Intelligent temperature control system for surimi product storage

Abstract

The present application relates to the technical field of temperature intelligent regulation, and particularly relates to a system for intelligent regulation of storage temperature of surimi products. The technical scheme comprises a multi-point temperature sensing network, a dynamic thermal response modeling unit, a differentiated temperature setting unit, a refrigeration coordination control unit, an abnormal event hierarchical regulation unit, a visual traceability unit and a self-learning optimization unit. The present application predicts the core temperature recovery curve of surimi products by fusing a physical model and a data-driven grey box model, dynamically divides the temperature control area according to the remaining shelf life, sets differentiated temperature targets and allowable fluctuation bands, coordinates and optimizes the compressor frequency, start-stop times and defrosting cycle under the model predictive control framework, performs hierarchical emergency regulation on abnormal events such as library door opening and equipment failure, and regularly adjusts the model parameters and control thresholds online through deep reinforcement learning, thereby reducing the storage temperature fluctuation, reducing the refrigeration energy consumption and equipment start-stop loss.

Description

Technical Field

[0001] This invention relates to the field of intelligent temperature control technology, and in particular to an intelligent temperature control system for storing surimi products. Background Technology

[0002] Surimi products are high-protein, high-moisture seafood products made primarily from fish meat through processes such as pounding, shaping, and heating. Their proteins are prone to denaturation during freezing, leading to decreased elasticity and deteriorated taste. Furthermore, the activity of residual microbial enzymes is extremely sensitive to temperature. Therefore, precise control of freezing storage temperature is a core technical aspect for ensuring the quality of surimi products and extending their shelf life.

[0003] Existing cold storage temperature control systems, when dealing with frozen products like surimi products—high in protein and moisture, with high thermal inertia, and extremely sensitive to temperature fluctuations—fail to achieve accurate prediction based on the product's real-time thermal state, dynamic temperature setting based on differentiated shelf-life requirements, and multi-objective collaborative optimization that comprehensively considers temperature accuracy, equipment wear and tear, and operational energy consumption within a unified control framework. Specifically: the physical model lacks precision, making it difficult to describe the complex nonlinear thermal response within the stack caused by packaging thermal resistance, gap convection, and ice crystal phase change; static constant temperature or zoning strategies cannot dynamically balance temperature targets and allowable fluctuation ranges based on the remaining shelf life of each batch of products; conventional model predictive control cost functions only focus on temperature deviation and power consumption, failing to incorporate mechanical wear and tear caused by frequent compressor start-ups and shutdowns and additional energy consumption introduced by defrosting operations into the unified optimization process; when faced with abnormal disturbances such as moisture intrusion from door openings or sudden refrigeration equipment failures, there is a lack of a closed-loop linkage mechanism for proactive identification, load estimation, and emergency control; furthermore, the control parameters or model parameters are essentially fixed after system commissioning, lacking the ability to autonomously evolve strategies based on long-term operational feedback.

[0004] The aforementioned shortcomings collectively make it difficult for existing systems to achieve a dynamic optimal balance among ensuring the quality and safety of surimi products, minimizing operating energy consumption, and maximizing equipment lifespan. Therefore, this application proposes an intelligent temperature control system for surimi product storage. Summary of the Invention

[0005] The purpose of this invention is to address the problem in the background art that existing systems are unable to achieve a dynamic optimal balance among ensuring the quality and safety of surimi products, minimizing operating energy consumption, and maximizing equipment lifespan, and to propose an intelligent temperature control system for surimi product storage.

[0006] The technical solution of the present invention: an intelligent temperature control system for storing surimi products, comprising:

[0007] The multi-point temperature sensing network consists of a structured sensor array deployed at multiple predetermined locations within the storage space, used to collect three-dimensional spatial temperature field data within the storage facility as well as environmental parameters inside and outside the storage facility. The dynamic thermal response modeling unit receives the output of the multi-point temperature sensing network, runs the gray box model, which integrates the equivalent heat conduction equation of surimi product stacking and the long short-term memory neural network. It takes stacking density, thermal resistance of packaging material, storage time, and external temperature and humidity as inputs and outputs the adaptive prediction value of the core temperature recovery curve of surimi product. The differentiated temperature setting unit, connected to the dynamic thermal response modeling unit, is used to dynamically divide the storage space into several control zones based on the predicted core temperature recovery curve and the remaining shelf life information of each batch of surimi products. It also generates an independent temperature setting target and allowable fluctuation band for each control zone, wherein the control zone nearing its expiration date is assigned a lower temperature setting target and a narrower allowable fluctuation band. The refrigeration coordination and control unit is used to receive the temperature set targets and allowable fluctuation ranges of each control zone, and solve a multi-objective optimization problem with temperature deviation, compressor start-stop frequency and defrost energy consumption as cost functions through a model predictive control framework, and generate coordinated control commands for compressor frequency sequence and defrost cycle. The abnormal event classification and control unit is used to identify the event type and disturbance intensity when abnormal temperature, equipment failure or door opening disturbance is detected, and to trigger the corresponding level of emergency control plan. In the response to door opening disturbance, the unit estimates the heat and humidity intrusion load based on the external temperature and humidity, postpones the defrosting cycle and increases the cooling capacity.

[0008] Optionally, the system may also include: The visual traceability unit receives and stores time-series data on temperature distribution, control instructions, and abnormal events, and generates a heat map of the evolution of the three-dimensional spatial temperature field and a cold chain compliance audit report. The self-learning optimization unit uses a deep reinforcement learning algorithm to construct a policy network with temperature maintenance accuracy, total energy consumption, and product quality loss index as reward signals. It periodically performs online iterative optimization on the parameters of the gray box model, the setting threshold of the differentiated temperature setting unit, and the control parameters of the refrigeration coordination control unit. The strategy network of the self-learning optimization unit continuously explores energy-saving and quality-preserving synergistic strategies during operation, such as utilizing low electricity prices for pre-cooling in the time dimension and optimizing airflow organization through shelf layout adjustments in the spatial dimension.

[0009] Optionally, the predetermined locations of the structured sensor array specifically include: the evaporator air outlet and return air outlet, the periphery of the refrigeration unit, the inside of the warehouse door, the middle of the crossbeams of each shelf layer, and the geometric center and surface apex of the surimi product stack.

[0010] Optionally, in the gray box model run by the dynamic thermal response modeling unit, the equivalent heat conduction equation is used to describe the macroscopic heat diffusion process inside the stack, the long short-term memory neural network is used to learn the dynamic residual between the actual measured value, and the structural parameters of the model are periodically updated through an online identification algorithm.

[0011] Optionally, the differentiated temperature setting unit divides the control zone in the following way: the area where the remaining shelf life of products is shorter than a first preset threshold is divided into a strict control zone, the area where the remaining shelf life of products is longer than a second preset threshold is divided into an energy-saving zone, and a wider allowable fluctuation band is generated for the energy-saving zone compared with the strict control zone.

[0012] Optionally, in the multi-objective optimization problem solved by the refrigeration coordination control unit, the constraints include the upper temperature limit of each control zone not exceeding the critical temperature for product quality deterioration, and the physical constraints of compressor frequency and defrost time interval.

[0013] Optionally, when the abnormal event classification and control unit detects a refrigeration equipment failure, it executes an emergency control plan including the following steps: re-solve the multi-objective optimization problem and allocate the refrigeration load of the failure equipment to nearby normal equipment; if the normal equipment cannot meet the load, it activates the backup cold storage device for temperature compensation.

[0014] Optionally, a batch information interface is also included, which is used to obtain the production date, shelf life, packaging density and stacking location information of each batch of surimi products from the warehouse management system and input them into the dynamic thermal response modeling unit and the differentiated temperature setting unit.

[0015] Optionally, the system is deployed based on a cloud-edge collaborative architecture: the dynamic thermal response modeling unit and the self-learning optimization unit are deployed on a cloud server; the multi-point temperature sensing network, the cooling coordination control unit, and the abnormal event hierarchical control unit are deployed on edge computing nodes.

[0016] Optionally, the dynamic thermal response modeling unit utilizes transfer learning technology to use the thermal response characteristic parameters learned from historical batches of surimi products as prior knowledge for the initial model of the new batch of products, thereby shortening the convergence time of online model identification.

[0017] Compared with the prior art, this application includes at least one of the following beneficial technical effects: By employing a gray-box model that integrates physical heat conduction equations and long short-term memory neural networks, the prediction accuracy of the core temperature recovery curve of surimi products stacking is significantly improved, taking into account both physical interpretability and data-driven compensation capabilities.

[0018] Based on real-time predicted temperature and remaining shelf life, the system dynamically divides the product into a strict control zone and an energy-saving zone. This allows for stricter temperature control on products nearing their expiration date and a more relaxed temperature range for products with ample shelf life, thus resolving the contradiction between quality and energy consumption that cannot be balanced by a fixed temperature setting.

[0019] Under the model predictive control framework, temperature deviation, compressor start-stop frequency and defrosting energy consumption are incorporated into the cost function for multi-objective optimization, which reduces equipment mechanical losses and additional energy consumption while ensuring temperature control accuracy.

[0020] When the door is opened and disturbance occurs, the load of damp heat intrusion is estimated by combining the external temperature and humidity and defrosting is delayed. When equipment fails, the load redistribution and cold storage device linkage are automatically executed to form a closed loop of active identification and emergency control, shortening the temperature deviation time.

[0021] By using deep reinforcement learning with temperature accuracy, total energy consumption, and quality loss index as reward signals, the system periodically optimizes model parameters and control strategies online, enabling it to autonomously adapt to seasonal changes, equipment degradation, and batch differences.

[0022] In summary, this invention predicts the core temperature recovery curve of surimi products stacking by integrating physical models and data-driven gray-box models. It dynamically divides temperature control zones based on the remaining shelf life and sets differentiated temperature targets and allowable fluctuation ranges. Under the model predictive control framework, it coordinates and optimizes compressor frequency, start-stop frequency, and defrost cycle. It performs graded emergency control for anomalies such as warehouse door opening and equipment failure. Furthermore, it periodically adjusts model parameters and control thresholds online through deep reinforcement learning, thereby reducing storage temperature fluctuations and lowering refrigeration energy consumption and equipment start-stop losses. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a smart temperature control system for storing surimi products. Detailed Implementation

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0025] Example like Figure 1 This embodiment provides an intelligent temperature control system for storing surimi products. The following details the system architecture, the specific construction and workflow of core components, and the complete working process.

[0026] I. System Overall Architecture and Deployment Method In this embodiment, the system is deployed based on a cloud-edge collaborative architecture to balance the demands of large-scale computing power with the real-time requirements of the control loop. Specifically: Edge layer: Deployed in industrial control computers and programmable logic controllers at the cold storage site, this layer includes a data acquisition front-end for a multi-point temperature sensing network, a refrigeration coordination and control unit, and a hierarchical control unit for abnormal events. These units are responsible for millisecond-level data acquisition, real-time control command issuance, and rapid response to emergency events.

[0027] Cloud Layer: Deployed on a server cluster in a remote data center. It includes dynamic thermal response modeling units, differentiated temperature setting units, visualization and traceability units, and self-learning optimization units. The cloud layer utilizes its powerful computing capabilities for complex model training, solving large-scale optimization problems, and persistent storage and analysis of historical data.

[0028] Data Interface: The batch information interface serves as the system's data entry point. It seamlessly integrates with the warehouse management system via a RESTful API or message queue to obtain real-time information on the production date, initial shelf life, packaging density, packaging material type, and three-dimensional stacking location of incoming batches within the warehouse.

[0029] II. Detailed Structure and Working Mechanism of Each Core Unit 1. A multi-point temperature sensing network, composed of a structured sensor array, is used to collect three-dimensional temperature fields, overcoming the limitation of traditional single-point measurements that cannot reflect the actual temperature distribution of surimi products. The predetermined locations are deployed strictly according to the refrigeration cycle airflow organization and shelf structure, specifically including: Evaporator side: Install an array of 3 thermocouples at the air outlet and return air outlet of the evaporator. Take the median as the effective value to sense the minimum supply air temperature and return air temperature of the refrigeration cycle, so as to calculate the actual effective cooling capacity output by the refrigeration unit.

[0030] Unit perimeter: PT100 temperature probes are installed near the three-phase windings of the compressor body and the oil sump of the crankcase to monitor the ambient temperature of the equipment. This signal serves as an important auxiliary signal in the equipment fault diagnosis algorithm to determine overload or oil shortage.

[0031] Inside the warehouse door: Fast-response thin-film thermistors are installed at three heights on the door frame: 0.5 meters, 1.5 meters, and 2.5 meters from the ground. The sampling frequency is set to 10 Hz. These thermistors are specifically designed to capture the intrusion front of humid and hot air formed at the moment the door is opened and its vertical distribution.

[0032] Shelf layer: A digital temperature sensor node is fixed in the middle of the beam of each shelf layer to form a regular vertical temperature gradient sampling grid.

[0033] Inside the product stack: For each surimi product stack, a probe-type temperature sensor with a food-grade 316L stainless steel shell is used. For regularly shaped stacks, insertion or contact measurements are performed at the geometric center of the stack and at the six surface vertices to directly obtain the thermal core temperature and boundary temperature of the product.

[0034] 2. The dynamic thermal response modeling unit receives the output of the multi-point temperature sensing network and runs a gray box model. This gray box model integrates the equivalent heat conduction equation of surimi product stacking with a long short-term memory neural network. It takes stacking density, packaging material thermal resistance, storage time, and external temperature and humidity as inputs and outputs an adaptive predicted value of the core temperature recovery curve of the surimi product. This unit is one of the intelligent cores of this system, running a gray box model specialized for the thermal behavior of surimi product stacking. Its structure and workflow are as follows: Model structure construction: Physical component (white box): The stacking of surimi products is abstracted as having an equivalent thermal conductivity. Equivalent specific heat capacity and equivalent density A cuboid. According to Fourier's law of heat conduction, in its three-dimensional form... ,in , This represents the temperature at any point within the stack. Indicates time, Indicates temperature The Laplace operator, The thermal diffusivity is used to establish a macroscopic partial differential equation for thermal diffusivity within the stack. The boundary conditions of the equation are the time-series temperature values ​​measured by sensors on the stack surface, and the initial condition is the initial temperature at the center of the stack. This part can explain and describe approximately 80%–85% of the dominant thermal behavior.

[0035] The data-driven part (black box) employs a Long Short-Term Memory (LSTM) neural network. This network has one input layer (receiving a 6-dimensional feature vector), two LSTM hidden layers (128 nodes each), and a fully connected output layer. The 6-dimensional input feature vector includes: stacking density, thermal resistance of the packaging material, storage time, ambient temperature, ambient relative humidity, and the prediction bias of the physical model at the previous time step. The network output is the dynamic residual between the predicted and measured values ​​of the physical model at the current time step. .

[0036] Model fusion and prediction: During system runtime, the physical equations and LSTM network are computed in parallel. At each sampling time, the final core temperature prediction output by the model is... . This represents the temperature value predicted solely by the physical model (heat conduction equation). When predicting future states, given the forecast values ​​of external temperature and humidity for a period of time (obtained from the meteorological service interface) and the cooling strategy plan, the model iteratively solves the physical equation and combines it with LSTM to extrapolate the future residuals, outputting a continuous "adaptive prediction value of the core temperature recovery curve of surimi products".

[0037] Online identification and adaptive update: Model parameters are periodically updated using an online identification algorithm. Specifically, an unscented Kalman filter algorithm is employed, using the equivalent thermal conductivity... Specific heat capacity It is used as a state variable for real-time estimation. Whenever new measurement data arrives, UKF first propagates the state distribution through Sigma point sampling, calculates the Kalman gain, and then uses the measured temperature value to estimate the equivalent thermal conductivity. and specific heat capacity A recursive correction is performed to make the physical model parameters approximate the actual thermal state changes of the stack caused by factors such as ice crystal recrystallization and moisture migration in real time.

[0038] Transfer learning acceleration mechanism: When a new batch of surimi is added to the database, the unit automatically retrieves the historical batch model file with the closest packaging material and stacking density from the model library. Using transfer learning, the weights of the first 6 layers of the old model's LSTM layer are frozen as prior knowledge for initializing the new model, and only the last 2 fully connected adaptation layers are randomly initialized. This allows the new batch model to converge quickly without starting from scratch, after inputting a small amount (50 sampling periods) of new batch feature data, reducing the online identification convergence time of the model from several hours to minutes.

[0039] 3. A differentiated temperature setting unit, connected to the dynamic thermal response modeling unit, is used to dynamically divide the storage space into several control zones based on the predicted core temperature recovery curve and the remaining shelf life information of each batch of surimi products. It generates an independent temperature setting target and allowable fluctuation range for each control zone, with control zones nearing their expiration date assigned lower temperature setting targets and narrower allowable fluctuation ranges. This unit breaks away from the traditional constant temperature setting mode, achieving differentiated zone control based on product shelf-life requirements. The specific implementation is as follows: Dynamic control area partitioning logic: Receives remaining shelf life information for each of the N surimi product stacks within the warehouse, obtained from the warehouse management system. , Indicates the first The remaining shelf life (days) of each stack. Two user-configurable thresholds are preset: the first preset threshold... 7 days and a second preset threshold The timeframe is 30 days. The system iterates through all stacks: Remaining shelf life The three-dimensional space grid where the stack is located is marked as "strictly controlled area".

[0040] Remaining shelf life The three-dimensional space grid where the stacks are located is marked as the "energy-saving zone".

[0041] The grid containing stacks between these two types is marked as the "standard zone". The grid is dynamically generated using a Voronoi diagram based on the center position of the stack.

[0042] Differentiated temperature target generation: For each control zone, robustly generate independent temperature target settings based on the most unfavorable scenario (i.e., the fastest recovery point) in the predicted temperature recovery curve of the stacks it contains. For example: Strictly controlled zone: The target temperature is set at -22℃, and the allowable fluctuation range is tightened to ±0.5℃. This can maximally inhibit the growth of ice crystals and the denaturation of proteins during freezing in the surimi. Although the energy consumption of the fan-cooled system is relatively high, it absolutely guarantees the edible quality and safety of products nearing their expiration date.

[0043] Energy-saving zone: The target is set at -18℃, with a permissible fluctuation range of ±2.0℃. This provides the refrigeration system with a huge range of flexibility, allowing the refrigeration unit to operate at a reduced frequency during peak electricity price periods, utilizing the significant thermal inertia of these deeply frozen products to absorb the small amount of heat transferred from the environment.

[0044] 4. A refrigeration coordination and control unit, used to receive the temperature setpoints and allowable fluctuation ranges of each control zone, and solve a multi-objective optimization problem with temperature deviation, compressor start-stop frequency, and defrost energy consumption as cost functions using a model predictive control framework, generating coordinated control commands for the compressor frequency sequence and defrost cycle; this unit uses a model predictive control framework to transform temperature control into a multi-objective optimization problem and solve it online, specifically: Optimize model construction: Predictive model: The output of the dynamic thermal response modeling unit is used as the internal model of the controller.

[0045] Cost function In the finite time domain The value is composed of a weighted sum of three parts.

[0046] in, This represents the total cost function value, which is the objective function to be minimized in the model predictive control framework. This represents the discrete time step index within the prediction time domain. , This represents the total number of steps in the prediction time domain. Indicates the first At a given time, the temperature value predicted by the dynamic thermal response model for a certain control area. Indicates the first At any given time, the target temperature setting value (unit: °C) for this control zone is dynamically generated by the differentiated temperature setting unit based on the product's shelf-life requirements; Indicates the entire prediction time domain The total number of times the compressor changes from a stopped state to a running state; Indicates the first The predicted defrosting energy consumption value at any given time. Temperature deviation penalty weight (extremely high, set to 100). The penalty weight for the number of compressor start-stop cycles is set to medium (10). The defrosting energy consumption penalty weight is relatively low, so it is set to 1.

[0047] Constraints: (1) Safety hard constraint: Temperature of all control regions at any prediction time All temperatures must not exceed the safe critical temperature determined based on the kinetic model of surimi product quality deterioration. (Set to -15℃).

[0048] (2) Physical constraints: frequency of variable frequency compressor The adjustment range is [30Hz, 75Hz]; the shortest running time after a single start is 3 minutes, and the shortest resting time after shutdown is 2 minutes; the minimum time interval between two consecutive defrost starts is 4 hours.

[0049] Online Solving and Rolling Execution: Each control cycle is set to 5 minutes. Using the latest multi-point temperature data and equipment status as initial conditions, the above optimization problem is solved online using sequential quadratic programming or heuristic particle swarm optimization to obtain an optimal compressor frequency sequence for the next 2 hours. And the optimal defrost start sequence. Only the first control action in the sequence is executed, which sets the compressor frequency for the current 5 minutes. When the next 5-minute cycle arrives, the solution is recalculated based on the latest state, forming a rolling time-domain optimization closed loop. This mechanism achieves conflict-free coordination of start / stop frequency, defrost cycle, and energy input for the first time.

[0050] 5. The abnormal event classification and control unit is used to identify the event type and disturbance intensity when abnormal temperature, equipment failure, or door opening disturbance is detected, and to trigger the corresponding level of emergency control plan. Specifically, when responding to door opening disturbance, it estimates the heat and humidity intrusion load based on external temperature and humidity, postpones the defrosting cycle, and increases the cooling capacity. This unit is specifically designed to handle sudden disturbances and prevent quality deterioration. Its identification and response process is as follows: Storage door opening disturbance recognition and response: When the door magnetic sensor signal changes from low to high, the system determines it as a door opening event. The unit immediately combines the data from the external temperature and humidity sensors, and applies the enthalpy difference formula:

[0051] in, This indicates the instantaneous heat load caused by the intrusion of moist heat. This indicates the effective area of ​​the warehouse door opening. This indicates the average airflow velocity at the opening. Indicates air density, The specific enthalpy of ambient air. This indicates the specific enthalpy of the air inside the storage chamber.

[0052] Real-time estimation of instantaneous heat load from damp heat intrusion Size and total enthalpy .when When the preset threshold is exceeded, a level-two warning is triggered, and a composite contingency plan is executed: (1) Force the frequency of the AC fan of the evaporator to be increased from 50Hz to 65Hz, forming a high-speed air curtain at the opening of the warehouse door, physically blocking the intrusion of external hot and humid air.

[0053] (2) Inject a feedforward compensation command into the refrigeration coordination control unit to instantly increase the compressor output cooling capacity target by 30% to offset the intruded heat load.

[0054] (3) Check the future defrosting plan. If the preset defrosting cycle will start within 15 minutes, then force the defrosting to be postponed until the door opening event ends and the internal temperature stabilizes again, so as to avoid the fatal superposition of defrosting heating and external heat load.

[0055] Fault identification and load redistribution of refrigeration equipment: When the system detects that the three-phase current imbalance of a refrigeration unit exceeds 20%, the compressor discharge pressure exceeds the safety threshold, or the suction pressure is too low for 10 seconds, it determines that the unit is faulty. The highest level emergency warning is immediately triggered, and the fault isolation and load redistribution contingency plan is executed. (1) Mark the control variables of the faulty equipment as "unavailable" in the solver and immediately restart the optimization solver of the refrigeration coordination control unit.

[0056] (2) Under the new constraints, the solver recalculates the optimal load allocation and automatically allocates the heat load originally borne by the faulty equipment to the normal refrigeration unit that is adjacent to it in the spatial topology and still has power margin.

[0057] (3) If the solver finds that even if the remaining normal equipment is running at full load (75Hz), it cannot meet the safety critical temperature constraints of all strictly controlled areas, the system will automatically issue a red alarm to the operator without waiting for manual instructions, and at the same time start the backup cold storage unit (containing eutectic salt phase change material) through digital relay linkage to perform emergency temperature compensation by slowly releasing cold.

[0058] 6. The Visual Traceability Unit receives and stores time-series data on temperature distribution, control commands, and abnormal events, generating a 3D thermal map of temperature field evolution and a cold chain compliance audit report. This unit transforms all process data into actionable visual information. It persistently stores historical data of all temperature measurement points, each control command, and abnormal event markers through a time-series database. The front end uses WebGL technology to render a 3D thermal map of the cold storage every half hour, dynamically evolving and playing back along a timeline. During periods of Level 1 or Level 2 warnings, the map automatically highlights these periods with red borders; clicking on them expands the detailed handling log for the event. Simultaneously, it includes built-in report generation templates compliant with audit requirements such as the "National Food Safety Standard for Cold Chain Food Temperature Control Management," allowing for one-click export of formal audit reports containing minute-by-minute temperature compliance rates, over-temperature duration statistics, and closed-loop records of abnormal event handling within a selected time period.

[0059] 7. A self-learning optimization unit employs a deep reinforcement learning algorithm to construct a policy network with temperature maintenance accuracy, total energy consumption, and product quality loss index as reward signals. This network periodically performs online iterative optimization of the parameters of the gray-box model, the setting thresholds of the differentiated temperature setting unit, and the control parameters of the refrigeration coordination control unit. This is a closed-loop mechanism for achieving long-term autonomous energy-saving evolution of the system and eliminating the need for manual experience. An intelligent agent based on a deep deterministic policy gradient algorithm is constructed.

[0060] State space: defined as an 8-dimensional vector, including: the average temperature of each control zone over the past 3 hours, the compressor operating frequency at the current moment, the shelf life distribution, the real-time electricity price for the next hour obtained from the power grid, and the predicted temperature of the external environment.

[0061] Action Space: This agent does not output low-level millisecond-level control quantities, but instead outputs parameter tuning instructions at the policy level, including: the fluctuation bandwidth value of the energy-saving zone in the differentiated temperature setting unit (continuously adjusted between 1.0℃ and 3.0℃), and the weighting coefficients in the cost function of the cooling coordination control unit. and .

[0062] Reward function design: This is a carefully constructed composite function R, calculated once every 4 hours for each control cycle.

[0063] Among them, the quality loss index is calculated offline based on the temperature history data of each stack using a fish surimi myosin freezing denaturation kinetic model based on the Arrhenius equation, accurately quantifying the unexpected loss during shelf life.

[0064] Strategy Optimization and Deployment: The DDPG agent continuously learns through trial and error in a cloud-based simulation sandbox synchronized with a real-time database, guided by the aforementioned states, actions, and rewards. After several weeks of training, it can explore collaborative energy-saving strategies that are difficult for manual maintenance personnel to detect. For example, after learning the pattern of peak electricity prices in the summer afternoon, the strategy will proactively pre-cool the critical area to -24°C (far below the set value) during the low electricity price period in the early morning, thus giving the system a longer thermal inertia buffer time during subsequent high electricity price periods, eliminating the need for frequent compressor starts. The validated optimal strategy network parameters will be deployed to the production environment regularly and smoothly.

[0065] III. Complete System Workflow During a complete operating cycle, the system operates according to the following steps: Data acquisition: The batch information interface obtains detailed information on the surimi products entering the warehouse; the multi-point temperature sensing network continuously scans the three-dimensional temperature field and equipment status of the entire cold storage.

[0066] Modeling and Prediction: The dynamic thermal response modeling unit uses transfer learning to quickly initialize the model of new batches of products and continuously performs online identification, outputting the predicted values ​​of the future core temperature recovery curves of all stacks in real time.

[0067] Decision-making and setting: Based on the forecast results and remaining shelf life, the differentiated temperature setting unit dynamically divides the cold storage into a strictly controlled zone and an energy-saving zone, generating differentiated temperature control targets.

[0068] Execution and Control: The refrigeration coordinated control unit performs multi-objective rolling time-domain optimization based on these objectives, and issues precise coordinated control commands to the compressor, fan and defrost heater.

[0069] Monitoring and Emergency Response: The abnormal event classification and control unit monitors the entire process. Once a disturbance or malfunction occurs, it immediately reconfigures the control problem or triggers emergency equipment to ensure temperature safety.

[0070] Assessment and Audit: The visual traceability unit records all data and generates 3D heat maps and audit reports.

[0071] Learning and Evolution: The self-learning optimization unit learns the best control strategy parameters from historical data and regularly optimizes the system's set thresholds and control weights, enabling the system to automatically adapt to seasonal changes, batch differences, and equipment aging during long-term operation, thereby achieving continuous optimization of energy consumption and quality.

[0072] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A smart temperature control system for storing surimi products, characterized in that, include: The multi-point temperature sensing network consists of a structured sensor array deployed at multiple predetermined locations within the storage space, used to collect three-dimensional spatial temperature field data within the storage facility as well as environmental parameters inside and outside the storage facility. The dynamic thermal response modeling unit receives the output of the multi-point temperature sensing network, runs the gray box model, which integrates the equivalent heat conduction equation of surimi product stacking and the long short-term memory neural network. It takes stacking density, thermal resistance of packaging material, storage time, and external temperature and humidity as inputs and outputs the adaptive prediction value of the core temperature recovery curve of surimi product. The differentiated temperature setting unit, connected to the dynamic thermal response modeling unit, is used to dynamically divide the storage space into several control zones based on the predicted core temperature recovery curve and the remaining shelf life information of each batch of surimi products. It also generates an independent temperature setting target and allowable fluctuation band for each control zone, wherein the control zone nearing its expiration date is assigned a lower temperature setting target and a narrower allowable fluctuation band. The refrigeration coordination and control unit is used to receive the temperature set targets and allowable fluctuation ranges of each control zone, and solve a multi-objective optimization problem with temperature deviation, compressor start-stop frequency and defrost energy consumption as cost functions through a model predictive control framework, and generate coordinated control commands for compressor frequency sequence and defrost cycle. The abnormal event classification and control unit is used to identify the event type and disturbance intensity when abnormal temperature, equipment failure or door opening disturbance is detected, and to trigger the corresponding level of emergency control plan. In the response to door opening disturbance, the unit estimates the heat and humidity intrusion load based on the external temperature and humidity, postpones the defrosting cycle and increases the cooling capacity.

2. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, The system also includes: The visual traceability unit receives and stores time-series data on temperature distribution, control instructions, and abnormal events, and generates a heat map of the evolution of the three-dimensional spatial temperature field and a cold chain compliance audit report. The self-learning optimization unit uses a deep reinforcement learning algorithm to construct a policy network with temperature maintenance accuracy, total energy consumption, and product quality loss index as reward signals. It periodically performs online iterative optimization on the parameters of the gray box model, the setting threshold of the differentiated temperature setting unit, and the control parameters of the refrigeration coordination control unit. The strategy network of the self-learning optimization unit continuously explores energy-saving and quality-preserving synergistic strategies during operation, such as utilizing low electricity prices for pre-cooling in the time dimension and optimizing airflow organization through shelf layout adjustments in the spatial dimension.

3. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, The predetermined locations of the structured sensor array specifically include: the evaporator air outlet and return air outlet, the periphery of the refrigeration unit, the inside of the warehouse door, the middle of the crossbeams of each shelf layer, and the geometric center and surface apex of the surimi product stack.

4. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, In the gray box model run by the dynamic thermal response modeling unit, the equivalent heat conduction equation is used to describe the macroscopic heat diffusion process inside the stack, the long short-term memory neural network is used to learn the dynamic residual between the actual measured value and the model, and the structural parameters of the model are periodically updated through an online identification algorithm.

5. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, The differentiated temperature setting unit divides the control zone in the following way: the area where the remaining shelf life of products is shorter than a first preset threshold is divided into a strict control zone, the area where the remaining shelf life of products is longer than a second preset threshold is divided into an energy-saving zone, and a wider allowable fluctuation band is generated for the energy-saving zone compared with the strict control zone.

6. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, In the multi-objective optimization problem solved by the refrigeration coordination control unit, the constraints include the upper temperature limit of each control zone not exceeding the critical temperature for product quality deterioration, as well as the physical constraints on the compressor frequency and defrost time interval.

7. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, When the abnormal event classification and control unit detects a refrigeration equipment failure, it executes an emergency control plan including the following steps: re-solving the multi-objective optimization problem and distributing the refrigeration load of the failure equipment to nearby normal equipment; If the normal equipment cannot meet the load, the backup cold storage device will be activated for temperature compensation.

8. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, It also includes a batch information interface, which is used to obtain the production date, shelf life, packaging density and stacking location information of each batch of surimi products from the warehouse management system, and input them into the dynamic thermal response modeling unit and the differentiated temperature setting unit.

9. The intelligent temperature control system for storing surimi products according to claim 2, characterized in that, The system is deployed based on a cloud-edge collaborative architecture: the dynamic thermal response modeling unit and the self-learning optimization unit are deployed on the cloud server; the multi-point temperature sensing network, the cooling coordination and control unit, and the abnormal event hierarchical control unit are deployed on the edge computing nodes.

10. The intelligent temperature control system for storing surimi products according to claim 1, characterized in that, The dynamic thermal response modeling unit utilizes transfer learning technology to use thermal response feature parameters learned from historical batches of surimi products as prior knowledge for the initial model of new batches of products, thereby shortening the convergence time of online model identification.

Images