Multimodal AIGC cold chain logistics information processing method and system

Abstract

The present application relates to the field of cold chain logistics information technology, and discloses a multimodal AIGC cold chain logistics information processing method and system, which includes: based on the logistics information AIGC model, feature extraction and feature enhancement of video data, temperature data and location data in the cold chain logistics system to obtain an enhanced feature matrix; digital twin modeling to obtain a digital twin model of the logistics system, and construct a state parameter space; deep reinforcement learning and information delay index calculation to generate a real-time data synchronization solution; construct a distribution network graph structure, and perform deep graph network coding on the distribution network graph structure to obtain a distribution network representation vector; perform multi-agent collaborative decision-making on the distribution network representation vector to obtain a cold chain distribution scheduling plan, and update the cold chain distribution scheduling plan to the logistics system digital twin model. In this way, the intelligent scheduling and precise control of the cold chain logistics system in complex scenarios are realized.

Description

Technical Field

[0001] The present application relates to the field of cold chain logistics information technology, and in particular to a multimodal AIGC cold chain logistics information processing method and system. Background Art

[0002] As an important part of the modern logistics system, cold chain logistics is of great significance in ensuring food safety, pharmaceutical transportation and other fields. With the development of Internet of Things technology, a large amount of multimodal data such as video surveillance, temperature sensing and location tracking are collected in real time, providing a data foundation for the intelligent management of cold chain logistics. However, traditional cold chain logistics information processing methods often separate and process data of different modes, making it difficult to fully explore the deep correlation between data, resulting in insufficient accuracy and real-time performance of distribution decisions.

[0003] The introduction of digital twin technology provides a new idea for real-time monitoring and prediction of cold chain logistics systems. However, due to the delay in data synchronization between physical entities and digital models, and the difficulty of the synchronization mechanism to adapt to complex and changing distribution scenarios, the actual application effect of the digital twin model of the logistics system is affected. Especially under the constraints of multiple objectives such as temperature control, delivery timeliness and energy consumption, how to achieve the reasonable scheduling of distribution resources has become an urgent problem to be solved. Summary of the invention

[0004] The present application provides a multimodal AIGC cold chain logistics information processing method and system, thereby realizing intelligent scheduling and precise control of the cold chain logistics system in complex scenarios.

[0005] The first aspect of the present application provides a multimodal AIGC cold chain logistics information processing method, the multimodal AIGC cold chain logistics information processing method comprising:

[0006] Based on the logistics information AIGC model, feature extraction and feature enhancement are performed on the video data, temperature data and location data in the cold chain logistics system to obtain an enhanced feature matrix;

[0007] Perform digital twin modeling based on the enhanced feature matrix to obtain a digital twin model of the logistics system and construct a state parameter space;

[0008] Perform deep reinforcement learning and information delay index calculation on the state parameter space, introduce temperature control parameters and refrigeration energy consumption parameters, and generate a real-time data synchronization solution;

[0009] Constructing a distribution network graph structure according to the real-time data synchronization scheme, and performing deep graph network coding on the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure to obtain a distribution network representation vector;

[0010] Multi-agent collaborative decision-making is performed on the distribution network representation vector, and a cold chain distribution scheduling plan is obtained based on multi-objective constraints such as temperature control accuracy, distribution timeliness and energy consumption. The cold chain distribution scheduling plan is then updated to the digital twin model of the logistics system.

[0011] The second aspect of the present application provides a multimodal AIGC cold chain logistics information processing system, the multimodal AIGC cold chain logistics information processing system comprising:

[0012] The feature enhancement module is used to extract and enhance the features of the video data, temperature data and location data in the cold chain logistics system based on the logistics information AIGC model to obtain an enhanced feature matrix;

[0013] A modeling module, used to perform digital twin modeling based on the enhanced feature matrix, obtain a digital twin model of the logistics system, and construct a state parameter space;

[0014] A calculation module, used to perform deep reinforcement learning and information delay index calculation on the state parameter space, and introduce temperature control parameters and refrigeration energy consumption parameters to generate a real-time data synchronization solution;

[0015] An encoding module, used to construct a distribution network graph structure according to the real-time data synchronization scheme, and perform deep graph network encoding on the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure to obtain a distribution network representation vector;

[0016] The updating module is used to perform multi-agent collaborative decision-making on the distribution network representation vector, obtain a cold chain distribution scheduling plan based on multi-objective constraints such as temperature control accuracy, distribution timeliness and energy consumption, and update the cold chain distribution scheduling plan to the digital twin model of the logistics system.

[0017] The third aspect of the present application provides an electronic device, 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 electronic device executes the above-mentioned multimodal AIGC cold chain logistics information processing method.

[0018] The fourth aspect of the present application provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned multimodal AIGC cold chain logistics information processing method.

[0019] Compared with the prior art, the present application has the following beneficial effects: the feature enhancement and fusion of multimodal data such as video, temperature and location are performed through the logistics information AIGC model, which significantly improves the data representation ability, and the conditional variational autoencoder is combined to realize the efficient integration of data of different modalities; the synchronization mechanism of deep reinforcement learning and the real-time data synchronization scheme are adopted to optimize the information transmission efficiency of the digital twin model of the logistics system and reduce the system latency; the key features of the distribution network are extracted through deep graph network coding and hierarchical node structure, and the attention mechanism enhances the quality of information transmission between nodes; the reasonable allocation of distribution tasks is realized based on multi-agent collaborative decision-making, and the decision optimization under multi-objective constraints balances performance indicators such as temperature control, delivery timeliness and energy consumption; the hierarchical decision-making architecture and closed-loop control mechanism are adopted to ensure the continuous optimization capability of the system and realize the intelligent scheduling and precise control of the cold chain logistics system in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0022] Figure 1 It is a flowchart of a multi-modal AIGC cold chain logistics information processing method provided by an embodiment of the present invention;

[0023] Figure 2 is a schematic block diagram of the structure of a multi-modal AIGC cold chain logistics information processing system provided by an embodiment of the present invention;

[0024] Figure 3 It is a schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0027] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0028] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the multimodal AIGC cold chain logistics information processing method in the embodiment of the present application includes:

[0029] Step 100: Based on the logistics information AIGC model, feature extraction and feature enhancement are performed on the video data, temperature data and location data in the cold chain logistics system to obtain an enhanced feature matrix;

[0030] It is understandable that the execution subject of the present application may be a multimodal AIGC cold chain logistics information processing system, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0031] Specifically, video data, temperature data and location data in the cold chain logistics environment are obtained through sensors and monitoring equipment. For the processing of video data, spatiotemporal feature extraction technology is used to extract dynamic and static information in the video and generate a video feature sequence. Spatiotemporal feature extraction can capture the distribution of objects in space and analyze their changes in time. For temperature data, wavelet transform is used for processing to effectively extract subtle changes in temperature signals and obtain temperature feature sequences. Wavelet transform can analyze data at different frequencies, thereby identifying instantaneous changes in temperature fluctuations, which helps to monitor the stability of the cold chain environment. For location data, trajectory smoothing and dimensionality reduction processing are used to eliminate noise and reduce data dimensions to obtain a clearer location information sequence. Trajectory smoothing is achieved through algorithms such as Kalman filtering to ensure the accuracy and consistency of location data. The video feature sequence, temperature feature sequence and location feature sequence are aligned in the time dimension to obtain an aligned feature sequence. The aligned feature sequence is standardized to eliminate the dimensional differences between different features, so that the data can be better accepted by the model, and a multimodal feature matrix is ​​obtained to fully reflect the status of the cold chain logistics system. The multimodal feature matrix is ​​input into the encoder layer of the logistics information AIGC model. The encoder layer design includes a multi-head self-attention module and a feedforward neural network. The multi-head self-attention module has 8 attention heads, each with a dimension of 64. The multi-head self-attention module captures the relationship between features by focusing on different feature parts and generates an attention feature matrix. The attention feature matrix is ​​input into the decoder layer. The decoder layer also includes a feedforward neural network and a cross-attention module. The cross-attention module is used to fuse features from different modalities, so that the decoder can generate richer output features. The feedforward neural network consists of two hidden layers with 2048 and 1024 nodes respectively. Through this deep structure, the network can better learn the complex mapping relationship between features and generate a decoding feature matrix. The decoding feature matrix is ​​input into the feature enhancement layer, which adopts a conditional variational autoencoder structure and generates latent variables through reparameterization techniques. Reparameterization techniques are an effective generative model method that enables the model to maintain differentiability to data distribution during training and improve training efficiency and effect. The features are reconstructed through the decoding network to obtain an enhanced feature matrix.

[0032] The decoded feature matrix is ​​input into the feature enhancement layer, and the matrix is ​​transformed in dimension. The feature dimension is mapped to 512-dimensional space through the fully connected layer to obtain the transformed feature matrix. The purpose of this process is to enhance the correlation and expression ability between features in a higher-dimensional space. The transformed feature matrix is ​​encoded and a three-layer convolutional neural network is used to extract features. The size of each convolution kernel is 3×3, the step size is 2, and the number of channels is set to 64, 128, and 256 respectively. Through this structure, the convolutional network can capture the spatial information of local features. At the same time, as the number of channels increases, the model can extract higher-level abstract features. After multiple layers of convolution processing, the encoded features are obtained. The probability distribution parameters of the encoded features are calculated. Through two parallel fully connected layers, the mean vector and variance vector are output respectively to form complete probability distribution parameters. The mean vector and variance vector are key components in the conditional variational autoencoder. The former determines the center position of the latent variable, and the latter reflects the distribution range of the latent variable. On this basis, the reparameterization technique is used to randomly sample noise vectors from the standard normal distribution and combine the mean vector and variance vector to generate latent variables. The reparameterization process ensures the differentiability of the model, so that the model parameters can be effectively optimized during the training process. The latent variables are concatenated with the conditional vector. The conditional vector consists of temperature control parameters, delivery time parameters, and energy consumption parameters, and is intended to integrate external control factors into the feature enhancement process. In this way, the latent variables can be combined with environmental conditions to form a conditional enhancement vector, ensuring that the model can fully consider the needs and limitations of different scenarios when reconstructing features. The conditional enhancement vector is upsampled and a three-layer deconvolution network is used to restore the feature size. The kernel size of each deconvolution layer is 4×4, the step size is 2, and the number of channels is 256, 128, and 64, respectively. The role of the deconvolution network is to gradually restore the latent features to higher spatial dimensions, thereby restoring the shape and information of the original features and ensuring that the reconstructed features contain enough detail information for subsequent fusion processing. After the reconstruction of the features is completed, residual connection processing is used to fuse the reconstructed features with the decoded feature matrix. Residual connection can effectively alleviate the gradient vanishing problem in deep networks, promote the smooth transmission of information flow, and make it easier for the model to learn complex feature mapping relationships. After feature fusion, fused features are obtained. The fused features are normalized to adjust the feature distribution through layer normalization. Layer normalization can improve the training efficiency and convergence speed of the model, so that the model maintains good performance when processing data of different scales and ranges. The normalized features are mapped to the original feature dimensions through a layer of fully connected network, and finally an enhanced feature matrix is ​​obtained.

[0033] Step 200: Perform digital twin modeling based on the enhanced feature matrix to obtain a digital twin model of the logistics system and construct a state parameter space;

[0034] Specifically, the enhanced feature matrix is ​​segmented according to the time window, and the time series data is divided into several smaller time periods, so as to extract meaningful time series features and statistical features in each data segment. The selection of each time window needs to be adjusted according to the specific application scenario and the frequency of data changes to ensure that the key dynamic information in the cold chain logistics system is captured. After extracting the time series features, the statistical features of each segment of data, such as mean, variance, extreme value, etc., are calculated to reflect the state changes in each time window. Through this process, a feature component sequence is obtained, including key feature information from each time window. The feature component sequence is screened, and the key features are retained according to the importance weights of the feature components. The importance of features is evaluated by various methods, such as model-based feature selection, information gain, correlation analysis, etc. Through these methods, the features that have the greatest impact on the state of the cold chain logistics system can be identified to form a target feature sequence. Based on the target feature sequence, the cold chain state model is constructed, involving three main aspects, namely, the temperature change law, the position change law, and the energy consumption change law. The modeling of these three aspects is the core of the digital twin technology, which can simulate the performance of the cold chain system under different conditions. The modeling of temperature variation uses time series prediction techniques, such as the autoregressive moving average (ARMA) model or the long short-term memory (LSTM) network, to capture the dynamic changes of temperature over time. At the same time, the location variation is realized through trajectory analysis and prediction algorithms, such as interpolation methods based on historical location data or dynamic time warping (DTW) technology, to optimize the delivery route and improve efficiency. The energy consumption variation is constructed through the energy consumption prediction model to identify the energy consumption characteristics under different states. The cold chain state model is digitally modeled to obtain the digital twin model of the logistics system. By combining real-time data streams and historical data, the digital twin model is dynamically updated to keep pace with the actual system. The output results of the digital twin model are state mapped to construct the corresponding relationship between the temperature state, location state and energy consumption state to form a state mapping matrix, which reflects the correlation between different state variables. The state mapping matrix is ​​supplemented with features to add delivery time constraints and task priority information. The delivery time constraint helps the model consider the urgency of time when scheduling, while the task priority information ensures that important tasks can be processed in a timely manner. By supplementing the features, the target state vector is obtained. A state transition relationship graph is constructed based on the target state vector. The nodes of the state transition relationship graph represent the state parameters, and the edges represent the correlation between the parameters. In this relationship graph, the node settings should cover all important state parameters, such as temperature, position, energy consumption, etc., and the edge weights are assigned by calculating the correlation between the state parameters. The state transition relationship graph can help identify the key state parameter subsets of the system and reveal the complex relationship between the parameters, thereby constructing a comprehensive state parameter space.

[0035] Step 300: Perform deep reinforcement learning and information delay index calculation on the state parameter space, introduce temperature control parameters and refrigeration energy consumption parameters, and generate a real-time data synchronization solution;

[0036] It should be noted that the state parameter space is divided into action space, temperature control, cooling power and data transmission time slot are set as action variables, and a reasonable value range is set for each action variable to form a complete action space. The state parameter space and action space are input into the policy network for training. The policy network is designed to contain three fully connected layers with 256, 128 and 64 nodes respectively. The choice of this structure is to balance the expressive power and computational efficiency of the model. The deeper network can capture the complex relationship between state and action. After obtaining the output through forward propagation, the policy network generates a set of action probability distributions, which reflects the selection probability of each action under a given state. The generated action probability distribution is sampled, and the temperature control action, cooling power action and data transmission time slot action are selected to obtain a specific action combination. While considering the environmental state, the output of the policy network is combined to select the optimal action combination. After executing the selected action combination, the timestamp and data update time of each state transition are recorded, and the information delay index value is calculated. Introducing the delay index as a reward signal helps guide model learning and optimize its decision-making process. The state parameter space, action combination and reward signal are input into the value function network to evaluate the value of the current state. The value function network adopts a fully connected structure, including two fully connected layers with 128 and 64 nodes respectively. The main function of the value function network is to estimate the expected benefits of taking a specific action under a given state, so as to provide feedback for the optimization strategy. By analyzing the reward signal, the network can learn the effectiveness of actions under different states, providing a basis for reinforcement learning for the policy network. The state value evaluation results are optimized for action. Constraints are constructed based on temperature control parameters and refrigeration energy consumption parameters to ensure that the decisions in the optimization process meet the actual cold chain operation constraints. The policy gradient method is used to update the network parameters, which can effectively adjust the parameters in the policy network to improve the quality of action selection. After multiple iterations, the optimized synchronization strategy is obtained. The optimized synchronization strategy is applied to real-time data transmission, and the data update time interval and transmission success rate are recorded to evaluate the synchronization performance indicators. The data update time interval refers to the timeliness of the system response during the data update process, while the transmission success rate reflects the probability of successful data transmission under specific conditions. By monitoring and analyzing these indicators, the potential bottlenecks and optimization space in the system can be effectively identified, and then the resource allocation can be adjusted to improve the overall system performance. Communication resources are allocated according to synchronization performance indicators to generate an effective real-time data synchronization solution. Reasonable resource allocation can ensure that the system can still maintain good data transmission performance under different workloads and environmental conditions, providing support for the intelligent management of cold chain logistics.

[0037] Step 400: construct a distribution network graph structure according to the real-time data synchronization solution, and perform deep graph network coding on the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure to obtain a distribution network representation vector;

[0038] Specifically, the data update time interval in the real-time data synchronization scheme is used as the data sampling period of the distribution network nodes. The temperature change sequence, energy consumption data sequence and distribution time sequence of each distribution point are periodically collected to obtain a synchronized data set. The distribution network nodes are constructed according to the data transmission time slots in the synchronized data set. The distribution points with a data transmission success rate higher than the preset threshold are set as master nodes, and the remaining distribution points are set as slave nodes to form a hierarchical node set. The hierarchical structure can effectively distinguish the importance of different distribution points in the network, ensure that the status of the master node is given priority during data transmission, and improve the reliability and efficiency of the entire distribution network. Based on the allocation of communication resources in the real-time data synchronization scheme, the hierarchical node set is connected. A connection relationship is established between the node pairs sharing the same transmission channel to obtain a network topology structure, which reflects the connection and interaction between different distribution points. A feature vector is constructed for each node in the network topology. The temperature control action, the cooling power action and the time slot allocation action are combined into node state features to more comprehensively describe the operating status of each node. In this way, a node feature set is formed. The edge features are constructed based on the node feature set, and the temperature difference coefficient, energy consumption correlation coefficient and timing matching degree between the node pairs are calculated to obtain the edge feature matrix. The node feature set and the edge feature matrix are input into the graph attention network for processing. In the graph attention network, the information transmission between nodes will be calculated based on the weight of the feature to obtain the attention feature map. The graph attention network can dynamically adjust the information flow between different nodes, set different aggregation weights based on the master-slave relationship of the nodes, and ensure that the features of important nodes can be fully utilized. After the weight calculation, the obtained attention feature map is aggregated, and the node features are updated through the message passing mechanism to form a new fusion feature matrix. The process of information aggregation can effectively integrate information from different nodes, improve the coordination ability of the entire distribution network, and make nodes more efficient and accurate in transmitting information. The fusion feature matrix is ​​globally pooled to compress the graph structure information into a vector representation of a fixed dimension to obtain the distribution network representation vector. Through global pooling, the dimension of the data can be reduced and important feature information can be retained.

[0039] Step 500: Perform multi-agent collaborative decision-making on the distribution network representation vector, obtain a cold chain distribution scheduling plan based on multi-objective constraints such as temperature control accuracy, distribution timeliness, and energy consumption, and update the cold chain distribution scheduling plan to the digital twin model of the logistics system.

[0040] Specifically, an agent decision environment is constructed based on the distribution network representation vector. Each distribution node is regarded as an independent agent, forming a dynamic agent interaction environment, so that each agent can share information and resources in interaction to achieve the goal of collaborative scheduling. The constructed agent interaction environment is input into the collaborative decision network. The network consists of a central agent and multiple executive agents, in which the central agent is responsible for global strategy generation, while the executive agent focuses on the execution of local tasks. The hierarchical decision architecture can effectively combine global decision-making with local execution to improve the response speed and efficiency of the entire distribution system. The central agent makes decisions based on global state information, while the executive agent makes flexible adjustments for specific distribution tasks to achieve more accurate scheduling. After establishing the hierarchical decision architecture, a multi-objective reward design is carried out to provide effective guidance for the decision-making process of the agent. A reward function is constructed based on temperature control accuracy, delivery timeliness and energy consumption. In this function, the weight of temperature control accuracy is set to 0.4, the weight of delivery timeliness is 0.3, and the weight of energy consumption is also 0.3, forming a target weight matrix. In this way, it is possible to ensure that each target is reasonably balanced during the optimization process and improve the overall scheduling effect. The target weight matrix is ​​input into the policy update module, and the action selection is performed through the dual Q network. The dual Q network consists of an online network and a target network, in which the online network is responsible for policy update and the target network is used for policy evaluation. The advantage of this structure is that it can effectively reduce the overfitting problem in the policy update process and improve the stability and efficiency of the learning process. Through continuous training of the online network, the agent can obtain more accurate action selection, thereby making better decisions in a complex distribution environment. The agent strategy set is collaboratively optimized. Based on the state parameters of the digital twin model of the logistics system, a communication protocol between agents is constructed to promote information sharing and collaboration between agents and improve the overall coordination of the system. In this process, the influence weights between agents are calculated through the attention mechanism to obtain the collaborative decision results. This mechanism can dynamically adjust the information flow according to the relative importance between agents, ensure the priority transmission of important information, and improve the efficiency of the entire scheduling process. According to the collaborative decision results, execution instructions are generated, which include distribution path planning, temperature control parameters, and energy consumption allocation schemes. The formulation of these execution instructions needs to comprehensively consider the status of each distribution node and dynamically adjust them based on real-time data to ensure effective execution in actual operation. Combining the execution instructions into a cold chain distribution scheduling plan and mapping it with a digital twin can achieve real-time monitoring and evaluation of the scheduling effect. Through digital twin mapping, the scheduling execution data can be fed back to the digital twin model of the logistics system in a timely manner to ensure that the system can self-adjust and optimize according to the real-time operating status.

[0041] In the embodiments of the present application, the features of multimodal data such as video, temperature and location are enhanced and fused through the logistics information AIGC model, which significantly improves the data representation ability, and the conditional variational autoencoder is combined to achieve efficient integration of data of different modalities; the synchronization mechanism of deep reinforcement learning and the real-time data synchronization scheme are used to optimize the information transmission efficiency of the digital twin model of the logistics system and reduce the system latency; the key features of the distribution network are extracted through deep graph network coding and hierarchical node structure, and the attention mechanism enhances the quality of information transmission between nodes; the reasonable allocation of distribution tasks is achieved based on multi-agent collaborative decision-making, and the decision optimization under multi-objective constraints balances performance indicators such as temperature control, delivery timeliness and energy consumption; the hierarchical decision-making architecture and closed-loop control mechanism are adopted to ensure the continuous optimization capability of the system and realize the intelligent scheduling and precise control of the cold chain logistics system in complex scenarios.

[0042] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0043] Obtain video data, temperature data, and location data in the cold chain logistics system;

[0044] Extract spatiotemporal features from video data to obtain a video feature sequence; perform wavelet transform on temperature data to obtain a temperature feature sequence; perform trajectory smoothing and dimensionality reduction on position data to obtain a position feature sequence;

[0045] The video feature sequence, the temperature feature sequence and the position feature sequence are aligned in the time dimension to obtain an aligned feature sequence, and the aligned feature sequence is standardized to obtain a multimodal feature matrix;

[0046] The multimodal feature matrix is ​​input into the encoder layer of the logistics information AIGC model. The encoder layer contains a multi-head self-attention module and a feedforward neural network. The multi-head self-attention module is set with 8 attention heads, and the dimension of each attention head is 64, and the attention feature matrix is ​​obtained.

[0047] The attention feature matrix is ​​input into the decoder layer of the logistics information AIGC model. The decoder layer includes a cross attention module and a feedforward neural network. The cross attention module is used for feature fusion. The feedforward neural network includes two hidden layers with 2048 and 1024 nodes respectively, and the decoding feature matrix is ​​obtained.

[0048] The decoded feature matrix is ​​input into the feature enhancement layer. The feature enhancement layer adopts the conditional variational autoencoder structure, generates latent variables through the reparameterization technique, and reconstructs the features through the decoding network to obtain the enhanced feature matrix.

[0049] Specifically, video data, temperature data, and location data are obtained. Video data captures the status of goods during cold chain transportation through surveillance cameras, temperature data monitors temperature changes in the cold chain environment through sensors in real time, and location data collects the geographic location of the transport vehicle through the GPS positioning system. Spatiotemporal feature extraction is performed on the video data to obtain a video feature sequence. The key features in each frame are extracted through a convolutional neural network (CNN), and the information in the time dimension is captured using a long short-term memory network (LSTM) to form a feature sequence that represents the temporal changes in the video. When processing temperature data, wavelet transform technology is used to obtain a temperature feature sequence. Wavelet transform can effectively extract local features in time series, provide the performance of temperature data at different frequencies, and define the temperature sequence. The wavelet transform of is:

[0050] ;

[0051] in, is the scale factor, is the translation factor, is the mother wavelet function. When processing the position data, the trajectory is smoothed to eliminate the noise in the position data, and then the position features are extracted through dimensionality reduction methods (such as principal component analysis) to obtain the position feature sequence. The video feature sequence, temperature feature sequence and position feature sequence are aligned in the time dimension to ensure that the data of different modes are consistent in time and avoid information confusion caused by time inconsistency. After alignment, the aligned feature sequence is standardized, which is achieved through the Z-score standardization formula:

[0052] ;

[0053] in, is the original eigenvalue, is the feature mean, is the feature standard deviation, and finally the multimodal feature matrix is ​​obtained. The multimodal feature matrix is ​​input into the encoder layer of the logistics information AIGC model. The encoder layer contains a multi-head self-attention module and a feedforward neural network. The multi-head self-attention module can process multiple feature subsets in parallel by setting 8 attention heads, each with a dimension of 64. By calculating the attention weight, different parts of the input features are weighted to more effectively extract important information. This process is expressed by the following formula:

[0054] ;

[0055] in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key. The output of this module is the attention feature matrix. The attention feature matrix is ​​input to the decoder layer. The decoder layer consists of a cross-attention module and a feedforward neural network. The cross-attention module is used for feature fusion and combines the contextual information from the encoder. The feedforward neural network contains two hidden layers with 2048 and 1024 nodes respectively. This design ensures that the model has sufficient expressive power to capture complex feature relationships. The output of the decoder is the decoding feature matrix. The decoding feature matrix is ​​input to the feature enhancement layer. This layer adopts the conditional variational autoencoder structure, generates latent variables through the reparameterization technique, and reconstructs the features through the decoding network to obtain the enhanced feature matrix. The reparameterization technique allows the sampling operation of random variables to be separated from the deterministic operation during training, so that the model can be optimized through backpropagation. The core formula of this process is:

[0056] ;

[0057] in, is a latent variable, is the mean, is the standard deviation, is the noise sampled from the standard normal distribution. In this way, the model can generate richer feature information and effectively enhance the representation ability of the original data.

[0058] In a specific embodiment, the step of executing the decoded feature matrix is ​​input into the feature enhancement layer, the feature enhancement layer adopts the conditional variational autoencoder structure, generates latent variables through the reparameterization technique, and reconstructs the features through the decoding network. The process of obtaining the enhanced feature matrix can specifically include the following steps:

[0059] The decoded feature matrix is ​​transformed in dimension, and the feature dimension is mapped to a 512-dimensional space through a fully connected layer to obtain a transformed feature matrix;

[0060] The transformed feature matrix is ​​encoded and features are extracted through a three-layer convolutional neural network. The convolution kernel size of each layer is 3×3, the step size is 2, and the number of channels is 64, 128, and 256 respectively, to obtain the encoded features;

[0061] The probability distribution parameters of the encoded features are calculated, and the mean vector and variance vector are output respectively through two parallel fully connected layers to obtain the probability distribution parameters. Reparameterized sampling is performed based on the probability distribution parameters, and a noise vector is randomly sampled from the standard normal distribution. The mean vector and variance vector are combined to generate latent variables.

[0062] The latent variables are concatenated with the conditional vector, which is composed of temperature control parameters, delivery time parameters and energy consumption parameters, to obtain the conditional enhancement vector;

[0063] The conditional enhancement vector is upsampled and the feature size is restored through a three-layer deconvolution network. The deconvolution kernel size of each layer is 4×4, the step size is 2, and the number of channels is 256, 128, and 64 respectively to obtain the reconstructed features;

[0064] The reconstructed features are processed by residual connection, and the reconstructed features are fused with the decoding feature matrix to obtain the fused features. The fused features are normalized, and the feature distribution is adjusted by layer normalization. After a layer of fully connected network, they are mapped to the original feature dimension to obtain the enhanced feature matrix.

[0065] Specifically, the decoding feature matrix is ​​transformed. The feature dimension is mapped to a 512-dimensional space through a fully connected layer to obtain a transformed feature matrix. By reducing the feature dimension through mapping, the computational efficiency and feature extraction capability of the subsequent model are improved. Assume that the decoding feature matrix is , the linear transformation after the fully connected layer is expressed as:

[0066] ;

[0067] in, is the transformed feature matrix, is the weight matrix of the fully connected layer, is the bias term. The transformed feature matrix needs to be encoded and features are extracted through a three-layer convolutional neural network. The convolution kernel size of each layer is set to 3×3, the step size is 2, and the number of channels is 64, 128, and 256 respectively. The convolution operation effectively extracts local features and increases the expressive power of the model by increasing the number of channels. For example, the convolution operation of the first layer is expressed as:

[0068] ;

[0069] After the output of the first layer, it is passed to the next layer for convolution, and finally the encoded features are obtained. After obtaining the encoded features, the probability distribution parameters are calculated. Through two parallel fully connected layers, the mean vectors are output respectively. and variance vector , these parameters are used in the subsequent reparameterization process. The calculation of the mean and variance is expressed as follows:

[0070] ;

[0071] ;

[0072] in, and are the fully connected networks for calculating the mean and variance respectively. After obtaining the mean vector and variance vector, the noise vector is randomly sampled from the standard normal distribution through the reparameterization technique. , combining the mean and variance to generate latent variables

[0073]

[0074] Among them, ⊙ represents element-by-element multiplication. This process allows the gradient to be effectively propagated, improving the training effect of the model. The latent variables are concatenated with the conditional vector. The conditional vector consists of temperature control parameters, delivery time parameters, and energy consumption parameters. The concatenation process is expressed as:

[0075] ;

[0076] Among them, [;] represents vector concatenation, is the obtained conditional enhancement vector. This step ensures that the enhanced features can effectively combine the original information with the external conditions and enhance the representation ability of the features. The conditional enhancement vector is upsampled to restore the feature size. This is achieved through a three-layer deconvolution network. The kernel size of the deconvolution layer is set to 4×4, the step size is 2, and the number of channels is 256, 128, and 64 respectively. The deconvolution operation can effectively restore the feature size, which is specifically expressed as:

[0077] Deconv , kernel size , stride=2, channels=256);

[0078] This process will continue and eventually the reconstructed features will be obtained. After the reconstructed features are obtained, residual connection processing is used to fuse the reconstructed features with the decoded feature matrix. Residual connection helps alleviate the gradient vanishing problem in deep networks and maintain the flow of information. The fusion operation is expressed as:

[0079] ;

[0080] in, is the fusion feature. After that, the fusion feature is normalized and the feature distribution is adjusted by layer normalization. The formula for layer normalization is:

[0081] ;

[0082] in, and They are the mean and standard deviation of the fused features, ensuring that the output features are on the same scale and enhancing the convergence speed of the model. The normalized features are mapped to the original feature dimensions through a fully connected network to obtain the enhanced feature matrix :

[0083]

[0084] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0085] The enhanced feature matrix is ​​segmented according to the time window, and the time series features and statistical features are extracted from each segment of data to obtain a feature component sequence. The feature component sequence is screened, and the key features are retained according to the importance weights of the feature components to obtain the target feature sequence.

[0086] Based on the target feature sequence, a cold chain status model is constructed, which includes the temperature change law, position change law and energy consumption change law;

[0087] Digitally model the cold chain state model to obtain the digital twin model of the logistics system, and perform state mapping on the output results of the digital twin model of the logistics system to build the corresponding relationship between temperature state, position state and energy consumption state to obtain the state mapping matrix;

[0088] Supplement the state mapping matrix with features, add delivery time constraints and task priority information, and obtain the target state vector;

[0089] A state transition relationship graph is constructed according to the target state vector. The nodes of the state transition relationship graph represent state parameters, and the edges represent the correlation between parameters. A subset of key state parameters is extracted based on the state transition relationship graph to construct a state parameter space.

[0090] Specifically, the enhanced feature matrix is ​​segmented according to the time window. Assume that the enhanced feature matrix , which is divided into time windows, the size of each window is Each time window The expression is:

[0091] ;

[0092] in, is the starting index of the time window, as changes, By processing each time window, we can extract the time series features and statistical features, such as mean, variance, and extreme value, and obtain the feature component sequence. The specific time series feature extraction is expressed as:

[0093] ;

[0094] Among them, mean, var, and Represent the mean, variance, maximum and minimum values ​​respectively. The resulting feature component sequence Contains the key information of each time window. Filter the feature component sequence and retain the key features by calculating the importance weight of each feature. Use methods such as information gain or Pearson correlation coefficient to evaluate the importance of features. Assume that the weight of each feature is calculated as , select weights exceeding a certain threshold The target feature sequence is obtained

[0095] ;

[0096] The cold chain state model is constructed based on the target feature sequence. The cold chain state model can characterize the dynamic characteristics of the cold chain system, including the temperature change law, position change law and energy consumption change law. For example, the temperature change law is obtained through time series analysis and expressed as follows:

[0097] ;

[0098] in, It's time The temperature, is the temperature change in the current time period. The cold chain state model is digitally modeled to generate a digital twin model of the logistics system. The model can reflect the operating status of the cold chain logistics in real time, and map the state through the output results, thereby constructing the corresponding relationship between the temperature state, position state and energy consumption state. Assume that the state mapping matrix is , define it as:

[0099] ;

[0100] in, , and Represent the state information of temperature, location and energy consumption respectively. The state mapping matrix is ​​supplemented with features, and the delivery time constraint and task priority information are added to form the target state vector The target state vector reflects the current state and takes into account time sensitivity and task urgency. The formula is:

[0101] ;

[0102] Among them, delivery time and task priority are added to the state mapping matrix in the form of vectors to ensure that the output of the model can reflect the complexity of actual operations. A state transition relationship graph is constructed based on the target state vector. The nodes of the state transition relationship graph represent state parameters, and the edges represent the correlation between parameters. Assuming that the graph is used Represent this relationship graph and represent the node set as and the edge set is represented as

[0103] ;

[0104] Among them, the node It can be state parameters such as temperature, position and energy consumption, while the edge It indicates the relationship between these parameters, for example, there is a negative correlation between temperature and energy consumption. By analyzing the state transition relationship diagram, a subset of key state parameters is extracted and the state parameter space is constructed. , helping the system to better understand and optimize the operation of cold chain logistics in actual operations.

[0105] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0106] The state parameter space is divided into action space, temperature control, cooling power and data transmission time slot are set as action variables, and the value range is set for each action variable to obtain the action space;

[0107] The state parameter space and action space are input into the policy network. The policy network contains three fully connected layers with 256, 128, and 64 nodes respectively, and the action probability distribution is obtained;

[0108] Sampling the action probability distribution, selecting the temperature control action, cooling power action and data transmission time slot action, obtaining the action combination, and executing the state transfer according to the action combination, recording the timestamp and data update time of each transfer, calculating the information delay index value, and obtaining the reward signal;

[0109] The state parameter space, action combination and reward signal are input into the value function network, which contains two fully connected layers with 128 and 64 nodes respectively, to obtain the state value evaluation result;

[0110] The state value evaluation results are optimized, constraints are constructed based on temperature control parameters and refrigeration energy consumption parameters, and the policy gradient method is used to update network parameters to obtain the optimized synchronization strategy.

[0111] The optimized synchronization strategy is applied to real-time data transmission, the data update time interval and transmission success rate are recorded, the synchronization performance index is obtained, and communication resources are allocated according to the synchronization performance index to generate a real-time data synchronization plan.

[0112] Specifically, the state parameter space is divided into action spaces. Temperature control, cooling power and data transmission time slot are set as the main action variables. The value range of each action variable is set by previous experience data or expert knowledge. For example, the value range of the temperature control action is The cooling power is between The data transmission time slot is set to By setting, we get a multi-dimensional action space , where each dimension corresponds to a specific action variable:

[0113] ;

[0114] After establishing the action space, the state parameter space and action space The input is sent to the policy network, which consists of three fully connected layers with 256, 128 and 64 nodes respectively. This structure can effectively learn the relationship between state and action and output the probability distribution of action. The probability distribution is expressed by the following formula:

[0115] ;

[0116] in, is the weight matrix of the fully connected layer, is the state parameter vector, and ReLU is the activation function. Through this network structure, the probability distribution of an action is output. The action probability distribution is sampled. According to the sampling results, the corresponding temperature control action, cooling power action, and data transmission time slot action are selected to form an action combination. After executing the state transition according to the action combination, record the timestamp of each state transition and data update time , and calculate the information delay index value , that is, the delay experienced during the state transfer process:

[0117] ;

[0118] in, is the new state transition time, is the old state time. Based on this, we get the reward signal , which is usually calculated in the following form:

[0119] ;

[0120] in, and Respectively represent the contribution of temperature control and cooling power to the overall efficiency, and the two are measured by a preset function. , Action Combination and reward signals Input into the value function network. The value function network consists of two fully connected layers, with 128 and 64 nodes respectively. Through this network, the state value is evaluated to obtain the state value

[0121] ;

[0122] In this formula, and is the weight matrix of the value function network. Through this step, the value of the current state can be evaluated and optimized in subsequent decisions. After obtaining the state value evaluation results, action optimization is performed. Constraints based on temperature control parameters and refrigeration energy consumption parameters help the system avoid unreasonable operations during the optimization process. By using the policy gradient method to update the network parameters, the optimized synchronization strategy is obtained. , the process is expressed by the following formula:

[0123] ;

[0124] in, is the learning rate, is the gradient of the objective function relative to the policy parameters. Through this update mechanism, the policy can be continuously improved to improve the real-time response capability of the system. When the optimized synchronization policy is applied to real-time data transmission, the data update time interval is recorded. and transmission success rate On this basis, the synchronization performance index is calculated This is done using the following formula:

[0125] ;

[0126] Communication resources are allocated according to synchronization performance indicators to generate a real-time data synchronization solution, which can improve the efficiency of data transmission and ensure that the cold chain system maintains optimal operation in a dynamically changing environment.

[0127] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0128] The data update time interval in the real-time data synchronization scheme is used as the data sampling period of the distribution network nodes, and the temperature change series, energy consumption data series and distribution time series of each distribution point are periodically collected to obtain a synchronized data set;

[0129] The nodes of the distribution network are constructed according to the data transmission time slots in the synchronization data set, and the distribution points with data transmission success rates higher than the preset threshold are set as master nodes, and the remaining distribution points are set as slave nodes to obtain a hierarchical node set;

[0130] Based on the communication resource allocation in the real-time data synchronization scheme, the connection edges of the hierarchical node set are constructed, and the connection relationship between the node pairs sharing the same transmission channel is established to obtain the network topology structure;

[0131] Construct feature vectors for each node in the network topology, combine temperature control actions, cooling power actions, and time slot allocation actions into node state features, obtain node feature sets, and construct edge features based on the node feature sets. Calculate the temperature difference coefficient, energy consumption correlation coefficient, and timing matching degree between node pairs to obtain edge feature matrices.

[0132] The node feature set and edge feature matrix are input into the graph attention network, the information transfer between nodes is weighted, the attention feature graph is obtained, and the information of the attention feature graph is aggregated. Different aggregation weights are set based on the master-slave relationship of the nodes, and the node features are updated through the message passing mechanism to obtain the fusion feature matrix.

[0133] A global pooling operation is performed on the fused feature matrix to compress the graph structure information into a vector representation of a fixed dimension to obtain a distribution network representation vector.

[0134] Specifically, the data update time interval is used as the data sampling period of the distribution network node. The temperature change series, energy consumption data series and distribution time series of each distribution point are periodically collected to build a complete synchronous data set. For example, assuming that the sampling period is set to once an hour, if the temperature change data of a distribution point in the past 24 hours is , the energy consumption data is , and the delivery sequence is , get the synchronized data set On this basis, the nodes of the distribution network are constructed according to the data transmission time slots in the synchronization data set. When the data transmission success rate of the distribution point is higher than a preset threshold (such as 70%), the distribution point is regarded as the master node; otherwise, it is a slave node. A hierarchical node set is formed. ,in Contains all master nodes, Contains all slave nodes. For example, if there are three distribution points with data transmission success rates of 80%, 60%, and 75% respectively, then the master node set is , and the slave node is Based on the allocation of communication resources in the real-time data synchronization scheme, the connection edges of the hierarchical node set are constructed. All node pairs sharing the same transmission channel are connected to obtain the topological structure of the network. ,in Represents a collection of nodes. represents the edge set. For example, if the node and Shared channel, and nodes Without sharing channels with other nodes, the network topology is expressed as After the network topology is constructed, the feature vector is constructed for each node. The state characteristics of the node are controlled by the temperature , Refrigeration power action and time slot allocation actions Combined to form a node feature set For example, the node The eigenvector of is represented as:

[0135] ;

[0136] in , and Node Similarly, edge features are constructed based on node feature sets to calculate the temperature difference coefficient between node pairs. , energy consumption correlation coefficient and timing matching , forming the edge feature matrix ,For example:

[0137] ;

[0138] in, Representation Node and The temperature difference coefficient between the two is calculated as follows:

[0139] ;

[0140] The node feature set and edge feature matrix are input into the graph attention network. The network can calculate the weights of information transfer between nodes to obtain the attention feature graph ,in and Represents the index of the node. The weight is calculated using the following formula:

[0141] ;

[0142] in, LeakyReLU Representation Node and The attention coefficient between and Is a node and By aggregating the attention feature graph, the node features are updated, especially by setting different aggregation weights based on the master-slave relationship. The node features are updated through the message passing mechanism to obtain the fused feature matrix The matrix retains the feature information of each node and includes its relationship with other nodes. A global pooling operation is performed on the fused feature matrix to compress the graph structure information into a vector representation of a fixed dimension, and finally a distribution network representation vector is obtained.

[0143] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0144] Based on the distribution network representation vector, an intelligent agent decision environment is constructed, each distribution node is set as an independent intelligent agent, and an intelligent agent interaction environment is constructed;

[0145] The agent interaction environment is input into the collaborative decision-making network, which includes a central agent and multiple executive agents. The central agent is responsible for global strategy generation, and the executive agent is responsible for local task execution, thus obtaining a hierarchical decision-making architecture.

[0146] A multi-objective reward design is performed on the hierarchical decision-making architecture. A reward function is constructed based on temperature control accuracy, delivery timeliness, and energy consumption. The weight of temperature control accuracy is 0.4, the weight of delivery timeliness is 0.3, and the weight of energy consumption is 0.3. The target weight matrix is ​​obtained.

[0147] The target weight matrix is ​​input into the policy update module, and the action selection is performed through the dual Q network. The dual Q network includes an online network and a target network. The online network is responsible for policy update, and the target network is responsible for policy evaluation to obtain the agent policy set.

[0148] Collaboratively optimize the agent strategy set, build a communication protocol between agents based on the state parameters of the digital twin model of the logistics system, calculate the influence weights between agents through the attention mechanism, and obtain collaborative decision-making results;

[0149] Execution instructions are generated based on the results of collaborative decision-making, including distribution route planning, temperature control parameters and energy consumption allocation plans. The execution instructions are combined into a cold chain distribution scheduling plan, and the cold chain distribution scheduling plan is digitally mapped. The scheduling execution data is fed back to the digital twin model of the logistics system.

[0150] Specifically, we construct an agent decision environment based on the distribution network representation vector and set each distribution node as an independent agent. We create an agent interaction environment in which the agents can work together and share information. For example, suppose there are three distribution nodes. and , defining them as agents and In this interactive environment, agents optimize their own delivery strategies by sharing state information, such as current temperature, energy consumption, and location data. The agent interaction environment is input into the collaborative decision-making network. The network consists of a central agent and multiple executing agents , where the central agent is responsible for global strategy generation, while the executive agent focuses on the execution of local tasks. The design of a hierarchical decision-making architecture helps improve the overall efficiency of the system. The central agent makes decisions based on global information, while the executive agent performs corresponding operations based on specific tasks. The formal expression of this architecture is:

[0151] ;

[0152] in, is the global strategy generated by the central agent, is the status information of the entire distribution network. When designing multi-objective rewards for hierarchical decision-making architecture, a reward function is constructed based on temperature control accuracy, delivery timeliness, and energy consumption. The weight of temperature control accuracy is set to 0.4, the weight of delivery timeliness is set to 0.3, and the weight of energy consumption is set to 0.3, and the target weight matrix is ​​obtained.

[0153] ;

[0154] The double Q network is used to select actions by inputting the target weight matrix into the policy update module. The double Q network contains an online network and target network , the online network is responsible for policy updating, while the target network is responsible for policy evaluation. The formula for action selection is:

[0155] ;

[0156] in, Indicates that at time step Select the action, is the time step status, is a possible set of actions. The strategy set of agents is collaboratively optimized. The communication protocol between agents is constructed based on the state parameters of the digital twin model of the logistics system. The influence weights between agents are calculated using the attention mechanism to determine the degree of mutual influence between different agents. For example, setting the agent For Agents The influence weight is , the influence weight calculated by the attention mechanism is expressed as:

[0157] ;

[0158] in, Representing an Agent right The attention score, is the set of all agents. In this way, the collaborative decision results of each agent are obtained. Based on the collaborative decision results, specific execution instructions are generated, including distribution path planning, temperature control parameters and energy consumption allocation plan. The combination of execution instructions forms a cold chain distribution scheduling plan, which is recorded as , which includes:

[0159] ;

[0160] in, Represents the delivery path, is the temperature control parameter, Energy consumption allocation plan. Digital twin mapping of cold chain distribution scheduling plan to ensure that scheduling execution data is fed back to the digital twin model of the logistics system to achieve closed-loop feedback. Through the above steps, each agent in the distribution network can effectively share information and allocate tasks under the framework of the collaborative decision-making network, and finally achieve an optimized cold chain distribution scheduling plan. For example, assuming that the temperature control accuracy does not meet the set standard in a certain period of time, the central agent adjusts the temperature control parameters in real time according to the collaborative decision-making results to ensure the safety and quality of the goods during the distribution process. The real-time data feedback mechanism between agents can also effectively reduce energy consumption and improve distribution efficiency.

[0161] The above describes the multimodal AIGC cold chain logistics information processing method in the embodiment of the present application. The following describes the multimodal AIGC cold chain logistics information processing system 10 in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the multimodal AIGC cold chain logistics information processing system 10 includes:

[0162] The feature enhancement module 11 is used to extract and enhance the features of the video data, temperature data and location data in the cold chain logistics system based on the logistics information AIGC model to obtain an enhanced feature matrix;

[0163] A modeling module 12 is used to perform digital twin modeling based on the enhanced feature matrix, obtain a digital twin model of the logistics system, and construct a state parameter space;

[0164] The calculation module 13 is used to perform deep reinforcement learning and information delay index calculation on the state parameter space, and introduce temperature control parameters and refrigeration energy consumption parameters to generate a real-time data synchronization solution;

[0165] The encoding module 14 is used to construct a distribution network graph structure according to the real-time data synchronization scheme, and perform deep graph network encoding on the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure to obtain a distribution network representation vector;

[0166] The updating module 15 is used to make multi-agent collaborative decisions on the distribution network representation vector, obtain the cold chain distribution scheduling plan based on the multi-objective constraints of temperature control accuracy, distribution timeliness and energy consumption, and update the cold chain distribution scheduling plan to the digital twin model of the logistics system.

[0167] Through the collaborative cooperation of the above components, the logistics information AIGC model is used to enhance and fuse the features of multimodal data such as video, temperature and location, which significantly improves the data representation ability. The conditional variational autoencoder is combined to achieve efficient integration of data of different modalities. The synchronization mechanism of deep reinforcement learning and the real-time data synchronization solution are used to optimize the information transmission efficiency of the digital twin model of the logistics system and reduce the system latency. The key features of the distribution network are extracted through deep graph network coding and hierarchical node structure, and the attention mechanism enhances the quality of information transmission between nodes. Based on multi-agent collaborative decision-making, the reasonable allocation of distribution tasks is achieved, and the decision optimization under multi-objective constraints balances performance indicators such as temperature control, delivery timeliness and energy consumption. The hierarchical decision-making architecture and closed-loop control mechanism are used to ensure the continuous optimization capability of the system and realize the intelligent scheduling and precise control of the cold chain logistics system in complex scenarios.

[0168] See also Figure 3 , Figure 3 This is a schematic block diagram of the structure of an electronic device 300 provided in an embodiment of the present application. The electronic device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected via a system bus 303, wherein the memory 302 may include a non-volatile storage medium and an internal memory.

[0169] The non-volatile storage medium can store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 can execute any of the above-mentioned multimodal AIGC cold chain logistics information processing methods.

[0170] The processor 301 is used to provide computing and control capabilities to support the operation of the entire electronic device 300 .

[0171] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned multimodal AIGC cold chain logistics information processing methods.

[0172] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the electronic device 300 involved in the present application scheme. The specific electronic device 300 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0173] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0174] It should be noted that technical personnel in the relevant field can clearly understand that, for the convenience and conciseness of description, the specific working process of the electronic device 300 described above can refer to the corresponding process of the aforementioned multimodal AIGC cold chain logistics information processing method, and will not be repeated here.

[0175] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the one or more processors implement the multimodal AIGC cold chain logistics information processing method provided in the embodiment of the present application.

[0176] The computer-readable storage medium may be an internal storage unit of the electronic device 300 in the aforementioned embodiment, such as a hard disk or memory of the electronic device 300. The computer-readable storage medium may also be an external storage device of the electronic device 300, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped with the electronic device 300.

[0177] 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.

[0178] 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 application is essentially or the part that contributes 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 to enable an electronic device (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 application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0179] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. 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 application.

Claims

1. A multimodal AIGC cold chain logistics information processing method, characterized in that: The method comprises: Based on the logistics information AIGC model, feature extraction and feature enhancement are performed on the video data, temperature data and location data in the cold chain logistics system to obtain an enhanced feature matrix; specifically, the following steps are performed: obtaining the video data, temperature data and location data in the cold chain logistics system; extracting spatiotemporal features from the video data to obtain a video feature sequence; performing wavelet transform processing on the temperature data to obtain a temperature feature sequence; performing trajectory smoothing and dimensionality reduction processing on the location data to obtain a location feature sequence; aligning the video feature sequence, the temperature feature sequence and the location feature sequence in the time dimension to obtain an aligned feature sequence, and normalizing the aligned feature sequence to obtain a multimodal feature matrix; and inputting the multimodal feature matrix into the logistics information AIGC. The encoder layer of the model includes a multi-head self-attention module and a feedforward neural network. The multi-head self-attention module is set with 8 attention heads, and the dimension of each attention head is 64, so as to obtain an attention feature matrix; the attention feature matrix is ​​input into the decoder layer of the logistics information AIGC model, the decoder layer includes a cross attention module and a feedforward neural network, the cross attention module is used for feature fusion, and the feedforward neural network includes two hidden layers, and the number of hidden layer nodes is 2048 and 1024 respectively, so as to obtain a decoding feature matrix; the decoding feature matrix is ​​input into the feature enhancement layer, and the feature enhancement layer adopts a conditional variational autoencoder structure, generates latent variables through reparameterization techniques, and reconstructs features through a decoding network to obtain an enhanced feature matrix; Perform digital twin modeling based on the enhanced feature matrix to obtain a digital twin model of the logistics system and construct a state parameter space; The state parameter space is subjected to deep reinforcement learning and information delay index calculation, and temperature control parameters and refrigeration energy consumption parameters are introduced to generate a real-time data synchronization scheme; specifically, the state parameter space is divided into an action space, temperature control, refrigeration power and data transmission time slot are set as action variables, and a value range is set for each action variable to obtain an action space; the state parameter space and the action space are input into a strategy network, the strategy network includes three fully connected layers, and the number of nodes is 256, 128, and 64, respectively, to obtain an action probability distribution; the action probability distribution is sampled, and temperature control action, refrigeration power action and data transmission time slot action are selected to obtain an action combination, and the state is executed according to the action combination. Transfer, record the timestamp and data update time of each transfer, calculate the information delay index value, and obtain a reward signal; input the state parameter space, the action combination and the reward signal into the value function network, the value function network includes two fully connected layers, and the number of nodes is 128 and 64 respectively, to obtain a state value evaluation result; optimize the state value evaluation result, construct constraint conditions based on temperature control parameters and refrigeration energy consumption parameters, and use the policy gradient method to update network parameters to obtain an optimized synchronization strategy; apply the optimized synchronization strategy to real-time data transmission, record the data update time interval and transmission success rate, obtain synchronization performance indicators, and allocate communication resources according to the synchronization performance indicators to generate a real-time data synchronization solution; Constructing a distribution network graph structure according to the real-time data synchronization scheme, and performing deep graph network coding on the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure to obtain a distribution network representation vector; Multi-agent collaborative decision-making is performed on the distribution network representation vector, and a cold chain distribution scheduling plan is obtained based on multi-objective constraints such as temperature control accuracy, distribution timeliness and energy consumption. The cold chain distribution scheduling plan is then updated to the digital twin model of the logistics system.

2. The multimodal AIGC cold chain logistics information processing method according to claim 1 is characterized in that: The decoding feature matrix is ​​input into the feature enhancement layer, the feature enhancement layer adopts the conditional variational autoencoder structure, generates latent variables through reparameterization techniques, and reconstructs features through the decoding network to obtain an enhanced feature matrix, including: Performing dimension transformation processing on the decoding feature matrix, mapping the feature dimension to a 512-dimensional space through a fully connected layer, and obtaining a transformed feature matrix; The transformation feature matrix is ​​encoded and features are extracted through a three-layer convolutional neural network, where the convolution kernel size of each layer is 3×3, the step size is 2, and the number of channels is 64, 128, and 256, respectively, to obtain encoded features; Calculate probability distribution parameters for the coding features, output a mean vector and a variance vector respectively through two parallel fully connected layers to obtain probability distribution parameters, perform reparameterization sampling based on the probability distribution parameters, randomly sample a noise vector from a standard normal distribution, and generate latent variables in combination with the mean vector and the variance vector; The latent variable is concatenated with a conditional vector, wherein the conditional vector is composed of a temperature control parameter, a delivery timeliness parameter, and an energy consumption parameter, to obtain a conditional enhancement vector; The conditional enhancement vector is upsampled, and the feature size is restored through a three-layer deconvolution network, where the deconvolution kernel size of each layer is 4×4, the step size is 2, and the number of channels is 256, 128, and 64 respectively, to obtain a reconstructed feature; The reconstructed features are subjected to residual connection processing, and the reconstructed features are fused with the decoding feature matrix to obtain fused features, and the fused features are normalized, and the feature distribution is adjusted through layer normalization, and mapped to the original feature dimension through a layer of fully connected network to obtain an enhanced feature matrix.

3. The multimodal AIGC cold chain logistics information processing method according to claim 2 is characterized in that: The digital twin modeling is performed based on the enhanced feature matrix to obtain a digital twin model of the logistics system and construct a state parameter space, including: The enhanced feature matrix is ​​segmented according to the time window, and the time series features and statistical features are extracted from each segment of data to obtain a feature component sequence, and the feature component sequence is screened, and key features are retained according to the importance weights of the feature components to obtain a target feature sequence; Based on the target feature sequence, a cold chain state model is constructed, wherein the cold chain state model includes temperature change rules, position change rules and energy consumption change rules; Digitally modeling the cold chain state model to obtain a digital twin model of the logistics system, and state mapping the output result of the digital twin model of the logistics system to construct a corresponding relationship between the temperature state, position state and energy consumption state to obtain a state mapping matrix; Supplementing the state mapping matrix with features, adding delivery time constraints and task priority information, and obtaining a target state vector; A state transfer relationship graph is constructed according to the target state vector, wherein the nodes of the state transfer relationship graph represent state parameters, and the edges represent the correlation between parameters. A key state parameter subset is extracted based on the state transfer relationship graph to construct a state parameter space.

4. The multimodal AIGC cold chain logistics information processing method according to claim 1 is characterized in that: The distribution network graph structure is constructed according to the real-time data synchronization scheme, and the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure are subjected to deep graph network coding to obtain a distribution network representation vector, including: The data update time interval in the real-time data synchronization scheme is used as the data sampling period of the distribution network node, and the temperature change series, energy consumption data series and distribution time series of each distribution point are periodically collected to obtain a synchronized data set; According to the data transmission time slots in the synchronization data set, nodes of the distribution network are constructed, distribution points with a data transmission success rate higher than a preset threshold are set as master nodes, and the remaining distribution points are set as slave nodes, to obtain a hierarchical node set; Based on the communication resource allocation in the real-time data synchronization scheme, the connection edges of the hierarchical node set are constructed, and a connection relationship is established between the node pairs sharing the same transmission channel to obtain a network topology structure; Constructing a feature vector for each node in the network topology, combining the temperature control action, the cooling power action and the time slot allocation action as a node state feature to obtain a node feature set, and constructing edge features based on the node feature set, calculating the temperature difference coefficient, energy consumption correlation coefficient and timing matching degree between node pairs, and obtaining an edge feature matrix; Input the node feature set and the edge feature matrix into the graph attention network, perform weight calculation on the information transfer between nodes to obtain an attention feature graph, perform information aggregation on the attention feature graph, set different aggregation weights based on the master-slave relationship of the nodes, update the node features through a message passing mechanism, and obtain a fusion feature matrix; A global pooling operation is performed on the fused feature matrix to compress the graph structure information into a vector representation of a fixed dimension to obtain a distribution network representation vector.

5. The multimodal AIGC cold chain logistics information processing method according to claim 4 is characterized in that: The multi-agent collaborative decision-making of the distribution network representation vector is performed to obtain a cold chain distribution scheduling plan based on multi-objective constraints of temperature control accuracy, distribution timeliness and energy consumption, and the cold chain distribution scheduling plan is updated to the digital twin model of the logistics system, including: Building an agent decision environment based on the distribution network representation vector, setting each distribution node as an independent agent, and building an agent interaction environment; The agent interaction environment is input into a collaborative decision-making network, wherein the collaborative decision-making network includes a central agent and multiple execution agents, wherein the central agent is responsible for generating global strategies and the execution agents are responsible for executing local tasks, thereby obtaining a hierarchical decision-making architecture; A multi-objective reward design is performed on the hierarchical decision-making architecture, and a reward function is constructed according to temperature control accuracy, delivery timeliness, and energy consumption. The weight of temperature control accuracy is 0.4, the weight of delivery timeliness is 0.3, and the weight of energy consumption is 0.3, and the target weight matrix is ​​obtained; The target weight matrix is ​​input into the strategy update module, and action selection is performed through a dual Q network, wherein the dual Q network includes an online network and a target network, wherein the online network is responsible for strategy update and the target network is responsible for strategy evaluation, thereby obtaining an agent strategy set; Collaboratively optimize the agent strategy set, build a communication protocol between agents based on the state parameters of the digital twin model of the logistics system, calculate the influence weights between agents through the attention mechanism, and obtain a collaborative decision-making result; According to the collaborative decision-making results, execution instructions are generated, including distribution route planning, temperature control parameters and energy consumption allocation plan. The execution instructions are combined into a cold chain distribution scheduling plan, and the cold chain distribution scheduling plan is digitally twin mapped, and the scheduling execution data is fed back to the digital twin model of the logistics system.

6. A multi-modal AIGC cold chain logistics information processing system, characterized in that: Used to execute the multimodal AIGC cold chain logistics information processing method according to any one of claims 1 to 5, the multimodal AIGC cold chain logistics information processing system comprises: The feature enhancement module is used to extract and enhance the features of the video data, temperature data and location data in the cold chain logistics system based on the logistics information AIGC model to obtain an enhanced feature matrix; A modeling module, used to perform digital twin modeling based on the enhanced feature matrix, obtain a digital twin model of the logistics system, and construct a state parameter space; A calculation module, used to perform deep reinforcement learning and information delay index calculation on the state parameter space, and introduce temperature control parameters and refrigeration energy consumption parameters to generate a real-time data synchronization solution; An encoding module, used to construct a distribution network graph structure according to the real-time data synchronization scheme, and perform deep graph network encoding on the temperature change sequence, energy consumption data sequence and distribution time sequence in the distribution network graph structure to obtain a distribution network representation vector; The updating module is used to perform multi-agent collaborative decision-making on the distribution network representation vector, obtain a cold chain distribution scheduling plan based on multi-objective constraints such as temperature control accuracy, distribution timeliness and energy consumption, and update the cold chain distribution scheduling plan to the digital twin model of the logistics system.

7. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory so that the electronic device executes the multimodal AIGC cold chain logistics information processing method as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the multimodal AIGC cold chain logistics information processing method as described in any one of claims 1-5 is implemented.

Images