Method and system for intelligently controlling temperature and humidity of vehicle-mounted cold chain and storage medium

Abstract

The invention relates to the technical field of vehicle-mounted temperature control, and discloses a method and system for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain and a storage medium. The method comprises the following steps: acquiring and preprocessing temperature and humidity data through a sensor; performing state evaluation based on a three-way decision model; constructing a temperature and humidity trend analysis model to predict a change trend; constructing a control decision system by adopting multi-agent reinforcement learning; executing a multi-mode control strategy; and adaptively updating the control system according to the execution effect. According to the application, the future temperature and humidity change trend is predicted according to the historical data and the real-time state, collaborative optimization control of temperature and humidity is realized, self-learning and continuous optimization capabilities are provided, and the accuracy, stability and energy utilization efficiency of cold chain temperature and humidity control are effectively improved.

Description

Technical Field

[0001] The present application relates to the field of vehicle-mounted temperature control technology, and in particular to a method, system and storage medium for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain. Background Art

[0002] At present, cold chain logistics is increasingly used in the fields of food, medicine, fresh produce, etc., and vehicle-mounted cold chain temperature and humidity control is a key link in cold chain logistics. The traditional vehicle-mounted cold chain temperature and humidity control method mainly relies on a simple threshold control strategy, that is, when the temperature or humidity exceeds the preset threshold, the corresponding refrigeration or dehumidification equipment is started, and when the temperature and humidity return to the preset range, the equipment is turned off. This control method is simple to implement and can meet basic needs in a stable environment. With the development of technology, some improved control methods such as proportional-integral-differential (PID) control and fuzzy control have also been applied to vehicle-mounted cold chain temperature and humidity control. Through more complex control algorithms, the accuracy and stability of temperature and humidity control are improved. In addition, remote monitoring systems based on Internet of Things technology have also been gradually applied to cold chain logistics, realizing real-time monitoring and abnormal alarm of the temperature and humidity status of the vehicle-mounted cold chain.

[0003] However, the existing methods for controlling temperature and humidity in vehicle-mounted cold chains still have many shortcomings. First, traditional methods such as threshold control and PID control lack foresight and cannot adjust the control strategy in advance according to future trends in temperature and humidity, resulting in control lag and difficulty in coping with complex and changeable transportation environments. Secondly, these methods usually regard temperature and humidity as independent control objects, ignoring the coupling relationship between the two, making it difficult to achieve coordinated optimization control of temperature and humidity. Thirdly, traditional control methods do not give enough consideration to energy consumption optimization, which often leads to frequent start and stop of the refrigeration system, increased energy consumption and reduced equipment life. In addition, existing methods generally lack adaptive learning capabilities, cannot continuously optimize control strategies based on historical operating data, and are difficult to adapt to the special needs of different cargo types and transportation environments. Finally, most existing methods lack intelligent means for handling abnormal situations. Once equipment failure or environmental changes occur, manual intervention is often required, making it difficult to ensure the continuity and reliability of cold chain quality. Summary of the invention

[0004] The present application provides a method, system and storage medium for intelligently controlling the temperature and humidity of the vehicle-mounted cold chain, which is used to predict future temperature and humidity change trends based on historical data and real-time status, realize collaborative optimization control of temperature and humidity, and have self-learning and continuous optimization capabilities, effectively improving the accuracy, stability and energy utilization efficiency of cold chain temperature and humidity control.

[0005] In the first aspect, the present application provides a method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain, and the method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain includes: collecting temperature and humidity data inside the vehicle and temperature and humidity data outside the vehicle through multiple temperature and humidity sensors in the vehicle, and processing the temperature and humidity data to obtain preprocessed temperature and humidity data; based on the preprocessed temperature and humidity data, using a three-way decision model to divide the temperature and humidity states in the vehicle to obtain a state evaluation result of the temperature and humidity in the vehicle; using the preprocessed temperature and humidity data and the state evaluation result, a temperature and humidity trend analysis model is constructed based on a long short-term memory network and an attention mechanism to obtain a temperature and humidity change trend; based on the state evaluation result and the temperature and humidity change trend, a control decision system is constructed using multi-agent deep reinforcement learning to obtain a temperature and humidity control decision; based on the temperature and humidity control decision, the control decision is divided into multiple operation modes, and a control strategy is executed in combination with the state evaluation result and the temperature and humidity change trend to obtain an execution effect; based on the execution effect, the control decision system is adaptively updated to obtain an optimized temperature and humidity control system.

[0006] In a second aspect, the present application provides a system for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain, the system for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain comprising: A processing module, used to collect temperature and humidity data inside the vehicle compartment and temperature and humidity data outside the vehicle compartment through multiple temperature and humidity sensors in the vehicle compartment, and process the temperature and humidity data to obtain pre-processed temperature and humidity data; A division module, for dividing the temperature and humidity state in the compartment by using a three-way decision model based on the preprocessed temperature and humidity data, and obtaining a state evaluation result of the temperature and humidity in the compartment; An analysis module, for using the preprocessed temperature and humidity data and the state assessment result to construct a temperature and humidity trend analysis model based on a long short-term memory network and an attention mechanism to obtain a temperature and humidity change trend; A construction module is used to construct a control decision system using multi-agent deep reinforcement learning according to the state evaluation result and the temperature and humidity change trend to obtain a temperature and humidity control decision; A control module, for dividing the temperature and humidity control decision into a plurality of operation modes based on the temperature and humidity control decision, and executing a control strategy in combination with the state evaluation result and the temperature and humidity change trend to obtain an execution effect; An updating module is used to adaptively update the control decision system according to the execution effect to obtain an optimized temperature and humidity control system.

[0007] In a third aspect, a device for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the device for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain executes the above-mentioned method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain.

[0008] In a fourth aspect, a computer-readable storage medium is provided, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned method for intelligently controlling the temperature and humidity of the vehicle-mounted cold chain.

[0009] In the technical solution provided by the present application, the present invention collects temperature and humidity data and performs preprocessing by arranging multiple temperature and humidity sensors in the car, thereby constructing a comprehensive and accurate data foundation, avoiding the one-sidedness of data caused by single-point sampling, and ensuring the reliability of subsequent analysis and decision-making. A three-way decision model is used to divide the temperature and humidity state in the car, and the state is accurately classified into positive domain, negative domain and boundary domain, which overcomes the limitations of traditional dichotomy judgment, can more carefully identify the critical situation of temperature and humidity state, and provide a clear state representation for precise control. The temperature and humidity trend analysis model constructed based on long short-term memory network and attention mechanism effectively captures the temporal characteristics and long-term dependencies of temperature and humidity data, automatically identifies key historical data points through attention mechanism, improves the prediction accuracy of temperature and humidity changes in complex environments, and makes the control system forward-looking, able to predict and respond to temperature and humidity anomalies in advance. A control decision system is constructed by multi-agent deep reinforcement learning, and temperature control and humidity control are decomposed into subtasks that work together. Through the collaborative mechanism between agents and the design of global reward function, the coupling problem in temperature and humidity control is solved, and the coordinated optimization and regulation of temperature and humidity are realized. The control decision is divided into multiple operation modes and the control strategy is executed based on the state evaluation results and the temperature and humidity change trend, which improves the adaptability of the control system to different working conditions. It can take the optimal control strategy for different demand scenarios such as normal, rapid temperature adjustment, rapid humidity adjustment, energy saving and emergency, improve the control accuracy and response speed, and reduce energy consumption. The mechanism of adaptive updating of the control decision system enables the system to have the ability of continuous learning and self-optimization, can mine the optimization space from the historical operation data, and continuously adjust and improve the control strategy according to the execution effect, thereby improving the robustness and long-term performance of the system. The present invention combines artificial intelligence algorithms with cold chain physical models for the specific application field of vehicle-mounted cold chain temperature and humidity control, gives full play to the advantages of deep reinforcement learning in complex decision-making problems, the expertise of long short-term memory networks in time series data processing, and the ability of attention mechanisms in key information identification. At the same time, by introducing physical constraints and multi-mode control architectures, it ensures that the algorithm output conforms to the actual physical laws and control requirements, realizes the substantial contribution of algorithm features to the solution, solves the complex scene control problems that are difficult to deal with by traditional methods, and significantly improves the accuracy, stability, foresight and energy efficiency of vehicle-mounted cold chain temperature and humidity control. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 This is a schematic diagram of an embodiment of a method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain in an embodiment of the present application; Figure 2 This is a schematic diagram of an embodiment of a system for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain in an embodiment of the present application; Figure 3 It is a schematic block diagram of the structure of a device for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present application provide a method, system and storage medium for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the method for intelligently controlling the temperature and humidity of the vehicle-mounted cold chain in the embodiment of the present application includes: Step S101, collecting temperature and humidity data inside the vehicle compartment and temperature and humidity data outside the vehicle compartment through multiple temperature and humidity sensors in the vehicle compartment, and processing the temperature and humidity data to obtain pre-processed temperature and humidity data; Step S102: Based on the pre-processed temperature and humidity data, a three-way decision model is used to divide the temperature and humidity status in the compartment to obtain a status evaluation result of the temperature and humidity in the compartment; Step S103: Using the pre-processed temperature and humidity data and the state evaluation results, a temperature and humidity trend analysis model is constructed based on the long short-term memory network and the attention mechanism to obtain the temperature and humidity change trend; Step S104: Based on the state evaluation results and the temperature and humidity change trends, a control decision system is constructed using multi-agent deep reinforcement learning to obtain temperature and humidity control decisions; Step S105: Based on the temperature and humidity control decision, the control decision is divided into multiple operation modes, and the control strategy is executed in combination with the state evaluation result and the temperature and humidity change trend to obtain the execution effect; Step S106: According to the execution effect, the control decision system is adaptively updated to obtain an optimized temperature and humidity control system.

[0014] It is understandable that the execution subject of the present application may be a system for intelligently controlling the temperature and humidity of the vehicle cold chain, or a terminal or a server, which is not limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0015] Specifically, the method for intelligently controlling the temperature and humidity of the vehicle-mounted cold chain of the present invention first collects the temperature and humidity data inside the compartment and the temperature and humidity data outside the compartment through multiple temperature and humidity sensors in the compartment, and then performs data preprocessing. In the specific implementation, 4 sensor nodes are set in the top area of ​​the compartment, 6 sensor nodes are set in the middle area, and 4 sensor nodes are set in the bottom area to form a three-dimensional monitoring network. The data collected by these sensors include real-time temperature and humidity values ​​at each point in the compartment, as well as information such as the ambient temperature, ambient humidity, vehicle speed, and door switch status outside the compartment. When preprocessing the collected raw data, the local anomaly factor algorithm is first used to detect outliers. The algorithm identifies outliers by calculating the local density ratio of the data point and its neighboring points. Then, the detected abnormal data and missing data are supplemented using a time series interpolation algorithm, and interpolation is performed based on valid data at adjacent time points. Finally, the sliding average method is used to smooth the data to reduce the impact of random noise.

[0016] Based on the preprocessed temperature and humidity data, the three-way decision model is used to divide the temperature and humidity state in the compartment, and the state evaluation result of the temperature and humidity in the compartment is obtained. The three-way decision model is based on the three-branch decision theory, and the temperature and humidity state in the compartment is divided into three states: positive domain (POS), negative domain (NEG) and boundary domain (BND). According to the optimal storage conditions of the transported goods, the temperature threshold and humidity threshold are determined. When the temperature and humidity of all points in the compartment are within the target range, the system state belongs to the positive domain, indicating an ideal state; when the temperature or humidity of a point exceeds the range by a certain value, the system state belongs to the negative domain, indicating an abnormal state; other situations belong to the boundary domain, indicating a critical state that needs to be regulated. At the same time, the spatial uniformity index of the temperature and humidity in the compartment is calculated, including the standard deviation of the temperature and humidity of each measuring point, and the uniformity evaluation value of the temperature and humidity distribution is obtained. The temperature and humidity state is combined with the uniformity evaluation value to generate a comprehensive state evaluation result as the input of the subsequent control decision.

[0017] Using the preprocessed temperature and humidity data and state evaluation results, a temperature and humidity trend analysis model is constructed based on the long short-term memory network and attention mechanism to obtain the temperature and humidity change trend. The model first divides the historical data according to the time window to form an input sequence containing multiple continuous sampling points. Each sequence contains features such as temperature vector, humidity vector, ambient temperature, ambient humidity, vehicle speed, door status, etc. After normalization, these time series feature data are input into a neural network composed of three layers of LSTM for processing. The first layer contains 128 neurons, the second layer contains 64 neurons, and the third layer contains 32 neurons. A discard layer is added between each layer to prevent overfitting. The time series features extracted by the LSTM network are weighted by the attention mechanism to automatically identify historical data points that have a greater impact on the prediction results. The weighted features are mapped to the output space through the fully connected layer to obtain the temperature and humidity change trends at multiple time points in the future. At the same time, the model adds the compartment heat transfer model as a physical constraint to ensure that the prediction results conform to physical laws.

[0018] According to the state evaluation results and the temperature and humidity change trends, a control decision system is constructed using multi-agent deep reinforcement learning to obtain the temperature and humidity control decision. The system decomposes the control task into two subtasks, temperature control and humidity control, which are completed by the temperature control agent and the humidity control agent respectively. The temperature control agent is responsible for adjusting the compressor power, fan speed and circulation damper opening; the humidity control agent is responsible for adjusting the desiccant wheel speed and the regeneration heater power. The state space of the control system includes the current temperature and humidity distribution in the car, ambient temperature and humidity, vehicle speed, door state, state evaluation results and temperature and humidity change trends. The reward function design of each agent takes into account the temperature and humidity deviation penalty, uniformity penalty and energy consumption penalty. In order to achieve collaboration between agents, a global reward function is designed, combining the rewards of each agent and the collaborative reward items. The agent adopts the Actor-Critic network architecture and is trained through the proximal policy optimization algorithm to generate the optimal control strategy.

[0019] Based on the temperature and humidity control decision, the control decision is divided into multiple operation modes, and the control strategy is executed in combination with the state evaluation results and the temperature and humidity change trend to obtain the execution effect. The operation modes include normal mode, fast temperature adjustment mode, fast humidity adjustment mode, energy-saving mode and emergency mode. Each operation mode corresponds to a different control target and actuator control strategy matrix. The actuator control adopts a hierarchical architecture. The high-level controller generates the control target of each subsystem according to the selected operation mode and the control action output by the intelligent agent; the low-level controller adopts the proportional integral control algorithm to ensure that each actuator accurately tracks the control target. After the system executes the control action, the control effect is monitored in real time, the real-time temperature and humidity change data are compared with the predicted temperature and humidity change trend, the temperature and humidity deviation value is calculated, and the execution effect of the control strategy is determined as the basis for subsequent adaptive updates.

[0020] According to the execution effect, the control decision system is adaptively updated to obtain the optimized temperature and humidity control system. This step first conducts in-depth mining of the collected historical operation data, and uses a hierarchical clustering algorithm to divide the historical operation data into multiple typical operation modes according to similarity. For each operation mode, key performance indicators are calculated, including temperature adjustment rate, humidity adjustment rate and energy utilization efficiency. Based on these performance evaluation data, the operating points with significant performance differences are identified, and the operating laws and potential optimization space of the control system are extracted. The temperature and humidity trend analysis model uses an incremental learning strategy to update parameters and add newly collected data to the training set. The control decision system maintains an experience buffer pool to store the state-action-reward sequence during operation. When the amount of data in the buffer pool reaches the threshold, the policy distillation technology is used to update the parameters. At the same time, the system also uses an anomaly detection method based on the isolation forest algorithm to monitor the operating parameters of the cold chain system in real time, and promptly detect and handle abnormal situations.

[0021] Taking the transportation of frozen meat as an example, the target temperature range in the carriage is set at -18℃ to -15℃, and the target humidity range is 85% to 90%. During a transportation process, the system detected that the temperature in the middle area of ​​the carriage gradually rose to -14℃, which was higher than the target upper limit, and the state assessment results were divided into boundary domains. The temperature and humidity trend analysis model predicts that if no measures are taken, the temperature in this area will rise to -12℃ within half an hour. Based on this prediction, the multi-agent control system automatically selects the fast temperature adjustment mode, increases the compressor power output, and adjusts the fan speed and the circulation damper opening to make more cold air flow to the central area. After 20 minutes of execution, the temperature in the central area dropped to -16℃, returning to the target range. The system records the state-action-reward sequence of this temperature adjustment process and updates the control strategy. The next time a similar situation occurs, the temperature can be controlled within the target range more quickly and energy-efficiently.

[0022] In the embodiment of the present application, the present invention collects temperature and humidity data and performs preprocessing by arranging multiple temperature and humidity sensors in the car, thereby constructing a comprehensive and accurate data foundation, avoiding the one-sidedness of data caused by single-point sampling, and ensuring the reliability of subsequent analysis and decision-making. A three-way decision model is used to divide the temperature and humidity state in the car, and the state is accurately classified into positive domain, negative domain and boundary domain, which overcomes the limitations of traditional dichotomy judgment, can more carefully identify the critical situation of temperature and humidity state, and provide a clear state representation for precise control. The temperature and humidity trend analysis model constructed based on long short-term memory network and attention mechanism effectively captures the temporal characteristics and long-term dependencies of temperature and humidity data, automatically identifies key historical data points through attention mechanism, improves the prediction accuracy of temperature and humidity changes in complex environments, and makes the control system forward-looking, able to predict and respond to temperature and humidity anomalies in advance. A control decision system is constructed by multi-agent deep reinforcement learning, and temperature control and humidity control are decomposed into collaborative subtasks. Through the collaborative mechanism between agents and the design of global reward function, the coupling problem in temperature and humidity control is solved, and the coordinated optimization and regulation of temperature and humidity are realized. The control decision is divided into multiple operation modes and the control strategy is executed based on the state evaluation results and the temperature and humidity change trend, which improves the adaptability of the control system to different working conditions. It can take the optimal control strategy for different demand scenarios such as normal, rapid temperature adjustment, rapid humidity adjustment, energy saving and emergency, improve the control accuracy and response speed, and reduce energy consumption. The mechanism of adaptive updating of the control decision system enables the system to have the ability of continuous learning and self-optimization, can mine the optimization space from the historical operation data, and continuously adjust and improve the control strategy according to the execution effect, thereby improving the robustness and long-term performance of the system. The present invention combines artificial intelligence algorithms with cold chain physical models for the specific application field of vehicle-mounted cold chain temperature and humidity control, gives full play to the advantages of deep reinforcement learning in complex decision-making problems, the expertise of long short-term memory networks in time series data processing, and the ability of attention mechanisms in key information identification. At the same time, by introducing physical constraints and multi-mode control architectures, it ensures that the algorithm output conforms to the actual physical laws and control requirements, realizes the substantial contribution of algorithm features to the solution, solves the complex scene control problems that are difficult to deal with by traditional methods, and significantly improves the accuracy, stability, foresight and energy efficiency of vehicle-mounted cold chain temperature and humidity control.

[0023] In a specific embodiment, the process of executing step S101 may specifically include the following steps: Four sensor nodes are set in the top area of ​​the carriage, six sensor nodes are set in the middle area, and four sensor nodes are set in the bottom area to form a three-dimensional monitoring network. The real-time temperature and humidity values ​​at each point in the carriage are collected to obtain the temperature and humidity data set in the carriage. An ambient temperature and humidity sensor is set outside the vehicle compartment to collect the external ambient temperature, ambient humidity, vehicle speed and door status, and obtain the vehicle compartment external data set; The local anomaly factor algorithm is used to detect outliers on the temperature and humidity data set inside the carriage and the data set outside the carriage, identifying data that differs too much from adjacent time points and obtaining abnormal labeled data; Based on the valid data at adjacent time points, the time series interpolation algorithm is applied to the abnormal labeled data to fill in the missing values ​​and obtain a complete data set; The complete data set is smoothed using the sliding average method to obtain the temperature and humidity data after noise reduction; The denoised temperature and humidity data are combined with cargo type information to record the cargo's physical and chemical properties and optimal storage temperature and humidity range, and preprocessed temperature and humidity data are obtained for temperature and humidity status assessment.

[0024] Specifically, four temperature and humidity sensor nodes are set in the top area of ​​the car, and these four nodes are located at the four corners of the top of the car, forming a rectangular distribution; six sensor nodes are set in the middle area, four of which are located in the central position of the four sides of the middle of the car, and the other two are located in the center of the middle space of the car, one is biased to the front and the other is biased to the back; four sensor nodes are set in the bottom area, distributed in the four corners of the bottom of the car, corresponding to the top nodes. This distribution method forms a three-dimensional monitoring network to ensure comprehensive monitoring of temperature and humidity conditions at different locations in the car. Each sensor node uses a high-precision temperature and humidity sensor with a temperature measurement accuracy of ±0.1℃ and a humidity measurement accuracy of ±2%RH. These sensors transmit the collected data to the on-board central processing unit via bus or wireless communication, and the data collection frequency is set to once every 30 seconds. The data collected by the sensors in the car include real-time temperature and humidity values ​​at each point.

[0025] At the same time, an environmental temperature and humidity sensor is set outside the car to collect the external environmental temperature and humidity. In addition, the vehicle speed information is obtained through the vehicle's CAN bus, and the door switch status is obtained through the door magnetic switch sensor. These data together constitute the external data set of the car, providing important environmental parameter information for the intelligent control system. The temperature and humidity data set inside the car and the data set outside the car together form the original data set, which needs to be further processed before it can be used for subsequent intelligent control decisions.

[0026] For the collected raw data, the local anomaly factor algorithm is first used to detect outliers. The local anomaly factor algorithm is a density-based anomaly detection method. Its core idea is to identify outliers by comparing the local density differences between data points and their neighborhoods. The specific steps include: first calculating the distance between data points, then determining the nearest neighbors of each data point, calculating the local reachable density of each data point, and finally calculating the local anomaly factor. For time series data, the data of multiple consecutive time points form a sliding window, and the distance and density relationship between each data point in the window and its neighboring points is calculated. When the local anomaly factor value of a data point is significantly higher than the threshold, it is marked as abnormal data. For example, when the temperature of a sensor suddenly changes by more than 3°C or the humidity changes by more than 10% in a short period of time, and this change is not observed on other sensors, the data point will be marked as abnormal.

[0027] For the detected abnormal data and missing data points, the time series interpolation algorithm is used to complete them. The time series interpolation algorithm is based on the valid data of adjacent time points for interpolation. Commonly used methods include linear interpolation, spline interpolation and polynomial interpolation. In this method, the cubic spline interpolation method is mainly used. This method not only considers the continuity of the data, but also the smoothness of the data. The specific operation is to use several valid data points before and after the abnormal marked data as reference points, construct a cubic spline function, and then calculate the interpolation result corresponding to the abnormal point. This method can effectively retain the trend and smoothness of temperature and humidity changes, and avoid the jagged changes that may be caused by simple linear interpolation.

[0028] After completing the missing value filling, the sliding average method is applied to the complete data set for smoothing to reduce the interference of random noise on the system judgment. The sliding average method takes the average of the data of the current time point and a certain number of time points before and after it as the new value of the current time point. In this method, the sliding window size is set to 5 sampling points, that is, the data of the current point and the two points before and after it are averaged. Through sliding average processing, data fluctuations in a short period of time can be effectively eliminated, and a more stable temperature and humidity change trend can be obtained.

[0029] Finally, the de-noised temperature and humidity data are combined with the cargo type information to record the cargo's physical and chemical properties and the optimal storage temperature and humidity range. The system has preset parameter libraries for different types of cargo, including the optimal storage temperature and humidity ranges for frozen meat, fresh fruits and vegetables, vaccines, and other different cargoes. According to the type of cargo currently being transported, the system automatically selects the corresponding parameter settings. This information, together with the de-noised temperature and humidity data, constitutes the pre-processed temperature and humidity data, providing a basis for subsequent status evaluation.

[0030] Taking the transportation of cold fresh food as an example, a cold chain vehicle transports a batch of fresh food that needs to be kept in an environment of 2°C to 8°C and a relative humidity of 75% to 85%. The 14 sensors in the carriage are installed according to the above distribution method, and data is collected every 30 seconds. During one operation, sensor 11 suddenly reported a temperature of 15°C at a certain moment, while other sensors still showed 4-6°C. The local anomaly factor algorithm calculates that the anomaly factor value of this point is much higher than the set threshold, so the point is marked as abnormal data. Subsequently, the time series interpolation algorithm is used to take the valid data points within two minutes before and after the sensor, and the temperature at this moment should be 5.4°C through cubic spline interpolation calculation. After completing the processing of all abnormal points, the entire data set is smoothed by the sliding average method to eliminate random fluctuations in a short period of time. Finally, the system combines the processed temperature and humidity data with the parameters of the fresh food currently being transported, records the optimal storage temperature range and humidity range, and provides an accurate data basis for subsequent temperature and humidity status evaluation and control decisions.

[0031] In a specific embodiment, the process of executing step S102 may specifically include the following steps: According to the optimal storage conditions of the transported goods, the temperature threshold and humidity threshold are set to obtain the target range of temperature and humidity; The pre-processed temperature and humidity data are analyzed. When the temperature of all points in the car is within the temperature threshold and the humidity is within the humidity threshold, the system state is divided into the positive domain and the ideal state identification is obtained; The pre-processed temperature and humidity data are analyzed. When the temperature or humidity of a point in the car exceeds the temperature threshold range or the humidity threshold range, the system state is divided into a negative domain and an abnormal state mark is obtained. The pre-processed temperature and humidity data are analyzed, and the states that do not belong to the positive domain and the negative domain are divided into boundary domains to obtain critical state identification; Calculate the spatial uniformity index of temperature and humidity in the carriage, including the standard deviation of temperature and humidity at each measuring point in the carriage, and obtain the evaluation value of temperature and humidity distribution uniformity; The ideal state identification, abnormal state identification or critical state identification is combined with the temperature and humidity distribution uniformity evaluation value to generate a comprehensive state evaluation result as an input for control decision-making.

[0032] Specifically, the temperature threshold and humidity threshold are set according to the optimal storage conditions of the transported goods to obtain the target range of temperature and humidity. Different types of cold chain goods have different optimal storage temperature and humidity conditions, which are usually determined by food safety standards or drug storage specifications. In practical applications, the corresponding temperature and humidity parameters are obtained by querying the cold chain goods database. For example, the optimal storage temperature range of frozen meat is -18°C to -15°C, and the humidity range is 85% to 90%; the optimal storage temperature range of fresh fruits and vegetables is 2°C to 8°C, and the humidity range is 90% to 95%; the optimal storage temperature range of vaccines is 2°C to 5°C, and the humidity range is 45% to 65%. These parameters are set as the control target range as a standard for status evaluation.

[0033] After obtaining the target range of temperature and humidity, the three-way decision model is used to divide the temperature and humidity status in the car. The three-way decision model is derived from the three-branch decision theory, which divides the decision space into three parts: positive domain, negative domain and boundary domain. In this method, the positive domain represents the ideal state, the negative domain represents the abnormal state, and the boundary domain represents the critical state. First, the preprocessed temperature and humidity data are analyzed to determine whether the temperature and humidity of each point in the car are within the target range. When the temperature of all measuring points in the car is within the set temperature threshold range, and the humidity is within the set humidity threshold range, the system state is divided into the positive domain, and an ideal state mark is generated. The specific judgment process is to traverse the temperature and humidity data of all sensor nodes in the car, and check whether the temperature of each node is within the temperature threshold range and the humidity is within the humidity threshold range. Only when the temperature and humidity data of all nodes meet the conditions, the ideal state mark is generated. Then continue to analyze the preprocessed temperature and humidity data. When the temperature or humidity of a point in the car exceeds a certain value of the set threshold range, the system state is divided into the negative domain, and an abnormal state mark is generated. "Exceeding a certain value" here means that the degree of exceeding the threshold range reaches the preset fault tolerance value. For example, if the temperature fault tolerance value is set to 0.5℃ and the humidity fault tolerance value is 3%, when the temperature of the measuring point is lower than the minimum temperature threshold minus 0.5℃, or higher than the maximum temperature threshold plus 0.5℃, or the humidity is lower than the minimum humidity threshold minus 3%, or higher than the maximum humidity threshold plus 3%, it is judged as an abnormal state. This design takes into account the actual needs of cold chain temperature and humidity control and avoids frequent state switching due to small fluctuations.

[0034] For states that are neither in the positive domain nor in the negative domain, that is, when the temperature and humidity of the measuring point exceed the target range but do not meet the abnormal state judgment criteria, they are divided into boundary domains and critical state identification is generated. The critical state is a state that requires attention but has not yet reached the abnormal level, and is also a state that the control system needs to adjust. The specific judgment process is to first exclude the conditions that meet the positive domain and negative domain conditions, and the remaining states are divided into boundary domains.

[0035] In addition to determining whether the absolute values ​​of temperature and humidity are within the target range, it is also necessary to calculate the spatial uniformity index of temperature and humidity in the car to evaluate the uniformity of temperature and humidity distribution. The spatial uniformity index is expressed by calculating the standard deviation of temperature and humidity at each measuring point in the car. The smaller the standard deviation, the more uniform the temperature and humidity distribution; the larger the standard deviation, the more uneven the temperature and humidity distribution. The process of calculating the standard deviation is to first find the average temperature of all measuring points, then calculate the square of the difference between the temperature of each measuring point and the average value, sum it up and divide it by the number of measuring points, and finally take the square root. The calculation method of humidity standard deviation is similar.

[0036] Combine the ideal state identification, abnormal state identification or critical state identification obtained above with the temperature and humidity distribution uniformity evaluation value to generate a comprehensive state evaluation result. The combination method is to encode the temperature and humidity state classification and uniformity evaluation results into a comprehensive state code. For example, a two-bit state code can be used, the first bit represents the temperature and humidity state classification (0 represents the positive domain, 1 represents the boundary domain, and 2 represents the negative domain), and the second bit represents the uniformity level (0 represents uniform, 1 represents relatively uniform, and 2 represents uneven). The comprehensive state evaluation result obtained in this way can fully reflect the state of the temperature and humidity environment in the car, and serve as an important input for subsequent control decisions.

[0037] Taking the cold chain transportation of vaccines as an example, the optimal storage temperature range of a batch of vaccines is 2℃ to 5℃, and the humidity range is 45% to 65%. The temperature tolerance is set to 0.5℃, and the humidity tolerance is 3%. During a transportation process, the pre-processed temperature and humidity data collected by 14 sensors show that the temperature of 12 sensors is between 2.3℃ and 4.7℃, and the humidity is between 48% and 62%, but the temperature of 2 sensors is 5.3℃ and 5.4℃, and the humidity is 58% and 60%. First, determine whether the positive domain condition is met. Because there are measurement points with a temperature exceeding 5℃, it does not belong to the positive domain. Then determine whether the negative domain condition is met. The highest temperature of 5.4℃ exceeds the upper limit of 5℃, but does not exceed the upper limit plus the tolerance value of 5.5℃, so it does not belong to the negative domain. Therefore, the current state is divided into a boundary domain and a critical state identifier is generated. Then calculate the temperature uniformity index. The average temperature of all measurement points is 3.8℃, and the standard deviation is 0.9℃; the average humidity is 55%, and the standard deviation is 4.2%. According to the set uniformity evaluation criteria, when the temperature standard deviation is less than 1.0℃ and the humidity standard deviation is less than 5%, the temperature and humidity distribution is determined to be uniform. Therefore, the uniformity evaluation value is "uniform". Finally, the critical state identification and uniformity evaluation value are combined to generate a comprehensive state evaluation result, which is coded as "10", indicating "critical state and uniform". This comprehensive evaluation result will serve as the input of the control decision system to guide the system to perform appropriate temperature and humidity control.

[0038] In a specific embodiment, the process of executing step S103 may specifically include the following steps: The preprocessed temperature and humidity data are divided into time windows, each of which contains multiple continuous sampling points to form an input sequence and obtain time series feature data; Perform feature normalization on the time series feature data, convert the temperature, humidity, and environmental parameters into a unified dimension, and obtain normalized feature data; The normalized feature data is input into a three-layer long short-term memory network for processing. The first layer contains 128 neurons, the second layer contains 64 neurons, and the third layer contains 32 neurons. A dropout layer is added between each layer to obtain a temporal feature representation. The attention mechanism is applied to the time series feature representation for weighting, and the importance of historical data points is sorted by calculating the weight coefficient to obtain the weighted feature representation; The weighted feature representation is mapped to the output space through the fully connected layer to obtain the predicted values ​​of temperature and humidity change trends at multiple time points in the future; The cabin heat transfer model is added as a physical constraint to the predicted value of the temperature and humidity change trend. The model parameters are optimized through the loss function composed of the mean square error loss and the physical model constraint regularization term to obtain the temperature and humidity change trend that conforms to the physical laws.

[0039] Specifically, based on the preprocessed temperature and humidity data and state assessment results, it is necessary to construct a temperature and humidity trend analysis model to accurately predict future temperature and humidity change trends. First, the preprocessed temperature and humidity data are divided according to time windows, and each time window contains multiple continuous sampling points to form an input sequence to obtain time series feature data. The time window refers to the range of data samples selected for analysis within a continuous period of time in time series analysis. In this method, the length of the time window is set to L sampling points, and there may or may not be overlap between adjacent windows. For example, if the sampling frequency is once every 30 seconds and the window length L is set to 20, each window contains 10 minutes of continuous data. For the data in each time window, the extracted features include the temperature vector, humidity vector, ambient temperature, ambient humidity, vehicle speed, door status, etc. in the car, which together constitute the input sequence for subsequent deep learning model training.

[0040] After the time series feature data is extracted, it is necessary to perform feature normalization processing on it, convert the temperature, humidity, and environmental parameters to a unified dimension, and obtain normalized feature data. Normalization is the process of converting features of different dimensions and different numerical ranges to the same scale, which helps to improve the stability and convergence speed of model training. There are usually two normalization methods: minimum-maximum normalization and Z-score standardization. Minimum-maximum normalization scales the feature values ​​to the interval [0,1], and Z-score standardization converts the features into a distribution with a mean of 0 and a standard deviation of 1. In this method, minimum-maximum normalization is used for temperature data, linear scaling is performed on humidity data (which is already in the range of 0-100%), and Z-score standardization is used for other features such as vehicle speed. After processing in this way, each dimension of the data has a similar numerical range, which is conducive to the training of neural networks.

[0041] After the normalized feature data is prepared, it is input into a three-layer long short-term memory network for processing. Long short-term memory network (LSTM) is a special recurrent neural network that can effectively process sequence data and capture long-term dependencies. LSTM solves the gradient vanishing problem in traditional recurrent neural networks by introducing memory units and gating mechanisms. In this method, a three-layer LSTM network is constructed, with the first layer containing 128 neurons, the second layer containing 64 neurons, and the third layer containing 32 neurons. A dropout layer is added between each layer, and the dropout rate is set to 0.2 to prevent overfitting. The dropout layer is a regularization technique that randomly sets the output of a part of the neurons to zero during the training process, forcing the network to learn a more robust feature representation. The output of the LSTM network after processing is the time series feature representation, which captures the time dependency and change law of the temperature and humidity data.

[0042] The time series feature representation extracted by LSTM is weighted by applying the attention mechanism, and the importance of historical data points is sorted by calculating the weight coefficient to obtain the weighted feature representation. The attention mechanism is a technology that allows the model to focus on the important parts of the input sequence, which is particularly effective in time series prediction. In this method, the dot product attention mechanism is used to calculate the dot product of the query vector and the key vector, and the weight coefficient is obtained by the softmax function, and then the value vector is weighted and summed. Specifically, the hidden state of LSTM is used as the key vector and the value vector, and the query vector can be the hidden state of the last time step or learned through additional parameters. Through the attention mechanism, the model can automatically identify historical data points that have a greater impact on the prediction results, and give them higher weights to improve the prediction accuracy.

[0043] The weighted feature representation is mapped to the output space through the fully connected layer to obtain the predicted values ​​of the temperature and humidity change trend at multiple time points in the future. The fully connected layer is a basic neural network layer that maps the input features to the output space through linear transformation and nonlinear activation function. In this method, the input of the fully connected layer is the weighted feature representation, and the output is the predicted value of the temperature and humidity in the car at the next K time points. The number of neurons in the output layer is K×N, where K is the number of predicted time steps and N is the number of sensor nodes in the car. ReLU is selected as the activation function of the fully connected layer to introduce nonlinearity and avoid the gradient vanishing problem.

[0044] Finally, the cabin heat transfer model is added as a physical constraint to the predicted value of the temperature and humidity change trend. The model parameters are optimized through the loss function composed of the mean square error loss and the physical model constraint regularization term to obtain the temperature and humidity change trend that conforms to the physical laws. The loss function is defined as follows:

[0045] in, represents the total loss function, represents the model parameters, M represents the number of training samples, represents the model's predicted value for the i-th sample, represents the true value of the i-th sample, represents the square of the Euclidean distance, is the weight coefficient, is the physical model constraint regularization term. The physical model constraint regularization term is based on the thermodynamic equation, taking into account factors such as heat conduction of the compartment wall, convection heat transfer of the airflow, and heat exchange caused by door opening and closing, and is defined as follows:

[0046] in, and They represent the temperature and humidity of the jth node at time t predicted by the model, and They represent the external environment temperature and humidity at time t respectively, represents the door state at time t, represents the vehicle speed at time t, Represents the cabin heat transfer model function based on thermodynamic equations. represents the physical model constraint regularization term, which is used to constrain the prediction results to conform to the physical laws. K represents the number of prediction time steps, that is, the predicted temperature and humidity values ​​at how many time points in the future, and N represents the number of sensor nodes in the car. By adding this regularization term, the model will respect the physical laws while learning the data pattern, avoiding prediction results that violate the principles of thermodynamics.

[0047] Taking the transportation of refrigerated medicines as an example, the medicines loaded in a refrigerated truck need to be stored in an environment of 2-8℃. The temperature and humidity data are continuously collected by 14 sensors in the car, once every 30 seconds, and one day of historical data is accumulated. First, the data is divided into groups of 20 sampling points. Each group of data contains 10 minutes of continuous records, and adjacent groups overlap by 10 points. For each group of data, the features such as temperature and humidity of the 14 sensors in the car, ambient temperature, ambient humidity, vehicle speed and door status are extracted. Then these features are normalized. The temperature data is converted to the [0,1] interval through minimum-maximum normalization, the humidity data is linearly scaled, and the vehicle speed is normalized by Z-score. The normalized data is input into a three-layer LSTM network, and is processed layer by layer with 128, 64, and 32 neurons. A dropout layer with dropout=0.2 is added between each layer to prevent overfitting. The time series features output by the LSTM are weighted by the attention mechanism to identify key historical data points. For example, data points after the door switch state changes will receive higher weights because the data at these moments are more important for predicting future temperature and humidity changes. The weighted features are mapped to the output space through the fully connected layer to predict the temperature and humidity changes at each sensor point in the next 20 minutes. Finally, by adding physical constraint regularization terms, ensure that the prediction results conform to the laws of thermodynamics. For example, when the prediction shows that the temperature in a certain area rises abnormally, the heat transfer model will check whether this conforms to the influence of external factors such as door opening or ambient temperature changes. If not, the model will correct the prediction results.

[0048] In a specific embodiment, the process of executing step S104 may specifically include the following steps: The control task is decomposed into two subtasks: temperature control and humidity control. The temperature control agent and humidity control agent are constructed respectively to form a multi-agent control framework. Define the control system state space, including the current temperature and humidity distribution in the car, ambient temperature and humidity, vehicle speed, door status, state evaluation results, and temperature and humidity change trends, and obtain the agent state representation; Define the action space of the temperature control agent including compressor power, fan speed and circulation damper opening, and the action space of the humidity control agent including desiccant wheel speed and regeneration heater power, and obtain the control action set that the agent can execute; Constructing a reward function for a temperature control agent including a temperature deviation penalty term, a temperature non-uniformity penalty term, and an energy consumption penalty term, and a reward function for a humidity control agent including a humidity deviation penalty term, a humidity non-uniformity penalty term, and an energy consumption penalty term, and obtaining the agent optimization target; Design a global reward function, combine the temperature control agent reward, humidity control agent reward and collaboration reward items, train through the proximal policy optimization algorithm, and obtain the agent policy network parameters; The state evaluation results and temperature and humidity change trends are input into the trained intelligent agent strategy network to generate the optimal control action sequence and obtain the temperature and humidity control decision.

[0049] Specifically, the control task is decomposed into two subtasks, temperature control and humidity control, and the temperature control agent and humidity control agent are constructed respectively to form a multi-agent control framework. Multi-agent reinforcement learning is a distributed decision-making method. Each agent is responsible for the decision of a specific subtask and achieves overall optimization through a coordination mechanism. In the on-board cold chain temperature and humidity control, temperature and humidity are two interrelated and relatively independent control objects, which are adjusted by different actuators. The temperature control agent is mainly responsible for adjusting the operating parameters of the refrigeration system, while the humidity control agent is responsible for adjusting the operating parameters of the dehumidification or humidification equipment. This decomposition design conforms to the physical structure of the cold chain system and is also convenient for flexible adjustment according to the special needs of different types of goods. The control system state space is defined, including the current temperature and humidity distribution in the car, ambient temperature and humidity, vehicle speed, door state, state evaluation results and temperature and humidity change trends, and the agent state representation is obtained. The state space is the basis for the agent to perceive the environment and contains all the key information required for decision-making. The current temperature and humidity distribution in the car is composed of real-time temperature and humidity data collected by 14 sensor nodes, reflecting the temperature and humidity conditions in different areas of the car; the ambient temperature and humidity data comes from the external sensors of the car, providing important information about external interference; the vehicle speed information is obtained through the vehicle CAN bus, reflecting the vehicle running status; the door status is obtained through the door magnetic switch sensor, recording the door switch status, which has an important impact on the temperature and humidity fluctuations; the state evaluation result is the classification of the temperature and humidity state of the car generated by the three-way decision model, including the identification of the positive domain, negative domain and boundary domain, as well as the uniformity evaluation value; the temperature and humidity change trend is the future temperature and humidity change predicted by the model constructed by the long short-term memory network and the attention mechanism. These information together constitute a complete state representation, providing a basis for the decision-making of the intelligent agent.

[0050] Define the action space of the temperature control agent including compressor power, fan speed and circulation damper opening, and the action space of the humidity control agent including desiccant wheel speed and regeneration heater power, and obtain the control action set that the agent can execute. The action space defines all possible actions that the agent can take. The compressor power controls the cooling capacity of the cabin, ranging from 0 to maximum power, usually expressed in percentage form; the fan speed controls the circulation speed of the cold air, affecting the temperature uniformity and adjustment rate, ranging from 0 to maximum speed; the circulation damper opening controls the direction of the cold air flow, which is used to adjust the temperature distribution in different areas, ranging from 0° to 90°. The desiccant wheel speed controls the rate of humidity reduction, ranging from 0 to maximum speed; the regeneration heater power controls the regeneration efficiency of the desiccant, ranging from 0 to maximum power. These parameters together constitute the control action set of the agent. By adjusting these parameters, the agent can achieve precise control of the temperature and humidity environment in the cabin.

[0051] The reward function of the temperature control agent includes temperature deviation penalty, temperature non-uniformity penalty and energy consumption penalty, and the reward function of the humidity control agent includes humidity deviation penalty, humidity non-uniformity penalty and energy consumption penalty, and the agent optimization target is obtained. The reward function is the core of reinforcement learning and defines the criteria for judging the quality of the agent's behavior. The temperature deviation penalty calculates the difference between the temperature of each point in the car and the target temperature. The smaller the difference, the smaller the penalty; the temperature non-uniformity penalty calculates the standard deviation of the temperature distribution in the car. The smaller the standard deviation, the smaller the penalty; the energy consumption penalty calculates the energy consumption caused by the control action. The smaller the energy consumption, the smaller the penalty. The reward function design of the humidity control agent is similar to that of the temperature control agent, and includes three corresponding penalty terms. Through the combination of these penalty terms, a comprehensive optimization target that balances the accuracy, uniformity and energy efficiency of temperature and humidity control is formed. The global reward function is designed, combining the temperature control agent reward, the humidity control agent reward and the collaborative reward term, and trained through the proximal policy optimization algorithm to obtain the agent policy network parameters. The global reward function is used to promote collaboration between agents, and includes three parts: temperature control reward, humidity control reward, and collaboration reward. The collaboration reward mainly considers the mutual influence between temperature and humidity control. For example, the cooling process will cause the relative humidity to increase, and the dehumidification process will cause the temperature to increase. The collaboration reward measures the effect of the collaboration between the two agents by evaluating the difference between the actual temperature and humidity and the predicted temperature and humidity. The proximal policy optimization algorithm is a classic policy gradient reinforcement learning algorithm with high sample efficiency and good stability. The algorithm avoids instability caused by excessive policy updates by limiting the difference between the new and old policies. During the training process, the experience replay buffer is used to store transfer samples, and each sample contains information about the state, action, reward, and next state. Through multiple iterative training, the parameters of the agent policy network are continuously optimized, and the optimal control strategy is finally learned.

[0052] The state evaluation results and the temperature and humidity change trends are input into the trained agent strategy network to generate the optimal control action sequence and obtain the temperature and humidity control decision. In the actual control process, the current temperature and humidity distribution in the car, the ambient temperature and humidity, the vehicle speed, the door status and other information are first obtained, and the current state representation is constructed by combining the state evaluation results and the temperature and humidity change trends obtained in the previous steps. Then the state representation is input into the strategy network of the trained temperature control agent and humidity control agent, and their respective control actions are output respectively. The temperature control agent outputs the control values ​​of the compressor power, the fan speed and the circulation damper opening, and the humidity control agent outputs the control values ​​of the desiccant wheel speed and the regeneration heater power. These control values ​​together constitute the temperature and humidity control decision, which is used to guide the subsequent actuator control.

[0053] Taking the cold chain transportation of vaccines as an example, a cold chain vehicle transports a batch of vaccines and needs to maintain a temperature of 2-5°C and a humidity of 45-65%. When the vehicle enters a mountain road, the external temperature suddenly drops and the humidity rises. The 14 sensors in the car monitor the temperature and humidity data in real time. The three-way decision model evaluates that the current state is in the boundary domain, the temperature distribution is uniform, but the temperature in some local areas is close to the lower limit. The temperature and humidity trend analysis model predicts that if no measures are taken, the temperature in some areas of the car will be lower than 2°C in 20 minutes, and the humidity will rise to more than 70%. At this time, this information is input into the multi-agent control decision system. Based on the state information, the temperature control agent calculates that the compressor power needs to be reduced to 30%, the fan speed is adjusted to 40%, and the circulation damper opening is set to 65° to reduce the cooling capacity and improve the cold air distribution. At the same time, the humidity control agent calculates that the desiccant wheel speed needs to be increased to 60%, and the regeneration heater power needs to be set to 50% to cope with the rising humidity. These control decisions take into account factors such as temperature and humidity change trends, energy consumption optimization, and equipment coordination. They can effectively respond to environmental changes, keep the temperature and humidity in the carriage within the ideal range, and ensure the safe transportation of vaccines.

[0054] In a specific embodiment, the process of executing step S105 may specifically include the following steps: According to the state evaluation results, the control strategy is divided into five operation modes: normal mode, rapid temperature adjustment mode, rapid humidity adjustment mode, energy-saving mode and emergency mode, and the operation mode set is obtained; For each operation mode in the operation mode set, a corresponding actuator control strategy matrix is ​​designed to map the temperature and humidity control decision to specific actuator control parameters to obtain an actuator control parameter set; A hierarchical control architecture is adopted to generate the control targets of each subsystem based on the operation mode set and temperature and humidity control decision, and obtain the subsystem control target set; Apply proportional-integral control algorithm to the subsystem control target set to generate actuator control signals, ensure that each actuator accurately tracks the control target and obtains control execution instructions; By controlling the execution instructions to adjust the compressor power, fan speed, circulation damper opening, desiccant wheel speed and regeneration heater power, the temperature and humidity adjustment operation is performed to obtain real-time temperature and humidity change data; Compare the real-time temperature and humidity change data with the temperature and humidity change trend, calculate the temperature and humidity deviation value, and determine the execution effect of the control strategy as the basis for adaptive updating.

[0055] Specifically, the control strategy is divided into five operation modes according to the state evaluation results: normal mode, fast temperature adjustment mode, fast humidity adjustment mode, energy-saving mode and emergency mode, and an operation mode set is obtained. The normal mode is applicable to the case where the temperature and humidity state in the cabin belongs to the positive domain. At this time, the temperature and humidity of all points in the cabin are within the target range. The control goal is to maintain the current temperature and humidity state and minimize energy consumption. The fast temperature adjustment mode is applicable to the case where the temperature in the cabin deviates from the target range but the humidity is within the target range. The control goal is to adjust the temperature to the target range as quickly as possible. The fast humidity adjustment mode is applicable to the case where the humidity in the cabin deviates from the target range but the temperature is within the target range. The control goal is to adjust the humidity to the target range as quickly as possible. The energy-saving mode is applicable to the case where the vehicle is parked or the temperature and humidity state in the cabin is good and it is predicted that there will be no significant changes in a short time. The control goal is to minimize the system energy consumption. The emergency mode is applicable to the case where the temperature and humidity in the cabin deviate seriously from the target range or are predicted to deviate seriously in a short time. The control goal is to adjust the temperature and humidity to a safe range in the shortest time. The mode selection process is based on a comprehensive judgment of the state evaluation results of the three-way decision model, the temperature and humidity uniformity evaluation value, and the temperature and humidity change trend, and the most suitable operation mode is automatically selected using rule matching or decision tree algorithm.

[0056] For each operation mode in the operation mode set, the corresponding actuator control strategy matrix is ​​designed to map the temperature and humidity control decision to specific actuator control parameters to obtain the actuator control parameter set. The control strategy matrix is ​​a mapping relationship table that converts high-level control decisions into low-level actuator control parameters. Different mapping rules are designed for different operation modes. For example, for the normal mode, the control strategy matrix will prioritize energy consumption minimization, appropriately reduce the compressor power and fan speed, and maintain a low desiccant wheel speed; for the rapid temperature adjustment mode, the control strategy matrix will prioritize the adjustment of the compressor power and fan speed, and adjust the circulation damper opening as needed to reach the target temperature at the fastest speed; for the rapid humidity adjustment mode, the control strategy matrix will prioritize the adjustment of the desiccant wheel speed and the regeneration heater power to quickly adjust the humidity; for the energy-saving mode, the control strategy matrix will adjust the compressor power curve in advance according to the predicted temperature and humidity change trend, avoid frequent start and stop, and reduce peak power consumption; for the emergency mode, the control strategy matrix will maximize the temperature and humidity adjustment capabilities at the same time, regardless of energy consumption factors, and adjust all actuators to the optimal working state. These mapping relationships are based on the control decisions obtained by multi-agent reinforcement learning, but further consider the specific needs of different operating modes. A hierarchical control architecture is adopted to generate the control objectives of each subsystem based on the operating mode set and the temperature and humidity control decision, and obtain the subsystem control objective set. The hierarchical control architecture is a control method that decomposes complex control problems into different levels, with the high level responsible for decision-making and the low level responsible for execution. In the vehicle-mounted cold chain temperature and humidity control, the high-level controller generates the control objectives of each subsystem based on the selected operating mode and the control action output by the agent; the low-level controller is responsible for ensuring that each actuator accurately tracks the control objectives. The subsystems include refrigeration system, fan system, damper system, drying system, etc., and each subsystem is responsible for controlling the corresponding actuator. The control objective of the refrigeration system is the set value of the compressor power; the control objective of the fan system is the set value of the fan speed; the control objective of the damper system is the set value of the circulation damper opening; the control objective of the drying system is the set value of the desiccant wheel speed and the regeneration heater power. These control objectives constitute the subsystem control objective set, which serves as the input of the low-level controller.

[0057] The proportional-integral control algorithm is applied to the subsystem control target set to generate actuator control signals, ensure that each actuator accurately tracks the control target, and obtain control execution instructions. The proportional-integral control algorithm is a classic feedback control algorithm that achieves accurate tracking of the target value through the combination of proportional and integral terms. The role of the proportional term is to give the corresponding control output according to the current error size, and the role of the integral term is to accumulate historical errors and eliminate static errors. For compressor speed control, the control law of the PI controller is to calculate the deviation between the current speed and the set speed, and then multiply the deviation by the proportional coefficient, plus the integral value of the deviation multiplied by the integral coefficient to obtain the control output. Similarly, for the fan speed, circulation damper opening, desiccant wheel speed and regeneration heater power, the corresponding PI controller is used for precise control. The parameters of the PI controller are obtained through offline optimization or online adaptive adjustment to meet the control requirements under different working conditions. In this way, high-level control decisions are accurately converted into specific control signals for each actuator to form control execution instructions.

[0058] By controlling the execution instructions to adjust the compressor power, fan speed, circulation damper opening, desiccant wheel speed and regeneration heater power, the temperature and humidity adjustment operation is performed to obtain real-time temperature and humidity change data. The compressor is the core component of the refrigeration system. The refrigeration capacity can be controlled by adjusting its power; the fan is responsible for the circulation and distribution of cold air, and the temperature uniformity and adjustment rate are affected by adjusting the speed; the circulation damper controls the flow of cold air, and the temperature distribution in different areas is affected by adjusting the opening; the desiccant wheel adjusts the humidity by absorbing or desorbing moisture, and the speed affects the humidity adjustment rate; the regeneration heater is responsible for the regeneration of the desiccant, and the power affects the dehumidification capacity of the desiccant. These actuators work together to achieve precise control of the temperature and humidity environment in the car. During the execution of the control operation, the sensor network in the car continuously collects real-time temperature and humidity data to reflect the execution effect of the control operation. After preprocessing, these real-time data form temperature and humidity change data for subsequent control effect evaluation. Compare the real-time temperature and humidity change data with the temperature and humidity change trend, calculate the temperature and humidity deviation value, and determine the execution effect of the control strategy as the basis for adaptive update. The temperature and humidity deviation value is the difference between the actual temperature and humidity and the predicted temperature and humidity, which reflects the accuracy and effectiveness of the control strategy execution. By calculating the temperature deviation and humidity deviation of each sensor point, the overall deviation distribution is obtained. The temperature deviation is calculated by subtracting the predicted temperature at the corresponding moment from the actual measured temperature; the humidity deviation is calculated in a similar way. These deviation data are statistically analyzed, and indicators such as the mean and variance are calculated to comprehensively evaluate the execution effect of the control strategy. If it is detected that the actual temperature and humidity changes are significantly different from the expected ones (for example, the temperature deviation exceeds ±0.5℃ or the humidity deviation exceeds ±5%RH), it indicates that the execution effect of the control strategy is not ideal and needs to be adjusted. These deviation data and their statistical analysis results serve as an important basis for subsequent adaptive updates, providing data support for the optimization of the control decision system.

[0059] Taking the cold chain transportation of fresh food as an example, a cold chain vehicle transports a batch of fresh food that needs to be kept in an environment of 0-4℃ and 85-95% humidity. The temperature and humidity data obtained through the temperature and humidity monitoring network in the car compartment are evaluated by the three-way decision model, and it is found that the temperature in the car compartment is within the normal range, but the humidity is low. The humidity of multiple sensor points drops to 80-83%. The state evaluation result is a boundary domain and the temperature and humidity distribution is relatively uniform. The temperature and humidity trend analysis predicts that if no measures are taken, the humidity will continue to drop. Based on this information, the control system automatically selects the rapid humidity adjustment mode and inputs this decision into the actuator control strategy matrix. For the rapid humidity adjustment mode, the parameters generated by the control strategy matrix are set as follows: reduce the desiccant wheel speed to the minimum value of 20%, turn off the regeneration heater (power 0%), appropriately reduce the compressor power to 60%, keep the fan speed at 70%, and adjust the damper opening to 45°. These parameters are passed to each lower-level controller as the subsystem control target. Through the PI control algorithm, each actuator accurately tracks these target values. For example, the PI controller of the desiccant wheel motor takes the deviation between the current speed and the target speed as input to generate the corresponding control voltage. After 30 minutes of operation, the real-time temperature and humidity data showed that the humidity had risen to 86-91%, and the temperature remained within the range of 1-3.5°C, which was basically consistent with the expected trend and the deviation was within the allowable range. This shows that the control strategy is executing well and no major adjustments are required. Only the desiccant wheel speed needs to be fine-tuned to maintain the current humidity level.

[0060] In a specific embodiment, the process of executing step S106 may specifically include the following steps: Conduct in-depth mining of the collected historical operation data, and use a hierarchical clustering algorithm to divide the historical operation data into multiple typical operation modes according to similarity, and obtain a typical operation mode set; Based on a set of typical operating modes, key performance indicators are calculated, including temperature adjustment rate, humidity adjustment rate and energy utilization efficiency, to obtain mode performance evaluation data; According to the model performance evaluation data, identify the operating points with significant performance differences, extract the operating rules and potential optimization space of the control system, and obtain the optimization direction; Adopting the incremental learning strategy, the data in the execution effect is added to the training set, the parameters of the temperature and humidity trend analysis model are updated, and the updated temperature and humidity trend analysis model is obtained; In actual operation, an experience buffer pool is maintained to store the state-action-reward sequence during operation. When the amount of data in the buffer pool reaches the threshold, the policy distillation technology is used to update the parameters of the control decision system to obtain the optimized control strategy. An anomaly detection method based on the isolation forest algorithm is used to monitor the operating parameters of the vehicle-mounted cold chain system in real time. When an abnormal pattern is detected, the diagnosis process is started and the control strategy is automatically adjusted to obtain an optimized temperature and humidity control system.

[0061] Specifically, the collected historical operation data is deeply mined, and the hierarchical clustering algorithm is used to divide the historical operation data into multiple typical operation modes according to similarity to obtain a typical operation mode set. The hierarchical clustering algorithm is a method of constructing a clustering hierarchy based on the distance or similarity between data points, which is divided into bottom-up agglomerative clustering and top-down divisive clustering. In the on-board cold chain temperature and humidity control, agglomerative hierarchical clustering is mainly used, and its processing steps include: first, each data point is regarded as a separate category, and the distance between each category is calculated; then the two categories with the closest distance are merged, and the distance between the merged category and other categories is recalculated; repeat this process until the termination condition is met. The distance metric uses the Euclidean distance, and the inter-class distance uses the average link method, that is, the distance between two categories is the average of the point-to-point distances between all categories. Clustering features include temperature distribution in the car, humidity distribution, ambient temperature, ambient humidity, vehicle speed, load conditions, door switch frequency, etc. Through cluster analysis, the historical operation data are divided into different typical operation modes, such as "full load operation under high temperature environment", "partial load under low temperature environment", "frequent door opening operation", etc. Each mode represents a specific combination of operating conditions. Based on the identified typical operation mode set, the key performance indicators under each mode are calculated, including temperature adjustment rate, humidity adjustment rate and energy utilization efficiency, to obtain the mode performance evaluation data. The temperature adjustment rate is defined as the change in compartment temperature per unit time, and the calculation method is the temperature change divided by the adjustment time; the definition and calculation method of the humidity adjustment rate are similar. The energy utilization efficiency is defined as the refrigeration / dehumidification amount generated per unit energy consumption, and the calculation method is the refrigeration / dehumidification amount divided by the product of system power and operation time. For each typical operation mode, the corresponding data segment is extracted, these key performance indicators are calculated, and statistical analysis is performed to obtain the average value, standard deviation, maximum value, minimum value and other statistics of each indicator. These statistics together constitute the mode performance evaluation data, reflecting the control effect and energy utilization of the cold chain system under different operation modes.

[0062] According to the mode performance evaluation data, the operating points with significant performance differences are identified, the operating rules and potential optimization space of the control system are extracted, and the optimization direction is obtained. The operating points with significant performance differences refer to the operating state points where the system performance indicators are significantly different from the average level under similar conditions under specific conditions. Identification methods include box plot analysis, Z score calculation, and outlier detection. By comparing the performance indicators under different operating modes, the operating modes with the best and worst performance are found, and the differences in conditions and control strategies are analyzed to extract the operating rules of the control system. For example, under what conditions the temperature adjustment rate is the fastest or the energy utilization efficiency is the highest, these rules can guide the optimization of the control strategy. At the same time, compare the performance differences in different time periods under the same operating mode to find potential optimization space. For example, under certain conditions, the energy consumption is too high but the temperature and humidity control effect is not ideal, indicating that there is room for improvement in the control strategy. Through this analysis, a clear optimization direction is obtained to provide guidance for subsequent model updates and strategy optimization.

[0063] Adopting the incremental learning strategy, the data in the execution effect is added to the training set, the parameters of the temperature and humidity trend analysis model are updated, and the updated temperature and humidity trend analysis model is obtained. Incremental learning is a learning method that does not require retraining the entire model, but updates the existing model based on new data. In this method, the newly collected data is added to the training set at regular intervals or when the prediction error exceeds the threshold, and only some parameters of the model are updated to reduce the computational overhead. The processing steps of incremental learning include: first, the latest execution effect data is selected and preprocessed in the same way as the original training data; then the processed data and a part of the historical data are combined into a new training batch; finally, this batch of data is used to update the model parameters, and generally only the weights of the last few layers are updated. The update frequency is automatically adjusted according to the prediction error: when the prediction error is continuously monitored to exceed the preset threshold (for example, the temperature error is greater than 0.5℃ or the humidity error is greater than 5%RH), the model update is triggered; when the prediction error continues to be lower than the threshold, the update frequency is reduced to reduce the computational burden. Through incremental learning, the temperature and humidity trend analysis model can continuously adapt to new operating data and improve prediction accuracy. In actual operation, the experience buffer pool is maintained to store the state-action-reward sequence during operation. When the amount of data in the buffer pool reaches the threshold, the policy distillation technology is used to update the parameters of the control decision system to obtain the optimized control strategy. The experience buffer pool is a data structure that stores the interactive experience of the intelligent agent. Each experience contains the current state, the action executed, the reward obtained, and the next state. In this method, each time a control decision is executed, the corresponding state-action-reward sequence is stored in the experience buffer pool. When the amount of data in the buffer pool reaches the set threshold (for example, 10,000 records), the policy update process is triggered. Policy distillation is a technology that transfers knowledge from one model (teacher model) to another model (student model). In this method, the current strategy is combined with the historical optimal strategy to generate an improved strategy that comprehensively considers historical experience and the characteristics of the new environment. The specific processing steps include: first, extracting data from the experience buffer pool, calculating the action distribution of the current strategy and the historical optimal strategy; then, updating the parameters of the current strategy with the goal of minimizing the KL divergence between the two distributions; and finally, applying the updated strategy to actual control. In this way, the control decision system can be continuously optimized and gradually approach the optimal control strategy.

[0064] An anomaly detection method based on the isolation forest algorithm is used to monitor the operating parameters of the vehicle-mounted cold chain system in real time. When an abnormal pattern is detected, the diagnosis process is started, the control strategy is automatically adjusted, and the optimized temperature and humidity control system is obtained. The isolation forest algorithm is an unsupervised anomaly detection method based on a tree structure. Its basic idea is that abnormal data points tend to be isolated faster in randomly constructed decision trees. The algorithm processing steps include: constructing multiple isolation trees, each tree divides the data points into different subspaces by randomly selecting features and split points; calculating the average path length required for each data point to be isolated. The shorter the path, the more likely the data point is an abnormal point. In this method, the operating parameters such as compressor current, evaporator temperature, condenser temperature, and system pressure are monitored in real time, and these parameters are input into the isolation forest model to calculate the anomaly score. When the anomaly score exceeds the threshold, the anomaly diagnosis process is triggered. The diagnosis process locates the cause of the anomaly based on a preset fault decision tree or a reasoning system based on a knowledge graph, and automatically adjusts the control strategy according to the diagnosis results. For example, when a refrigerant leak is detected, the system will reduce the compressor operating frequency and increase the fan speed to minimize the impact on the cold chain quality; at the same time, it will send early warning information to maintenance personnel and provide fault diagnosis and handling suggestions.

[0065] Taking the cold chain transportation of fresh milk as an example, a cold chain vehicle transports fresh milk for a long time, and in-depth mining is carried out by collecting half a year of operation data. First, the hierarchical clustering algorithm is used to analyze the historical data, and the data is divided into typical operation modes such as full load in summer, full load in winter, partial load in summer, partial load in winter, and frequent door opening operation. For each mode, the key performance indicators are calculated: for example, in the full load mode in summer, the temperature adjustment rate is an average of 1.2℃ per hour, the humidity adjustment rate is an average of 3.5% per hour, and the energy utilization efficiency is 2.8 kWh / ton. By comparing the performance data in different modes, it is found that the energy utilization efficiency in the full load mode in summer is significantly lower than that in other modes, which indicates that there is room for optimization of the control strategy under high temperature environment and full load conditions. Further analysis found that under this working condition, the compressor power often reaches the peak but maintains for a short time, and frequent start and stop lead to reduced energy efficiency. Based on this finding, the optimization direction is determined: improve the compressor control strategy under full load conditions in summer and reduce frequent start and stop. Subsequently, the recent operation data is applied to the incremental learning of the temperature and humidity trend analysis model, update the model parameters, and improve the prediction accuracy of temperature and humidity changes under high temperature conditions in summer. At the same time, the experience buffer pool is maintained during operation. After collecting enough state-action-reward sequences under full load conditions in summer, the policy distillation technology is applied to update the control decision system and generate a new control strategy: pre-adjust the fan speed, delay the temperature rise in the car, and reduce the start-up frequency of the compressor. While implementing the new strategy, the isolation forest algorithm continuously monitors the system operating parameters to ensure that the system can adjust the control strategy in time under abnormal circumstances. For example, when abnormal fluctuations in the compressor current are detected, the operating frequency is immediately reduced to avoid damage to the equipment.

[0066] The above describes the method for intelligently controlling the temperature and humidity of the vehicle cold chain in the embodiment of the present application. The following describes the system for intelligently controlling the temperature and humidity of the vehicle cold chain in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the system for intelligently controlling the temperature and humidity of the vehicle cold chain includes: A processing module is used to collect temperature and humidity data inside the vehicle compartment and temperature and humidity data outside the vehicle compartment through multiple temperature and humidity sensors in the vehicle compartment, and process the temperature and humidity data to obtain pre-processed temperature and humidity data; A division module is used to divide the temperature and humidity status in the compartment based on the preprocessed temperature and humidity data using a three-way decision model to obtain a status evaluation result of the temperature and humidity in the compartment; An analysis module, for using the preprocessed temperature and humidity data and the state assessment result to construct a temperature and humidity trend analysis model based on a long short-term memory network and an attention mechanism to obtain a temperature and humidity change trend; A construction module is used to construct a control decision system using multi-agent deep reinforcement learning according to the state evaluation result and the temperature and humidity change trend to obtain a temperature and humidity control decision; A control module, for dividing the temperature and humidity control decision into a plurality of operation modes based on the temperature and humidity control decision, and executing a control strategy in combination with the state evaluation result and the temperature and humidity change trend to obtain an execution effect; An updating module is used to adaptively update the control decision system according to the execution effect to obtain an optimized temperature and humidity control system.

[0067] Through the collaborative cooperation of the above-mentioned components, the present invention collects temperature and humidity data and performs preprocessing by arranging multiple temperature and humidity sensors in the car, constructs a comprehensive and accurate data foundation, avoids the one-sidedness of data caused by single-point sampling, and ensures the reliability of subsequent analysis and decision-making. A three-way decision model is used to divide the temperature and humidity state in the car, and the state is accurately classified into positive domain, negative domain and boundary domain, which overcomes the limitations of traditional dichotomy judgment, can more carefully identify the critical situation of temperature and humidity state, and provide a clear state representation for precise control. The temperature and humidity trend analysis model constructed based on long short-term memory network and attention mechanism effectively captures the temporal characteristics and long-term dependencies of temperature and humidity data, automatically identifies key historical data points through attention mechanism, improves the prediction accuracy of temperature and humidity changes in complex environments, and makes the control system forward-looking, able to predict and respond to temperature and humidity anomalies in advance. A control decision system is constructed by multi-agent deep reinforcement learning, and temperature control and humidity control are decomposed into collaborative subtasks. Through the collaborative mechanism between agents and the design of global reward function, the coupling problem in temperature and humidity control is solved, and the collaborative optimization and regulation of temperature and humidity are realized. The control decision is divided into multiple operation modes and the control strategy is executed based on the state evaluation results and the temperature and humidity change trend, which improves the adaptability of the control system to different working conditions. It can take the optimal control strategy for different demand scenarios such as normal, rapid temperature adjustment, rapid humidity adjustment, energy saving and emergency, improve the control accuracy and response speed, and reduce energy consumption. The mechanism of adaptive updating of the control decision system enables the system to have the ability of continuous learning and self-optimization, can mine the optimization space from the historical operation data, and continuously adjust and improve the control strategy according to the execution effect, thereby improving the robustness and long-term performance of the system. The present invention combines artificial intelligence algorithms with cold chain physical models for the specific application field of vehicle-mounted cold chain temperature and humidity control, gives full play to the advantages of deep reinforcement learning in complex decision-making problems, the expertise of long short-term memory networks in time series data processing, and the ability of attention mechanisms in key information identification. At the same time, by introducing physical constraints and multi-mode control architectures, it ensures that the algorithm output conforms to the actual physical laws and control requirements, realizes the substantial contribution of algorithm features to the solution, solves the complex scene control problems that are difficult to deal with by traditional methods, and significantly improves the accuracy, stability, foresight and energy efficiency of vehicle-mounted cold chain temperature and humidity control.

[0068] above Figure 2 From the perspective of modular functional entities, the system for intelligently controlling the temperature and humidity of the vehicle cold chain in an embodiment of the present invention is described in detail. From the perspective of hardware processing, the device for intelligently controlling the temperature and humidity of the vehicle cold chain in an embodiment of the present invention is described in detail.

[0069] Figure 3It is a structural schematic diagram of a device for intelligently controlling the temperature and humidity of a vehicle cold chain provided by an embodiment of the present invention. The device 300 for intelligently controlling the temperature and humidity of a vehicle cold chain may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the device 300 for intelligently controlling the temperature and humidity of a vehicle cold chain. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the device 300 for intelligently controlling the temperature and humidity of a vehicle cold chain to implement the steps of the above-mentioned method for intelligently controlling the temperature and humidity of a vehicle cold chain.

[0070] The device 300 for intelligently controlling vehicle-mounted cold chain temperature and humidity may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The device structure shown for intelligently controlling the temperature and humidity of the vehicle-mounted cold chain does not constitute a limitation on the device for intelligently controlling the temperature and humidity of the vehicle-mounted cold chain provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0071] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain.

[0072] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0073] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or partly contributed to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a device for intelligently controlling the temperature and humidity of the vehicle cold chain (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0074] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain, characterized in that: The method comprises: The temperature and humidity data inside the vehicle and the temperature and humidity data outside the vehicle are collected by using multiple temperature and humidity sensors inside the vehicle, and the temperature and humidity data are processed to obtain pre-processed temperature and humidity data; Based on the pre-processed temperature and humidity data, a three-way decision model is used to divide the temperature and humidity status in the compartment to obtain a status evaluation result of the temperature and humidity in the compartment; Using the preprocessed temperature and humidity data and the state assessment result, a temperature and humidity trend analysis model is constructed based on a long short-term memory network and an attention mechanism to obtain a temperature and humidity change trend; According to the state evaluation results and the temperature and humidity change trends, a control decision system is constructed using multi-agent deep reinforcement learning to obtain temperature and humidity control decisions; Based on the temperature and humidity control decision, the control decision is divided into multiple operation modes, and the control strategy is executed in combination with the state evaluation result and the temperature and humidity change trend to obtain the execution effect; According to the execution effect, the control decision system is adaptively updated to obtain an optimized temperature and humidity control system.

2. The method for intelligently controlling vehicle-mounted cold chain temperature and humidity according to claim 1, characterized in that: The method collects the temperature and humidity data inside the vehicle compartment and the temperature and humidity data outside the vehicle compartment through multiple temperature and humidity sensors inside the vehicle compartment, and processes the temperature and humidity data to obtain pre-processed temperature and humidity data, including: Four sensor nodes are set in the top area of ​​the carriage, six sensor nodes are set in the middle area, and four sensor nodes are set in the bottom area to form a three-dimensional monitoring network. The real-time temperature and humidity values ​​at each point in the carriage are collected to obtain the temperature and humidity data set in the carriage. An ambient temperature and humidity sensor is set outside the vehicle compartment to collect the external ambient temperature, ambient humidity, vehicle speed and door status, and obtain a data set of the vehicle compartment exterior; Using a local anomaly factor algorithm to perform outlier detection on the in-car temperature and humidity data set and the external car data set, identifying data that differs too much from adjacent time points, and obtaining abnormal marked data; Based on the valid data at adjacent time points, a time series interpolation algorithm is applied to the abnormally marked data to complete the missing values ​​and obtain a complete data set; Applying a sliding average method to smooth the complete data set to obtain temperature and humidity data after noise reduction; The de-noised temperature and humidity data are combined with the cargo type information to record the cargo's physical and chemical properties and the optimal storage temperature and humidity range, and the pre-processed temperature and humidity data are obtained for temperature and humidity status assessment.

3. The method for intelligently controlling vehicle-mounted cold chain temperature and humidity according to claim 1, characterized in that: Based on the pre-processed temperature and humidity data, a three-way decision model is used to divide the temperature and humidity state in the compartment to obtain a state evaluation result of the temperature and humidity in the compartment, including: According to the optimal storage conditions of the transported goods, the temperature threshold and humidity threshold are set to obtain the target range of temperature and humidity; Analyzing the preprocessed temperature and humidity data, when the temperature of all points in the vehicle compartment is within the range of the temperature threshold and the humidity is within the range of the humidity threshold, dividing the system state into a positive domain, and obtaining an ideal state identification; Analyzing the pre-processed temperature and humidity data, when the temperature or humidity of a point in the vehicle compartment exceeds the range of the temperature threshold or a certain value of the range of the humidity threshold, dividing the system state into a negative domain, and obtaining an abnormal state mark; Analyzing the preprocessed temperature and humidity data, dividing the states that do not belong to the positive domain and the negative domain into boundary domains, and obtaining critical state identifiers; Calculate the spatial uniformity index of temperature and humidity in the carriage, including the standard deviation of temperature and humidity at each measuring point in the carriage, and obtain the evaluation value of temperature and humidity distribution uniformity; The ideal state identifier or the abnormal state identifier or the critical state identifier is combined with the temperature and humidity distribution uniformity evaluation value to generate a comprehensive state evaluation result as an input for control decision.

4. The method for intelligently controlling vehicle-mounted cold chain temperature and humidity according to claim 1, characterized in that: The pre-processed temperature and humidity data and the state evaluation result are used to construct a temperature and humidity trend analysis model based on a long short-term memory network and an attention mechanism to obtain a temperature and humidity change trend, including: Dividing the preprocessed temperature and humidity data into time windows, each time window containing a plurality of continuous sampling points, forming an input sequence, and obtaining time series feature data; Performing feature normalization processing on the time series feature data, converting temperature, humidity, and environmental parameters into a unified dimension, and obtaining normalized feature data; Input the normalized feature data into a three-layer long short-term memory network for processing, wherein the first layer contains 128 neurons, the second layer contains 64 neurons, and the third layer contains 32 neurons, and a discard layer is added between each layer to obtain a time series feature representation; Applying an attention mechanism to weight the time series feature representation, and sorting the importance of historical data points by calculating weight coefficients to obtain a weighted feature representation; The weighted feature representation is mapped to the output space through a fully connected layer to obtain the predicted values ​​of temperature and humidity change trends at multiple time points in the future; A cabin heat transfer model is added to the predicted value of the temperature and humidity change trend as a physical constraint, and the model parameters are optimized through a loss function composed of a mean square error loss and a physical model constraint regularization term to obtain a temperature and humidity change trend that conforms to physical laws.

5. The method for intelligently controlling vehicle-mounted cold chain temperature and humidity according to claim 1, characterized in that: According to the state evaluation result and the temperature and humidity change trend, a control decision system is constructed by using multi-agent deep reinforcement learning to obtain a temperature and humidity control decision, including: The control task is decomposed into two subtasks: temperature control and humidity control. The temperature control agent and humidity control agent are constructed respectively to form a multi-agent control framework. Define the control system state space, including the current temperature and humidity distribution in the car, the ambient temperature and humidity, the vehicle speed, the door state, the state evaluation result and the temperature and humidity change trend, and obtain the agent state representation; Define the action space of the temperature control agent including compressor power, fan speed and circulation damper opening, and the action space of the humidity control agent including desiccant wheel speed and regeneration heater power, and obtain the control action set that the agent can execute; Constructing a reward function for a temperature control agent including a temperature deviation penalty term, a temperature non-uniformity penalty term, and an energy consumption penalty term, and a reward function for a humidity control agent including a humidity deviation penalty term, a humidity non-uniformity penalty term, and an energy consumption penalty term, and obtaining the agent optimization target; Design a global reward function, combine the temperature control agent reward, humidity control agent reward and collaboration reward items, train through the proximal policy optimization algorithm, and obtain the agent policy network parameters; The state evaluation result and the temperature and humidity change trend are input into the trained intelligent agent strategy network to generate an optimal control action sequence and obtain a temperature and humidity control decision.

6. The method for intelligently controlling vehicle-mounted cold chain temperature and humidity according to claim 1, characterized in that: Based on the temperature and humidity control decision, the control decision is divided into multiple operation modes, and the control strategy is executed in combination with the state evaluation result and the temperature and humidity change trend to obtain the execution effect, including: According to the state evaluation result, the control strategy is divided into five operation modes: normal mode, rapid temperature adjustment mode, rapid humidity adjustment mode, energy saving mode and emergency mode, to obtain an operation mode set; For each operation mode in the operation mode set, a corresponding actuator control strategy matrix is ​​designed, and the temperature and humidity control decision is mapped into specific actuator control parameters to obtain an actuator control parameter set; Adopting a hierarchical control architecture, generating control targets of each subsystem based on the operation mode set and the temperature and humidity control decision, and obtaining a subsystem control target set; Applying a proportional-integral control algorithm to the subsystem control target set to generate actuator control signals, ensure that each actuator accurately tracks the control target, and obtain control execution instructions; The control execution instruction adjusts the compressor power, fan speed, circulation damper opening, desiccant wheel speed and regeneration heater power, performs temperature and humidity adjustment operations, and obtains real-time temperature and humidity change data; The real-time temperature and humidity change data are compared with the temperature and humidity change trend, the temperature and humidity deviation value is calculated, and the execution effect of the control strategy is determined as a basis for adaptive updating.

7. The method for intelligently controlling vehicle-mounted cold chain temperature and humidity according to claim 1, characterized in that: According to the execution effect, the control decision system is adaptively updated to obtain an optimized temperature and humidity control system, including: Conduct in-depth mining of the collected historical operation data, and use a hierarchical clustering algorithm to divide the historical operation data into multiple typical operation modes according to similarity, and obtain a typical operation mode set; Based on the typical operation mode set, key performance indicators are calculated, including temperature adjustment rate, humidity adjustment rate and energy utilization efficiency, to obtain mode performance evaluation data; According to the performance evaluation data of the model, the operating points with significant performance differences are identified, the operating rules and potential optimization space of the control system are extracted, and the optimization direction is obtained; Adopting an incremental learning strategy, adding the data in the execution effect to a training set, updating the parameters of the temperature and humidity trend analysis model, and obtaining an updated temperature and humidity trend analysis model; In actual operation, an experience buffer pool is maintained to store the state-action-reward sequence during operation. When the amount of data in the buffer pool reaches a threshold, the policy distillation technology is used to update the parameters of the control decision system to obtain an optimized control strategy. An anomaly detection method based on the isolation forest algorithm is used to monitor the operating parameters of the vehicle-mounted cold chain system in real time. When an abnormal pattern is detected, the diagnosis process is started and the control strategy is automatically adjusted to obtain an optimized temperature and humidity control system.

8. A system for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain, characterized in that: Used to implement the method for intelligently controlling the temperature and humidity of a vehicle cold chain according to any one of claims 1 to 7, the system for intelligently controlling the temperature and humidity of a vehicle cold chain comprises: A processing module, used to collect temperature and humidity data inside the vehicle compartment and temperature and humidity data outside the vehicle compartment through multiple temperature and humidity sensors in the vehicle compartment, and process the temperature and humidity data to obtain pre-processed temperature and humidity data; A division module, for dividing the temperature and humidity state in the compartment by using a three-way decision model based on the preprocessed temperature and humidity data, and obtaining a state evaluation result of the temperature and humidity in the compartment; An analysis module, for using the preprocessed temperature and humidity data and the state assessment result to construct a temperature and humidity trend analysis model based on a long short-term memory network and an attention mechanism to obtain a temperature and humidity change trend; A construction module is used to construct a control decision system using multi-agent deep reinforcement learning according to the state evaluation result and the temperature and humidity change trend to obtain a temperature and humidity control decision; A control module, for dividing the temperature and humidity control decision into a plurality of operation modes based on the temperature and humidity control decision, and executing a control strategy in combination with the state evaluation result and the temperature and humidity change trend to obtain an execution effect; An updating module is used to adaptively update the control decision system according to the execution effect to obtain an optimized temperature and humidity control system.

9. A device for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the method for intelligently controlling the temperature and humidity of a vehicle-mounted cold chain as described in any one of claims 1 to 7.

Images