Integrated system for controlling opening and closing of cold storage door and control method

Abstract

The invention relates to the technical field of cold storage door control, and discloses a cold storage door opening and closing control method which comprises the following steps: acquiring multi-mode sensing information related to operation of a cold storage door system; driving a digital twinborn model based on the multi-mode sensing information, and analyzing the current state and predicting the future state of the cold storage door system by using the digital twinborn model; based on the multi-mode sensing information and the analysis and prediction result of the digital twinborn model, a deep reinforcement learning algorithm is adopted to generate a cooperative control instruction for the cold storage door opening and closing motion and the operation of at least one energy-saving device; and executing the cooperative control instruction. Energy flow is accurately simulated and predicted by using a high-fidelity digital twinborn model, and the actual traffic demand is accurately judged in combination with multi-modal perception, so that a deep reinforcement learning agent can generate a cooperative control instruction with the lowest comprehensive energy consumption as a target, and active intelligent energy saving is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of cold storage door control, and in particular to an integrated system and a control method for opening and closing control of cold storage doors. Background Art

[0002] Cold storage, an essential facility in modern logistics, food processing, biomedicine, and other industries, maintains a low-temperature environment to ensure the quality and safety of stored goods. Cold storage doors serve as the primary channel for material exchange between the cold storage and the external environment. The performance of their opening and closing controls directly impacts the cold storage's insulation, energy consumption, operational efficiency, and internal operational safety.

[0003] Traditional cold storage door control methods are relatively simple, relying on manual operation or automatic triggering mechanisms based on single sensors. For example, these methods use infrared sensors, geomagnetic induction coils, or simple pushbutton switches to control the door's opening and closing. While these methods achieve a certain degree of automation, their limitations are becoming increasingly apparent when dealing with complex and changing real-world scenarios.

[0004] Existing technologies still have significant room for improvement in the intelligent control of cold storage door opening and closing. Current control logic often struggles to accurately perceive and understand the true intentions of traffic flow. For example, it cannot effectively distinguish between traffic flow intended for passage and traffic flow merely passing through. This results in unnecessary door opening or excessively prolonged door opening, resulting in significant cooling loss and energy waste. Furthermore, these control systems lack the ability to effectively perceive and adapt to changing environmental factors (such as internal and external temperature differences and humidity fluctuations) and the evolving state of the door itself (such as frost formation, structural deformation, and component wear). Consequently, their preset fixed control parameters struggle to maintain optimal performance over long periods of operation, often leading to low operating efficiency and increased energy consumption.

[0005] Furthermore, current cold storage door control systems generally rely on a passive response when monitoring and maintaining the door's health. These systems lack the ability to provide early warning and lifespan predictions for potential risks in key components such as the door structure, sealing performance, and drive mechanism. Maintenance is typically delayed until a significant failure or performance degradation occurs. This not only impacts the cold storage's normal operations but also increases the costs of unplanned downtime and repairs. Furthermore, existing control strategies fail to adequately optimize the door's motion trajectory to reduce mechanical impact and wear, limiting their contribution to extending the door's service life.

[0006] In terms of system coordination and comprehensive optimization, existing cold storage door controls mostly operate as isolated subsystems, failing to effectively connect and coordinate with other energy-saving devices within the cold storage or the broader building energy efficiency management system. This lack of system-level optimization limits the further exploration of the cold storage's overall energy-saving potential. Furthermore, while basic safety measures (such as anti-pinch sensors) are widely used, existing technologies remain insufficient to address more complex security scenarios, provide higher levels of proactive safety assurance, and adapt to dynamically changing risks.

[0007] Therefore, it is of great practical significance and application value to develop an advanced control method and integrated system that can integrate multimodal perception, intelligent decision-making and collaborative control to achieve efficient, energy-saving, safe and long-life operation of the cold storage door system. Summary of the Invention

[0008] In response to the shortcomings of the existing technology, the present invention provides an integrated system and control method for the opening and closing control of cold storage doors, which solves the problems of low energy utilization efficiency, low operating efficiency, insufficient safety and reliability, lack of predictive maintenance capabilities, and difficulty in adapting to complex and changeable working conditions in the existing cold storage door control technology.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for controlling the opening and closing of a cold storage door, comprising the following steps: Acquiring multimodal perception information related to the operation of the cold storage door system; Driving a digital twin model based on the multimodal perception information, and using the digital twin model to analyze the current state of the cold storage door system and predict its future state; Based on the multimodal perception information and the analysis and prediction results of the digital twin model, a deep reinforcement learning algorithm is used to generate collaborative control instructions for the opening and closing movements of the cold storage door and the operation of at least one energy-saving device; and the collaborative control instructions are executed.

[0010] Preferably, the acquiring of multimodal perception information includes: Acquire target information within the passage area of the cold storage door through at least one millimeter-wave radar sensor, the target information including the position and speed of the target, and determine the target's passage intention based on the target information; The surface temperature distribution of the cold storage door body or door frame is obtained by at least one infrared thermal imager, and the frosting state is identified based on the surface temperature distribution; and the structural deformation information of the cold storage door body or door frame is obtained by at least one laser monitoring device.

[0011] Preferably, the acquiring of multimodal perception information further includes: The operating parameters of the millimeter-wave radar sensor, the infrared thermal imager or the laser monitoring equipment are dynamically adjusted according to the confidence level of the passage intention evaluated based on the target information and the door operation risk level evaluated based on the frosting state and the structural deformation information.

[0012] Preferably, the digital twin model at least includes a mutually coupled thermodynamic model and a kinematic and dynamic model; The use of the digital twin model to analyze the current state of the cold storage door system and predict the future state further includes: evaluating the health status of key components of the cold storage door system and predicting their remaining service life.

[0013] Preferably, it further includes: comparing the simulation output of the digital twin model with the actual sensor measurement value obtained through the multimodal perception information, and dynamically calibrating the key parameters of the digital twin model based on the comparison result.

[0014] Preferably, the generating of collaborative control instructions by using a deep reinforcement learning algorithm includes: Using the multimodal perception information and the analysis and prediction results of the digital twin model as state inputs of the deep reinforcement learning algorithm; The collaborative control instructions output by the deep reinforcement learning algorithm include door motion parameters for controlling the opening and closing movements of the cold storage door and scheduling instructions for regulating at least one energy-saving device; and the deep reinforcement learning algorithm is optimized based on a multi-objective reward function that takes into account energy consumption, traffic efficiency, operational safety and system life.

[0015] Preferably, the use of a deep reinforcement learning algorithm to generate collaborative control instructions includes: using the multimodal perception information and the analysis and prediction results of the digital twin model as state inputs of the deep reinforcement learning algorithm; the collaborative control instructions output by the deep reinforcement learning algorithm include door motion parameters for controlling the opening and closing movement of the cold storage door and scheduling instructions for regulating at least one energy-saving device, and the door motion parameters include at least opening and closing timing, opening amplitude, operating speed curve and holding time; and the deep reinforcement learning algorithm is optimized based on a multi-objective reward function that takes into account energy consumption, traffic efficiency, operation safety and system life.

[0016] Preferably, the at least one energy-saving device comprises a phase change material gradient energy storage module and / or a door body kinetic energy recovery device integrated into the door body or door frame of the cold storage door; The execution of the collaborative control instruction further includes: regulating the working state of the phase change material gradient energy storage module, storing cold energy through solidification phase change when the cold storage door is closed, and absorbing external heat through melting phase change when the cold storage door is open, so as to utilize the latent heat characteristics of the phase change material to buffer the heat exchange during the opening and closing process of the door body; and / or managing the kinetic energy recovery and reuse of the door body kinetic energy recovery device during the deceleration or braking process of the cold storage door body, specifically, during the deceleration or braking process of the door body, driving the door body drive motor to work in the power generation mode, converting the door body kinetic energy into electrical energy and storing it in an energy storage unit, and using the electrical energy stored in the energy storage unit to assist in driving the motor during the next door body start-up or acceleration stage.

[0017] Preferably, further comprising: Collect actual operating performance data of the cold storage door system after executing the collaborative control instructions; the data includes traffic efficiency, energy consumption indicators and manual intervention operation events; and use the actual operating effect data to continuously optimize the control strategy of the deep reinforcement learning algorithm and the model parameters of the digital twin model.

[0018] An integrated system for cold storage door opening and closing control, including: A multimodal perception module, used to obtain multimodal perception information related to the operation of the cold storage door system; a digital twin and simulation module configured to receive the multimodal sensing information, drive a digital twin model, and utilize the digital twin model to analyze the current state of the cold storage door system and predict its future state; An intelligent decision-making and control module is configured to generate coordinated control instructions for the opening and closing movement of the cold storage door and the operation of at least one energy-saving device using a deep reinforcement learning algorithm based on the multimodal perception information and the analysis and prediction results of the digital twin and simulation module; an actuator configured to receive and execute the coordinated control instruction to control the opening and closing of the cold storage door and the operation of the at least one energy-saving device; The integrated system further includes a redundant control module and a mechanical emergency device; when the intelligent decision-making and control module fails, the redundant control module automatically takes over and performs basic opening and closing control; when the system is completely powered off, the mechanical emergency device can be operated to realize manual opening and closing of the cold storage door to ensure the high reliability and safety of the system.

[0019] The present invention provides an integrated system and control method for opening and closing control of cold storage doors. It has the following beneficial effects: 1. This invention uses a deep reinforcement learning algorithm to precisely control the opening and closing of cold storage doors, and coordinates the use of phase-change energy storage modules and kinetic energy recovery devices to minimize cold leakage and unnecessary energy consumption by the drive motor. A high-fidelity digital twin model accurately simulates and predicts energy flow during cold storage door operation. Combined with multimodal perception to accurately assess actual traffic demand, the deep reinforcement learning agent generates coordinated control instructions aimed at minimizing overall energy consumption, achieving proactive and intelligent energy conservation for the cold storage door system.

[0020] 2. The integrated system of the present invention uses real-time multimodal sensing technology (specific measures include but are not limited to millimeter-wave radar) to accurately identify the intention, status, and type of passing targets. This rich sensory information is combined with the digital twin model's rapid prediction of the door's future dynamics to drive the deep reinforcement learning algorithm to dynamically adjust and optimize the cold storage door's opening and closing strategies (such as opening timing, amplitude, and holding time). The key to this invention is the deep integration of real-time sensing, accurate prediction, and intelligent decision-making, which enables the cold storage door's response to accurately match actual traffic needs, thereby effectively shortening unnecessary waiting time, enabling the rapid and smooth passage of people and goods, and improving the traffic volume per unit time and the overall operating efficiency of the cold storage.

[0021] 3. The present invention utilizes multimodal sensing methods such as infrared thermal imaging to monitor frost and laser to monitor structural deformation to obtain key information that may affect the safe operation of the cold storage door in real time, and analyzes the potential impact of this information on the door structure and motion performance through a digital twin model. When formulating a control strategy, the deep reinforcement learning algorithm will take the safe operation status of the system as an important optimization goal. The point of the invention is that through a multi-dimensional, forward-looking risk perception and assessment mechanism, combined with the intelligent risk avoidance decision-making ability of deep reinforcement learning, the cold storage door system can actively avoid potential collision risks, jam risks, and failures caused by frost or deformation, thereby significantly improving the safety of operation and the overall reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is one of the system flow charts of the present invention; Figure 2 This is a schematic diagram of the digital twin model process of the present invention; Figure 3 This is the second system flow chart of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] Please see the attached Figure 1 -Attached Figure 3 The embodiment of the present invention provides an integrated system and control method for opening and closing control of a cold storage door, including: S1. Acquiring multimodal perception information related to the operation of the cold storage door system; S2. Driving a digital twin model based on the multimodal perception information, and using the digital twin model to analyze the current state of the cold storage door system and predict its future state; S3. Based on the multimodal perception information and the analysis and prediction results of the digital twin model, a deep reinforcement learning algorithm is used to generate coordinated control instructions for the opening and closing movement of the cold storage door and the operation of at least one energy-saving device; S4. Execute the collaborative control instruction.

[0025] The following is a detailed description of each step: S1. Acquiring multimodal perception information related to the operation of the cold storage door system This step aims to collect various types of information related to the system's operating status, environmental conditions, and traffic targets in real time and comprehensively through the deployment of multiple sensors around the cold storage door system and on the door body, providing a data basis for subsequent digital twin modeling and intelligent decision-making.

[0026] 1.1 Millimeter-wave radar target perception and intention recognition: Deployment and Scanning: Deploy one or more 77GHz millimeter-wave radar sensors on the inside and outside of the cold storage door or above the door frame. Configure their field of view, scanning period, and waveform parameters to cover the critical access area near the cold storage door. The radar continuously transmits a frequency-modulated continuous wave (FMCW) or other similar waveform and receives reflected signals from the environment and targets.

[0027] Point cloud processing and target detection: The raw radar point cloud data is preprocessed, including noise filtering, such as using the statistical outlier removal (SOR) algorithm, and ground point cloud filtering, such as using the random sampling consensus (RANSAC) algorithm to fit the ground. Subsequently, a density-based spatial clustering algorithm, such as the density-based spatial clustering application with noise (DBSCAN) algorithm, is used to cluster the filtered point cloud and segment it into point cloud clusters representing different potential targets.

[0028] Target Tracking and State Estimation: For each identified target point cloud cluster, its geometric and motion features are extracted. The target state (including 3D position, 3D velocity, and 3D acceleration) is estimated and tracked using a Kalman filter (KF) or its variants, such as the extended Kalman filter (EKF) or the unscented Kalman filter (UKF). Data association algorithms, such as the Hungarian algorithm or the Joint Probabilistic Data Association (JPDA) algorithm, are used to address the data association problem in multi-target tracking.

[0029] Passage intention prediction: Based on the target's real-time motion status, historical trajectory, and prior knowledge of the scene (such as door location, work area division, and common passage paths), a motion model such as a constant velocity (CV) model or a constant acceleration (CA) model is used, or a more complex learning model such as a trajectory prediction and intention classification model based on a hidden Markov model (HMM) or a recurrent neural network (RNN) is used to predict the target's passage intention probability in the short future. For example, it can distinguish whether the target intends to enter the cold storage, leave the cold storage, or simply cross the door area.

[0030] 1.2 Infrared thermal imaging frost status monitoring: Deployment and Temperature Measurement: Uncooled focal plane infrared thermal imagers are deployed in key areas of cold storage doors prone to frost formation, such as around the door frame seal and thermal bridge areas on the door panels. Appropriate emissivity parameters are set based on the material properties of the surface being measured. The thermal imager captures infrared radiation from the surface in real time and converts it into a two-dimensional temperature distribution map.

[0031] Image Processing and Frost Identification: Preprocess the acquired thermal images, such as performing noise filtering like median filtering. This can be done using temperature thresholds, such as identifying pixels below a specific dew point or a preset critical frost temperature as potential frost areas. Alternatively, an edge detection algorithm (such as the Canny operator) can be combined with a region growing algorithm to segment out abnormally low temperature areas within the image, indicating potential frost areas.

[0032] Frost Quantification: Count the number of pixels in the segmented low-temperature anomaly area and convert this to the actual frost coverage area based on camera calibration parameters (which define the actual size represented by a unit pixel). The severity of frost can be preliminarily assessed by analyzing the average or minimum temperature in the low-temperature area, combined with the ambient temperature and humidity.

[0033] Quantification and classification of frost conditions: We have defined a quantitative assessment process for frost conditions. First, we use infrared image processing technology to accurately calculate the actual frost coverage area. Then, combined with the average temperature of the frosted area (which indirectly reflects the thickness of the frost layer), we classify the severity of frost into at least three actionable risk levels: mild, moderate, and severe. For example: Light frost: The frosted area is less than 5% of the total area, and the average temperature is not significantly lower than the warehouse temperature. It may be just a thin frost.

[0034] Moderate frost: The frosted area is between 5% and 20%, affecting the sealing of the door and posing a potential operational risk.

[0035] Severe frost: Frost area is greater than 20%, or critical locations (such as guide rails and sealing strips) experience severe low temperatures, which may cause the door to become stuck or damaged.

[0036] Control strategy: This risk level isn't just an isolated metric; it's a crucial input to intelligent control systems. We've clearly stated in the invention summary that "the deep reinforcement learning algorithm considers the risk level as part of the state input when generating collaborative control instructions, and adopts differentiated control strategies based on the risk level." In case of light frost: the system may only record the status and optionally send a low-priority maintenance reminder to the operation and maintenance personnel.

[0037] During moderate frost, the DRL algorithm automatically adjusts speed and torque limits when optimizing door movement, achieving a more gentle operation and preventing damage to the drive mechanism due to excessive frost resistance. Simultaneously, the system automatically triggers linked heating devices (such as door frame heaters) for active defrosting, and provides this information to the digital twin model to predict future risk developments.

[0038] In the event of heavy frost, the system will identify the situation as high-risk. The DRL algorithm may prioritize safety, for example by limiting the maximum door opening range or, in some extreme cases, refusing to execute the opening command. It will also immediately issue a high-level alert to the backend management system and operations personnel, requiring manual intervention and inspection, thus completely avoiding major equipment failures or safety accidents caused by frost.

[0039] 1.3 Laser structure deformation monitoring: Deployment and Benchmark Measurement: Install multi-point laser displacement sensors at key locations on the door, such as the corners and center of the door panel, and key structural points on the door frame. Alternatively, use a fixed, periodic laser scanner to scan the door surface. During system installation and commissioning, measure and record the initial 3D coordinates or point cloud model of the door and door frame in ideal, no-load, uniform temperature conditions as a benchmark.

[0040] Real-time deformation measurement: The laser displacement sensor periodically outputs the displacement value of each monitoring point relative to the reference position. The laser scanner periodically obtains the real-time 3D point cloud of the door surface.

[0041] Deformation analysis: For multi-point laser displacement sensors, the system directly obtains the precise displacement of each measuring point.

[0042] For laser scanners, the system uses a point cloud registration algorithm (such as the iterative closest point ICP algorithm) to align and compare the point cloud collected in real time with the pre-stored digital reference model, thereby calculating the deviation vector of each corresponding point or area.

[0043] Based on this discrete position deviation data, the system generates a current overall deformation model of the door body by constructing a deformation field or employing surface fitting methods. Analysis of this model accurately quantifies the overall macroscopic deformation patterns and magnitudes of the door body, such as overall warpage, torsion angle, and changes in flatness. These quantitative deformation metrics serve as key indicators for assessing the door body's structural health and operational risks.

[0044] Deformation severity classification: The laser monitoring system obtains accurate and quantified door deformation data. Based on this data, the system compares it with the multi-level safety thresholds preset in the digital twin model in real time to automatically classify the severity of the deformation. For example: Normal: The door deformation is within the design tolerance. This is usually caused by minor temperature changes or normal operating stress and does not affect safety and sealing.

[0045] Warning level: Deformation values exceed the normal range but have not yet reached the danger threshold. This may indicate early problems such as loose bolts, local material fatigue or slight misalignment of the guide rail.

[0046] Danger level: The deformation value exceeds the standard seriously, indicating that there is a significant risk of the door body getting stuck, derailing or serious sealing failure.

[0047] Dynamic linkage with control strategies: This classification result is immediately used as an important basis for deep reinforcement learning (DRL) decision-making. The system will adopt completely different control strategies based on different levels, achieving a leap from "passive response" to "active management": At Normal level: The system operates normally according to the optimal efficiency and energy consumption strategy. All deformation data is continuously recorded and used to analyze long-term health trends.

[0048] At the alert level, the DRL algorithm automatically switches to a "protective operating mode." It proactively reduces the door's maximum operating speed and acceleration / deceleration to minimize impact and stress on the deformed structure during operation, preventing the problem from worsening. Simultaneously, the system sends a medium-priority maintenance recommendation to the backend, prompting operators to conduct an inspection at their convenience, truly achieving "predictive maintenance."

[0049] At the critical level, safety becomes the top priority. The system immediately enters "Safety Lockout Mode." The DRL will refuse any automatic opening and closing commands and may only allow very low-speed "manual inching" by authorized personnel for inspection. Most importantly, the system will immediately trigger the highest level of audible and visual alarms and send an emergency alert to all relevant management personnel, effectively locking the system before a catastrophic failure occurs, ensuring the absolute safety of personnel and cargo.

[0050] 1.4 Other auxiliary perception information collection: Deploy temperature and humidity sensors on both sides of the cold storage door to collect ambient temperature and relative humidity in real time.

[0051] Door status sensors can be integrated, such as using a door magnetic switch to detect the fully closed state of the door, and using an encoder to accurately detect the open position of the door.

[0052] 1.5 Multimodal Perception Data Fusion and Adaptive Scheduling: Data synchronization and alignment: Synchronize heterogeneous data from different sensors based on their precise timestamps. Leverage the pre-calibrated relative pose relationships (extrinsic parameters) between sensors to unify all sensor data into the same world coordinate system or portal coordinate system.

[0053] State fusion: Use appropriate multi-sensor data fusion technologies, such as Kalman filter-based fusion, Bayesian inference, evidence theory (DS) fusion, or task-specific deep learning fusion networks to integrate various processed sensory information (target state, frost state, deformation data, environmental parameters, etc.) into a unified, high-confidence system-level comprehensive state vector.

[0054] Adaptive scheduling based on "intention-risk": Evaluate the confidence level of the passing intention.

[0055] Assess the risk level of door operation.

[0056] Dynamically adjust perception parameters: A central perception and decision-making management module will dynamically adjust the operating parameters of each perception subsystem and the optimization objectives of the DRL algorithm based on the real-time assessment of the "confidence of passage intention" and "door operation risk level".

[0057] For example, when the door operation risk level increases (such as when moderate frost or warning-level deformation is detected), the system will implement a coordinated response strategy: Enhanced perception: Immediately increase the scanning frequency of infrared thermal imagers on frosted areas, or increase the scanning density of laser scanners on deformed areas to obtain more detailed and timely risk evolution data to support the accurate prediction of the digital twin model.

[0058] Active Intervention: At the same time, the system immediately initiates or adjusts appropriate countermeasures. For example, in the event of frost, the system automatically activates or increases the power of the door frame heating wire for active defrosting. In the event of deformation, the alarm system is activated to issue a maintenance warning.

[0059] Adjust decision weights: The optimization objectives of the deep reinforcement learning (DRL) algorithm will change dynamically, temporarily reducing the optimization weight of "traffic efficiency" and instead making "operational safety" and "reducing equipment loss" the primary goals, such as actively reducing the door's operating speed and acceleration.

[0060] On the contrary, when the system runs smoothly and the risk is low, the sampling frequency and data processing load of each sensor can be appropriately reduced, and the optimization goals can be refocused on energy consumption and efficiency to achieve a balance and optimization of the overall operating cost of the system.

[0061] S2. Driving a digital twin model based on the multimodal perception information, and using the digital twin model to analyze the current state of the cold storage door system and predict its future state; This step aims to build and run a high-fidelity digital twin that is synchronized with the physical cold storage door system in real time. This digital twin model uses sensory data acquired by the S1 as input to simulate the system's multi-physics dynamic behavior, enabling in-depth analysis of the system's internal state, prediction of future evolution trends, and assessment of the system's health and remaining useful life.

[0062] It is worth emphasizing that the main purpose of evaluating the health status and remaining service life is not to serve as a rigid criterion for whether to execute the door opening command to ensure that access needs can be met at all times, but rather as an advanced predictive maintenance and intelligent adjustment method. Specifically, the evaluation results are mainly used in two aspects: Triggering predictive maintenance and alerts: When the digital twin model identifies a declining health trend for key components (such as motors, reducers, and door panels) or their predicted remaining service life falls below a preset threshold, the system automatically sends different levels of alerts or maintenance reminders to the backend management platform. This enables the operations team to proactively intervene and perform planned maintenance or prepare spare parts, effectively avoiding unexpected failures and reducing repair costs and downtime losses.

[0063] Adjusting Intelligent Control Strategies: This assessment result is fed into the deep reinforcement learning algorithm as a state input. When the health status is poor, the DRL algorithm automatically selects protective operating strategies that minimize component impact and wear when generating control instructions. For example, the system automatically and smoothly reduces the door's maximum operating speed and acceleration and deceleration. The door will still open and close normally to facilitate passage, but its operation will be more gentle. This minimizes component aging and extends its actual service life while ensuring normal operation.

[0064] 2.1 Digital Twin Model Composition: Geometric model: The geometric model accurately describes the key components of the cold storage door, including the door panel, door frame, sealing strip, phase change material (PCM) layer and drive mechanism.

[0065] Physical model: This model includes at least a coupled thermodynamic model and a kinematic and kinetic model, specifically: Thermodynamic model: Control equation: Fourier heat conduction partial differential equation is used to describe the temperature distribution inside the door body, including the PCM layer Changes over time and space: ; in: is the density of the material in kg / m 3 is the specific heat capacity of the material, in J / (kg Degree, the unit is K or ℃; is time, unit is s; is the divergence operator; is the thermal conductivity of the material, in W / (m·K); is the temperature gradient; is the heat source generation rate per unit volume, in W / m 3 .

[0066] Phase Change Material (PCM) Model: Enthalpy of PCM (unit is J / kg) is a function of temperature. For the PCM layer, the internal heat source term The following can be contributed: ; in: is the density of PCM material, in kg / m 3; is the enthalpy of the PCM material in J / kg, which is a function of temperature and includes latent heat in the phase change temperature range; is the rate of change of PCM enthalpy with time.

[0067] Boundary conditions: Convective heat transfer: The convection heat transfer between the inner and outer surfaces of the door and the air follows Newton's law of cooling: ); in: is the thermal conductivity of the door surface material, in W / (m K); Along the surface normal direction The temperature gradient is in K / m; is the convective heat transfer coefficient, unit is W / ( K), this coefficient is related to factors such as door movement speed and air flow state; The temperature of the door surface, in K or °C; is the temperature of the surrounding air in K or ℃.

[0068] Radiative heat transfer: The radiative heat transfer between the door surface and the surrounding environment follows the Stefan-Boltzmann law: ; in is the emissivity of the door surface, which is a dimensionless parameter and ranges from 0 to 1; is the Stefan-Boltzmann constant, which is approximately W / (m 2 ·K 4 ) is the absolute temperature of the door surface, in K; is the average absolute temperature of the surrounding surface, in K.

[0069] Air infiltration and heat leakage: When the door is open or not sealed tightly, the air infiltration through the door gap causes cold or heat exchange, the size of which is related to the geometry of the gap, the pressure difference between the inside and outside, and the air temperature and humidity.

[0070] Kinematic and dynamic models: For a sliding door, its dynamic equation can be simplified as: ; in: is the effective mass of the door body, in kg; For the door body in time Acceleration in m / s 2 ; To drive the motor at time The driving force provided, in N; is the friction force, in N, is the door speed , contact surface temperature and the function of frost condition Frost; Other loads acting on the door body, such as unbalanced gravity or wind pressure, etc., the unit is N; is the air resistance, in N, and is a function of the door speed.

[0071] For a revolving door, its dynamic equation can be simplified as: ; in: The moment of inertia of the door body around the axis of rotation, in kg·m 2 ; For the door body in time Angular acceleration, in rad / s 2 ; To drive the motor at time The driving torque provided is in N·m; is the friction torque, in N·m, is the angular velocity of the door , contact surface temperature and the function of frost condition Frost; is the other load moment acting on the door body, in N·m; is the air resistance torque, in N·m, and is a function of the door angular velocity.

[0072] Multiphysics Coupling Interfaces: Thermally affected movement: The frosting state calculated by the thermodynamic model will update the friction coefficient in the dynamic model in real time. The thermal expansion and contraction of the door due to temperature changes may also affect the clearance between the door and the door frame, further affecting friction or the risk of jamming.

[0073] Motion affects heat: The movement of the door will change the boundary conditions in the thermodynamic model, such as changing the air infiltration rate and updating the convective heat transfer coefficient due to the forced convection enhancement caused by the movement.

[0074] 2.2 Real-time drive and state synchronization of digital twin models: The perception state vector acquired and fused in S1 and the actual feedback signal of the actuator are used as the real-time input or boundary conditions of the digital twin model.

[0075] Coupled physical models are efficiently solved using numerical methods (such as the finite element method (FEM), the finite difference method (FDM), or the finite volume method (FVM). These numerical methods are the core computational engines driving the physical models within the digital twin (such as thermodynamics and structural mechanics models). They use sensor data acquired by S1 as real-time boundary conditions and input loads to solve the partial differential equations that describe the physical behavior of the system. Simultaneously, the model's calculation results are compared and verified with other sensor data acquired by S1 to dynamically calibrate the model parameters, thereby updating the state variables within the digital twin model in real time, keeping them synchronized with the state of the physical entity with high precision.

[0076] 2.3 Status analysis and future prediction based on digital twins: In-depth analysis of the current status: The digital twin model can provide internal state information that is difficult or too costly to measure directly with physical sensors, such as the temperature gradient and cold storage capacity inside the PCM layer, the stress distribution of the door structure under specific loads, and the actual compression state of the sealing strip under deformation and low temperature.

[0077] Prediction of future state evolution: Based on the current system state and predicted external disturbances (such as the ambient temperature change trend in the future and the predicted traffic target sequence), the digital twin model can be deduced forward to predict the state evolution of the system in the future (such as minutes to hours). For example, it can predict the internal temperature recovery of the door after multiple consecutive openings, the exhaustion time of the PCM, the growth rate of frost, the potential peak of cooling loss, etc.

[0078] 2.4 System health status assessment and remaining useful life (RUL) prediction: This component is one of the core functions of the digital twin model for predictive maintenance and intelligent control. It continuously evaluates the health status of key components and predicts their remaining useful life, providing decision support for smarter and more economical full lifecycle management.

[0079] Build a health index (HI) system: Define a series of quantifiable health indicators for key components of the cold storage door system (such as the drive motor, transmission mechanism, sealing strips, PCM materials, and the door structure itself). These indicators are derived from the simulation output of the digital twin model (such as internal stress, accumulated fatigue value, and phase change cycle number) and actual sensor measurement data (such as motor current and vibration signals).

[0080] Integration of Damage Accumulation and Aging Models: Within the digital twin model, corresponding damage accumulation models or material aging models are integrated for each key component. For example, for transmission mechanisms, a linear fatigue accumulation damage model based on Miner's law can be used; for PCM materials, a performance degradation model based on the number of phase change cycles can be established. These models continuously calculate the degree of component damage accumulation based on the system's operating load history.

[0081] RUL prediction and its relationship with the control system: Based on the currently calculated health indicators and damage accumulation, the system uses data-driven or a combination of physical models and data-driven methods to predict the remaining useful life (RUL) of each key component. The RUL prediction results are directly and closely related to this control system, mainly reflected in the following two aspects: As a direct input for predictive maintenance: The predicted RUL value is the core basis for triggering maintenance alerts at different levels. For example, when the predicted RUL of a component falls below the set "warning threshold," the system automatically generates a preventive maintenance work order; when it falls below the "danger threshold," an emergency replacement alert is issued. This ensures that maintenance activities are based on actual conditions and are prepared, rather than blindly scheduled repairs or costly post-failure repairs.

[0082] As a dynamic regulator of intelligent control strategies: RUL prediction results are an important state input to the deep reinforcement learning (DRL) algorithm. When the system predicts that the RUL of one or more key components is low, the DRL algorithm's built-in multi-objective reward function will automatically increase the weight of the goal of "extending life" when performing decision optimization. This will guide the algorithm to generate a more "mild" protective control strategy. For example, without seriously affecting traffic efficiency, it will actively select smoother acceleration and deceleration curves and avoid extreme speed operation, thereby reducing the impact and wear on aging components. In this way, the system can actively and intelligently extend the service life of the equipment while ensuring its core functions. This is the core manifestation of the correlation between RUL prediction and control systems.

[0083] 2.5 Dynamic calibration and fidelity maintenance of digital twin models: Error monitoring and calibration trigger: Continuously compare the simulation output of the digital twin model with the real-time measurement values of the corresponding physical sensors in S1. It can be expressed as: ; in: For in time The model prediction error; For in time The actual measurement value of the physical sensor; For digital twin models in time Based on model parameters The predicted output value of A set of key adjustable parameters in the digital twin model. When the error between the two continuously exceeds a preset threshold, the model calibration procedure is triggered.

[0084] Parameter identification and optimization: Use system identification techniques or optimization algorithms, such as batch least squares, recursive least squares (RLS), particle swarm optimization (PSO) or genetic algorithm (GA), to adjust key uncertain parameters in the digital twin model (such as material thermophysical properties, convective heat transfer coefficient, friction model parameters, PCM phase change characteristic parameters, etc.) online or offline to minimize the error between model predictions and actual measurements, thereby continuously improving the fidelity of the model.

[0085] S3. Based on the multimodal perception information and the analysis and prediction results of the digital twin model, a deep reinforcement learning (DRL) algorithm is used to generate collaborative control instructions for the opening and closing movement of the cold storage door and the operation of at least one energy-saving device. This step is the core decision-making brain of the entire intelligent control system. It transforms the aforementioned perception and modeling prediction capabilities into specific, optimized action instructions. Its complete closed-loop control logic places the deep reinforcement learning (DRL) algorithm in a continuous optimization cycle of "perception-modeling-decision-execution-feedback" to achieve a dynamic balance between the cold storage system's energy efficiency, traffic efficiency, operational safety, and system lifespan.

[0086] The closed-loop logic of this intelligent control can be summarized as the following interrelated links: State Construction: The DRL agent obtains comprehensive state information from two sources: Direct perception information (from S1): includes the passing intention identified by millimeter-wave radar, the frost level analyzed by the infrared thermal imager, the deformation level measured by the laser scanner, and the real-time position and speed returned by the motor encoder.

[0087] Deep analysis and prediction information (from S2): The digital twin model provides in-depth information that cannot be directly measured, such as stress distribution within the door, health assessment of key components, and remaining useful life (RUL) prediction. These two pieces of information are integrated to form a high-dimensional "state vector" that describes the current internal and external environment of the system.

[0088] Decision Generation (Action): The DRL agent (at its core, a deep neural network) receives the aforementioned "state vector" and, based on its long-term learned strategy, outputs an "action vector" containing multiple coordinated actions. This action vector is the coordinated control instruction, specifically including: Door movement parameters: such as precise opening and closing timing, optimal opening range, optimized operating speed curve and appropriate opening holding time.

[0089] Energy-saving device scheduling instructions: such as controlling the working status of the PCM module, managing the charging and discharging of the kinetic energy recovery device, etc.

[0090] Instruction execution (Execution): The collaborative control instructions generated by S3 are sent to the underlying actuators (such as PLC, inverter, servo drive), accurately driving the cold storage door and related energy-saving devices to complete the specified physical actions.

[0091] Reward Evaluation: After the command is executed, the system will immediately evaluate the effect of this action. The reward function is a comprehensive evaluation system with multiple objectives. It will quantify the benefits or costs of this action in the following aspects: Energy consumption: Estimate the actual energy consumption of this operation using a power meter or model.

[0092] Efficiency: Evaluate whether traffic is flowing smoothly and waiting times are being shortened through timing or radar tracking.

[0093] Safety and Lifespan: Combined with digital twin analysis, the impact of the operation on component health and structural stability is assessed. These evaluations are weighted and summed to produce a scalar "reward value."

[0094] Policy Optimization: The core learning mechanism of the DRL algorithm is to continuously adjust the weight parameters of its internal neural network based on the complete set of "state-action-reward" experiences, using algorithms such as Q-Learning and PolicyGradient. The goal is to ensure that actions with higher expected rewards are taken in similar states in the future.

[0095] Through the repeated cycles of the above five links, the DRL agent can continuously learn and evolve from its interaction with the environment (physical world + digital twin). Its control strategy therefore has a high degree of adaptability and continuous optimization capabilities, forming a complete and intelligent control closed loop.

[0096] 3.1DRL Agent Construction: Algorithm selection: Choose an appropriate DRL algorithm based on the characteristics of the action space (continuous or discrete). For example, for tasks with continuous action spaces, such as those requiring precise control of door speed curves and energy-saving device parameters, algorithms such as Deep Deterministic Policy Gradient (DDPG), Double Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC) can be used.

[0097] State Space Definition: DRL agent at each decision moment Received status input It should fully reflect the current status and future trends of the system, including at least: Multimodal perception information from S1 (such as traffic target status, environmental parameters, frost status, deformation data); Digital twin model analysis and prediction results from S2 (such as precise door status, motor load and temperature, PCM status, cooling loss rate, component health indicators and RUL, load forecast); system operation context information (such as global operation mode, current time).

[0098] Action Space Definition: Output action of a DRL agent It is a control parameter or a group of control parameters used to coordinate the control of door movement and energy-saving devices, including at least: door movement control parameters (such as speed curve parameters, maximum opening amplitude, holding time); scheduling instructions for at least one energy-saving device (such as the start and stop and power setting of the PCM active control unit, or the start / stop threshold and recovery intensity of KERS).

[0099] Multi-objective reward function Design: DRL agent at each decision moment Obtained from the environment Instant rewards It can be a weighted sum of multiple sub-reward items: ; in: For in time Total instant rewards; These are the weight coefficients of each sub-reward, both dimensionless parameters. These weights can be dynamically adjusted based on the current operating mode or system status; This is a sub-reward item related to energy consumption. It is negatively correlated with the power consumption and cooling loss of the drive motor, and positively correlated with kinetic energy recovery. This is a sub-reward item related to traffic efficiency. It is negatively correlated with the traffic waiting time and the total time of a single opening and closing cycle. This is a sub-reward item related to operational safety, which imposes negative penalties on potential collision risks and door movement impact; This is a sub-reward item related to system life, which imposes negative penalties on behaviors that accelerate component aging or damage; This is a sub-reward item related to user comfort. It is optional and positively correlated to the smoothness of door movement.

[0100] 3.2 DRL training and strategy optimization based on digital twins: Build a simulation environment: Encapsulate the calibrated and validated digital twin model in S2 into a simulation platform that conforms to standard reinforcement learning environment interfaces, such as the OpenAlGym interface. Offline training: The DRL agent conducts a large number of rounds of interactive exploration and learning in the constructed digital twin simulation environment. By continuously sampling experience and utilizing the optimization mechanism of the selected DRL algorithm to update the parameters of the neural network, the goal is to maximize the expected cumulative discounted reward: ; in: is the expectation operator; is the discount factor, is a dimensionless parameter, , used to balance the importance of immediate rewards and future rewards; For the next time step During the training process, techniques such as experience replay, target network, and noise exploration can be used to improve the stability and efficiency of learning.

[0101] 3.3 Cross-condition transfer learning and online adaptive adjustment: Build a knowledge base and identify working conditions: For the various typical operating conditions that cold storage doors may encounter (such as temperature differences in different seasons, traffic flow differences between peak and off-peak periods, and specific cargo loading and unloading operation modes), the DRL model parameters and corresponding digital twin model configurations for these conditions can be obtained through pre-offline training or fine-tuning, and stored in the knowledge base.

[0102] Model migration and initialization: When a significant change in the operating conditions is detected, the DRL model and digital twin model parameters corresponding to the existing operating conditions that are most similar to the current new operating conditions are selected from the knowledge base as the initial model under the new operating conditions to accelerate learning convergence.

[0103] Online fine-tuning and rapid adaptation: After deploying the offline trained DRL policy model to the actual cold storage door control system, the system continuously collects real-world interaction data during operation. Leveraging this online data, the DRL model parameters can be fine-tuned online in small batches to accommodate potential model deviations between the real physical environment and the digital twin simulation environment, and to cope with slow changes in the environment or system characteristics.

[0104] S4. Execute the collaborative control instruction This step is responsible for accurately and reliably converting the optimized collaborative control instructions generated by the DRL agent in S3 into the actual actions of the various actuators of the physical cold storage door system, thereby realizing closed-loop control of the door opening and closing process and the operation of the energy-saving device.

[0105] 4.1 Precise execution of door movement: The central controller (such as an embedded industrial computer or PLC) receives the door motion control parameters from the DRL agent.

[0106] The controller interprets and converts these high-level instructions into control signals that the underlying motor driver (such as a frequency converter or servo drive) can understand.

[0107] The motor driver precisely controls the speed and torque of the drive motor (such as a three-phase asynchronous motor or a permanent magnet synchronous motor), and drives the door body through a transmission mechanism (such as a gear rack, chain, or synchronous belt) to smoothly and efficiently open and close the door according to the optimal trajectory planned by the DRL.

[0108] During the movement, encoders and other position or speed sensors feed back the actual movement state of the door to the controller, forming a closed-loop control of position or speed to ensure the accuracy of the movement.

[0109] 4.2 Coordinated operation and control of at least one energy-saving device: Coordinated regulation of phase change material (PCM) gradient energy storage modules: For PCM modules that include active control units, the central controller controls the working status of these units according to the PCM scheduling instructions output by the DRL.

[0110] For purely passive PCM modules, DRL indirectly maximizes the energy-saving benefits of PCM by optimizing the overall opening and closing behavior of the door.

[0111] Coordinated management of the door kinetic energy recovery system (KERS): The kinetic energy recovery system of this invention is designed to capture and reuse the huge kinetic energy generated by heavy cold storage doors during deceleration and braking. Its specific implementation method is as follows: Kinetic Energy Recovery Mechanism: When the door needs to slow down or brake, the servo drive system switches the control command from "drive mode" to "regenerative braking mode." The door's inertia then pulls the motor inward, instantly transforming it into a generator. This process efficiently converts the door's mechanical kinetic energy into electrical energy while simultaneously generating a smooth braking torque for steady deceleration.

[0112] Basic components: A typical kinetic energy recovery unit mainly includes: Servo motor: It can be used as a high-efficiency generator itself.

[0113] Four-quadrant (or regenerative) servo drive: As the core power electronic converter, it can not only transmit electrical energy to the motor, but also recover the electrical energy generated by the motor in regenerative braking mode to its internal DC bus.

[0114] Energy storage unit: Used to store electrical energy recovered from the DC bus.

[0115] Selection of energy storage unit and low temperature adaptability: Considering the low temperature working environment of cold storage, the selection of energy storage unit is crucial.

[0116] Supercapacitor Array: This invention prefers supercapacitor arrays as energy storage units. Their significant advantages include excellent low-temperature performance (maintaining high efficiency even at -40°C), extremely high charge and discharge power density (capable of instantaneously absorbing and releasing large currents), and an extremely long cycle life (up to millions of cycles), making them ideal for applications requiring frequent door opening and closing and requiring short-term, high-power energy recovery.

[0117] Low-temperature battery packs: If a battery pack is used as the energy storage unit, necessary low-temperature protection measures must be implemented to address performance degradation and safety risks in low temperatures. Specific measures include installing the battery pack in a separate compartment with heating and insulation functions, monitoring the cell temperature through a battery management system (BMS), and preheating the battery before charging to ensure safe operation and performance in low-temperature environments.

[0118] DRL Energy Optimization Management: The DRL agent plays a central role in overall energy management. It not only determines the activation of energy recovery during braking but, more importantly, continuously monitors the state of charge (SoC) of the energy storage unit. During the subsequent door-opening acceleration phase, the DRL determines, based on the energy storage status, whether and to what extent to utilize recovered energy to assist in motor starting, thereby minimizing the instantaneous power drawn from the grid. This creates a complete "deceleration-energy storage-acceleration-assistance" energy closed loop, enabling efficient, on-site energy recycling.

[0119] 4.3 System operation effect feedback and continuous optimization to achieve closed-loop learning: During and after the actuator executes the control instructions, the system will continue to collect actual operating performance data through the S1 perception module and the actuator's own feedback sensors, including actual energy consumption, opening and closing time, waiting time for passage, safety events, PCM temperature changes, KERS recovered power, etc.

[0120] These real operating data are recorded and used to: DRL agents are fine-tuned online to continuously improve the performance of the strategy in real-world environments.

[0121] The digital twin model is continuously calibrated to continually improve its simulation accuracy and predictive accuracy.

[0122] This closed-loop process of "perception-modeling and prediction-decision-execution-feedback" is constantly iterating, enabling the entire integrated system to have the ability of continuous learning and adaptive optimization, maintaining efficient, energy-saving, safe and reliable operation for a long time, and adapting to changes in the environment and its own working conditions. While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for controlling the opening and closing of a cold storage door, characterized in that: The following steps are involved: Acquiring multimodal perception information related to the operation of the cold storage door system; Driving a digital twin model based on the multimodal perception information, and using the digital twin model to analyze the current state of the cold storage door system and predict its future state; Based on the multimodal perception information and the analysis and prediction results of the digital twin model, a deep reinforcement learning algorithm is used to generate collaborative control instructions for the opening and closing movements of the cold storage door and the operation of at least one energy-saving device; and the collaborative control instructions are executed.

2. The method for controlling the opening and closing of a cold storage door according to claim 1, characterized in that: The acquiring of multimodal perception information includes: Acquire target information within the passage area of the cold storage door through at least one millimeter-wave radar sensor, the target information including the position and speed of the target, and determine the target's passage intention based on the target information; The surface temperature distribution of the cold storage door body or door frame is obtained by at least one infrared thermal imager, and the frosting state is identified based on the surface temperature distribution; and the structural deformation information of the cold storage door body or door frame is obtained by at least one laser monitoring device.

3. The method for controlling the opening and closing of a cold storage door according to claim 2, characterized in that: The acquiring of multimodal perception information further includes: The operating parameters of the millimeter-wave radar sensor, the infrared thermal imager or the laser monitoring equipment are dynamically adjusted according to the confidence level of the passage intention evaluated based on the target information and the door operation risk level evaluated based on the frosting state and the structural deformation information.

4. The method for controlling the opening and closing of a cold storage door according to claim 1, characterized in that: The digital twin model at least includes a mutually coupled thermodynamic model and a kinematic and dynamic model; The use of the digital twin model to analyze the current state of the cold storage door system and predict the future state further includes: evaluating the health status of key components of the cold storage door system and predicting their remaining service life.

5. The method for controlling the opening and closing of a cold storage door according to claim 4, characterized in that: Further including: The simulation output of the digital twin model is compared with the actual sensor measurement value obtained through the multimodal perception information, and the key parameters of the digital twin model are dynamically calibrated based on the comparison result.

6. The method for controlling the opening and closing of a cold storage door according to claim 1, characterized in that: The use of a deep reinforcement learning algorithm to generate collaborative control instructions includes: Using the multimodal perception information and the analysis and prediction results of the digital twin model as state inputs of the deep reinforcement learning algorithm; The collaborative control instructions output by the deep reinforcement learning algorithm include door motion parameters for controlling the opening and closing movements of the cold storage door and scheduling instructions for regulating at least one energy-saving device; and the deep reinforcement learning algorithm is optimized based on a multi-objective reward function that takes into account energy consumption, traffic efficiency, operational safety and system life.

7. The method for controlling the opening and closing of a cold storage door according to claim 6, characterized in that: The training of the deep reinforcement learning algorithm is performed based on the digital twin model, and has the ability to achieve cross-working condition transfer learning and online adaptive adjustment by loading pre-stored model parameters for different working conditions.

8. The method for controlling the opening and closing of a cold storage door according to claim 1, characterized in that: The at least one energy-saving device includes a phase change material gradient energy storage module and / or a door body kinetic energy recovery device integrated into the door body or door frame of the cold storage door; The executing collaborative control instruction further includes: Regulate the working state of the phase change material gradient energy storage module to utilize the latent heat characteristics of the phase change material to buffer the heat exchange during the opening and closing process of the door body; and / or manage the kinetic energy recovery and reuse of the door body kinetic energy recovery device during the deceleration or braking process of the cold storage door body.

9. The method for controlling the opening and closing of a cold storage door according to claim 1, characterized in that: Further including: Collecting actual operating effect data of the cold storage door system after executing the collaborative control instruction; And use the actual operation effect data to continuously optimize the control strategy of the deep reinforcement learning algorithm and the model parameters of the digital twin model.

10. An integrated system for controlling the opening and closing of a cold storage door, according to the method for controlling the opening and closing of a cold storage door according to any one of claims 1 to 9, characterized in that: include: a multimodal perception module configured to obtain multimodal perception information related to the operation of the cold storage door system; a digital twin and simulation module configured to receive the multimodal sensing information, drive a digital twin model, and utilize the digital twin model to analyze the current state of the cold storage door system and predict its future state; An intelligent decision-making and control module is configured to generate coordinated control instructions for the opening and closing movement of the cold storage door and the operation of at least one energy-saving device using a deep reinforcement learning algorithm based on the multimodal perception information and the analysis and prediction results of the digital twin and simulation module; The execution mechanism is configured to receive and execute the collaborative control instruction to control the opening and closing of the cold storage door and the operation of the at least one energy-saving device.

Images