A meat product thawing energy consumption conditioning internet of things monitoring and early warning system

Abstract

The present application relates to a kind of conditioning meat product thawing energy consumption internet of things monitoring early warning system, specifically to meat product thawing energy consumption monitoring field, the scheme is fused by data assimilation sparse observation and physical model, real-time reconstruction thawing library high-precision three-dimensional physical field, significantly improve the integrity and accuracy of state perception, and then, based on this full-field state, multi-objective predictive control rolling optimization is carried out, while ensuring thawing uniformity and process stability, the minimization of energy consumption is realized, online correction mechanism continuously self-adapting adjustment model parameter, effectively against model drift and environmental disturbance, maintain long-term control accuracy, cloud cooperation then gather global operation data, constantly evolve physical model and optimization strategy, and the knowledge safety is fed back to each edge node, form local rapid response and global continuous optimization of synergistic enhancement closed loop, finally achieve accurate, energy-saving, adaptive high-quality thawing control.

Description

Technical Field

[0001] This invention relates to the field of energy consumption monitoring for thawing meat products, and more specifically, to an Internet of Things (IoT) monitoring and early warning system for energy consumption monitoring of thawing prepared meat products. Background Technology

[0002] In the processing of prepared meat products in large-scale food processing enterprises, the slow-freezing and thawing of raw meat is a crucial preliminary step. A typical thawing workshop usually has dozens or even hundreds of independent thawing chambers or rooms. These facilities operate continuously to process large quantities of raw materials. The thawing process is essentially a complex heat and mass transfer process, and its efficiency and uniformity are affected by the coupling of multiple factors. Inside the thawing chamber, the temperature, humidity, and flow velocity fields of the air are not uniformly distributed. The initial core temperature, geometric dimensions, bulk density, and arrangement of the meat products also vary. These internal factors collectively lead to spatial inconsistencies in the heat exchange rate. At the same time, external environmental conditions such as temperature fluctuations in the workshop environment caused by day-night cycles, seasonal climate changes, and subtle disturbances in power grid parameters can penetrate the insulation structure of the chamber or directly affect the performance of the refrigeration unit, thereby interfering with the stable operation of the thawing process. Currently, the industry generally relies on feedback control based on fixed temperature measurement points to manage this process. In existing technologies, monitoring and control systems for thawing energy consumption often employ a limited number of temperature sensors installed at key locations within the thawing space. The average temperature of these monitoring points or the temperature of a representative point is calculated as the basis for determining the entire thawing process. Only when this measured value reaches a preset threshold does the control system issue start / stop or power adjustment commands to the refrigeration unit, heating device, fan, and other actuators. This control strategy, which relies on "point" measurements to reflect the "field" state, has fundamental flaws. First, the sparse sensor network cannot capture the true and continuous temperature distribution in the three-dimensional space within the thawing chamber, making it difficult to identify localized overheating areas and low-temperature dead zones. Overheating can lead to denaturation of meat proteins, increased juice loss, and accelerated microbial proliferation. Low-temperature dead zones prevent the center of the meat from reaching the required thawing endpoint temperature. Secondly, due to the inertia of heat transfer and the limitations of sensor distribution, the temperature information obtained by the system lags significantly behind the actual thermal state changes. Control actions based on such lagging information inevitably deviate from the optimal trajectory. This forces refrigeration or heating equipment to frequently operate in uneconomical start-stop cycles or to operate for extended periods outside the optimal energy efficiency point. This not only causes unnecessary energy loss but also makes it difficult to accurately determine the thawing endpoint, ultimately leading to inconsistent product quality and high production costs. Therefore, how to achieve real-time and accurate perception and prediction of the temperature field across the entire thawing warehouse, and implement forward-looking and precise energy regulation accordingly, has become a key technological bottleneck for improving product quality and reducing production energy consumption. Summary of the Invention

[0003] This invention addresses the technical problems existing in the prior art by providing an IoT monitoring and early warning system for the energy consumption of thawing processed meat products. The system utilizes a data assimilation module, a predictive control module, an online correction module, and a cloud-based collaborative module for data interaction with various edge devices to solve the problems mentioned in the background.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a physical sensor network set up in the thawing chamber to obtain physical quantities of the thawing process, a data assimilation module deployed on the edge side of the thawing chamber, a predictive control module, an online correction module, and a cloud collaboration module that interacts with each edge side for data. Data assimilation module: In response to real-time observation data collected by the physical sensor network, it calls the preset thawing process order reduction model and combines it with the data assimilation algorithm. It uses real-time observation data to correct the current state output of the thawing process order reduction model, thereby estimating the physical field state of the entire domain in the thawing warehouse at the current moment and outputting the corrected high-fidelity spatial state vector. Predictive control module: In response to the corrected high-fidelity spatial state vector output by the data assimilation module, based on the reduced-order model of the thawing process and the corrected high-fidelity spatial state vector as the initial state, it predicts the physical field evolution, thawing completion and cumulative energy consumption at multiple future time points under different control command sequences. It also generates the optimal control command to be executed at the current time and sends it to the actuator by solving a model predictive control problem that incorporates multi-objective optimization of energy consumption, temperature uniformity and control stability. Online calibration module: In response to the new real-time observation data collected by the physical sensor network at the next moment, the prediction deviation vector between the corrected high-fidelity spatial state vector, the optimal control command to be executed at the current moment, and the predicted physical field value at the next moment calculated by the thawing process reduction model, the module dynamically adjusts the key physical parameter vector contained in the thawing process reduction model through an online parameter estimation algorithm, and synchronizes the updated key physical parameter vector of the thawing process reduction model to the data assimilation module and the predictive control module. The cloud-based collaborative module periodically aggregates the corrected high-fidelity spatial state vectors, the optimal control command to be executed at the current moment, new real-time observation data, and the key physical parameter vectors of the updated thawing process reduction model from various edge sides. Based on big data and machine learning technologies, it uses the aggregated corrected high-fidelity spatial state vectors, the optimal control command to be executed at the current moment, the new real-time observation data, and the key physical parameter vectors of the updated thawing process reduction model to globally train and optimize the general thawing process high-fidelity model and the multi-objective weight coefficients used in the model predictive control problem. The optimized thawing process high-fidelity model and multi-objective weight coefficients are then sent as update data to the corresponding edge-side data assimilation module, predictive control module, and online correction module to update their preset thawing process reduction model and the multi-objective weight coefficients used in the model predictive control problem. In a preferred embodiment, the specific process by which the data assimilation module corrects the current state output of the thawing process reduction model using real-time observation data is as follows: First, the data assimilation module receives real-time observation data reflecting the physical quantities of the thawing process collected at time t from a physical sensor network consisting of temperature sensors, humidity sensors, and wind speed sensors deployed at multiple spatial locations within the thawing chamber. The module then performs validity verification, outlier removal, and dimension normalization on the received real-time observation data to form a standardized observation vector at time t. At the same time, a pre-set thawing process reduction model is invoked. The thawing process reduction model is defined by a basis function matrix consisting of a set of orthogonal basis functions that are pre-deployed from the cloud collaborative module through the intrinsic orthogonal decomposition method or pre-stored and used to characterize the main spatial distribution features of the physical field in the thawing chamber. The thawing process reduction model performs state evolution calculations in a low-dimensional space spanned by orthogonal basis functions. The high-dimensional state, which is used to generate the corrected high-fidelity space state vector in the entire thawing library, is expressed as a linear combination of the basis function matrix and a time-varying low-dimensional coefficient vector. The data assimilation module uses the low-dimensional coefficient vector of the previous optimal analysis field, which is the output of the previous assimilation cycle and serves as the optimal estimate of the previous time step, to calculate the current low-dimensional coefficient vector of the background field, which is used to characterize the prior estimate of the physical field at time t, through the dynamic equation of the thawing process reduction model. Then, the current high-dimensional background physical field state vector is obtained through a linear combination of the basis function matrix and the current low-dimensional coefficient vector of the background field.

[0005] In a preferred embodiment, after obtaining the current high-dimensional background physical field state vector, the data assimilation module performs a state fusion operation in the low-dimensional space based on regularized variational assimilation, specifically: The data assimilation module constructs a minimization objective function in a low-dimensional space spanned by orthogonal basis functions, and constructs a state estimation optimization problem based on the minimization objective function to obtain the optimal low-dimensional coefficient vector of the analysis field at time t. The minimization objective function consists of three terms. The first term is the observation fitting term, which measures the weighted difference between the standardized observation vector at time t and the predicted observation value extracted from the candidate state vector of the high-dimensional physical field obtained by a linear combination of the low-dimensional coefficient vector and the basis function matrix as optimization variables through an observation operator. The calculation of this weighted difference uses the inverse matrix of the observation error covariance matrix as the weight. The observation error covariance matrix is ​​a diagonal matrix used to characterize the uncertainty of sensor measurement error. The second term is the background constraint term, which measures the weighted difference between the low-dimensional coefficient vector, which is the optimization variable, and the low-dimensional coefficient vector of the background field at the current moment. The weighted difference is calculated using the inverse of a background error covariance matrix as the weight. This background error covariance matrix is ​​a matrix obtained in low-dimensional space based on historical forecast error statistics, used to characterize the uncertainty of the forecast error of the reduced-order model during the thawing process. The third term is the physical regularization term, which is used to introduce prior physical constraints into the candidate state vector of the high-dimensional physical field. This term is obtained by weighting the output of a physical constraint regularization function through an adjustable regularization parameter. By solving the state estimation optimization problem, the optimal low-dimensional coefficient vector of the analysis field is obtained under the joint constraints of the standardized observation vector, the low-dimensional coefficient vector of the background field at the current time, and the physical constraint regularization term.

[0006] In a preferred embodiment, after obtaining the low-dimensional coefficient vector of the optimal analysis field, the data assimilation module performs optimal state reconstruction and output operations, specifically as follows: The data assimilation module performs matrix multiplication on the low-dimensional coefficient vector of the optimal analysis field and the basis function matrix of the reduced-order model of the thawing process, reconstructing the high-dimensional physical field state vector at time t, which serves as the corrected high-fidelity spatial state vector.

[0007] In a preferred embodiment, the predictive control module, using the corrected high-fidelity spatial state vector as the initial state, specifically predicts the physical field evolution, thawing completion rate, and cumulative energy consumption process at multiple future time points under different control command sequences as follows: First, the predictive control module receives the corrected high-fidelity spatial state vector at time t from the data assimilation module, and projects the corrected high-fidelity spatial state vector into a low-dimensional coefficient space spanned by the column vectors of the basis function matrix defined by the thawing process reduction model, thereby obtaining a low-dimensional coefficient vector to describe the current system state at time t. Meanwhile, a predictive control time domain containing N consecutive control cycles is defined, and the control command vector for each control cycle within this predictive control time domain is parameterized. Each control command vector is a real number vector containing multiple components, which correspond to the set values ​​of different actuators in the defrosting chamber. The actuators include fans that regulate wind speed, valves that regulate refrigerant flow, and devices that regulate heating power. The control command vector within each control cycle is regarded as a constant decision variable to be optimized and which remains unchanged within the cycle. All these decision variables together constitute a complete sequence of control commands that spans the entire predictive control time domain.

[0008] In a preferred embodiment, the specific process of generating optimal control commands by solving a model predictive control problem incorporating multi-objective optimization is as follows: Based on the dynamic evolution equation of the reduced-order model of the thawing process, the low-dimensional coefficient vector obtained by projection is used as the initial condition, and the parameterized control command sequence is used as the input. In the digital domain, the complete evolution trajectory of the physical field in the thawing library, the thawing completion index calculated based on the physical field, and the cumulative energy consumption calculated by the correlation between the internal state of the model and the energy flow are predicted. The model predictive control problem is constructed as a constrained nonlinear optimization problem. Its decision variables are parameterized control command sequences, its objective function is a multi-objective composite cost function to be constructed, and its constraints include the dynamic constraints described by the dynamic evolution equation of the reduced-order model of the thawing process, the physical upper and lower bound constraints of each component of the control command vector, and the process safety constraints of the thawing process. To evaluate the merits of different control command sequences, a multi-objective composite cost function is constructed. This function consists of a weighted sum of four cost terms. Each of the four cost terms in the multi-objective composite cost function contains adjustable parameters. These adjustable parameters together constitute the multi-objective weight coefficients used in the model predictive control problem. The multi-objective weight coefficients specifically include the following trainable parameter set: all elements of the positive semi-definite weight matrix applied to the energy efficiency cost term, the global scaling factor of the discrete spatial gradient operator matrix applied to the uniformity cost term, the stationarity weight coefficient in the control stationarity cost term, and the thermodynamic cost weight coefficient in the thermodynamic irreversibility cost term. The first term is the energy efficiency cost term, which is characterized by calculating the weighted inner product of the low-dimensional coefficient vectors at each moment in the prediction time domain under a pre-defined positive semi-definite weight matrix used to correlate the energy change rate of the system. The second term is the uniformity cost term, which is characterized by the sum of squared L2 norms of the spatial gradients of the high-dimensional temperature field reconstructed from the low-dimensional coefficient vector and the basis function matrix at each time point in the prediction time domain. This gradient calculation is achieved through a preset discrete spatial gradient operator matrix based on the thawing library computing grid. Its overall contribution to the multi-objective composite cost function is adjusted by a global scaling factor. The third term is the control stationarity cost term, which is characterized by calculating the sum of squared L2 norms of the changes in adjacent control command vectors in the predicted time domain and multiplying it by a stationarity weighting coefficient used to penalize drastic fluctuations in control actions. The fourth term is the thermodynamic irreversible cost term, which uses a thermodynamic cost weighting coefficient to weight and sum the predicted thermodynamic entropy yield values ​​calculated based on the predicted temperature field gradient distribution and velocity field distribution at each time point in the prediction time domain. The model predictive control problem is to find a sequence of control commands that minimizes a multi-objective composite cost function while satisfying all constraints. To achieve real-time solution of optimization problems with nonlinear dynamic constraints under edge computing resource limitations, a sequential quadratic programming method based on sensitivity analysis is used for numerical solution, satisfying constraints including model dynamic equations, physical upper and lower bounds of each component of the control command vector, and process safety constraints during the thawing process. Specifically, this includes: First, the dynamic evolution equations of the reduced-order model of the thawing process are used to perform forward integration on the system determined by the initial conditions and candidate control command sequences to obtain the nominal state trajectory prediction over the entire predictive control time domain. Next, using automatic differentiation techniques, in forward or reverse modes, the gradient information of the multi-objective composite cost function and all process safety constraint functions with respect to the parameterized control command sequence is calculated. The gradient information characterizes the first-order change sensitivity of the objectives and constraints relative to the control decision. Simultaneously, the approximate Hessian matrix of the multi-objective composite cost function is calculated, which characterizes its second-order curvature. Then, in each... In the next optimization iteration, based on the gradient information and the approximate Hessian matrix at the current iteration point, the original nonlinear programming problem is approximately constructed as a quadratic programming subproblem with a quadratic objective function and linearization constraints. Subsequently, the effective set method is used to solve the quadratic programming subproblem to obtain the search direction and corresponding step size that can reduce the objective function. Finally, iterative updates are performed along the search direction, and it is determined whether the current solution satisfies the preset convergence condition. If it does not satisfy the condition, the aforementioned steps are repeated until the convergence condition is satisfied, thereby obtaining the control command sequence that minimizes the multi-objective composite cost function, denoted as the optimal control command sequence at time t. Based on the rolling time-domain execution principle of model predictive control, only the first control instruction vector corresponding to the start time t of the current control cycle is extracted from the optimal control instruction sequence at time t, and used as the optimal control instruction to be executed at the current time, and sent to the actuator immediately.

[0009] In a preferred embodiment, the specific process of dynamically adjusting the key physical parameter vectors contained in the thawing process reduction model through an online parameter estimation algorithm in the online correction module is as follows: First, the online calibration module responds to the calibrated high-fidelity spatial state vector output by the data assimilation module at time t and the optimal control command to be executed at the current time generated by the predictive control module at time t. It takes the calibrated high-fidelity spatial state vector and the optimal control command to be executed at the current time as inputs and calls the thawing process reduced-order model. This model is parameterized by a key physical parameter vector, which contains multiple parameters used to characterize the physical properties of heat transfer, mass transfer and flow processes in the thawing chamber, including the equivalent convective heat transfer coefficient, turbulent viscosity coefficient and meat equivalent thermal diffusivity. It performs a forward calculation to obtain a predicted value of the physical field state in the thawing chamber at the next time, i.e., time t+1, which is denoted as the model predicted physical field at the next time. Then, in response to the new real-time observation data actually collected by the physical sensor network at time t+1, an observation operator consistent with the observation operator defined in the data assimilation module is used to extract the physical quantity values ​​corresponding to the sensor spatial location from the model-predicted physical field at the next time step, which is recorded as the predicted observation value at time t+1. Finally, the predicted observation value at time t+1 is subtracted from the new real-time observation data, and the difference is the prediction bias vector at time t+1.

[0010] In a preferred embodiment, the online correction module performs an incremental parameter update operation based on regularized recursive Bayesian learning, specifically, based on the predicted bias vector: The online calibration module maintains and dynamically updates a parameter estimation error covariance matrix; Meanwhile, the online calibration module dynamically calculates an adaptive observation weight matrix based on the observation variance of each sensor obtained from the historical observation data collected by the physical sensor network and the prediction deviation vector at the current moment. The value of each diagonal element of the adaptive observation weight matrix is ​​obtained by weighting and fusing the corresponding component of the historical observation variance of the corresponding sensor and the prediction deviation vector at the current moment through a forgetting factor. Based on the prediction bias vector, the key physical parameter vector at the current time t, the parameter estimation error covariance matrix, and the adaptive observation weight matrix, the online correction module constructs a parameter update objective function. The parameter update objective function is a weighted sum of the following three terms: The first term is the instantaneous fitting term, which is calculated as follows: First, using a parameterized unfreezing process reduction model that explicitly depends on the key physical parameter vector to be determined, based on the corrected high-fidelity spatial state vector at time t and the optimal control command to be executed at the current time, the observation value at time t+1 is predicted. Then, the weighted L2 norm squared of the difference between this predicted observation value and the new real-time observation data at time t+1 is calculated. The second term is the parameter prior constraint term, which is calculated by: calculating the weighted L2 norm square of the difference between the key physical parameter vector to be determined and the key physical parameter vector at the current time t. The third term is the physical feasible region regularization term, which is calculated as follows: evaluate a physical feasible region regularization function and multiply it by a dynamically adjustable adaptive regularization coefficient. The physical feasible region regularization function is based on the preset physical reasonable value range of each component in the key physical parameter vector. For each component in the updated key physical parameter vector that exceeds its corresponding range, calculate the square of the excess amount and accumulate it. The online calibration module obtains the updated key physical parameter vector by solving an optimization problem that minimizes the objective function of parameter update. At the same time, based on the recursive Bayesian learning framework, the online calibration module uses a preset forgetting factor, a prediction bias vector, and the Jacobian matrix of the key physical parameter vector output by the parameter unfreezing process reduction model to recursively update the parameter estimation error covariance matrix. Finally, the online calibration module synchronizes the updated key physical parameter vectors to the data assimilation module and the prediction control module to replace the original parameters in the reduced-order model during the unfreezing process within these two modules, thereby completing the online self-calibration of the model parameters.

[0011] In a preferred embodiment, the specific process of global training and optimization of the general high-fidelity model for the thawing process and the multi-objective weight coefficients used in the model predictive control problem in the cloud collaboration module is as follows: First, the cloud-based collaborative module performs spatiotemporal alignment and automatic quality assessment on heterogeneous time-series data that periodically converges from multiple edge sides. The time-series data consists of data packets from each edge side. Each data packet contains a set of data converged from the same edge side at time t. This set of data is defined as a data sample, which includes a corrected high-fidelity spatial state vector from the edge-side data assimilation module, the optimal control command to be executed at the current time from the edge-side predictive control module, new real-time observation data from the edge-side physical sensor network, and the updated key physical parameter vector of the thawing process reduction model from the edge-side online correction module. The spatiotemporal alignment process resamples and aligns each time series data in all aggregated data samples according to a unified global timestamp and a preset standardized time interval. The automatic quality assessment process calculates a quality confidence score for each data sample. The calculation process for the quality confidence score is as follows: First, a physical consistency score is calculated, which is obtained by using the squared difference norm between the new real-time observation data in the data sample and the predicted observations based on the corrected high-fidelity spatial state vector, the optimal control command to be executed at the current moment, and the updated key physical parameter vector in the data sample. Second, a distribution typicality score is calculated, which is 1 minus a divergence metric. The divergence metric is used to calculate the difference between the data distribution of the edge-side thawing library corresponding to the data sample in the recent time window and the global data distribution formed by the aggregated data of all edge sides. Finally, the physical consistency score and the distribution typicality score are multiplied to obtain the final quality confidence score of the data sample. This score is used to screen out high-quality data samples that are physically consistent and representative of the operating conditions. Then, the cloud-based collaborative module uses the processed high-quality data samples to globally train and optimize the general high-fidelity model of the thawing process. This training process employs a physical information transfer learning method, where the overall loss function is defined as the sum of the weighted losses of all high-quality data samples. The loss of each data sample consists of two terms added together and multiplied by the quality confidence score of that data sample as a weight. Of these two losses, the first term is a data fitting term, which measures the square of the difference between the observation predicted by the general high-fidelity model of the thawing process based on the corrected high-fidelity spatial state vector input from the data sample and the optimal control command to be executed at the current moment, and the actual new real-time observation data in the data sample. The second term is a physical residual regularization term, which acts on the general high-fidelity model base through a physical residual regularization function. The function, obtained from the predicted internal physical field under the same input, calculates the norm of the discrete residuals of the predicted physical field that violates the preset physical conservation law, and then multiplies it by a positive coefficient used to balance data fitting and physical constraints. The training process adopts a meta-gradient descent optimization strategy, which includes two levels: inner loop optimization and outer loop meta-update. In the inner loop, the model parameters are updated with gradients on multiple subsets of data samples used to simulate different typical working conditions, which are divided from the high-quality data samples after quality assessment. In the outer loop, the combined performance of these multiple model variants updated in the inner loop is evaluated on a validation set independently divided from the high-quality data samples for verifying model performance, and the meta-gradient of the validation loss relative to the hyperparameters of the inner loop optimization process itself is calculated. This meta-gradient is used to dynamically adjust the optimization strategy of the inner loop.

[0012] In a preferred embodiment, the process of updating the preset thawing process reduction model and the multi-objective weighting coefficients used in the model predictive control problem specifically involves: First, the cloud-based collaborative module performs multi-objective Pareto front mining and federated meta-learning based on data samples that have undergone spatiotemporal alignment and quality assessment, in order to optimize the multi-objective weight coefficients used in the model predictive control problem. The specific operation of the multi-objective Pareto front mining is as follows: each data sample is regarded as a candidate control strategy generated by the edge-side predictive control module in actual operation, and the performance value of the strategy on multiple preset performance objectives is calculated. The performance objectives include predicted energy consumption, predicted temperature uniformity, predicted control stability, and predicted thermodynamic irreversibility. Subsequently, in the multi-objective space composed of these performance objectives, the Pareto front is identified through clustering algorithm and non-dominated ranking algorithm. That is, the set of elite strategies that are not surpassed by any other candidate strategies on all performance objectives. The specific operation of the federated meta-learning is as follows: a policy meta-network with trainable parameters is trained. This network takes as input a macroscopic working condition feature vector extracted from the corrected high-fidelity spatial state vector in the data samples, and outputs a recommended multi-objective weight coefficient vector. The training process aims to minimize an overall loss function consisting of two terms: the first term is the policy distance term, which calculates the expected value of the norm square of the difference between the recommended weight coefficients output by the policy meta-network and the corresponding weight coefficients extracted or inversely derived from the elite policy set of the Pareto front; the second term is the output smoothing term, which calculates the variance of the output weights of the policy meta-network under similar input features and multiplies it by a preset regularization coefficient. Using the federated learning framework, each edge calculates the gradient update of the policy meta-network locally based on its own data samples, and only the encrypted gradients are uploaded to the cloud for secure aggregation, thereby updating the parameters of the policy meta-network. Then, the cloud collaboration module performs versioning and differentiated secure distribution of models and strategies. This operation first includes model distillation and lightweighting: the general high-fidelity model of the thawing process, which has been globally trained and optimized, is used to generate a customized, lightweight new-generation thawing process reduced-order model through knowledge distillation technology. This new-generation reduced-order model includes a new basis function matrix and corresponding dynamic equation parameters. At the same time, using the optimized policy meta-network, a set of contextualized multi-objective weight coefficient lookup tables is generated for the edge-side thawing library. Secondly, versioning and compatibility management are implemented: a unique version number is assigned to each customized model update package and multi-objective weight coefficient lookup table, and its compatibility mapping relationship with the hardware configuration and software version of each edge side is recorded. Finally, through an encrypted communication channel, the customized model update package and weight coefficient lookup table are differentially distributed to the corresponding edge-side data assimilation module, prediction control module, and online correction module.

[0013] The beneficial effects of this invention are as follows: This scheme reconstructs a high-precision three-dimensional physical field within the thawing chamber in real time by assimilating and fusing sparse observations and physical models, significantly improving the completeness and accuracy of state perception. Furthermore, based on this overall state, multi-objective predictive control is rolled and optimized, minimizing energy consumption while ensuring thawing uniformity and process stability. The online correction mechanism continuously and adaptively adjusts model parameters, effectively combating model drift and environmental disturbances, maintaining long-term control accuracy. Cloud collaboration gathers global operating data, continuously evolves the physical model and optimization strategies, and securely feeds knowledge back to each edge node, forming a synergistic enhancement closed loop of rapid local response and continuous global optimization, ultimately achieving precise, energy-saving, and adaptive high-quality thawing control. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0017] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application. Example

[0018] This embodiment provides, for example Figure 1-2 The present invention discloses an IoT monitoring and early warning system for energy consumption during the thawing of prepared meat products, which includes: a physical sensor network installed in the thawing warehouse to acquire physical quantities of the thawing process, a data assimilation module deployed at the edge of the thawing warehouse, a predictive control module, an online correction module, and a cloud collaboration module for data interaction with each edge. Data assimilation module: In response to real-time observation data collected by the physical sensor network, it calls the preset thawing process order reduction model and combines it with the data assimilation algorithm. It uses real-time observation data to correct the current state output of the thawing process order reduction model, thereby estimating the physical field state of the entire domain in the thawing warehouse at the current moment and outputting the corrected high-fidelity spatial state vector. Predictive control module: In response to the corrected high-fidelity spatial state vector output by the data assimilation module, based on the reduced-order model of the thawing process and the corrected high-fidelity spatial state vector as the initial state, it predicts the physical field evolution, thawing completion and cumulative energy consumption at multiple future time points under different control command sequences. It also generates the optimal control command to be executed at the current time and sends it to the actuator by solving a model predictive control problem that incorporates multi-objective optimization of energy consumption, temperature uniformity and control stability. Online calibration module: In response to the new real-time observation data collected by the physical sensor network at the next moment, the prediction deviation vector between the new data and the prediction value of the physical field at the next moment calculated by the reduced-order model of the thawing process based on the corrected high-fidelity spatial state vector, the optimal control command to be executed at the current moment, and the prediction value of the physical field at the next moment, the module dynamically adjusts the key physical parameter vector contained in the reduced-order model of the thawing process through an online parameter estimation algorithm to reduce the model prediction error, and synchronizes the updated key physical parameter vector of the reduced-order model of the thawing process to the data assimilation module and the prediction control module; The cloud-based collaborative module periodically aggregates calibrated high-fidelity spatial state vectors, the optimal control command to be executed at the current moment, new real-time observation data, and the key physical parameter vectors of the updated thawing process reduction model from various edge sides. Based on big data and machine learning technologies, it uses the aggregated calibrated high-fidelity spatial state vectors, the optimal control command to be executed at the current moment, the new real-time observation data, and the key physical parameter vectors of the updated thawing process reduction model to globally train and optimize the general thawing process high-fidelity model and the multi-objective weight coefficients used in the model predictive control problem. The optimized thawing process high-fidelity model and multi-objective weight coefficients are then sent as update data to the corresponding edge-side data assimilation module, predictive control module, and online calibration module to update their preset thawing process reduction model and the multi-objective weight coefficients used in the model predictive control problem.

[0019] In this embodiment, it is specifically necessary to explain the process by which the data assimilation module corrects the current state output of the thawing process reduction model using real-time observation data: First, the data assimilation module receives real-time observation data reflecting the physical quantities of the thawing process collected at time t from a physical sensor network consisting of temperature sensors, humidity sensors, and wind speed sensors deployed at multiple spatial locations within the thawing chamber. The module performs validity verification, outlier removal, and dimensional normalization on the received real-time observation data to form a standardized observation vector for time t. Validity verification includes determining whether the data is within a physically reasonable range, such as temperature values ​​between -5°C and 20°C, humidity values ​​between 10% and 100% relative humidity, and wind speed values ​​between 0 and 5 m / s. Data points outside these ranges are considered outliers and are removed. Dimensional normalization uses a minimum-maximum scaling method to map the original readings of different sensors to the [0,1] interval to eliminate dimensional influence. Simultaneously, a pre-set thawing process order reduction model is invoked. This model is defined by a basis function matrix consisting of a set of orthogonal basis functions, which are pre-deployed from the cloud collaborative module via the intrinsic orthogonal decomposition method or pre-stored, and are used to characterize the main spatial distribution features of the physical field within the thawing chamber. The intrinsic orthogonal decomposition method performs statistical analysis on a large amount of snapshot data of the flow field and temperature field within the thawing chamber obtained from high-fidelity computational fluid dynamics simulations, and extracts the top few modes with the highest energy proportions as orthogonal basis functions. These basis functions can capture the main spatial coherence structure of the physical field within the thawing chamber. The columns of the basis function matrix are composed of these orthogonal basis function vectors. The thawing process's order reduction model performs state evolution calculations in a low-dimensional space spanned by orthogonal basis functions. It expresses the high-dimensional state of the entire thawing region, used to generate the final corrected high-fidelity spatial state vector, as a linear combination of the basis function matrix and a time-varying low-dimensional coefficient vector. This means that the complete, high-dimensional physical field state can be approximated with high precision through a linear combination of a few low-dimensional coefficient vectors and a fixed basis function matrix, thus greatly reducing real-time computational complexity. The data assimilation module utilizes the low-dimensional coefficient vector of the optimal analysis field from the previous assimilation cycle, which serves as the optimal estimate from the previous time step, to perform the order reduction model of the thawing process. The dynamic equations of the model are calculated to obtain the low-dimensional coefficient vector of the background field at the current time, which is used to characterize the prior estimate of the physical field at the current time t. Then, the high-dimensional background physical field state vector at the current time is obtained by linear combination of the basis function matrix and the low-dimensional coefficient vector of the background field at the current time. The high-dimensional background physical field state vector at the current time is an intermediate state for generating the corrected high-fidelity space state vector. The dynamic equations of the model of the thawing process are usually a low-dimensional set of ordinary differential equations or state-space equations. The coefficient matrix is ​​obtained by Galerkin projection of the high-dimensional control equations onto the basis function space. This equation describes the evolution of the low-dimensional coefficients over time. After obtaining the current high-dimensional background physical field state vector, the data assimilation module performs a state fusion operation in the low-dimensional space based on regularized variational assimilation, specifically: The data assimilation module constructs a minimization objective function in a low-dimensional space spanned by orthogonal basis functions, and then constructs a state estimation optimization problem based on this minimization objective function to obtain the optimal low-dimensional coefficient vector of the analysis field at time t. The minimization objective function consists of three terms. The first term is the observation fitting term, which measures the weighted difference between the standardized observation vector at time t and the predicted observations extracted from the candidate state vector of the high-dimensional physical field obtained by a linear combination of the low-dimensional coefficient vector and the basis function matrix as optimization variables through an observation operator. The difference is calculated using the inverse of the observation error covariance matrix as weights. The observation error covariance matrix is ​​usually set as a diagonal matrix, and the elements on its diagonal are the estimated values ​​of the measurement error variance of each sensor. These estimates can be obtained based on the sensor's factory accuracy calibration data and long-term operation statistics. For example, the error variance of the temperature sensor is set to 0.25 (°C)², the humidity sensor to 4 (%RH)², and the wind speed sensor to 0.01 (m / s)². This observation error covariance matrix is ​​a diagonal matrix used to characterize the uncertainty of the sensor's measurement error. The second term is the background constraint term, which measures the weighted difference between the low-dimensional coefficient vector used as the optimization variable and the low-dimensional coefficient vector of the background field at the current moment. The weighted difference is calculated using the inverse of the background error covariance matrix as the weight. This background error covariance matrix is ​​a matrix obtained in low-dimensional space based on historical forecast error statistics or flow-dependent methods, used to characterize the uncertainty of the reduced-order model forecast error during the thawing process. The background error covariance matrix reflects the uncertainty of the reduced-order model forecast. One method is to use error samples between the background field and the final analysis field in the historical assimilation period for statistical estimation. Another "flow-dependent" method can use ensemble Kalman filtering techniques to estimate its covariance in real time by maintaining a set of low-dimensional coefficient vectors. This matrix is ​​usually not diagonal, reflecting the correlation between forecast errors of different modes in low-dimensional space. The third term is the physical regularization term, which is used to introduce prior physical constraints into the candidate state vector of the high-dimensional physical field. This term is obtained by weighting the output of a physical constraint regularization function through an adjustable regularization parameter. The regularization parameter is a positive scalar whose value is determined through cross-validation or empirically, and its typical value ranges from 0.01 to 1.0. It is used to balance the strength of the fitting of the observation data and the constraint of the smoothness of the physical field. The larger the value of the regularization parameter, the stronger the requirement for the smoothness of the physical field. The physical constraint regularization function acts on the candidate state vector of the high-dimensional physical field. Its function construction penalizes spatial fluctuations in the physical field that do not conform to predetermined physical laws. Its form is the squared L2 norm of the vector obtained by multiplying a discrete Laplace operator matrix by the candidate state vector of the high-dimensional physical field. The discrete Laplace operator matrix is ​​constructed based on the computational grid of the thawing library (such as the grid of the finite volume method or finite element method) and is used for the Laplace operator (second-order spatial derivative) for numerically calculating the distribution of the physical field. This term penalizes drastic and discontinuous spatial changes in the physical field, because the actual fluid temperature and velocity fields are usually spatially smooth. This suppresses non-physical spurious fluctuations caused by observation noise or model errors, and improves the rationality of state estimation in sensor sparse regions. By solving the state estimation optimization problem, the optimal low-dimensional coefficient vector of the analysis field is obtained under the joint constraints of the standardized observation vector, the low-dimensional coefficient vector of the background field at the current time, and the physical constraint regularization term. The state estimation optimization problem is an unconstrained or convex optimization problem with simple boundary constraints. It is usually solved efficiently using numerical optimization algorithms such as gradient descent, conjugate gradient, or quasi-Newton methods because the variable dimension of the low-dimensional space is much lower than that of the original high-dimensional space, ensuring the feasibility of real-time solution. After obtaining the low-dimensional coefficient vector of the optimal analysis field, the data assimilation module performs optimal state reconstruction and output operations, specifically: The data assimilation module performs matrix multiplication on the low-dimensional coefficient vector of the optimal analysis field and the basis function matrix of the reduced-order model of the thawing process, reconstructing a high-dimensional physical field state vector at time t, which serves as the corrected high-fidelity spatial state vector. This matrix multiplication is the process of mapping the optimal representation of the low-dimensional space back to the original high-dimensional physical space. The resulting corrected high-fidelity spatial state vector is a high-dimensional vector, whose elements correspond to one or more physical quantities (such as temperature, humidity, and velocity components in each direction) at each grid point in the thawing chamber's computational grid. This corrected high-fidelity spatial state vector is directly passed to the predictive control module as the initial state upon which the predictive control module bases its rolling optimization control. This vector provides much richer information than sparse observation data, namely, the physically reasonable and complete three-dimensional temperature, humidity, and airflow velocity distribution inferred at the current moment within the entire thawing chamber. This allows the predictive control module to predict future states based on more accurate initial conditions, thereby formulating more precise and energy-efficient control strategies.

[0020] In this embodiment, it is specifically noted that in the predictive control module, using the corrected high-fidelity spatial state vector as the initial state, the prediction of the physical field evolution, thawing completion, and cumulative energy consumption process under different control command sequences at multiple future time points is as follows: First, the predictive control module receives the corrected high-fidelity space state vector at time t from the data assimilation module. It then projects this corrected high-fidelity space state vector onto a low-dimensional coefficient space spanned by the column vectors of the basis function matrix defined by the thawing process reduction model. This results in a low-dimensional coefficient vector describing the system state at time t. This projection operation is achieved by solving a least-squares problem or directly using the pseudo-inverse of the basis function matrix, ensuring that the representation in the low-dimensional space best approximates the original high-dimensional state. This low-dimensional coefficient vector and the corrected high-fidelity space state vector are correlated through a linear transformation relationship involving multiplication of the low-dimensional coefficient vector with the basis function matrix. The obtained low-dimensional coefficient vector serves as the initial condition for subsequent optimization calculations. Simultaneously, a predictive control time domain is defined, encompassing N consecutive control cycles. The value of N must balance predictive foresight with computational real-time performance, and is typically set based on the dominant time constant of the thawing process, for example, between 20 and 50 control cycles. The control cycle length is usually consistent with the data sampling period, such as 10 to 60 seconds. The control command vector for each control cycle within this predictive control time domain is parameterized. Each control command vector is a real-number vector containing multiple components, each corresponding to the setpoint of a different actuator within the thawing chamber. The actuators include at least a fan for regulating wind speed, a valve for regulating refrigerant flow, and a device for regulating heating power. The control command vector within each control cycle is considered as a constant decision variable to be optimized, remaining unchanged within the cycle. All these decision variables together constitute a complete sequence of control commands spanning the entire predictive control time domain. The specific process of generating optimal control commands by solving a model predictive control problem incorporating multi-objective optimization is as follows: Based on the dynamic evolution equation of the reduced-order model of the thawing process, the low-dimensional coefficient vector obtained by projection is used as the initial condition, and the parameterized control command sequence is used as the input. In the digital domain, the complete evolution trajectory of the physical field in the thawing library, the thawing completion index calculated based on the physical field, and the cumulative energy consumption calculated by the correlation between the internal state of the model and the energy flow are predicted. The model predictive control problem is constructed as a constrained nonlinear optimization problem. Its decision variables are parameterized control command sequences, its objective function is a multi-objective composite cost function to be constructed, and its constraints include at least the dynamic constraints described by the dynamic evolution equation of the reduced-order model of the thawing process, the physical upper and lower bound constraints of each component of the control command vector, and the process safety constraints of the thawing process. To evaluate the merits of different control command sequences, a multi-objective composite cost function is constructed. This function consists of a weighted sum of four cost terms. Each of the four cost terms in the multi-objective composite cost function contains adjustable parameters. These adjustable parameters together constitute the multi-objective weight coefficients used in the model predictive control problem. The multi-objective weight coefficients specifically include the following set of trainable or optimizable parameters: all elements of the positive semi-definite weight matrix applied to the energy efficiency cost term, the global scaling factor or internal weight elements of the discrete spatial gradient operator matrix applied to the uniformity cost term, the stationarity weight coefficient in the control stationarity cost term, and the thermodynamic cost weight coefficient in the thermodynamic irreversibility cost term. The initial values ​​of the weight coefficients for each term are determined by domain experience or offline simulation and can be dynamically updated by the cloud-based collaborative module. The first term is the energy efficiency cost term, which is represented by the weighted inner product of the low-dimensional coefficient vectors at each moment in the prediction time domain under a pre-set positive semi-definite weight matrix used to correlate the rate of change of system energy. This positive semi-definite weight matrix is ​​obtained by identifying the mapping relationship between the state space of the reduced-order model and the system energy consumption (such as compressor power and heater power integral). This representation method indirectly reflects the trend of system energy consumption. The second term is the uniformity cost term, which is characterized by the sum of squared L2 norms of the spatial gradients of the high-dimensional temperature field reconstructed from the low-dimensional coefficient vector and the basis function matrix at each time point in the prediction time domain. It is used to measure the uniformity of thawing. The gradient calculation is implemented through a preset discrete spatial gradient operator matrix based on the thawing library computing grid. The structure of the discrete spatial gradient operator matrix (such as the difference scheme) is preset, but its overall contribution to the multi-objective composite cost function is adjusted by a global scaling factor, or the gradient components in different spatial directions can be assigned different weight elements. These scaling factors or weight elements are part of the multi-objective weight coefficients. The third term is the control stationarity cost term, which is characterized by calculating the sum of squared L2 norms of the changes in adjacent control command vectors in the predicted time domain and multiplying it by a stationarity weighting coefficient used to penalize drastic fluctuations in control actions, in order to protect the actuator. The fourth term is the thermodynamic irreversible cost term. It uses a thermodynamic cost weighting coefficient to perform a weighted summation of the predicted thermodynamic entropy production rate calculated based on the predicted temperature field gradient distribution and flow velocity field distribution at each time point in the prediction time domain. The predicted thermodynamic entropy production rate is calculated based on the local equilibrium assumption and is obtained by numerically integrating the predicted temperature gradient field and velocity gradient field using the classical heat transfer entropy production and flow dissipation entropy production formulas. This term is used to guide the process towards the path optimization that minimizes irreversible losses from a thermodynamic perspective. The model predictive control problem is to find a sequence of control commands that minimizes a multi-objective composite cost function while satisfying all constraints. To satisfy constraints including model dynamic equations, physical upper and lower bounds of each component of the control command vector, and process safety constraints during the thawing process (which at least include the maximum permissible temperature difference threshold between different spatial locations within the thawing chamber—the maximum permissible temperature difference threshold is set according to the process requirements of prepared meat products, for example, not exceeding 5°C) to prevent local overheating or overcooling from damaging quality), a sequential quadratic programming method based on sensitivity analysis is used for numerical solution to achieve real-time solving of optimization problems with nonlinear dynamic constraints under edge-side computational resource constraints. Specifically, this includes: First, the dynamic evolution equations of the reduced-order model of the thawing process are used to perform forward integration on the system determined by the initial conditions and candidate control command sequences to obtain the nominal state trajectory prediction over the entire predictive control time domain. Next, using automatic differentiation techniques, in forward or backward modes, the gradient information of the multi-objective composite cost function and all process safety constraint functions with respect to the parameterized control command sequence is calculated. The gradient information characterizes the first-order change sensitivity of the objectives and constraints relative to the control decision. Simultaneously, the approximate Hessian matrix of the multi-objective composite cost function is calculated. The approximate Hessian matrix is ​​typically iteratively updated using quasi-Newton methods such as BFGS or SR1 to approximate the true Hessian matrix while ensuring positive definiteness, avoiding the high cost of directly calculating the second derivative. The approximate Hessian matrix characterizes its second-order curvature. Then, in each optimization iteration, based on the gradient information and the approximate Hessian matrix at the current iteration point, the original nonlinear programming problem is approximately constructed as a problem with a second-order curvature. The problem is a quadratic programming subproblem with a submorphic objective function and linearization constraints. Then, the effective set method or interior point method is used to efficiently solve this subproblem, obtaining the search direction and corresponding step size that reduces the objective function. Finally, the solution is iteratively updated along this search direction, and it is determined whether the current solution meets the preset convergence conditions. Convergence conditions include the objective function decreasing by less than a first threshold, the change in decision variables decreasing by less than a second threshold, or the number of optimization iterations reaching an upper limit. The first threshold can be set according to the magnitude of the cost function, for example, between 1e-4 and 1e-6. The second threshold can be set according to the resolution of the control command, for example, the change in the control quantity is less than 0.1% of its range. The upper limit of the number of optimization iterations is usually set to 20 to 100 times to ensure that the calculation is completed within a single control cycle. If the conditions are not met, the aforementioned steps are repeated until the convergence conditions are met, thus obtaining the control command sequence that minimizes the multi-objective composite cost function, denoted as the optimal control command sequence at time t. Based on the rolling time-domain execution principle of model predictive control, from the optimal control command sequence at time t, only the first control command vector corresponding to the start time t of the current control cycle is extracted as the optimal control command to be executed at the current time and immediately sent to the actuator. At the start time t+1 of the next control cycle, the predictive control module will re-execute the process of obtaining the initial low-dimensional coefficient vector and constructing the control command sequence based on the corrected high-fidelity spatial state vector newly output by the data assimilation module at time t+1, as well as the process of constructing a multi-objective composite cost function, applying constraints, and performing numerical solutions to generate a new optimal control command sequence. The optimal control command corresponding to time t+1 in the new sequence will be extracted again. This process is repeated to form a closed-loop feedback optimization control. This closed-loop mechanism ensures that the control strategy can continuously adapt to the dynamic changes and external disturbances of the thawing process, and use the latest high-precision state estimation to correct future control actions. Thus, while maintaining the computational complexity of the algorithm, continuous optimization control of thawing energy consumption and quality is achieved.

[0021] In this embodiment, it is specifically necessary to explain the process by which the key physical parameter vectors in the thawing process reduction model are dynamically adjusted using an online parameter estimation algorithm in the online correction module: First, in response to the corrected high-fidelity spatial state vector output by the data assimilation module at time t and the optimal control command to be executed at the current time generated by the predictive control module at time t, the online calibration module uses the corrected high-fidelity spatial state vector and the optimal control command to be executed at the current time as inputs to call the thawing process reduced-order model. This model is parameterized by a key physical parameter vector, which is a set of adjustable coefficients within the model. The values ​​of these coefficients directly affect the model's simulation accuracy of physical processes such as heat transfer and flow. The key physical parameter vector contains... Several parameters characterizing the physical properties of heat transfer, mass transfer and flow processes within the thawing chamber include the equivalent convective heat transfer coefficient, turbulent viscosity coefficient, and equivalent thermal diffusivity of the meat. These parameters have clear physical meanings. For example, the equivalent convective heat transfer coefficient comprehensively reflects the intensity of convective heat transfer between the air and the surface of the meat. Its typical initial value is determined by empirical formulas or preliminary experiments, for example, in the range of 5 to 50 W / (m²·K). A forward calculation is performed to obtain a predicted value of the physical field state within the thawing chamber at the next time step, i.e., time t+1, which is denoted as the model predicted physical field at that next time step. Then, in response to the new real-time observation data actually collected by the physical sensor network at time t+1, an observation operator consistent with the observation operator defined in the data assimilation module is used to extract the physical quantity values ​​corresponding to the sensor spatial location from the model-predicted physical field at the next time step, which is recorded as the predicted observation value at time t+1. Finally, the predicted observation value at time t+1 is subtracted from the new real-time observation data, and the difference is the prediction deviation vector at time t+1. This prediction deviation vector not only contains the systematic error caused by the inaccuracy of the model parameters, but may also contain random measurement noise. It is the core data that triggers the online learning of model parameters and is the direct input signal that drives subsequent parameter adjustments. Based on the predicted bias vector, the online correction module performs incremental parameter update operations based on regularized recursive Bayesian learning, specifically: The online calibration module maintains and dynamically updates a parameter estimation error covariance matrix. The initialization of this matrix can be based on the uncertainty of the prior knowledge of the parameters, for example, it can be set as a diagonal matrix, with the diagonal elements being several times the variance of the initial estimated values ​​of each parameter. Simultaneously, the online calibration module dynamically calculates an adaptive observation weight matrix based on the observation variance of each sensor obtained from the historical observation data collected by the physical sensor network and the prediction deviation vector at the current moment. This matrix is ​​a diagonal matrix used to quantify the relative reliability of the observation values ​​of each sensor. The value of each diagonal element of the adaptive observation weight matrix is ​​obtained by weighting and fusing the corresponding component of the historical observation variance of the corresponding sensor and the prediction deviation vector at the current moment through a forgetting factor between 0 and 1. The forgetting factor is used to control the weight balance between historical information and current information, and is usually set between 0.9 and 0.99. For example, 0.95 means that the weight of historical information is about 95%. The historical observation variance of each sensor can be calculated and updated online through a sliding time window (e.g., data from the most recent 24 hours). Based on the prediction bias vector, the key physical parameter vector at time t, the parameter estimation error covariance matrix, and the adaptive observation weight matrix, the online correction module constructs a parameter update objective function. The parameter update objective function is a weighted sum of the following three terms: The first term is the instantaneous fitting term, which is calculated as follows: First, using a parameterized unfreezing process reduction model that explicitly depends on the key physical parameter vector to be determined, based on the corrected high-fidelity spatial state vector at time t and the optimal control command to be executed at the current time, the observation value at time t+1 is predicted. Then, the weighted L2 norm squared of the difference between this predicted observation value and the new real-time observation data at time t+1 is calculated, where the weight is the inverse of the adaptive observation weight matrix. This term forces the parameter update to proceed in a direction that can reduce the prediction bias of the latest step. The second term is the parameter prior constraint term, which is calculated as follows: the weighted L2 norm squared of the difference between the key physical parameter vector to be determined and the key physical parameter vector at the current time t, where the weight is the inverse matrix of the parameter estimation error covariance matrix; this term is based on the Bayesian framework and uses the uncertainty information of the current parameter estimation to regularize the magnitude of the parameter update, so as to prevent unreasonable and drastic jumps in parameters due to noise from a single observation. The third term is the physical feasible region regularization term, which is calculated as follows: a physical feasible region regularization function is evaluated and multiplied by a dynamically adjustable adaptive regularization coefficient. The physical feasible region regularization function is based on the preset reasonable value range of each component in the key physical parameter vector. For example, the reasonable range of the equivalent convective heat transfer coefficient can be set to [1, 100] W / (m²·K), and the reasonable range of the equivalent thermal diffusivity of meat can be set to [1e-7, 5e-7] m² / s. For each component in the updated key physical parameter vector that exceeds its corresponding range, the square of the excess is calculated and accumulated. The adaptive regularization coefficient can be dynamically adjusted according to the changing trend during the parameter update iteration process. For example, when the parameter approaches the boundary multiple times in succession, the coefficient can be appropriately increased to strengthen the constraint. The online calibration module obtains the updated key physical parameter vector by solving an optimization problem that minimizes the objective function of parameter update. This optimization problem can be efficiently solved using nonlinear least squares optimization algorithms such as the Gauss-Newton method and the Levenberg-Marquardt method. Simultaneously, based on the recursive Bayesian learning framework, the online calibration module utilizes a preset forgetting factor (between 0 and 1), a prediction bias vector, and parameterization. The Jacobian matrix, which is the model's sensitivity matrix to the parameters, can be efficiently calculated using automatic differentiation techniques or adjoint methods to calculate the Jacobian matrix of the reduced-order model output regarding the key physical parameter vector during the thawing process. This allows for recursive updating of the parameter estimation error covariance matrix. The update rule reflects the reduction in parameter estimation uncertainty after obtaining new observation data. Furthermore, the forgetting factor enables the algorithm to gradually "forget" the influence of old data, thereby tracking the slow time-varying characteristics of the parameters and reflecting the reduction in parameter estimation uncertainty after incorporating new observation information. Finally, the online calibration module synchronizes the updated key physical parameter vector to the data assimilation module and the predictive control module to replace the corresponding original parameters in the reduced-order model during the unfreezing process within these two modules, thereby completing the online self-calibration of the model parameters. Specifically, the updated key physical parameter vector and the recursively updated parameter estimation error covariance matrix are sent synchronously to the data assimilation module and the predictive control module as update information. This synchronization process is usually implemented through message passing or shared memory mechanisms between modules within the edge computing node to ensure low latency. After receiving the above update information, the data assimilation module replaces the key physical parameter vector contained in its internally preset reduced-order model during the unfreezing process with the received updated key physical parameter vector, and uses the updated parameter estimation error covariance matrix, for example, in subsequent... In the assimilation calculation, the covariance matrix can be used as input for the prior uncertainty of the parameters, improving the accuracy of state estimation. This allows the state estimation process to be based on the updated and more accurate model during the assimilation period at time t+1. Simultaneously, after receiving the updated information, the predictive control module also replaces the key physical parameter vectors contained in its internal thawing process reduction model used for predictive control with the received updated key physical parameter vectors. This ensures that during the control period at time t+1, the predictive model used to construct and solve the model predictive control problem is based on this updated version. This guarantees that the generation of control commands always relies on the model that is closest to the actual process dynamics, thus maintaining excellent control performance even when the model drifts, and generating optimal control commands that rely on more accurate model predictions.

[0022] In this embodiment, the specific process of global training and optimization of the general high-fidelity model for the thawing process and the multi-objective weight coefficients used in the model predictive control problem in the cloud collaboration module is as follows: First, the cloud-based collaboration module performs spatiotemporal alignment and automatic quality assessment on the heterogeneous time-series data that periodically converges from multiple edge sides. The time-series data consists of data packets from each edge side. Each data packet contains a set of data converged from the same edge side at time t. This set of data is defined as a data sample, which includes the corrected high-fidelity spatial state vector from the edge side's data assimilation module, the optimal control command to be executed at the current time from the edge side's predictive control module, new real-time observation data from the edge side's physical sensor network, and the updated key physical parameter vector of the thawing process reduction model from the edge side's online correction module. The spatiotemporal alignment process resamples and aligns each time series data point in all aggregated data samples according to a unified global timestamp and a preset standardized time interval. The standardized time interval is typically aligned with the edge-side control cycle or data reporting cycle, for example, 60 seconds, to ensure that data from different edge sides and at different times are on comparable time bases. The automatic quality assessment process calculates a quality confidence score for each data sample. The calculation process is as follows: First, a physical consistency score is calculated. The preset mathematical function can be an exponential decay function, for example, score = exp(-difference norm squared / normalization constant). The normalization constant is set according to the typical measurement error range of the sensor. Its value is obtained by applying a preset mathematical function to the new real-time observation data in the data sample and the corrected high-fidelity spatial state vector based on the same data sample, the optimal control command to be executed at the current time, and using the updated key physical parameter vector in the data sample, and the squared difference norm between the predicted observations. This mathematical function ensures that the difference... The smaller the value, the higher the score, and the score is mapped between 0 and 1. Next, a distribution typicality score is calculated, which is 1 minus a divergence metric. The divergence metric is calculated using Kullback-Leibler divergence or Jensen-Shannon divergence to quantify the deviation of a single edge-side data distribution from the global distribution. The recent time window is set, for example, the most recent 24 hours. The divergence metric is used to calculate the difference between the data distribution of the edge-side thawing library corresponding to the data sample within the recent time window and the global data distribution formed by the aggregated data from all edge sides. The smaller the divergence metric, the higher the distribution typicality score, and the score is mapped between 0 and 1. Finally, the physical consistency score is multiplied by the distribution typicality score to obtain the final quality confidence score of the data sample. Data samples with a final quality confidence score lower than a preset threshold (e.g., 0.3) can be directly discarded and not participated in subsequent training. This score is used to weight data samples in subsequent model training to select high-quality data samples that are physically consistent and representative of the operating conditions. Then, the cloud-based collaborative module uses the processed high-quality data samples to globally train and optimize the general high-fidelity model of the thawing process. This training process employs a physical information transfer learning method, where the overall loss function is defined as the sum of the weighted losses of all high-quality data samples. The loss for each data sample consists of two terms added together and multiplied by the quality confidence score of that data sample as a weight. The first term of the two losses is a data fitting term, used to measure the general high-fidelity model of the thawing process based on the corrected high-fidelity spatial state vector input from that data sample and the current... The first term is the square of the difference between the observed value predicted by the optimal control command to be executed at any given time and the actual new real-time observed data in the data sample; the second term is the physical residual regularization term, which is obtained by applying a physical residual regularization function to the internal physical field predicted by the general high-fidelity model based on the same input. The physical conservation laws include at least discrete forms of mass conservation, momentum conservation, and energy conservation. The physical residual is the discrete residual of the control equation on the computational grid. This function calculates the norm of the discrete residual of the predicted physical field that violates the preset physical conservation laws, and then multiplies it by a factor used for balance. The positive coefficient between data fitting and physical constraints, determined through cross-validation (e.g., between 0.001 and 0.1), is an adjustable parameter of the loss function. The training process employs a meta-gradient descent optimization strategy, comprising two levels: inner loop optimization and outer loop meta-update. In the inner loop, model parameters are updated with gradients on multiple subsets of data samples selected from high-quality data samples after quality assessment, simulating different typical operating conditions. Each inner loop uses different learning rates or optimizer settings to simulate different adaptation strategies. In the outer loop… In this study, the combined performance of multiple model variants updated by the inner loop is evaluated on a validation set independently partitioned from high-quality data samples for validating model performance. The meta-gradient of the validation loss with respect to the hyperparameters of the inner loop optimization process itself is calculated. The hyperparameters include the learning rate of the inner loop, the gradient pruning threshold, etc. This meta-gradient is used to dynamically adjust the optimization strategy of the inner loop, so that the model can adapt to the multi-condition characteristics of the variable and thawing process more quickly and robustly. This meta-learning mechanism gives the high-fidelity model the "meta" ability to quickly adapt to new conditions or new thawing library configurations. The specific process of updating the pre-defined thawing process reduction model and the multi-objective weighting coefficients used in the model predictive control problem is as follows: First, the cloud collaboration module performs multi-objective Pareto front mining and federated meta-learning based on data samples that have undergone spatiotemporal alignment and quality assessment, in order to optimize the multi-objective weight coefficients used in the model predictive control problem. The specific operation of multi-objective Pareto front mining is as follows: Each data sample is regarded as a candidate control strategy generated by the edge-side predictive control module in actual operation. The performance value of the strategy on multiple preset performance objectives is calculated. The performance objectives include predicted energy consumption, predicted temperature uniformity, predicted control stability, and predicted thermodynamic irreversibility. Then, in the multi-objective space composed of these performance objectives, the Pareto front is identified by clustering algorithm and non-dominated ranking algorithm. The clustering algorithm can be K-means or DBSCAN, which is used to initially group the strategies in the objective space. The non-dominated ranking is used to select the Pareto optimal solution from each group, that is, the set of elite strategies that are not comprehensively surpassed by any other candidate strategy on all performance objectives. The specific operation of federated meta-learning is as follows: A policy meta-network with trainable parameters is trained. This network takes as input a macro-condition feature vector extracted from the corrected high-fidelity spatial state vector in the data samples. Macro-condition features include, but are not limited to, the average temperature inside the thawing warehouse, the temperature standard deviation, the average wind speed, and the external ambient temperature. The output is a vector of recommended multi-objective weight coefficients. The training process aims to minimize an overall loss function consisting of two terms: the first term is the policy distance term, which calculates the difference between the recommended weight coefficients output by the policy meta-network and the corresponding weight coefficients extracted or inversely derived from the elite policy set of the Pareto front. The first term is the expected value of the squared norm of the different inputs; the second term is the output smoothing term, which calculates the variance of the output weights of the policy meta-network under similar input features and multiplies it by a preset regularization coefficient; the regularization coefficient is used to balance the accuracy and stability of weight recommendation and prevent drastic fluctuations in output, and can be set to, for example, 0.01; to protect the data privacy of each edge side and aggregate global wisdom, a federated learning framework is adopted, specifically the FedAvg or FedProx algorithm. Each edge side calculates the gradient update of the policy meta-network locally based on its own data samples, and only uploads the encrypted gradient to the cloud for secure aggregation, thereby updating the parameters of the policy meta-network; Then, the cloud collaboration module performs versioning and differentiated secure distribution of models and strategies. This operation first includes model distillation and lightweighting: the general high-fidelity model of the thawing process, which has been globally trained and optimized, is distilled using knowledge distillation technology. For the specific geometry and configuration of each edge-side thawing library, knowledge distillation adopts a teacher-student learning framework, using the optimized general high-fidelity model as the teacher model and the lightweight reduced-order model to be generated as the student model. By minimizing the KL divergence or mean square error of the outputs of the two, the knowledge of the teacher model is transferred to the student model, generating a customized, lightweight new generation of thawing process reduced-order model. This new generation of reduced-order model includes a new basis function matrix and corresponding dynamic equation parameters. At the same time, using the optimized policy meta-network, the typical working condition features of the thawing library are input into the policy meta-network to obtain the corresponding weight coefficients. These (feature, weight) pairs are stored as a lookup table to generate a set of contextualized multi-objective weight coefficient lookup tables for the edge-side thawing library. Secondly, versioning and compatibility management are implemented: a unique version number is assigned to each customized model update package and multi-objective weight coefficient lookup table. The version number can be in the format of "major version number.minor version number.revision number" and is accompanied by a timestamp. The compatibility mapping relationship between the version number and the hardware configuration and software version of each edge side is recorded. The compatibility check includes verifying whether the API version and computing power of the edge side modules meet the requirements of the new model. Finally, through encrypted communication channels, such as using TLS / SSL protocols, the customized model update package and weight coefficient lookup table are differentially distributed to the corresponding edge-side data assimilation module, predictive control module, and online correction module. After verifying the integrity and compatibility of the data package, these edge-side modules perform verifications including digital signature verification and version compatibility checks. During preset system maintenance windows or low-load periods, they atomically replace their internally pre-built unfreezing process downgraded models and multi-objective weight coefficients. Atomic operations ensure that the update process either succeeds completely or rolls back completely, avoiding inconsistencies in the system during the update process. This completes the transformation and upgrade from cloud-based global knowledge to edge-side local execution capabilities. This closed-loop mechanism achieves continuous collaborative optimization of "edge operation - cloud evolution - edge update," enabling the system to continuously improve the performance of each local node using global data intelligence while protecting data privacy. This is key to improving the long-term adaptability and overall energy efficiency of this invention.

[0023] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0024] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0025] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0026] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0027] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0028] 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 both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0029] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An IoT-based monitoring and early warning system for energy consumption during the thawing of prepared meat products, characterized in that, Specifically, it includes: A physical sensor network set up inside the thawing chamber to acquire physical quantities of the thawing process, a data assimilation module, a predictive control module, an online calibration module, and a cloud collaboration module that interacts with each edge side for data exchange; Data assimilation module: In response to real-time observation data collected by the physical sensor network, it calls the preset thawing process order reduction model and combines it with the data assimilation algorithm. It uses real-time observation data to correct the current state output of the thawing process order reduction model, thereby estimating the physical field state of the entire domain in the thawing warehouse at the current moment and outputting the corrected high-fidelity spatial state vector. Predictive control module: In response to the corrected high-fidelity spatial state vector output by the data assimilation module, based on the reduced-order model of the thawing process and the corrected high-fidelity spatial state vector as the initial state, it predicts the physical field evolution, thawing completion and cumulative energy consumption at multiple future time points under different control command sequences. It also generates the optimal control command to be executed at the current time and sends it to the actuator by solving a model predictive control problem that incorporates multi-objective optimization of energy consumption, temperature uniformity and control stability. Online calibration module: In response to the new real-time observation data collected by the physical sensor network at the next moment, the prediction deviation vector between the corrected high-fidelity spatial state vector, the optimal control command to be executed at the current moment, and the predicted physical field value at the next moment calculated by the thawing process reduction model, the module dynamically adjusts the key physical parameter vector contained in the thawing process reduction model through an online parameter estimation algorithm, and synchronizes the updated key physical parameter vector of the thawing process reduction model to the data assimilation module and the predictive control module. The cloud-based collaborative module periodically aggregates calibrated high-fidelity spatial state vectors, the optimal control command to be executed at the current moment, new real-time observation data, and the key physical parameter vectors of the updated thawing process reduction model from various edge sides. Based on big data and machine learning technologies, it uses the aggregated calibrated high-fidelity spatial state vectors, the optimal control command to be executed at the current moment, the new real-time observation data, and the key physical parameter vectors of the updated thawing process reduction model to globally train and optimize the general thawing process high-fidelity model and the multi-objective weight coefficients used in the model predictive control problem. The optimized thawing process high-fidelity model and multi-objective weight coefficients are then sent as update data to the corresponding edge-side data assimilation module, predictive control module, and online calibration module to update their preset thawing process reduction model and the multi-objective weight coefficients used in the model predictive control problem.

2. The IoT monitoring and early warning system for energy consumption of thawing processed meat products according to claim 1, characterized in that: The specific process by which the data assimilation module corrects the current state output of the reduced-order model during the thawing process using real-time observation data is as follows: First, the data assimilation module receives real-time observation data reflecting the physical quantities of the thawing process collected at time t from a physical sensor network consisting of temperature sensors, humidity sensors, and wind speed sensors deployed at multiple spatial locations within the thawing chamber. The module then performs validity verification, outlier removal, and dimension normalization on the received real-time observation data to form a standardized observation vector at time t. At the same time, a pre-set thawing process reduction model is invoked. The thawing process reduction model is defined by a basis function matrix consisting of a set of orthogonal basis functions that are pre-deployed from the cloud collaborative module through the intrinsic orthogonal decomposition method or pre-stored and used to characterize the main spatial distribution features of the physical field in the thawing chamber. The thawing process reduction model performs state evolution calculations in a low-dimensional space spanned by orthogonal basis functions. The high-dimensional state, which is used to generate the corrected high-fidelity space state vector in the entire thawing library, is expressed as a linear combination of the basis function matrix and a time-varying low-dimensional coefficient vector. The data assimilation module uses the low-dimensional coefficient vector of the previous optimal analysis field, which is the output of the previous assimilation cycle and serves as the optimal estimate of the previous time step, to calculate the current low-dimensional coefficient vector of the background field, which is used to characterize the prior estimate of the physical field at time t, through the dynamic equation of the thawing process reduction model. Then, the current high-dimensional background physical field state vector is obtained through a linear combination of the basis function matrix and the current low-dimensional coefficient vector of the background field.

3. The IoT monitoring and early warning system for energy consumption in thawing prepared meat products according to claim 2, characterized in that: After obtaining the current high-dimensional background physical field state vector, the data assimilation module performs a state fusion operation in the low-dimensional space based on regularized variational assimilation, specifically: The data assimilation module constructs a minimization objective function in a low-dimensional space spanned by orthogonal basis functions, and constructs a state estimation optimization problem based on the minimization objective function to obtain the optimal low-dimensional coefficient vector of the analysis field at time t. The minimization objective function consists of three terms. The first term is the observation fitting term, which measures the weighted difference between the standardized observation vector at time t and the predicted observation value extracted from the candidate state vector of the high-dimensional physical field obtained by a linear combination of the low-dimensional coefficient vector and the basis function matrix as optimization variables through an observation operator. The calculation of this weighted difference uses the inverse matrix of the observation error covariance matrix as the weight. The observation error covariance matrix is ​​a diagonal matrix used to characterize the uncertainty of sensor measurement error. The second term is the background constraint term, which measures the weighted difference between the low-dimensional coefficient vector, which is the optimization variable, and the low-dimensional coefficient vector of the background field at the current moment. The weighted difference is calculated using the inverse of a background error covariance matrix as the weight. This background error covariance matrix is ​​a matrix obtained in low-dimensional space based on historical forecast error statistics, used to characterize the uncertainty of the forecast error of the reduced-order model during the thawing process. The third term is the physical regularization term, which is used to introduce prior physical constraints into the candidate state vector of the high-dimensional physical field. This term is obtained by weighting the output of a physical constraint regularization function through an adjustable regularization parameter. By solving the state estimation optimization problem, the optimal low-dimensional coefficient vector of the analysis field is obtained under the joint constraints of the standardized observation vector, the low-dimensional coefficient vector of the background field at the current time, and the physical constraint regularization term.

4. The IoT monitoring and early warning system for energy consumption of thawing processed meat products according to claim 3, characterized in that: After obtaining the low-dimensional coefficient vector of the optimal analysis field, the data assimilation module performs optimal state reconstruction and output operations, specifically: The data assimilation module performs matrix multiplication on the low-dimensional coefficient vector of the optimal analysis field and the basis function matrix of the reduced-order model of the thawing process, reconstructing the high-dimensional physical field state vector at time t, which serves as the corrected high-fidelity spatial state vector.

5. A IoT monitoring and early warning system for energy consumption in thawing prepared meat products according to claim 4, characterized in that: In the predictive control module, using the corrected high-fidelity spatial state vector as the initial state, the prediction of the physical field evolution, thawing completion rate, and cumulative energy consumption process under different control command sequences at multiple future time points is as follows: First, the predictive control module receives the corrected high-fidelity spatial state vector at time t from the data assimilation module, and projects the corrected high-fidelity spatial state vector into a low-dimensional coefficient space spanned by the column vectors of the basis function matrix defined by the thawing process reduction model, thereby obtaining a low-dimensional coefficient vector to describe the current system state at time t. Meanwhile, a predictive control time domain containing N consecutive control cycles is defined, and the control command vector for each control cycle within this predictive control time domain is parameterized. Each control command vector is a real number vector containing multiple components, which correspond to the set values ​​of different actuators in the defrosting chamber. The actuators include fans that regulate wind speed, valves that regulate refrigerant flow, and devices that regulate heating power. The control command vector within each control cycle is regarded as a constant decision variable to be optimized and which remains unchanged within the cycle. All these decision variables together constitute a complete sequence of control commands that spans the entire predictive control time domain.

6. The IoT monitoring and early warning system for energy consumption in thawing prepared meat products according to claim 5, characterized in that: The specific process of generating optimal control commands by solving a model predictive control problem incorporating multi-objective optimization is as follows: Based on the dynamic evolution equation of the reduced-order model of the thawing process, the low-dimensional coefficient vector obtained by projection is used as the initial condition, and the parameterized control command sequence is used as the input. In the digital domain, the complete evolution trajectory of the physical field in the thawing library, the thawing completion index calculated based on the physical field, and the cumulative energy consumption calculated by the correlation between the internal state of the model and the energy flow are predicted. The model predictive control problem is constructed as a constrained nonlinear optimization problem. Its decision variables are parameterized control command sequences, its objective function is a multi-objective composite cost function to be constructed, and its constraints include the dynamic constraints described by the dynamic evolution equation of the reduced-order model of the thawing process, the physical upper and lower bound constraints of each component of the control command vector, and the process safety constraints of the thawing process. To evaluate the merits of different control command sequences, a multi-objective composite cost function is constructed. This function consists of a weighted sum of four cost terms. Each of the four cost terms in the multi-objective composite cost function contains adjustable parameters. These adjustable parameters together constitute the multi-objective weight coefficients used in the model predictive control problem. The multi-objective weight coefficients specifically include the following trainable parameter set: all elements of the positive semi-definite weight matrix applied to the energy efficiency cost term, the global scaling factor of the discrete spatial gradient operator matrix applied to the uniformity cost term, the stationarity weight coefficient in the control stationarity cost term, and the thermodynamic cost weight coefficient in the thermodynamic irreversibility cost term. The first term is the energy efficiency cost term, which is characterized by calculating the weighted inner product of the low-dimensional coefficient vectors at each moment in the prediction time domain under a pre-defined positive semi-definite weight matrix used to correlate the energy change rate of the system. The second term is the uniformity cost term, which is characterized by the sum of squared L2 norms of the spatial gradients of the high-dimensional temperature field reconstructed from the low-dimensional coefficient vector and the basis function matrix at each time point in the prediction time domain. This gradient calculation is achieved through a preset discrete spatial gradient operator matrix based on the thawing library computing grid. Its overall contribution to the multi-objective composite cost function is adjusted by a global scaling factor. The third term is the control stationarity cost term, which is characterized by calculating the sum of squared L2 norms of the changes in adjacent control command vectors in the predicted time domain and multiplying it by a stationarity weighting coefficient used to penalize drastic fluctuations in control actions. The fourth term is the thermodynamic irreversible cost term, which uses a thermodynamic cost weighting coefficient to weight and sum the predicted thermodynamic entropy yield values ​​calculated based on the predicted temperature field gradient distribution and velocity field distribution at each time point in the prediction time domain. The model predictive control problem is to find a sequence of control commands that minimizes a multi-objective composite cost function while satisfying all constraints. To achieve real-time solution of optimization problems with nonlinear dynamic constraints under edge computing resource limitations, a sequential quadratic programming method based on sensitivity analysis is used for numerical solution, satisfying constraints including model dynamic equations, physical upper and lower bounds of each component of the control command vector, and process safety constraints during the thawing process. Specifically, this includes: First, the dynamic evolution equations of the reduced-order model of the thawing process are used to perform forward integration on the system determined by the initial conditions and candidate control command sequences to obtain the nominal state trajectory prediction over the entire predictive control time domain. Next, using automatic differentiation techniques, in forward or reverse modes, the gradient information of the multi-objective composite cost function and all process safety constraint functions with respect to the parameterized control command sequence is calculated. The gradient information characterizes the first-order change sensitivity of the objectives and constraints relative to the control decision. Simultaneously, the approximate Hessian matrix of the multi-objective composite cost function is calculated, which characterizes its second-order curvature. Then, in each... In the next optimization iteration, based on the gradient information and the approximate Hessian matrix at the current iteration point, the original nonlinear programming problem is approximately constructed as a quadratic programming subproblem with a quadratic objective function and linearization constraints. Subsequently, the effective set method is used to solve the quadratic programming subproblem to obtain the search direction and corresponding step size that can reduce the objective function. Finally, iterative updates are performed along the search direction, and it is determined whether the current solution satisfies the preset convergence condition. If it does not satisfy the condition, the aforementioned steps are repeated until the convergence condition is satisfied, thereby obtaining the control command sequence that minimizes the multi-objective composite cost function, denoted as the optimal control command sequence at time t. Based on the rolling time-domain execution principle of model predictive control, only the first control instruction vector corresponding to the start time t of the current control cycle is extracted from the optimal control instruction sequence at time t, and used as the optimal control instruction to be executed at the current time, and sent to the actuator immediately.

7. The IoT monitoring and early warning system for energy consumption in thawing prepared meat products according to claim 6, characterized in that: In the online correction module, the specific process of dynamically adjusting the key physical parameter vectors contained in the thawing process reduction model through an online parameter estimation algorithm is as follows: First, the online calibration module responds to the calibrated high-fidelity spatial state vector output by the data assimilation module at time t and the optimal control command to be executed at the current time generated by the predictive control module at time t. It takes the calibrated high-fidelity spatial state vector and the optimal control command to be executed at the current time as inputs and calls the thawing process reduced-order model. This model is parameterized by a key physical parameter vector, which contains multiple parameters used to characterize the physical properties of heat transfer, mass transfer and flow processes in the thawing chamber, including the equivalent convective heat transfer coefficient, turbulent viscosity coefficient and meat equivalent thermal diffusivity. It performs a forward calculation to obtain a predicted value of the physical field state in the thawing chamber at the next time, i.e., time t+1, which is denoted as the model predicted physical field at the next time. Then, in response to the new real-time observation data actually collected by the physical sensor network at time t+1, an observation operator consistent with the observation operator defined in the data assimilation module is used to extract the physical quantity values ​​corresponding to the sensor spatial location from the model-predicted physical field at the next time step, which is recorded as the predicted observation value at time t+1. Finally, the predicted observation value at time t+1 is subtracted from the new real-time observation data, and the difference is the prediction bias vector at time t+1.

8. The IoT monitoring and early warning system for energy consumption of thawing processed meat products according to claim 7, characterized in that: Based on the predicted bias vector, the online correction module performs incremental parameter update operations based on regularized recursive Bayesian learning, specifically: The online calibration module maintains and dynamically updates a parameter estimation error covariance matrix; Meanwhile, the online calibration module dynamically calculates an adaptive observation weight matrix based on the observation variance of each sensor obtained from the historical observation data collected by the physical sensor network and the prediction deviation vector at the current moment. The value of each diagonal element of the adaptive observation weight matrix is ​​obtained by weighted fusion of the corresponding component in the historical observation variance of the corresponding sensor and the prediction deviation vector at the current moment through a forgetting factor. Based on the prediction bias vector, the key physical parameter vector at the current time t, the parameter estimation error covariance matrix, and the adaptive observation weight matrix, the online correction module constructs a parameter update objective function. The parameter update objective function is a weighted sum of the following three terms: The first term is the instantaneous fitting term, which is calculated as follows: First, using a parameterized unfreezing process reduction model that explicitly depends on the key physical parameter vector to be determined, based on the corrected high-fidelity spatial state vector at time t and the optimal control command to be executed at the current time, the observation value at time t+1 is predicted. Then, the weighted L2 norm squared of the difference between this predicted observation value and the new real-time observation data at time t+1 is calculated. The second term is the parameter prior constraint term, which is calculated by: calculating the weighted L2 norm square of the difference between the key physical parameter vector to be determined and the key physical parameter vector at the current time t. The third term is the physical feasible region regularization term, which is calculated as follows: evaluate a physical feasible region regularization function and multiply it by a dynamically adjustable adaptive regularization coefficient. The physical feasible region regularization function is based on the preset physical reasonable value range of each component in the key physical parameter vector. For each component in the updated key physical parameter vector that exceeds its corresponding range, calculate the square of the excess amount and accumulate it. The online calibration module obtains the updated key physical parameter vector by solving an optimization problem that minimizes the objective function of parameter update. At the same time, based on the recursive Bayesian learning framework, the online calibration module uses a preset forgetting factor, a prediction bias vector, and the Jacobian matrix of the key physical parameter vector output by the parameter unfreezing process reduction model to recursively update the parameter estimation error covariance matrix. Finally, the online calibration module synchronizes the updated key physical parameter vectors to the data assimilation module and the prediction control module to replace the original parameters in the reduced-order model during the unfreezing process within these two modules, thereby completing the online self-calibration of the model parameters.

9. The IoT monitoring and early warning system for energy consumption of thawing processed meat products according to claim 8, characterized in that: The specific process of global training and optimization of the general high-fidelity model for the thawing process and the multi-objective weight coefficients used in model predictive control problems in the cloud collaboration module is as follows: First, the cloud-based collaborative module performs spatiotemporal alignment and automatic quality assessment on heterogeneous time-series data that periodically converges from multiple edge sides. The time-series data consists of data packets from each edge side. Each data packet contains a set of data converged from the same edge side at time t. This set of data is defined as a data sample, which includes a corrected high-fidelity spatial state vector from the edge-side data assimilation module, the optimal control command to be executed at the current time from the edge-side predictive control module, new real-time observation data from the edge-side physical sensor network, and the updated key physical parameter vector of the thawing process reduction model from the edge-side online correction module. The spatiotemporal alignment process resamples and aligns each time series data in all aggregated data samples according to a unified global timestamp and a preset standardized time interval. The automatic quality assessment process calculates a quality confidence score for each data sample. The calculation process for the quality confidence score is as follows: First, a physical consistency score is calculated, which is obtained by using the squared difference norm between the new real-time observation data in the data sample and the predicted observations based on the corrected high-fidelity spatial state vector, the optimal control command to be executed at the current moment, and the updated key physical parameter vector in the data sample. Second, a distribution typicality score is calculated, which is 1 minus a divergence metric. The divergence metric is used to calculate the difference between the data distribution of the edge-side thawing library corresponding to the data sample in the recent time window and the global data distribution formed by the aggregated data of all edge sides. Finally, the physical consistency score and the distribution typicality score are multiplied to obtain the final quality confidence score of the data sample. This score is used to screen out high-quality data samples that are physically consistent and representative of the operating conditions. Then, the cloud-based collaborative module uses the processed high-quality data samples to globally train and optimize the general high-fidelity model of the thawing process. This training process employs a physical information transfer learning method, where the overall loss function is defined as the sum of the weighted losses of all high-quality data samples. The loss of each data sample consists of two terms added together and multiplied by the quality confidence score of that data sample as a weight. Of these two losses, the first term is a data fitting term, which measures the square of the difference between the observation predicted by the general high-fidelity model of the thawing process based on the corrected high-fidelity spatial state vector input from the data sample and the optimal control command to be executed at the current moment, and the actual new real-time observation data in the data sample. The second term is a physical residual regularization term, which acts on the general high-fidelity model base through a physical residual regularization function. The function, obtained from the predicted internal physical field under the same input, calculates the norm of the discrete residuals of the predicted physical field that violates the preset physical conservation law, and then multiplies it by a positive coefficient used to balance data fitting and physical constraints. The training process adopts a meta-gradient descent optimization strategy, which includes two levels: inner loop optimization and outer loop meta-update. In the inner loop, the model parameters are updated with gradients on multiple subsets of data samples used to simulate different typical working conditions, which are divided from the high-quality data samples after quality assessment. In the outer loop, the combined performance of these multiple model variants updated in the inner loop is evaluated on a validation set independently divided from the high-quality data samples for verifying model performance, and the meta-gradient of the validation loss relative to the hyperparameters of the inner loop optimization process itself is calculated. This meta-gradient is used to dynamically adjust the optimization strategy of the inner loop.

10. The IoT monitoring and early warning system for energy consumption of thawing processed meat products according to claim 9, characterized in that: The specific process of updating the pre-defined thawing process reduction model and the multi-objective weighting coefficients used in the model predictive control problem is as follows: First, the cloud-based collaborative module performs multi-objective Pareto front mining and federated meta-learning based on data samples that have undergone spatiotemporal alignment and quality assessment, in order to optimize the multi-objective weight coefficients used in the model predictive control problem. The specific operation of the multi-objective Pareto front mining is as follows: each data sample is regarded as a candidate control strategy generated by the edge-side predictive control module in actual operation, and the performance value of the strategy on multiple preset performance objectives is calculated. The performance objectives include predicted energy consumption, predicted temperature uniformity, predicted control stability, and predicted thermodynamic irreversibility. Subsequently, in the multi-objective space composed of these performance objectives, the Pareto front is identified through clustering algorithm and non-dominated ranking algorithm. That is, the set of elite strategies that are not surpassed by any other candidate strategies on all performance objectives. The specific operation of the federated meta-learning is as follows: a policy meta-network with trainable parameters is trained. This network takes as input a macroscopic working condition feature vector extracted from the corrected high-fidelity spatial state vector in the data samples, and outputs a recommended multi-objective weight coefficient vector. The training process aims to minimize an overall loss function consisting of two terms: the first term is the policy distance term, which calculates the expected value of the norm square of the difference between the recommended weight coefficients output by the policy meta-network and the corresponding weight coefficients extracted or inversely derived from the elite policy set of the Pareto front; the second term is the output smoothing term, which calculates the variance of the output weights of the policy meta-network under similar input features and multiplies it by a preset regularization coefficient. Using the federated learning framework, each edge calculates the gradient update of the policy meta-network locally based on its own data samples, and only the encrypted gradients are uploaded to the cloud for secure aggregation, thereby updating the parameters of the policy meta-network. Then, the cloud collaboration module performs versioning and differentiated secure distribution of models and strategies. This operation first includes model distillation and lightweighting: the general high-fidelity model of the thawing process, which has been globally trained and optimized, is used to generate a customized, lightweight new-generation thawing process reduced-order model through knowledge distillation technology. This new-generation reduced-order model includes a new basis function matrix and corresponding dynamic equation parameters. At the same time, using the optimized policy meta-network, a set of contextualized multi-objective weight coefficient lookup tables is generated for the edge-side thawing library. Secondly, versioning and compatibility management are implemented: a unique version number is assigned to each customized model update package and multi-objective weight coefficient lookup table, and its compatibility mapping relationship with the hardware configuration and software version of each edge side is recorded. Finally, through an encrypted communication channel, the customized model update package and weight coefficient lookup table are differentially distributed to the corresponding edge-side data assimilation module, prediction control module, and online correction module.

Images