeding process, the measured salinity value of the brine is collected in real time. When the salinity deviation exceeds the preset threshold, the adaptive feedback adjustment mechanism is activated. The subsequent feeding control strategy is dynamically corrected according to the deviation between the salinity prediction value and the measured salinity value to form a closed-loop salinity control. After the salinity control reaches the standard, a linkage signal is sent to the production line conveying module to realize the continuous operation of the braising process. Where ω1 and ω2 are preset weighting coefficients; E d E is the sum of squared prediction errors of the network. w Let be the sum of squared network weights, and α and β be the regularization parameters for dynamic adaptive adjustment under the Bayesian framework.

[0009] Preferably, in step S1, the hidden layer activation function of the three-layer feedforward Bayesian regularized neural network framework adopts a hyperbolic tangent sigmoid function, and the output layer activation function adopts a linear activation function; the input layer also has bias nodes, and the hidden layer and output layer each have independent bias terms. The temperature, salinity, and pH values ​​corresponding to the input layer nodes in step S1 are real-time collected values ​​after time window smoothing, with a time window width of 10-30 seconds and a smoothing method of sliding window arithmetic mean. In step S2, the formula for calculating the mean relative error (MRE) of salinity prediction in the historical working condition dataset is: Where n is the total number of test samples, y i Let be the measured salinity value of the i-th sample. Let be the predicted salinity value of the i-th sample; the batch-to-batch salinity variation coefficient CV is the ratio of the standard deviation to the mean of the salinity mean of N consecutive production batches, where N ≥ 5. The genetic algorithm in step S2 uses real-number encoding, encoding the number of hidden layer nodes, the weight matrix from the input layer to the hidden layer, the weight matrix from the hidden layer to the output layer, the hidden layer threshold vector, and the output layer threshold vector into a chromosome; the chromosome length is: ; where n in =3, where n is the number of nodes in the input layer. h n is the number of hidden layer nodes. out This represents the number of nodes in the output layer. In step S2, the genetic algorithm, during each generation of evolution, uses a roulette wheel selection operator to select individuals to enter the next generation based on their fitness values. It also uses a single-point crossover operator with a crossover probability of 0.8 to exchange genes between selected parent individuals and a random mutation operator with a mutation probability of 0.1 to randomly perturb the individual's genes. When the change in fitness values ​​is lower than a preset threshold for 10 consecutive generations, the evolution is terminated early, and the current optimal individual is output as the globally optimal initial parameter combination. In step S3, the regularization parameters α and β under the Bayesian regularization framework are automatically updated after each iteration according to the Bayesian criterion. , Where γ is the number of effective network parameters; n' is the total number of training samples; γ is calculated based on the Hessian matrix under the Gaussian-Newton approximation, obtained by performing eigenvalue decomposition on the Hessian matrix after each Levenberg-Marquardt iteration; the number of effective network parameters γ gradually decreases and converges to a stable value during the training process, indicating that the network complexity is effectively controlled. The iteration step size of the Levenberg-Marquardt algorithm in step S3 is controlled by an adaptive adjustment factor λ: when the objective function G decreases after each iteration, λ is divided by a preset factor θ (θ>1) to accelerate the convergence speed; when the objective function G increases, λ is multiplied by θ to reduce the step size and recalculate the current iteration; the initial value of the adaptive adjustment factor λ is set to 0.001, and θ is set to 10. The specific method for mapping the control instruction set to control parameters in step S4 is as follows: based on the real-time collected weight W of the brine and the water replenishment ΔWw, feed replenishment ΔWm, and salt replenishment ΔWs output by the model, combined with the current target salinity S of the brine. target and measured salinity S current The control quantity of each actuator is calculated according to the following formula: Number of feedings K = Q / q unit Where Q is the total demand for water, feed, or salt replenishment, q unit This refers to the rated single-feed capacity of the single-feeding mechanism; Flow valve opening O=O base ×(Q / t set ) / v rated O base As the reference opening, t set To preset the feeding time, v rated Rated flow rate; Pump running time t=Q / (O×v) rated ); The feeding sequence is performed in the order of adding water first, then adding materials, and finally adding salt, with preset time intervals between each process.

[0010] The specific judgment logic for activating the adaptive feedback adjustment mechanism when the salinity deviation exceeds the preset threshold in step S4 is as follows: Let the predicted salinity value at the current time t be... The measured salinity value is S(t), and the deviation is... When |e(t)|>0.5% for more than 3 consecutive sampling periods, it is determined that the salinity deviation exceeds the limit, triggering the adaptive feedback adjustment mechanism; the sampling period is 1~5s.

[0011] As a preferred option, ω1 and ω2 are preset weight coefficients determined by the entropy weight method. Before the genetic algorithm starts, a set of objective weights is calculated using historical working condition datasets and algorithms such as the entropy weight method, which serve as the initial values ​​of ω1 and ω2.

[0012] Preferably, the historical working condition dataset in step S2 is constructed using the following method: S01. Collect the real-time temperature, real-time salinity, and pH value of the braising liquid during the braising process as input feature variables of the salinity prediction model. Collect the historical measured values ​​of water replenishment, material replenishment, and salt replenishment as target label variables for training the salinity prediction model. At the same time, collect the weight of the braising liquid, raw material type, material input, braising time, and braising liquid loss as auxiliary process parameters. S02. Noise and outliers are removed from the collected multi-source data. A linear normalization algorithm is used to map data of different dimensions to the same order of magnitude. The data is divided into training set, validation set and test set according to a preset ratio to obtain a standardized historical working condition dataset. The training set, validation set, and test set each contain three input features: temperature, salinity, and pH value, as well as corresponding target labels for water replenishment, material replenishment, and salt replenishment. The auxiliary process parameters are not directly used as inputs to the neural network, but are used for the condition classification and screening of training samples and as the basis for conversion from the control instruction set to the control parameter mapping stage in step S4.

[0013] Preferably, the loss of brine in step S01 includes the moisture loss due to evaporation during the braising process and the loss due to brine adhesion caused by the absorption of raw materials; the moisture loss is calculated by combining the weight difference of the brine before and after braising with the braising time, and the calculation formula is as follows: L evap =(W start -W end ) / t cook Among them, W start W represents the weight of the braising liquid at the start of the braising process. end The weight of the braising liquid at the end of the braising process, t cook The total braising time is denoted as ; the amount of loss due to adhesion of the braising liquid is obtained by multiplying the amount of raw materials fed into the preset unit raw material adhesion coefficient.

[0014] The linear normalization algorithm described in step S02 maps each input feature variable Xi to the [0,1] interval according to the following formula: ; Among them, X i Let X be the original value of the i-th feature variable. i,min and X i,maxThese are the minimum and maximum values ​​of the feature variable in the training set, respectively; the target label variable is also standardized using the same linear normalization method, and the model output value needs to be inversely normalized to obtain the actual predicted values ​​of water replenishment, feed replenishment, and salt replenishment.

[0015] Preferably, the preset ratio in step S02 is: the data is randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1; wherein the training set is used for learning model parameters, the validation set is used to evaluate the effect and adjust hyperparameters during training, and the test set is used to verify the true performance of the model on unknown data after training is completed.

[0016] Preferably, the specific parameters of the genetic algorithm in step S2 are set as follows: the population size is set to 50, the maximum number of generations is set to 100, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and an elite retention strategy is adopted to ensure that the best individual is not eliminated; the number of hidden layer nodes, the initial weights and the initial thresholds are encoded using real number encoding, and the optimization is iteratively sought through roulette wheel selection, single-point crossover and random mutation operations. After 50 generations of evolution, the fitness tends to stabilize, and the globally optimal initial parameter combination is output.

[0017] Preferably, the specific parameters for fine training of the Bayesian regularized neural network in step S3 are set as follows: the maximum number of iterations is set to 200, and the convergence error is 1×10⁻⁶. -5 The hidden layer activation function adopts a hyperbolic tangent sigmoid function, and the output layer activation function adopts a linear activation function. During training, the Bayesian framework dynamically and adaptively adjusts the values ​​of α and β, and the number of effective network parameters gradually converges to a stable value as the training process progresses. When the validation set error no longer decreases, training automatically stops, and a trained salinity prediction model is generated.

[0018] Preferably, the number of hidden layer nodes in the Bayesian regularized neural network is determined by global optimization of the genetic algorithm combined with control experiments. After traversing and verifying multiple node ranges, the number of hidden layer nodes is determined to be 8 by global optimization of the genetic algorithm, thus achieving the optimal balance between model complexity and prediction accuracy.

[0019] Preferably, the adaptive feedback adjustment mechanism in step S4 specifically includes: (1) The model parameters of the salinity prediction model are dynamically corrected according to the deviation between the predicted salinity value and the measured salinity value. When the deviation continues to exceed the standard, the local second iteration training of the salinity prediction model is triggered. (2) Construct a feeding execution error compensation model, and adjust the feeding sequence and pump running time in real time according to the flow valve opening deviation and pump response delay data to compensate for the control error caused by the response lag of the actuator.

[0020] Preferably, in the feeding execution error compensation model, the theoretical feeding amount for the k-th feeding instruction is set to Q. cmd (k), the actual amount of material fed is Q act (k), execution error is The execution error is estimated online using an exponentially weighted moving average method. Where ρ is the smoothing factor, 0 < ρ < 1; the compensation amount for the (k+1)th feeding instruction is... The revised feeding instruction is Q. cmd '(k+1)=Q cmd (k+1)+Q comp (k+1). The specific triggering condition for the local secondary iteration training of the salinity prediction model is as follows: when the average relative error (MRE) of salinity prediction for each of the five consecutive production batches exceeds a preset error threshold of 5%, local secondary iteration training is triggered. The local secondary iteration training uses the current model parameters as initial values ​​and retrains the model using the working condition data of the most recent 30 batches under the Bayesian regularization framework. The upper limit of the number of iterations is set to 50, and the convergence error is set to 1×10. -4 .

[0021] Preferably, the preset threshold in step S4 is a salinity deviation exceeding ±0.5%; when the salinity deviation exceeds the preset threshold, sensor data drifts, or feeding execution error exceeds the standard, an audible and visual warning is triggered and the adaptive feedback adjustment mechanism is activated.

[0022] Preferably, in step S4, the detection logic for sensor data drift is as follows: Assume the salinity sequence collected by the salinity sensor within n consecutive sampling periods (n≥10, sampling period is 1~5s) is {S(tk), …, S(t-1), S(t)}. Calculate the average sliding window rate of change ΔS / Δt of this sequence. If, under conditions of stable brine temperature (temperature change rate ≤0.5℃ / min) and stable pH value (pH change rate ≤0.05 / min), the average sliding window rate of change ΔS / Δt of the salinity sequence continuously exceeds 0.1% / min and exhibits a monotonically increasing or decreasing trend for more than 5 sampling periods, and the deviation of this trend from the baseline value of salinity change rate within the same temperature range under historical normal operating conditions (obtained from historical operating condition data statistics) exceeds 3 times the standard deviation, then it is determined to be sensor data drift. When a salinity deviation exceeding a preset threshold, sensor data drift, or feeding execution error exceeding the standard is detected, an audible and visual warning is triggered, and an adaptive feedback adjustment mechanism is initiated.

[0023] Preferably, the control parameters in step S4 include the number of feeding operations, the opening degree of the flow valve, the pump running time, and the feeding sequence. When the feeding mechanism is driven to perform water replenishment, material replenishment, and salt replenishment operations according to the control parameters, the PLC motion controller receives the control parameters output by the industrial computer and drives the multi-channel automatic feeding mechanism to complete the quantitative water replenishment, material replenishment, and salt replenishment operations according to the preset sequence.

[0024] This invention further claims a brine salinity intelligent prediction and control system, wherein the system employs the aforementioned multi-source data fusion-based intelligent brine salinity prediction method to achieve dual-objective optimization and adaptive closed-loop control of brine product salinity, comprising: The multi-source data acquisition module includes a temperature sensor, a salinity sensor, a pH sensor, and a weight sensor, which are used to collect the real-time temperature, real-time salinity, pH value, and weight of the braising liquid during the braising process. The data preprocessing module is used to perform noise and outlier removal and linear normalization on the collected multi-source data. The salinity prediction model building module includes a high-performance industrial computer encapsulating a fusion algorithm of genetic algorithm and Bayesian regularized neural network, used to generate a trained salinity prediction model. The control instruction generation and execution module includes a PLC motion controller and a multi-channel automatic quantitative feeding mechanism, which is used to map the control instruction set into control parameters that can be executed by the feeding system and drive the feeding mechanism to perform water replenishment, material replenishment and salt replenishment operations. The adaptive closed-loop feedback control module is used to dynamically correct the subsequent feeding control strategy based on the deviation between the salinity prediction value and the actual salinity value, thus forming a salinity closed-loop control. The output of the multi-source data acquisition module is connected to the input of the data preprocessing module. The output of the data preprocessing module is connected to the input of the control command generation and execution module. The feedback output of the control command generation and execution module is connected to the input of the adaptive closed-loop feedback control module. The correction signal output of the adaptive closed-loop feedback control module is connected to the feedback input of the salinity prediction model construction module and the control command generation and execution module.

[0025] Preferably, the temperature sensor in the multi-source data acquisition module is a PT100 platinum resistance temperature sensor, installed in the middle and bottom of the side wall of the braising pot, with a measurement range of 0~150℃ and an accuracy of ±0.1℃; the salinity sensor is a conductivity type salinity sensor, installed in the lower middle part of the side wall of the braising pot, with a measurement range of 0~30% and an accuracy of ±0.5%; the pH sensor is a glass electrode type pH sensor, installed in the upper part of the side wall of the braising pot, with a measurement range of 0~14 and an accuracy of ±0.01; and the weight sensor is a weighing sensor, installed below the support base of the braising pot, with a measurement range of 0~2000kg and an accuracy of ±0.5kg. The multi-channel automatic quantitative feeding mechanism includes a water replenishment channel, a material replenishment channel, and a salt replenishment channel. Each channel is independently equipped with a metering pump, a flow regulating valve, and a pulse counter. The metering pump is a peristaltic pump or a diaphragm pump, the flow regulating valve is an electric regulating valve, and the pulse counter is used to provide real-time feedback on the actual feeding amount. The feeding port of each channel is located at the top of the braising pot, and a splash guard and a disperser are provided at the feeding port. The high-performance industrial computer and the PLC motion controller communicate in real time via the industrial Ethernet protocol, with a communication cycle of 100~500ms. The industrial computer has a built-in data storage module for storing historical operating data, model parameters, control commands, and system operation logs. The industrial computer also has a built-in visual monitoring interface for displaying the salinity prediction curve, feeding progress, operating parameter values, and system warning information in real time. The system also includes a remote monitoring terminal, which is connected to the high-performance industrial computer via a wireless communication module for remote monitoring and parameter configuration of the braising salinity control process. The wireless communication module supports Wi-Fi, 4G, or 5G communication protocols.

[0026] The salinity sensor is pre-calibrated with a specific brine matrix and temperature compensation to eliminate the interference of non-NaCl soluble substances such as amino acids, organic acids, and flavor nucleotides in the brine on the conductivity detection value. At the same time, the original conductivity signals measured at different temperatures are uniformly converted into equivalent salinity values ​​at the reference temperature, ensuring that the salinity detection results are only related to the true NaCl concentration.

[0027] The present invention has at least the following beneficial effects: Firstly, this invention constructs a fitness function with the dual objectives of minimizing the mean relative error (MRE) of salinity prediction and the coefficient of variation (CV) of salinity between batches. It uses a genetic algorithm to globally search and determine the initial weights and thresholds of the neural network, thus overcoming the defect of traditional neural networks being prone to getting trapped in local optima due to random initialization of initial parameters. This significantly improves the generalization ability and prediction accuracy of the salinity prediction model. Secondly, this invention further introduces a Bayesian regularization framework based on the coarse optimization of the genetic algorithm, with the objective function G=αE. d +βEw The network is finely trained, and the regularization parameters α and β are dynamically and adaptively adjusted through the Bayesian framework to automatically balance the fitting accuracy and model complexity, effectively suppress overfitting, and ensure the prediction stability and robustness of the model under complex brine conditions. Thirdly, this invention combines the brine quality control instruction set output by the salinity prediction model with the real-time collected brine weight, and maps it into control parameters such as the number of feedings, flow valve opening, pump running time, and feeding sequence that the feeding system can execute. This realizes the automatic conversion of prediction results into execution instructions without manual intervention, and greatly improves the automation level and control response speed of braising processing.

[0028] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0029] Figure 1 This is a Bayesian regularized neural network framework described in one embodiment of the present invention; Figure 2 This is the training result of the salinity prediction model described in another embodiment of the present invention; Figure 3 This is a graph showing the prediction accuracy and generalization ability of the salinity prediction model in another embodiment of the present invention. Figure 4 This is a diagram illustrating the decision performance of the salinity prediction model described in another embodiment. Figure 5 This is a flowchart illustrating the intelligent prediction and control system for brine salinity according to another embodiment of the present invention. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0031] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0032] The example illustrates the process of salinity prediction and control of livestock and poultry meat braised products on an industrial continuous braising production line. However, the application of this invention is not limited to this and can also be applied to other types of braising products with intelligent salinity control scenarios.

[0033] 1. System Composition and Network Structure Design The core hardware components of this invention system include: a temperature sensor (PT100 platinum resistance temperature sensor, measurement range 0~150℃, accuracy ±0.1℃), a salinity sensor (conductivity type salinity sensor, measurement range 0~30%, accuracy ±0.5%), a pH sensor (glass electrode type pH sensor, measurement range 0~14, accuracy ±0.01%), a weight sensor (weighing sensor, measurement range 0~2000kg, accuracy ±0.5kg), a data synchronization acquisition instrument, a high-performance industrial computer, a PLC motion controller, a multi-channel automatic quantitative feeding mechanism (including a water supply channel, a material supply channel, and a salt supply channel, each channel is independently equipped with a metering pump, a flow regulating valve, and a pulse counter), and a visual interactive terminal.

[0034] The system adopts a modular design, consisting of a multi-source data preprocessing module, an initial parameter global optimization module, a neural network prediction module, and an adaptive feedback control module. A real-time data communication link is established between these modules via an industrial computer. Real-time data communication between the industrial computer and the PLC motion controller is achieved via the industrial Ethernet protocol, with a communication cycle of 100-500ms.

[0035] The salinity prediction model uses a three-layer feedforward Bayesian regularized neural network as its basic architecture, such as... Figure 1 As shown, the input layer has three nodes, corresponding to the real-time collected brine temperature, salinity, and pH value, respectively; the output layer has three nodes, corresponding to the water replenishment, feed replenishment, and salt replenishment, respectively. The number of hidden layer nodes was determined through global optimization using a genetic algorithm combined with multiple sets of control experiments. After traversing and verifying across multiple node ranges, the number of hidden layer nodes was determined to be 8. At this point, the difference between the model training error and the testing error was minimized, and the prediction correlation coefficient was maximized, achieving the optimal balance between model complexity and prediction accuracy. The activation function for the hidden layer is a hyperbolic tangent sigmoid function, and the activation function for the output layer is a linear activation function, to adapt to the industrial real-time prediction scenario.

[0036] 2. Multi-source data acquisition and preprocessing When raw materials are transferred to the braising station via the production line, multi-source sensors deployed on the braising pot trigger synchronous signal acquisition, outputting real-time multi-dimensional operating data of the braising liquid, including braising temperature (°C), real-time salinity (%), pH value, and braising liquid weight (kg). Simultaneously, auxiliary process parameters such as raw material type, feed amount, braising time, and braising liquid loss are also collected.

[0037] The collected multi-source data first underwent noise and outlier removal. Specifically, a statistical thresholding method was used to identify and remove outliers exceeding three times the standard deviation, and missing values ​​were filled using linear interpolation. Subsequently, a linear normalization algorithm was used to map data of different dimensions to the [0,1] interval. The normalization formula is as follows: ;where X iLet X be the original value of the i-th feature variable. i,min and X i,max These are the minimum and maximum values ​​of the feature variable in the training set, respectively.

[0038] After data cleaning and normalization, the data is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used to evaluate performance and adjust hyperparameters during training, and the test set is used to verify the model's true performance on unknown data after training is complete.

[0039] 3. Construction and optimization of salinity prediction model Based on a standardized historical working condition dataset, an optimized prediction model is constructed using a salinity prediction algorithm that integrates genetic algorithms and Bayesian regularized neural networks. The model aims to minimize the salinity prediction error and the salinity fluctuation rate between batches.

[0040] (I) Global coarse optimization using genetic algorithm To address the issue of traditional neural networks easily getting trapped in local optima due to random initialization of initial parameters, a genetic algorithm is first employed to solve for the globally optimal initial parameters within the parameter optimization space. Real-number encoding is used to encode the network's initial weights, thresholds, and the number of hidden layer nodes. The number of hidden layer nodes, the weight matrix from the input layer to the hidden layer, the weight matrix from the hidden layer to the output layer, the hidden layer threshold vector, and the output layer threshold vector are encoded into a chromosome with a length of: , where n in =3 represents the number of nodes in the input layer, n h n is the number of hidden layer nodes. out =3 represents the number of output layer nodes.

[0041] The population size is set to 50, the maximum number of generations is set to 100, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and an elite retention strategy is adopted to ensure that the best individuals are not eliminated.

[0042] A fitness function is constructed with the dual objectives of minimizing the mean relative error (MRE) of salinity prediction and the coefficient of variation (CV) of salinity between batches: Where F is the fitness value of the genetic algorithm, and a larger value indicates better overall model performance; ω1 and ω2 are preset weight coefficients; MRE is the mean relative error of salinity prediction, calculated using the following formula: CV is the coefficient of variation of salinity among N consecutive production batches, calculated using the following formula: σ is the standard deviation of batch salinity, and μ is the average salinity of batch.

[0043] Evolution is terminated early when the fitness value changes less than a preset threshold over 10 consecutive generations, using roulette wheel selection, single-point crossover, and random mutation operations to output the globally optimal initial parameter combination.

[0044] (II) Refined Training of Bayesian Regularized Neural Networks The optimal initial parameter combination, optimized by the genetic algorithm, is assigned to a three-layer feedforward Bayesian regularized neural network as the starting state for network training. Within the Bayesian regularization framework, the goal is to minimize the objective function G = αE. d + βE w To achieve the training objective, the Levenberg-Marquardt algorithm is used for iterative convergence training of the network. Where E... d E is the sum of squared prediction errors of the network. w Let be the sum of squared network weights, and α and β be the regularization parameters for dynamic adaptive adjustment under the Bayesian framework.

[0045] The regularization parameters α and β are automatically updated after each iteration according to the Bayesian criterion: , Where γ is the number of effective network parameters and n' is the total number of training samples. γ is calculated based on the Hessian matrix under the Gaussian-Newton approximation, obtained by eigenvalue decomposition of the Hessian matrix after each Levenberg-Marquardt iteration. The number of effective network parameters γ gradually decreases during training and converges to a stable value, indicating that the network complexity is effectively controlled.

[0046] The iteration step size of the Levenberg-Marquardt algorithm is controlled by an adaptive adjustment factor λ: when the objective function G decreases after each iteration, λ is divided by a preset factor θ (θ=10) to accelerate convergence; when the objective function G increases, λ is multiplied by θ to reduce the step size and recalculate the current iteration. The initial value of the adaptive adjustment factor λ is set to 0.001.

[0047] The maximum number of iterations is set to 200, and the convergence error is set to 1×10⁻⁶. -5 Training automatically stops when the validation set error no longer decreases, generating a completed salinity prediction model.

[0048] Model training results are as follows Figure 2 As shown: After training begins, the error decreases rapidly, with the largest decrease occurring in the first 20 rounds, indicating that the initial weights provided by the genetic algorithm have positioned the network well. From round 20 to round 73, the error continues to decrease slowly, while the values ​​of α and β are dynamically adjusted, and the number of effective parameters in the network gradually converges. After round 73, the error stops decreasing, and training automatically stops. The final root mean square error on the training set is 0.012, and the average relative error is less than 2%.

[0049] 4. Generation and mapping of control commands The real-time collected temperature, salinity, and pH data are input into the trained salinity prediction model. The model outputs a set of brine quality control instructions, including water replenishment ΔWw, material replenishment ΔWm, and salt replenishment ΔWs.

[0050] The control command set, combined with the real-time collected weight W of the brine, is mapped and transformed into control parameters executable by the feeding system. Specifically, based on the target salinity S of the current brine... target and measured salinity S current The control quantity of each actuator is calculated according to the following formula: Number of feedings K = Q / q unit Where Q is the total demand for water, feed, or salt replenishment, q unit This refers to the rated single-feed capacity of the single-feeding mechanism; Flow valve opening O=O base ×(Q / t set ) / v rated O base As the reference opening, t set To preset the feeding time, v rated Rated flow rate; Pump running time t=Q / (O×v) rated ); The feeding sequence is performed in the order of adding water first, then adding materials, and finally adding salt, with preset time intervals between each process.

[0051] Compare the predicted values ​​with the actual recorded values ​​(e.g.) Figure 3 As shown in the figure, the Pearson correlation coefficient was calculated, and the correlation coefficients of the output variables reached 0.985, 0.982 and 0.987 respectively, all above 0.98, indicating that there is a very strong linear positive correlation between the predicted value and the true value; the error distribution histogram shows that the absolute error of more than 95% of the samples falls within ±0.05 (after normalization), proving that the model has high prediction accuracy.

[0052] The PLC motion controller receives control parameters from the industrial computer and drives the multi-channel automatic feeding mechanism to complete the quantitative water replenishment, material replenishment, and salt replenishment operations according to a preset sequence. The feeding port of each channel is located on the top of the braising pot, and the feeding port is equipped with a splash baffle and a disperser to ensure uniform feeding.

[0053] 5. Real-time feedback correction and closed-loop control The industrial computer transmits model prediction parameters, brine quality control instruction sets, real-time sensor data, and feeding execution status to a visual interactive terminal. The interface displays real-time salinity prediction curves, feeding progress, operating parameter values, and system warning information. Operators can monitor the system's operating status in real time through the interface and manually adjust parameters, achieving full-process visual control.

[0054] During the feeding process, sensors collect real-time measurements of the brine salinity. Let the predicted salinity at the current time t be... The measured salinity value is S(t), and the deviation is... When |e(t)|>0.5% for more than 3 consecutive sampling periods (sampling period is 1~5s), it is determined that the salinity deviation exceeds the limit, and an audible and visual warning is immediately triggered and the adaptive feedback adjustment mechanism is started.

[0055] The adaptive feedback adjustment mechanism specifically includes the following two aspects: (1) Dynamic correction of model parameters: The model parameters of the salinity prediction model are dynamically corrected based on the deviation between the predicted and measured salinity values. When the average relative error (MRE) of salinity prediction for each of the five consecutive production batches exceeds the preset error threshold of 5%, the local secondary iteration training of the salinity prediction model is triggered. The local secondary iteration training uses the current model parameters as the initial values ​​and retrains the model using the working condition data of the most recent 30 batches under the Bayesian regularization framework. The upper limit of the number of iterations is set to 50, and the convergence error is set to 1×10. -4 .

[0056] 2) Material feeding execution error compensation: Construct a material feeding execution error compensation model, assuming the theoretical material feeding amount for the k-th material feeding instruction is Q. cmd (k), the actual amount of material fed is Q act (k), execution error is e q (k)=Q cmd (k)-Q act (k); The execution error is estimated online using the exponentially weighted moving average method: Where ρ is the smoothing factor (0 < ρ < 1); the compensation amount for the (k+1)th feeding instruction. The revised feeding instruction is Q. cmd '(k+1)=Q cmd (k+1)+Q comp (k+1).

[0057] Once the salinity is controlled to the target level, the industrial computer sends a linkage signal to the production line's conveyor module to enable continuous operation of the braising process.

[0058] Intelligent decision verification under typical working conditions, such as Figure 4As shown, three representative salinity conditions were selected to test the rationality of the model's decision-making. In the first condition, the current salinity was very close to the target salinity, and the model gave a small amount of salt replenishment, while the water and material replenishment remained stable with only minor adjustments, consistent with the "stable maintenance" production logic. In the second condition, the current salinity was significantly lower than the lower limit of the process requirements, and the model output a command to significantly increase the amount of salt replenishment, while appropriately reducing the amount of water replenishment to avoid dilution, demonstrating the model's ability to quickly correct deviations. In the third condition, the current salinity was high, and the model chose not to continue replenishing salt, but instead significantly increased the amount of water replenishment for dilution adjustment, while slightly reducing the amount of material replenishment to balance the total volume. Under all three conditions, the model provided strategies that were consistent with common production knowledge and quantitatively reasonable, demonstrating the rationality of the model's decision-making scheme and its ability to quickly correct deviations.

[0059] The overall workflow is as follows Figure 5 As shown, this method can replace the traditional manual experience-based control mode, keeping the salinity prediction error within 5% and the batch salinity fluctuation rate below 3%, significantly improving the intelligence and automation level of braised product processing and the consistency of product batches.

[0060] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for intelligent prediction of brine salinity based on multi-source data fusion, characterized in that, include: S1. Define the network architecture of the salinity prediction model. The network architecture adopts a three-layer feedforward Bayesian regularized neural network framework. Its input layer nodes correspond to the real-time collected temperature, salinity, and pH value, and the output layer nodes correspond to the water replenishment, feed replenishment, and salt replenishment. S2. Based on historical operating condition datasets, with the dual objectives of minimizing the mean relative error (MRE) of salinity prediction and minimizing the coefficient of variation (CV) of salinity between batches, a fitness function F is constructed, using the network architecture with parameters to be determined as the evaluation object: Using a genetic algorithm to maximize the evolutionary objective with the fitness function F, a global search is performed to determine the number of hidden layer nodes, initial weights, and initial thresholds of the network architecture in S1, thereby obtaining the globally optimal combination of initial parameters. S3. Assign the globally optimal initial parameter combination obtained in step S2 to the network architecture described in step S1 as the starting state for network training, under the Bayesian regularization framework, to minimize the objective function G=αE. d +βE w As the training objective, the Levenberg-Marquardt algorithm is used to iteratively converge the training of the network; During training, Bayesian regularization automatically balances fitting accuracy and model complexity to suppress overfitting, generating a trained salinity prediction model. S4. Input the real-time collected temperature, salinity, and pH values ​​into the salinity prediction model trained in step S3. The model outputs a brine quality control instruction set including water replenishment, material replenishment, and salt replenishment. The control instruction set is combined with the real-time collected brine weight mapping to convert it into control parameters that the feeding system can execute. The control parameters include the number of feedings, the opening degree of the flow valve, the pump running time, and the feeding sequence. The feeding mechanism is driven to perform water replenishment, material replenishment, and salt replenishment operations according to the control parameters. During the feeding process, the brine salinity is collected in real time. When the salinity deviation is detected to exceed the preset threshold, the adaptive feedback adjustment mechanism is activated. The subsequent feeding control strategy is dynamically corrected according to the deviation between the salinity prediction value and the salinity measurement value to form a closed-loop salinity control. Once the salinity is controlled to the target, a linkage signal is sent to the production line conveying module to realize the continuous operation of the braising process. Where ω1 and ω2 are preset weighting coefficients; E d E is the sum of squared prediction errors of the network. w Let be the sum of squared network weights, and α and β be the regularization parameters for dynamic adaptive adjustment under the Bayesian framework.

2. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The historical working condition dataset in step S2 is constructed using the following method: S01. Collect the real-time temperature, real-time salinity, and pH value of the braising liquid during the braising process as input feature variables of the salinity prediction model. Collect the historical measured values ​​of water replenishment, material replenishment, and salt replenishment as target label variables for training the salinity prediction model. At the same time, collect the weight of the braising liquid, raw material type, material input, braising time, and braising liquid loss as auxiliary process parameters. S02. Noise and outliers are removed from the collected multi-source data. A linear normalization algorithm is used to map data of different dimensions to the same order of magnitude. The data is divided into training set, validation set and test set according to a preset ratio to obtain a standardized historical working condition dataset. The training set, validation set, and test set each contain three input features: temperature, salinity, and pH value, as well as corresponding target labels for water replenishment, material replenishment, and salt replenishment. The auxiliary process parameters are not directly used as inputs to the neural network, but are used for the condition classification and screening of training samples and as the basis for conversion from the control instruction set to the control parameter mapping stage in step S4.

3. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 2, characterized in that, The preset ratio mentioned in step S02 is as follows: the data is randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1; wherein, the training set is used for learning model parameters, the validation set is used to evaluate the effect and adjust hyperparameters during training, and the test set is used to verify the true performance of the model on unknown data after training is completed.

4. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The specific parameters of the genetic algorithm in step S2 are set as follows: the population size is set to 50, the maximum number of generations is set to 100, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and an elite retention strategy is adopted to ensure that the best individual is not eliminated. The number of hidden layer nodes, the initial weights and the initial thresholds are encoded using real number encoding. The algorithm iteratively seeks optimization through roulette wheel selection, single-point crossover and random mutation operations. After several generations of evolution, the fitness tends to stabilize, and the globally optimal initial parameter combination is output.

5. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The specific parameters for fine training of the Bayesian regularized neural network in step S3 are set as follows: the maximum number of iterations is set to 200, and the convergence error is 1×10⁻⁶. -5 The hidden layer activation function adopts a hyperbolic tangent sigmoid function, and the output layer activation function adopts a linear activation function. During training, the Bayesian framework dynamically and adaptively adjusts the values ​​of α and β, and the number of effective network parameters gradually converges to a stable value as the training process progresses. When the validation set error no longer decreases, training automatically stops, and a trained salinity prediction model is generated.

6. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The number of hidden layer nodes in the Bayesian regularized neural network was determined by global optimization using a genetic algorithm combined with comparative experiments. After traversing and verifying multiple node ranges, the number of hidden layer nodes was determined to be 8 by global optimization using a genetic algorithm, achieving the optimal balance between model complexity and prediction accuracy.

7. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The adaptive feedback adjustment mechanism described in step S4 specifically includes: (1) The model parameters of the salinity prediction model are dynamically corrected according to the deviation between the predicted salinity value and the measured salinity value. When the deviation continues to exceed the standard, the local second iteration training of the salinity prediction model is triggered. (2) Construct a feeding execution error compensation model, and adjust the feeding sequence and pump running time in real time according to the flow valve opening deviation and pump response delay data to compensate for the control error caused by the response lag of the actuator.

8. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The preset threshold mentioned in step S4 is a salinity deviation exceeding ±0.5%; when the salinity deviation exceeds the preset threshold, sensor data drift, or feeding execution error exceeds the standard, an audible and visual warning is triggered and the adaptive feedback adjustment mechanism is activated.

9. The intelligent prediction method for brine salinity based on multi-source data fusion as described in claim 1, characterized in that, The control parameters mentioned in step S4 include the number of feedings, the opening degree of the flow valve, the pump running time, and the feeding sequence. When the feeding mechanism is driven to perform water replenishment, material replenishment, and salt replenishment operations according to the control parameters, the PLC motion controller receives the control parameters output by the industrial computer and drives the multi-channel automatic feeding mechanism to complete the quantitative water replenishment, material replenishment, and salt replenishment operations according to the preset sequence.

10. A brine salinity intelligent prediction and control system, characterized in that, The system employs the intelligent prediction method for brine salinity based on multi-source data fusion as described in any one of claims 1 to 9 to achieve dual-objective optimization and adaptive closed-loop control of brine salinity, comprising: The multi-source data acquisition module includes a temperature sensor, a salinity sensor, a pH sensor, and a weight sensor, which are used to collect the real-time temperature, real-time salinity, pH value, and weight of the braising liquid during the braising process. The data preprocessing module is used to perform noise and outlier removal and linear normalization on the collected multi-source data. The salinity prediction model building module includes a high-performance industrial computer encapsulating a fusion algorithm of genetic algorithm and Bayesian regularized neural network, used to generate a trained salinity prediction model. The control instruction generation and execution module includes a PLC motion controller and a multi-channel automatic quantitative feeding mechanism, which is used to map the control instruction set into control parameters that can be executed by the feeding system and drive the feeding mechanism to perform water replenishment, material replenishment and salt replenishment operations. The adaptive closed-loop feedback control module is used to dynamically correct the subsequent feeding control strategy based on the deviation between the salinity prediction value and the actual salinity value, thus forming a salinity closed-loop control. The output of the multi-source data acquisition module is connected to the input of the data preprocessing module. The output of the data preprocessing module is connected to the input of the control command generation and execution module. The feedback output of the control command generation and execution module is connected to the input of the adaptive closed-loop feedback control module. The correction signal output of the adaptive closed-loop feedback control module is connected to the feedback input of the salinity prediction model construction module and the control command generation and execution module.

Images