Whole-process digital management system for fresh food supply chain

Abstract

This invention relates to the field of fresh food supply chain management technology, and more particularly to a digital management system for the entire fresh food supply chain process. The system includes: a data acquisition module for real-time acquisition of status data of fresh products throughout the entire supply chain process; a model building module for iterative training of a fusion prediction network framework to output a fresh product demand prediction model; an inventory control scheme generation module for dynamically adjusting inventory turnover parameters of fresh products to obtain an inventory control scheme; a route optimization scheme generation module for collaborative analysis of transportation route planning parameters to obtain a transportation route optimization scheme; and an anomaly monitoring module for monitoring and analyzing the flow status of fresh products, generating anomaly handling information when the monitoring results meet preset anomaly conditions. This invention enables real-time dynamic monitoring and collaborative optimization of the entire fresh food supply chain process, improving the overall operational efficiency of the supply chain.

Description

Technical Field

[0001] The present invention relates to the technical field of fresh food supply chain management, and particularly to a full-process digital management system for fresh food supply chain. Background Art

[0002] Fresh food products have inherent characteristics such as perishability, strong timeliness, and high circulation requirements, which pose huge challenges to the management of fresh food supply chain. In the traditional mode, each link of the fresh food supply chain, such as procurement, warehousing, transportation, distribution, etc., often relies on manual operations and decentralized information systems, making it difficult to form a real-time dynamic monitoring and collaborative optimization mechanism covering the whole process. In the existing technologies, the monitoring means of fresh food supply chain mainly rely on local data collection within the link, such as the temperature and humidity records of the warehousing environment, vehicle positioning information, etc. However, these data usually exist in isolation and cannot achieve real-time fusion and dynamic feedback across links. Once there is a cold chain break or abnormal fluctuation, it is often only possible to trace back afterwards, unable to give early warnings and intervene in a timely manner, resulting in increased cargo damage. At the same time, inventory management mostly adopts static thresholds or empirical judgments, lacking real-time linkage analysis of the demand at the sales terminal, in-transit inventory, and warehousing status, resulting in a serious lag in inventory turnover regulation and making it difficult to balance the risks of inventory backlog and out-of-stock shortages. In terms of transportation route planning, traditional methods are mostly based on fixed routes or the personal experience of drivers, unable to comprehensively consider multi-dimensional dynamic factors such as real-time road conditions, vehicle status, and cargo temperature control requirements, resulting in increased transportation mileage, rising energy consumption, and a decline in distribution on-time rate. In addition, the information interaction between the various participants in the fresh food supply chain is not smooth, and systems such as orders, procurement, warehousing, and transportation are fragmented from each other, unable to form collaborative optimization decisions. The overall operation of the supply chain lacks the ability of dynamic adjustment and responds slowly to market fluctuations or emergencies. Demand forecasting is also inaccurate due to the inability to effectively integrate multi-source data such as historical sales, weather changes, and seasonal factors, further exacerbating the mismatch between procurement plans and market demand, resulting in an increase in inventory costs and losses.

[0003] Chinese Patent Publication No. CN118966964A discloses a data analysis method and system for fresh food supply chain based on big data and deep learning, including: obtaining multiple fresh food categories to be monitored, and obtaining the basic index data of each fresh food category to be monitored; in the case that the fresh food supply chain state represented by the basic index data is abnormal, taking the corresponding fresh food category to be monitored as the target fresh food category; obtaining the key indicators and multiple alternative influencing factors of the target fresh food category, and taking the influence dependence relationship between the key indicators and the target influencing factor determined from the multiple alternative influencing factors as the data analysis result of the fresh food supply chain of the target fresh food category. This solution only conducts ex-post anomaly analysis and influencing factor identification based on basic index data, unable to achieve real-time dynamic monitoring and collaborative optimization of the whole process of fresh food supply chain, resulting in a lag in inventory turnover regulation and unreasonable transportation route planning, reducing the overall operation efficiency of the supply chain. Summary of the Invention

[0004] To address this, the present invention provides a digital management system for the entire fresh food supply chain, which overcomes the problem that existing technologies cannot achieve real-time dynamic monitoring and collaborative optimization of the entire fresh food supply chain, resulting in lagging inventory turnover control, unreasonable transportation route planning, and reduced overall supply chain operating efficiency.

[0005] To achieve the above objectives, the present invention provides a digital management system for the entire fresh food supply chain, comprising: The data acquisition module is used to acquire real-time status data of fresh products throughout the entire fresh food supply chain through IoT devices, and to preprocess the status data to obtain a fused dataset containing timestamps and location identifiers. The fused dataset includes real-time sales data, real-time inventory data, temperature data, humidity data, and environmental image data. The model building module is used to construct a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network. It iteratively trains the fusion prediction network framework based on a historical fresh food supply chain dataset containing sales volume labeling information and inventory labeling information. When the prediction accuracy output by the trained fusion prediction network framework exceeds a preset prediction accuracy threshold, training stops, and a fresh food product demand prediction model is output. The historical fresh food supply chain dataset contains historical sales volume data, historical inventory data, historical temperature data, historical humidity data, and historical environmental image data of the same type as the fusion dataset. The sales volume labeling information is used to identify the actual sales volume data in the historical fresh food supply chain dataset, and the inventory labeling information is used to identify the actual inventory data in the historical fresh food supply chain dataset. The inventory control scheme generation module is used to input the fused dataset into the fresh product demand forecasting model for predictive analysis, obtain demand forecast results including the predicted demand amount and the fresh product inventory adjustment amount, and dynamically adjust the inventory turnover parameters of fresh products based on the demand forecast results and temperature data to obtain an inventory control scheme including replenishment time points and replenishment quantities. The route optimization scheme generation module is used to analyze the storage environment parameters based on the inventory control scheme and humidity data to obtain the storage environment control instructions, and to perform collaborative analysis on the transportation route planning parameters based on the storage environment control instructions and the inventory control scheme to obtain the transportation route optimization scheme. The anomaly monitoring module is used to monitor and analyze the circulation status of fresh products through the environmental image data and transportation route optimization scheme, obtain circulation anomaly monitoring results, and when the circulation anomaly monitoring results meet the preset anomaly conditions, generate anomaly handling information containing anomaly type identifier and anomaly location information, and send the anomaly handling information to the corresponding mobile terminal.

[0006] The technical principle of this application lies in the following: Real-time sales data, real-time inventory data, temperature, humidity, and environmental images are collected from the entire fresh food supply chain via IoT devices. After preprocessing the status data, a fused dataset containing timestamps and location identifiers is generated to ensure the real-time nature and integrity of the data. A fused prediction network framework, including a temporal convolutional network and an attention mechanism network, is constructed. This fused prediction network framework is iteratively trained using a historical fresh food supply chain dataset containing sales and inventory labeling information until the prediction accuracy reaches the target, outputting a fresh food demand prediction model. The fused dataset is input into this fresh food demand prediction model for analysis to obtain demand prediction results. Inventory turnover parameters are dynamically adjusted based on temperature data to generate an inventory control plan. Furthermore, humidity data is combined with warehousing environment analysis to optimize transportation routes in conjunction with the inventory control plan, generating an optimized transportation route plan. The circulation status is monitored through environmental images and optimized routes. In case of anomalies, disposal information is generated and sent to the corresponding mobile terminal, achieving full-process collaborative management and control.

[0007] Compared with the prior art, the beneficial effects of this application are as follows: By acquiring real-time status data of fresh produce throughout the entire supply chain using IoT devices and preprocessing this data to obtain a fused dataset containing timestamps and location identifiers, real-time collection and standardized processing of data across the entire fresh produce supply chain are achieved. This provides a high-quality data foundation for subsequent accurate forecasting and optimization, avoiding the limitations of existing technologies that rely solely on basic indicator data for post-event analysis. A fused prediction network framework incorporating temporal convolutional networks and attention mechanisms is constructed and iteratively trained based on historical fresh produce supply chain datasets to output a fresh produce demand forecasting model. This fully leverages the processing capabilities of temporal convolutional networks for time-series data and the focusing capabilities of attention mechanisms for key features, significantly improving the accuracy and robustness of demand forecasting and addressing the low prediction accuracy issue of existing technologies. By inputting the fused dataset into the fresh produce demand forecasting model for predictive analysis, demand forecast results including predicted demand volume and inventory adjustment volume for fresh produce are obtained. Based on the demand forecast results and temperature data, inventory turnover parameters for fresh produce are adjusted. Dynamic adjustments are made to obtain an inventory control plan that includes replenishment timing and quantity, achieving dynamic optimization of inventory turnover and precise replenishment, effectively reducing the risk of inventory backlog and stockouts. By analyzing warehousing environment parameters based on the inventory control plan and humidity data, warehousing environment control instructions are obtained. Furthermore, based on the warehousing environment control instructions and the inventory control plan, transportation route planning parameters are analyzed collaboratively to obtain a transportation route optimization plan. This achieves synergistic optimization of the warehousing environment and transportation routes, improving warehousing preservation and transportation efficiency, and reducing fresh produce spoilage. The circulation status of fresh products is monitored and analyzed using environmental image data and the transportation route optimization plan, resulting in abnormal circulation monitoring results. When the abnormal circulation monitoring results meet preset abnormal conditions, abnormal handling information containing abnormality type identifiers and abnormality location information is generated and sent to the corresponding mobile terminal. This enables real-time monitoring and rapid handling of abnormal fresh produce circulation, improving supply chain transparency and emergency response capabilities, thereby comprehensively enhancing the intelligent management level and operational efficiency of the fresh produce supply chain.

[0008] Furthermore, in the model building module, a fusion prediction network framework comprising a temporal convolutional network and an attention mechanism network is constructed, including: A temporal convolutional network component is constructed based on a preset inflation factor sequence and a preset convolutional kernel size. The temporal data in the historical fresh food supply chain dataset is processed by the temporal convolutional network component, and local temporal features of the temporal data under different inflation rates are extracted to obtain a local temporal feature set containing multi-scale time dependencies. A multi-head self-attention mechanism network component is constructed based on a preset number of attention heads and a preset feature dimension. The multi-head self-attention mechanism network component is then used to perform a linear transformation on each feature vector in the local temporal feature set to generate a query matrix, a key matrix, and a value matrix. The query matrix and key matrix are sequentially subjected to dot product, scaling, and softmax normalization to obtain the attention weight matrix. The value matrix is ​​then weighted and summed according to the attention weight matrix to obtain the context feature vector of each attention head. The context feature vector of each attention head is then sequentially concatenated and linearly transformed to obtain the enhanced temporal features containing global context information. A feature fusion layer is constructed based on the local temporal feature set and the enhanced temporal features. The local temporal feature set and the enhanced temporal features are concatenated through the feature fusion layer to obtain a concatenated feature vector. The concatenated feature vector is then input into a fully connected network for nonlinear transformation to obtain a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network.

[0009] This solution combines temporal convolutional networks with multi-head self-attention mechanisms to fully capture multi-scale local features and global dependencies in fresh food supply chain time-series data, improving the comprehensiveness and effectiveness of feature extraction. Multi-dilation factor convolution enables feature mining at different time scales, enhancing the model's ability to model long-term and short-term time-series patterns. Attention mechanisms weight key information, reducing redundant data interference and improving feature expression accuracy. Feature fusion and fully connected transformations construct a fusion prediction framework, effectively improving the accuracy and stability of fresh food supply chain time-series prediction.

[0010] Furthermore, in the model building module, the fusion prediction network framework is iteratively trained based on a historical fresh food supply chain dataset containing sales volume and inventory labeling information. When the prediction accuracy output by the trained fusion prediction network framework exceeds a preset prediction accuracy threshold, training stops, and a fresh food product demand prediction model is output, including the following steps: The historical fresh food supply chain dataset is divided according to a preset dataset division ratio to obtain a data training set for model parameter updates and a data validation set for model performance evaluation. The training data set is input into the fusion prediction network framework. The input features in the training data set are forward propagated according to the preset loss function to obtain the sales forecast and inventory forecast of fresh products. The parameters of the temporal convolutional network and the attention mechanism network in the fusion prediction network framework are updated by gradient based on the back propagation algorithm. When the number of iterations reaches the preset maximum number of iterations, the parameter update is stopped. The prediction accuracy of the current fusion prediction network framework is verified based on the data validation set. When the verified prediction accuracy is greater than the preset prediction accuracy threshold, the fusion prediction network framework corresponding to the current iteration round is determined as the fresh product demand prediction model.

[0011] In this solution, by dividing the data training set and the data validation set, the model parameter update and performance evaluation are separated, thereby improving the model's generalization ability. Through forward propagation and backpropagation iterative training, the parameters of the temporal convolution and attention network are precisely optimized to enhance the model's fitting effect. The training termination is controlled by a preset accuracy threshold to ensure that the model output is stable and reliable. By combining sales and inventory labeling information for training, the demand forecast for fresh products can be accurately achieved.

[0012] Furthermore, the mathematical expression of the preset loss function is: In the formula, This represents the value of the loss function. This represents the total number of training samples. This represents the sales labeling information for the i-th training sample. This represents the sales prediction value of the i-th training sample output by the fusion prediction network framework. The weighting coefficients representing the sales forecast error. This represents the inventory labeling information for the i-th training sample. This represents the inventory prediction value of the i-th training sample output by the fusion prediction network framework. The weighting coefficients representing inventory forecasting errors. This represents the set of trainable parameters in the fusion prediction network framework. The square of the L2 norm of the trainable parameter set is represented. This represents the regularization coefficient.

[0013] In this solution, the prediction accuracy is improved by setting a weighted loss that includes sales volume and inventory. By introducing L2 regularization constraint parameters, the risk of overfitting is reduced, the generalization ability and stability of the model are enhanced, and the reliable results of the fusion prediction network are ensured.

[0014] Furthermore, in the inventory control plan generation module, the fused dataset is input into the fresh produce demand forecasting model for predictive analysis, yielding demand forecast results that include both the predicted demand for fresh produce and the adjusted inventory levels for fresh produce, including: The real-time sales data and real-time inventory data in the fused dataset, as well as at least one of the temperature data, humidity data, and environmental image data in the fused dataset, are used as input data and fed into the fresh produce demand prediction model. Temporal features are extracted through a temporal convolutional network to obtain local temporal features. The local temporal features are then enhanced with global contextual features through a multi-head self-attention network to obtain enhanced fused features. The enhanced fused features are then fed into a fully connected output layer for linear mapping and nonlinear activation to obtain the predicted demand for fresh produce. The initial inventory adjustment is calculated based on the difference between the predicted demand for fresh produce and the current inventory, combined with the preset inventory volatility coefficient and safety stock level. The initial inventory adjustment amount is compared with the current available storage capacity upper limit threshold in the warehouse management system. When the initial inventory adjustment amount is greater than the current available storage capacity upper limit threshold, the current available storage capacity upper limit threshold is used as the fresh product inventory adjustment amount. When the initial inventory adjustment amount is less than or equal to the current available storage capacity upper limit threshold, the initial inventory adjustment amount is used as the fresh product inventory adjustment amount. The fresh product demand forecast amount and the fresh product inventory adjustment amount are combined and packaged into a demand forecast result.

[0015] In this solution, by inputting the fused dataset into the fresh produce demand forecasting model, a temporal convolutional network is used to extract local temporal features, and a multi-head self-attention network is used to enhance global contextual features to obtain accurate demand forecasts. Based on the difference between the demand forecast and the current inventory, the initial adjustment is calculated by combining the inventory volatility coefficient and the safety stock level. The final inventory adjustment is determined by comparing it with the current available storage capacity upper limit threshold, generating a demand forecast result that includes the demand forecast and the inventory adjustment, effectively improving forecast accuracy.

[0016] Furthermore, the initial inventory adjustment amount The mathematical expression is: In the formula, Represents the demand difference response coefficient. This indicates the forecast for demand for fresh produce. Indicates the current inventory level. Indicates the sensitivity coefficient to demand fluctuations. This represents the total number of historical prediction cycles. This represents the demand forecast for the t-th historical period. This represents the average of historical demand forecasts.

[0017] This solution sets demand difference response coefficient and demand fluctuation sensitivity coefficient, and combines demand forecast, current inventory and historical forecast fluctuations to accurately and dynamically adjust the initial inventory of fresh products, improve the matching degree of supply and demand, reduce the risk of inventory backlog and shortage, and enhance the stability and rationality of inventory control.

[0018] Furthermore, in the inventory control plan generation module, the inventory turnover parameters of fresh produce are dynamically adjusted based on the demand forecast results and temperature data to obtain an inventory control plan that includes replenishment timing and replenishment quantity, including: Based on the fresh produce inventory adjustment amount in the demand forecast results, the net demand is calculated by combining the current inventory and the inventory in transit. The net demand is then matched with the preset order quantity model to obtain the initial replenishment quantity. Based on the initial replenishment quantity and the inventory turnover target, the theoretical replenishment time point is calculated. The theoretical replenishment time point is the moment when the inventory drops to the safety stock level. The temperature data is compared with the optimal storage temperature range corresponding to the fresh product category to obtain the temperature offset. The dynamic remaining shelf life under the current temperature condition is calculated based on the temperature offset and the preset shelf life decay function. The theoretical replenishment time point is forward or backward corrected based on the dynamic remaining shelf life to obtain the replenishment time point. The replenishment time point is matched with the supplier's delivery cycle. When the replenishment time point conflicts with the delivery cycle, the initial replenishment quantity is adjusted to obtain the replenishment quantity. An initial inventory control plan is generated based on the replenishment time point and replenishment quantity. The initial inventory control plan is then compared with the workload data in the warehouse operation scheduling system for collision detection. When it is detected that the warehouse operation load corresponding to the replenishment time point exceeds the preset operation capacity threshold, the replenishment time point is shifted forward or backward to an available time period when the operation load is lower than the preset operation capacity threshold, thus obtaining an inventory control plan that includes the replenishment time point and replenishment quantity.

[0019] This solution combines demand forecasting, inventory data, and a pre-set order quantity model to determine theoretical replenishment information. Temperature data is introduced to correct product shelf life and replenishment timing. Collision detection is implemented by matching the warehousing operation load to accurately determine the replenishment time and quantity, improve the turnover efficiency of fresh food inventory, reduce the risk of loss, and ensure the smooth and orderly operation of warehousing.

[0020] Furthermore, in the path optimization scheme generation module, the warehouse environment parameters are analyzed based on the inventory control scheme and humidity data to obtain warehouse environment control instructions, including: Based on the replenishment time and quantity in the inventory control plan, and combined with historical inventory data, the product density and space occupancy rate of each storage zone in the future preset period are predicted to obtain the storage load prediction data. The humidity data is then analyzed to determine the deviation between the humidity data and the optimal humidity range corresponding to the fresh product category to obtain the current humidity deviation value. The warehouse load prediction data and the current humidity deviation value are input into a preset fuzzy logic controller to reason about the priority of warehouse environment regulation, generate regulation weight coefficients for different temperature and humidity control areas, and perform joint optimization analysis on the operating parameters of refrigeration units, humidification equipment and ventilation equipment based on the regulation weight coefficients to obtain a comprehensive environmental regulation parameter set including target temperature value, target humidity value, equipment operating power and operating time. The comprehensive environmental control parameter set is compared with the real-time operating status parameters of each environmental control device to generate a sequence of device control instructions containing device identifiers, control mode switching instructions and parameter setting values. The sequence of device control instructions is then output to the centralized controller of the warehouse environment as a warehouse environment control instruction.

[0021] This solution combines inventory control plans with historical data to predict the load of storage zones, introduces humidity data to obtain deviation values, uses fuzzy logic reasoning to adjust priorities, optimizes the operating parameters of multiple devices, accurately generates equipment control commands, achieves dynamic adaptation to the storage environment, improves the accuracy of temperature and humidity control and equipment operating efficiency, reduces fresh produce spoilage, and ensures a stable and reliable storage environment.

[0022] Furthermore, in the anomaly monitoring module, the circulation status of fresh produce is monitored and analyzed using the environmental image data and transportation route optimization scheme to obtain anomaly monitoring results, including: The environmental image data is analyzed by frame sequence and target detection using an image feature extraction network to obtain the surface condition features and location distribution features of fresh products. Based on the path node information and time information in the transportation route optimization scheme, the surface state characteristics and location distribution characteristics of the fresh products are spatiotemporally aligned to obtain a flow state characteristic sequence associated with the transportation route. An anomaly degree analysis is performed on each feature point in the flow status feature sequence by using preset anomaly discrimination rules to obtain an anomaly degree score. When the anomaly degree score exceeds a preset anomaly threshold, the feature point corresponding to the preset anomaly threshold is marked as a flow anomaly monitoring result.

[0023] In this solution, environmental image data is analyzed through image feature extraction network to obtain the status and location information of fresh products. Combined with transportation route information, spatiotemporal alignment is achieved to construct a flow status feature sequence. Based on anomaly rules, accurate scoring and judgment are made to identify flow anomalies in a timely manner, improve the real-time performance and accuracy of monitoring, reduce fresh product spoilage, and ensure the safety and controllability of the entire transportation process. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of the digital management system for the entire fresh food supply chain according to an embodiment of the present invention. Detailed Implementation

[0025] The following detailed description illustrates the specific implementation method: like Figure 1 As shown, it is a structural diagram of the digital management system for the entire fresh food supply chain according to an embodiment of the present invention, including: The data acquisition module is used to acquire real-time status data of fresh products throughout the entire fresh food supply chain through IoT devices, and to preprocess the status data to obtain a fused dataset containing timestamps and location identifiers. The fused dataset includes real-time sales data, real-time inventory data, temperature data, humidity data, and environmental image data. The model building module is used to construct a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network. It iteratively trains the fusion prediction network framework based on a historical fresh food supply chain dataset containing sales volume labeling information and inventory labeling information. When the prediction accuracy output by the trained fusion prediction network framework exceeds a preset prediction accuracy threshold, training stops, and a fresh food product demand prediction model is output. The historical fresh food supply chain dataset contains historical sales volume data, historical inventory data, historical temperature data, historical humidity data, and historical environmental image data of the same type as the fusion dataset. The sales volume labeling information is used to identify the actual sales volume data in the historical fresh food supply chain dataset, and the inventory labeling information is used to identify the actual inventory data in the historical fresh food supply chain dataset. The inventory control scheme generation module is used to input the fused dataset into the fresh product demand forecasting model for predictive analysis, obtain demand forecast results including the predicted demand amount and the fresh product inventory adjustment amount, and dynamically adjust the inventory turnover parameters of fresh products based on the demand forecast results and temperature data to obtain an inventory control scheme including replenishment time points and replenishment quantities. The route optimization scheme generation module is used to analyze the storage environment parameters based on the inventory control scheme and humidity data to obtain the storage environment control instructions, and to perform collaborative analysis on the transportation route planning parameters based on the storage environment control instructions and the inventory control scheme to obtain the transportation route optimization scheme. The anomaly monitoring module is used to monitor and analyze the circulation status of fresh products through the environmental image data and transportation route optimization scheme, obtain circulation anomaly monitoring results, and when the circulation anomaly monitoring results meet the preset anomaly conditions, generate anomaly handling information containing anomaly type identifier and anomaly location information, and send the anomaly handling information to the corresponding mobile terminal.

[0026] In this embodiment, the data acquisition module acquires real-time status data of fresh products throughout the entire fresh food supply chain via IoT devices. This includes deploying multiple sensors and smart terminals at various nodes throughout the process to obtain real-time status data of fresh products. In the warehousing stage, high-precision temperature and humidity sensors are deployed at key locations such as the four corners, center, and air vents inside the cold storage to continuously collect temperature and humidity data. In the transportation stage, refrigerated trucks are equipped with integrated IoT data acquisition nodes that combine temperature and humidity sensors, GPS positioning modules, and cameras. The cameras are positioned to capture panoramic views of the goods or specific indicator labels through observation windows, recording environmental image data in real time. These devices utilize microcontrollers such as STM32 as core processors and actively read data from various sensors at a preset frequency (e.g., every minute or dynamically adjusted according to a strategy). The collected raw status data is then transmitted in real-time to a cloud server via wireless communication modules such as 4G, 5G, or NB-IoT using protocols such as MQTT or HTTP, thus completing the acquisition of status data throughout the entire process from production, warehousing, transportation to delivery.

[0027] The preprocessing of the state data to obtain a fused dataset containing timestamps and location identifiers includes: First, filtering the raw state data transmitted by IoT devices to remove outliers exceeding preset threshold ranges in temperature and humidity data, and blurry or invalid frames in environmental image data, ensuring the accuracy of the raw state data. Then, noise reduction and frame extraction processing are performed on the filtered environmental image data to retain valid image information that clearly reflects the environmental state of fresh produce; simultaneously, temperature and humidity data are standardized to unify the data format. Finally, for each set of processed temperature, humidity, and environmental image data, the timestamp and location identifier at the time of acquisition are associated, and the three types of data are integrated according to the timestamp and location identifier to form a fused dataset containing timestamps and location identifiers, and synchronously associated with temperature, humidity, and environmental image data. The preset threshold ranges include temperature threshold ranges, humidity threshold ranges, and image clarity thresholds. For example, the temperature threshold range is 0℃~4℃, the humidity threshold range is 85%RH~95%RH, and the environmental image data is judged as valid frames with a clarity score ≥30 and a signal-to-noise ratio ≥15dB.

[0028] Specifically, in the model building module, a fusion prediction network framework comprising a temporal convolutional network and an attention mechanism network is constructed, including: A temporal convolutional network component is constructed based on a preset inflation factor sequence and a preset convolutional kernel size. The temporal data in the historical fresh food supply chain dataset is processed by the temporal convolutional network component, and local temporal features of the temporal data under different inflation rates are extracted to obtain a local temporal feature set containing multi-scale time dependencies. A multi-head self-attention mechanism network component is constructed based on a preset number of attention heads and a preset feature dimension. The multi-head self-attention mechanism network component is then used to perform a linear transformation on each feature vector in the local temporal feature set to generate a query matrix, a key matrix, and a value matrix. The query matrix and key matrix are sequentially subjected to dot product, scaling, and softmax normalization to obtain the attention weight matrix. The value matrix is ​​then weighted and summed according to the attention weight matrix to obtain the context feature vector of each attention head. The context feature vector of each attention head is then sequentially concatenated and linearly transformed to obtain the enhanced temporal features containing global context information. A feature fusion layer is constructed based on the local temporal feature set and the enhanced temporal features. The local temporal feature set and the enhanced temporal features are concatenated through the feature fusion layer to obtain a concatenated feature vector. The concatenated feature vector is then input into a fully connected network for nonlinear transformation to obtain a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network.

[0029] In this embodiment, the preset inflation factor sequence is typically determined based on the time span and periodic characteristics of the time-series data. For example, for daily data with weekly periodicity, the sequence [1, 2, 4, 8] can be used to cover dependencies from near to far. The preset convolution kernel size needs to balance the length of local patterns and computational efficiency, and is often set to 3 or 5 to ensure that short-term fluctuations in adjacent time steps can be captured. The preset number of attention heads is usually set to 8 or 12, enabling the model to focus on dependencies of different dimensions in parallel from multiple representation subspaces. The preset feature dimension needs to match the model capacity and data scale, and is often set to 64 or 128 dimensions to ensure that each attention head can effectively perform linear transformations of the query, key, and value matrices and feature extraction within that dimension.

[0030] The temporal convolutional network component receives temporal data from a historical fresh food supply chain dataset as input. First, it constructs convolutional layers with different dilation rates based on a preset kernel size (e.g., set to 3) and a preset dilation factor sequence (e.g., set to [1, 2, 4, 8]). During processing, the temporal data is sequentially fed into these convolutional layers. Each layer slides along the temporal dimension using a convolutional kernel of the preset kernel size, while simultaneously performing convolution operations by skipping fixed time steps according to the dilation rate corresponding to the current layer. For example, when the dilation rate is 2, the time points covered by the convolutional kernel are spaced one step apart, thereby expanding the receptive field. Through this layer-by-layer convolutional operation, the temporal convolutional network component can systematically scan the entire temporal data and generate a preliminary convolutional feature map. After the temporal convolutional network component completes the convolutional processing, it extracts the corresponding local temporal features from each convolutional layer with a specific dilation rate. Specifically, for each dilation rate in the preset dilation factor sequence [1, 2, 4, 8], the temporal convolutional network component generates a set of feature maps. These feature maps capture the local patterns of the time series under that dilation rate. For example, the convolutional layer with a dilation rate of 1 extracts the subtle fluctuation features between adjacent time steps; while the convolutional layer with a dilation rate of 8 extracts the long-term trend features between time steps that are far apart. These local temporal features output from layers with different dilation rates are collected to form a local temporal feature set. The multi-head self-attention mechanism network component first receives a local temporal feature set from the temporal convolutional network component. Based on a preset number of attention heads (e.g., 8) and a preset feature dimension (e.g., 64 dimensions), the multi-head self-attention mechanism network component initializes three independent, learnable weight matrices for each attention head, corresponding to the query linear transformation, key linear transformation, and value linear transformation, respectively. The multi-head self-attention mechanism network component inputs each feature vector in the local temporal feature set into the three linear transformation layers of each attention head. Specifically, a feature vector is multiplied by the query weight matrix of the first attention head to generate a row vector of the query matrix of that attention head; simultaneously, the feature vector is multiplied by the key weight matrix of the same head to generate a row vector of the key matrix; and multiplied by the value weight matrix to generate a row vector of the value matrix. This process is repeated for all feature vectors and all attention heads, ultimately generating a complete query matrix, a key matrix, and a value matrix for each attention head. The rows of these matrices correspond to different time steps of the input sequence. The multi-head self-attention mechanism network component performs matrix multiplication on the query matrix and key matrix generated by each attention head. That is, it calculates the dot product similarity between the query vector at each time step in the query matrix and the key vector at each time step in the key matrix to obtain an initial attention score matrix.To prevent the dot product from becoming too large and causing the gradient of the softmax function to vanish, the initial attention score matrix is ​​divided by a scaling factor, which is usually the square root of the key vector dimension (for example, if the preset feature dimension is 64, then the scaling factor is 8). The scaled attention score matrix is ​​then input into the softmax function for normalization, so that the sum of all elements in each row of the scaled attention score matrix is ​​1, resulting in the attention weight matrix. Each row in the attention weight matrix represents the attention weight distribution of one time step to all other time steps.

[0031] After obtaining the attention weight matrix, the multi-head self-attention mechanism network component performs matrix multiplication on the attention weight matrix and the value matrix. Specifically, each row of the attention weight matrix is ​​used as a weight vector, and it is weighted and summed with each column of the value matrix to generate the context feature vector corresponding to that time step. This operation is performed on all rows of the attention weight matrix to obtain the context feature vectors for all time steps of the attention head. The multi-head self-attention mechanism network component collects the context feature vectors of all time steps calculated by each attention head. For each time step in the input sequence, the multi-head self-attention mechanism network component concatenates the context feature vectors obtained by each attention head at that time step to form a long vector with a dimension equal to (the preset number of attention heads multiplied by the preset feature dimension). For example, if the preset number of attention heads is 8 and the preset feature dimension is 64, the dimension of the concatenated vector is 512. Then, the concatenated vector is projected through an additional, learnable linear transformation layer (usually a fully connected network) to transform the dimension of the concatenated vector back to the same dimension as the original features, or to the dimension required by the subsequent part of the model. The output obtained after this linear transformation is the enhanced temporal feature containing global context information. The feature fusion layer receives two inputs: a local temporal feature set output from the temporal convolutional network component and an enhanced temporal feature output from the multi-head self-attention mechanism network component. The local and enhanced temporal features are aligned in the time step dimension. The feature fusion layer concatenates the local and enhanced temporal feature vectors from the first time step to form a fusion vector. This operation is repeated for each time step of the input sequence in the feature fusion layer, ultimately generating a concatenated feature vector. This concatenated feature vector is then input into a fully connected network for further processing. This fully connected network is typically composed of multiple linear layers and non-linear activation functions (such as ReLU) stacked alternately. The concatenated feature vector first undergoes an affine transformation through the first linear layer to change its feature dimension. Then, non-linearity is introduced through the activation function to enhance the model's expressive power. Subsequently, it passes through multiple similar linear and activation function layers, performing feature extraction and dimensionality compression layer by layer. Finally, the last layer of the fully connected network maps the features to the target output dimension, resulting in the fusion prediction network framework.

[0032] Specifically, in the model building module, the fusion prediction network framework is iteratively trained based on a historical fresh food supply chain dataset containing sales volume and inventory labeling information. When the prediction accuracy output by the trained fusion prediction network framework is greater than a preset prediction accuracy threshold, training stops, and a fresh food product demand prediction model is output, including the following steps: The historical fresh food supply chain dataset is divided according to a preset dataset division ratio to obtain a data training set for model parameter updates and a data validation set for model performance evaluation. The training data set is input into the fusion prediction network framework. The input features in the training data set are forward propagated according to the preset loss function to obtain the sales forecast and inventory forecast of fresh products. The parameters of the temporal convolutional network and the attention mechanism network in the fusion prediction network framework are updated by gradient based on the back propagation algorithm. When the number of iterations reaches the preset maximum number of iterations, the parameter update is stopped. The prediction accuracy of the current fusion prediction network framework is verified based on the data validation set. When the verified prediction accuracy is greater than the preset prediction accuracy threshold, the fusion prediction network framework corresponding to the current iteration round is determined as the fresh product demand prediction model.

[0033] Specifically, the mathematical expression for the preset loss function is: In the formula, This represents the value of the loss function. This represents the total number of training samples. This represents the sales labeling information for the i-th training sample. This represents the sales prediction value of the i-th training sample output by the fusion prediction network framework. The weighting coefficients representing the sales forecast error. This represents the inventory labeling information for the i-th training sample. This represents the inventory prediction value of the i-th training sample output by the fusion prediction network framework. The weighting coefficients representing inventory forecasting errors. This represents the set of trainable parameters in the fusion prediction network framework. The square of the L2 norm of the trainable parameter set is represented. This represents the regularization coefficient.

[0034] In this embodiment, the preset dataset partitioning ratio is typically determined based on the total sample size of the historical fresh food supply chain dataset. Generally, 70% to 80% of the data is used as the training set, and the remaining 20% ​​to 30% as the validation set. The specific value of the preset dataset partitioning ratio must ensure that the training set contains sufficiently diverse samples to support the fusion prediction network framework in fully learning features, while the validation set can effectively represent the overall data distribution for reliable evaluation. The preset maximum number of iterations is determined based on the convergence of the fusion prediction network framework on the training set. It is usually stopped when the error on the validation set no longer decreases. The specific value of the preset maximum number of iterations can be set by observing the change curve of the preset loss function during the training process. It is generally set in the range of 100 to 500 iterations to ensure that the parameters of the temporal convolutional network and the attention mechanism network are fully updated and to avoid overfitting. The preset prediction accuracy threshold is determined based on business needs and tolerance for prediction accuracy, and is usually set between 85% and 95%. The specific value of the preset prediction accuracy threshold needs to take into account the actual application scenario of fresh product demand prediction. If high accuracy is required, a higher threshold such as 95% should be set, and if a certain error is allowed, a lower threshold such as 85% should be set to ensure that the fresh product demand prediction model meets the actual application requirements.

[0035] The input features from the training data set are fed into the fusion prediction network framework for forward propagation calculation to obtain the demand forecast values ​​for fresh products, including sales forecast values ​​and inventory forecast values. Then, the errors between the sales forecast values ​​and the sales label information, and the errors between the inventory forecast values ​​and the inventory label information are calculated according to the preset loss function L, and a regularization term is added. Next, the gradient of the preset loss function L with respect to the parameters of the temporal convolutional network and the attention mechanism network is calculated using the chain rule. The gradient is backpropagated from the output layer to the input layer, and the parameters are updated layer by layer. During the update, an optimizer such as Adam or SGD is used to adjust the parameter update step size according to the learning rate, so that the preset loss function L gradually decreases. This process is repeated until the number of iterations reaches the preset maximum number of iterations, thus completing the gradient update of the parameters of the temporal convolutional network and the attention mechanism network. After each iteration or every few iterations, the data validation set is input into the current fusion prediction network framework for forward propagation calculation to obtain the fresh product demand prediction value for each sample in the data validation set, including the sales forecast value and the inventory forecast value. Then, the sales forecast value is compared with the sales label information, the number of correctly predicted samples is counted, and the prediction accuracy is calculated. The prediction accuracy can be calculated as the proportion of samples where the error between the sales forecast value and the sales label information is within the allowable range, or as the combined accuracy of sales and inventory. The calculated prediction accuracy is compared with the preset prediction accuracy threshold. If the prediction accuracy of the current fusion prediction network framework is greater than the preset prediction accuracy threshold, training is stopped, and the fusion prediction network framework corresponding to the current iteration is determined as the fresh product demand prediction model; otherwise, training continues until the preset maximum number of iterations is reached.

[0036] Specifically, in the inventory control plan generation module, the fused dataset is input into the fresh produce demand forecasting model for predictive analysis, resulting in demand forecasting results that include the predicted demand for fresh produce and the adjusted inventory levels for fresh produce, including: The real-time sales data and real-time inventory data in the fused dataset, as well as at least one of the temperature data, humidity data, and environmental image data in the fused dataset, are used as input data and fed into the fresh produce demand prediction model. Temporal features are extracted through a temporal convolutional network to obtain local temporal features. The local temporal features are then enhanced with global contextual features through a multi-head self-attention network to obtain enhanced fused features. The enhanced fused features are then fed into a fully connected output layer for linear mapping and nonlinear activation to obtain the predicted demand for fresh produce. The initial inventory adjustment is calculated based on the difference between the predicted demand for fresh produce and the current inventory, combined with the preset inventory volatility coefficient and safety stock level. The initial inventory adjustment amount is compared with the current available storage capacity upper limit threshold in the warehouse management system. When the initial inventory adjustment amount is greater than the current available storage capacity upper limit threshold, the current available storage capacity upper limit threshold is used as the fresh product inventory adjustment amount. When the initial inventory adjustment amount is less than or equal to the current available storage capacity upper limit threshold, the initial inventory adjustment amount is used as the fresh product inventory adjustment amount. The fresh product demand forecast amount and the fresh product inventory adjustment amount are combined and packaged into a demand forecast result.

[0037] Specifically, the initial inventory adjustment amount The mathematical expression is: In the formula, Represents the demand difference response coefficient. This indicates the forecast for demand for fresh produce. Indicates the current inventory level. Indicates the sensitivity coefficient to demand fluctuations. This represents the total number of historical prediction cycles. This represents the demand forecast for the t-th historical period. This represents the average of historical demand forecasts.

[0038] In this embodiment, real-time sales data and real-time inventory data from the fused dataset, as well as at least one of temperature data, humidity data, and environmental image data from the fused dataset, are used as input data and fed into a temporal convolutional network. The temporal convolutional network consists of multiple stacked causal convolutional layers. Each convolutional layer uses a fixed-size convolutional kernel that slides along the time dimension to perform convolution operations on the local receptive field of the input sequence, capturing the dependencies between adjacent time steps. By introducing an expanded convolution mechanism, the receptive field is increased layer by layer, enabling the temporal convolutional network to extract a longer range of temporal patterns. Each convolutional layer is usually followed by batch normalization and a non-linear activation function such as ReLU to enhance feature representation capabilities. After layer-by-layer abstraction by the multi-layer temporal convolutional network, the final output is a local temporal feature containing local patterns in the time dimension. Subsequently, the local temporal features are input into a multi-head self-attention network. First, three different linear transformations map the features into a query matrix, a key matrix, and a value matrix. The multi-head self-attention mechanism divides the query, key, and value into multiple heads, each independently calculating its attention weight. This attention score is obtained by the dot product of the query and the key, normalized by softmax, and then weighted and summed with the value to capture global dependencies in different subspaces. The outputs of all heads are concatenated and subjected to a linear transformation to obtain an enhanced fused feature incorporating the global context. The enhanced fused feature is first converted into a fixed-length feature vector through global average pooling or flattening, and then input into a fully connected output layer. This fully connected output layer typically consists of multiple stacked fully connected layers. The first layer maps the features to the hidden dimension through a linear transformation and follows a non-linear activation function such as ReLU for non-linear mapping to improve the model's expressive power. The last fully connected layer converts the hidden representation into the target dimension, i.e., the predicted demand for fresh produce, corresponding to the predicted demand value at future time steps, ultimately yielding the predicted demand for fresh produce.

[0039] Specifically, in the inventory control plan generation module, the inventory turnover parameters of fresh produce are dynamically adjusted based on the demand forecast results and temperature data to obtain an inventory control plan that includes replenishment timing and replenishment quantity, including: Based on the fresh produce inventory adjustment amount in the demand forecast results, the net demand is calculated by combining the current inventory and the inventory in transit. The net demand is then matched with the preset order quantity model to obtain the initial replenishment quantity. Based on the initial replenishment quantity and the inventory turnover target, the theoretical replenishment time point is calculated. The theoretical replenishment time point is the moment when the inventory drops to the safety stock level. The temperature data is compared with the optimal storage temperature range corresponding to the fresh product category to obtain the temperature offset. The dynamic remaining shelf life under the current temperature condition is calculated based on the temperature offset and the preset shelf life decay function. The theoretical replenishment time point is forward or backward corrected based on the dynamic remaining shelf life to obtain the replenishment time point. The replenishment time point is matched with the supplier's delivery cycle. When the replenishment time point conflicts with the delivery cycle, the initial replenishment quantity is adjusted to obtain the replenishment quantity. An initial inventory control plan is generated based on the replenishment time point and replenishment quantity. The initial inventory control plan is then compared with the workload data in the warehouse operation scheduling system for collision detection. When it is detected that the warehouse operation load corresponding to the replenishment time point exceeds the preset operation capacity threshold, the replenishment time point is shifted forward or backward to an available time period when the operation load is lower than the preset operation capacity threshold, thus obtaining an inventory control plan that includes the replenishment time point and replenishment quantity.

[0040] In this embodiment, the preset order quantity model is obtained by collecting historical demand data, single order cost, and unit inventory holding cost for fresh produce, and then substituting these data into the basic formula for economic order quantity (EOQ). Specifically, the average demand for this fresh produce category over a historical period is first calculated as the demand rate D. The fixed cost incurred in placing each order with the supplier is calculated as the order cost S. The storage, loss, and other costs incurred per unit of goods per unit of time are calculated as the unit holding cost H. The basic formula for EOQ is: , This represents the Economic Order Quantity (EOQ). The EOQ value calculated using this basic formula is the preset EOQ model. The preset shelf-life decay function is derived based on the Arrhenius equation and is used to describe the effect of temperature on the rate of quality decay of fresh products. The mathematical expression of the preset shelf-life decay function is: In the formula, This represents absolute temperature, and the unit is Kelvin. Indicates absolute temperature The reaction rate constant at the given time, Represents the frequency factor. Represents an exponential function. Indicates the activation energy of the reaction. Denotes the ideal gas constant, and the reaction rate constant is... By integrating over time and substituting the initial quality value, the basis for calculating the dynamic remaining shelf life can be obtained. By comparing the standard decay rate at standard storage temperature with the actual decay rate at actual storage temperature, the impact of temperature deviation on the remaining shelf life can be quantified. The preset operational capacity threshold is determined by collecting historical operational load data from the warehouse operation scheduling system and performing statistical analysis. The specific method for determining the preset operational capacity threshold is as follows: extract the actual operational load values ​​for the same time period each day within a preset time period, calculate the average and standard deviation of these historical load data, and use the sum of the average and the standard deviation of a preset multiple as the preset operational capacity threshold for that period. Alternatively, the percentile method can be used, where historical load data are arranged in ascending order, and the load value corresponding to the 90th percentile is taken as the preset operational capacity threshold. This ensures that the warehouse operational capacity can cover 90% of the historical peak demand, while avoiding setting the threshold too low, which would lead to frequent triggering of replenishment time adjustments.

[0041] The calculated net demand is compared with the Economic Order Quantity (EOQ) output by the preset Economic Order Quantity (EOQ) model. If the net demand is exactly equal to or an integer multiple of EOQ, the net demand is directly used as the initial replenishment quantity. If the net demand is less than EOQ, EOQ is used as the initial replenishment quantity. If the net demand is greater than EOQ but not an integer multiple of EOQ, the smallest integer multiple of EOQ greater than the net demand is used as the initial replenishment quantity. Subsequently, real-time temperature data of the fresh produce storage environment is collected, and the difference between this data and the optimal storage temperature range corresponding to the product category is calculated to obtain the temperature offset. The current absolute temperature is then used as the basis for further calculations. Substituting the values ​​into the preset shelf-life decay function, the actual decay rate constant at the current temperature is calculated. Simultaneously, the standard decay rate constant k0 corresponding to the median temperature T0 of the optimal storage temperature range is calculated. The quotient of the standard remaining shelf life multiplied by the standard decay rate constant k0 and divided by the actual decay rate constant k is used to obtain the dynamic remaining shelf life under the current temperature conditions. The calculated dynamic remaining shelf life is compared with the expected remaining shelf life corresponding to the theoretical replenishment time. If the dynamic remaining shelf life is less than the expected remaining shelf life, it indicates that the temperature deviation is causing accelerated quality deterioration, and the theoretical replenishment time needs to be adjusted forward, i.e., replenishing earlier to avoid the product exceeding its shelf life when inventory drops to the safety stock level. If the dynamic remaining shelf life is greater than the expected remaining shelf life, it indicates that the storage conditions are better than the standard state, and the theoretical replenishment time can be adjusted backward, i.e., replenishing is postponed to reduce inventory holding costs. The adjustment range is proportional to the difference between the dynamic remaining shelf life and the expected remaining shelf life. The system obtains the supplier's fixed delivery cycle for this fresh produce category. The revised replenishment time point is compared with the corresponding arrival time window of the delivery cycle. If the replenishment time point falls within the arrival time window covered by the delivery cycle, no adjustment to the initial replenishment quantity is needed. If the replenishment time point conflicts with the delivery cycle (i.e., delivery cannot be arranged at that time), the replenishment time point is adjusted to the nearest available delivery time. Simultaneously, the corresponding demand is recalculated based on the adjusted replenishment interval days. This demand is then matched again with the preset economic order quantity model to obtain the adjusted replenishment quantity, ensuring that the replenishment quantity matches the new replenishment time point. The replenishment time point and replenishment quantity are combined to form an initial inventory control plan record. Using the replenishment time point in this initial inventory control plan record as a query condition, the system calls the warehouse operation scheduling system interface to obtain the scheduled operational load data for the corresponding time period. The obtained operational load data is compared with a preset operational capacity threshold. If the operational load data is lower than or equal to the preset operational capacity threshold, the collision detection passes; if the operational load data is higher than the preset operational capacity threshold, the collision detection fails, indicating that the warehouse operational capacity is insufficient at that replenishment time point. When a collision detection fails, the system scans the workload data in the warehouse operation scheduling system sequentially forward and backward from the original replenishment time point. It finds the first available time period where the workload is lower than the preset operational capacity threshold and uses it as the new replenishment time point. This new replenishment time point is then matched with the supplier's delivery cycle. If there is a conflict with the delivery cycle, the scan continues until an available time period that simultaneously meets the conditions of workload being lower than the preset operational capacity threshold and delivery cycle is found. This available time period is then determined as the replenishment time point in the final inventory control plan. The replenishment quantity is adjusted based on the time difference between the new and original replenishment time points to ensure that the replenishment quantity matches the new replenishment time point, thereby generating the final inventory control plan that includes the replenishment time point and the replenishment quantity.

[0042] Specifically, in the path optimization scheme generation module, the warehouse environment parameters are analyzed based on the inventory control scheme and humidity data to obtain warehouse environment control instructions, including: Based on the replenishment time and quantity in the inventory control plan, and combined with historical inventory data, the product density and space occupancy rate of each storage zone in the future preset period are predicted to obtain the storage load prediction data. The humidity data is then analyzed to determine the deviation between the humidity data and the optimal humidity range corresponding to the fresh product category to obtain the current humidity deviation value. The warehouse load prediction data and the current humidity deviation value are input into a preset fuzzy logic controller to reason about the priority of warehouse environment regulation, generate regulation weight coefficients for different temperature and humidity control areas, and perform joint optimization analysis on the operating parameters of refrigeration units, humidification equipment and ventilation equipment based on the regulation weight coefficients to obtain a comprehensive environmental regulation parameter set including target temperature value, target humidity value, equipment operating power and operating time. The comprehensive environmental control parameter set is compared with the real-time operating status parameters of each environmental control device to generate a sequence of device control instructions containing device identifiers, control mode switching instructions and parameter setting values. The sequence of device control instructions is then output to the centralized controller of the warehouse environment as a warehouse environment control instruction.

[0043] In this embodiment, the future preset time period is set to the next 4 hours based on the inventory turnover rate to match the replenishment cycle and prediction accuracy. The optimal humidity range is determined according to the preservation standards of different fresh product categories. For example, the optimal humidity range for leafy vegetables is 85%–95%, and for root vegetables it is 80%–90%. The input variable of the preset fuzzy logic controller, the storage load prediction data, has a domain of discourse set to 0–100%, the current humidity deviation value has a domain of discourse set to -1–1, and the output variable, the control weight coefficient, has a domain of discourse set to 0–1. The fuzzy rules are formulated based on expert experience, and the precise weights are obtained by defuzzification using Mamdani inference and the centroid method.

[0044] The departure time window for the transportation task is determined based on the replenishment time point in the inventory control plan, and the vehicle load requirement for the transportation task is determined based on the replenishment quantity. Environmental control requirements for fresh products during transportation are determined based on the dehumidification or humidification operation in the warehouse environment control instructions. Historical environmental data on candidate transportation routes is queried according to the environmental control requirements to obtain an environmental suitability score for each candidate transportation route. Candidate transportation routes that meet the load and time constraints are then selected based on the vehicle load requirement and departure time window. The environmental suitability score, vehicle load requirement, and departure time window are input into a multi-objective path planning algorithm to optimize the selected candidate transportation routes and obtain an optimized transportation route plan. The multi-objective path planning algorithm uses a path planning comprehensive cost function for analysis. The mathematical expression of the path planning comprehensive cost function is: In the formula, This represents the overall cost of path planning. Indicates the number of road segments in the path. This represents the distance of the k-th road segment. This represents the weighting factor for distance cost (e.g., 0.2). This represents the estimated travel time for the k-th road segment. A weighting factor representing time cost (e.g., 0.4). This represents the humidity value of the k-th road segment. Indicates the target humidity value. A weighting factor (e.g., 0.4) representing the cost of environmental adaptability. This indicates the replenishment quantity for the current transportation task. Indicates the vehicle's maximum load capacity. This represents the weighting factor (e.g., 100) for the overload penalty. This represents the function that takes the maximum value. If the result is greater than 0, then take that value. Value; if the If the value is less than or equal to 0, then 0 is taken; the candidate transportation routes are traversed and calculated according to the route planning comprehensive cost function to obtain the route planning comprehensive cost of each candidate transportation route. The route planning comprehensive cost is compared with the preset cost threshold, and the candidate transportation route with the minimum route planning comprehensive cost is selected as the transportation route optimization scheme.

[0045] Specifically, in the anomaly monitoring module, the circulation status of fresh produce is monitored and analyzed using the environmental image data and transportation route optimization scheme to obtain anomaly monitoring results, including: The environmental image data is analyzed by frame sequence and target detection using an image feature extraction network to obtain the surface condition features and location distribution features of fresh products. Based on the path node information and time information in the transportation route optimization scheme, the surface state characteristics and location distribution characteristics of the fresh products are spatiotemporally aligned to obtain a flow state characteristic sequence associated with the transportation route. An anomaly degree analysis is performed on each feature point in the flow status feature sequence by using preset anomaly discrimination rules to obtain an anomaly degree score. When the anomaly degree score exceeds a preset anomaly threshold, the feature point corresponding to the preset anomaly threshold is marked as a flow anomaly monitoring result.

[0046] In this embodiment, when performing frame sequence analysis and target detection on the environmental image data using an image feature extraction network, the continuously acquired environmental image data is first organized into a frame sequence according to the acquisition time order. The image feature extraction network uses a convolutional neural network as its backbone structure and integrates features from adjacent layers through a cross-layer feature fusion module to alleviate multi-scale feature differences. During the encoding stage, the image feature extraction network extracts deep semantic features and shallow detail features from the environmental image data layer by layer. High-level features are extracted using a semantic awareness feature extraction module to uncover the location distribution features of fresh produce, while low-level features are recovered using a detail-aware context attention module to restore the surface state features of fresh produce, including visual attributes such as color, texture, and integrity. During the decoding stage, the image feature extraction network fuses multi-scale features and uses a target detection head to locate and classify fresh produce instances in each frame image, ultimately outputting the surface state feature description vector and location distribution feature coordinates of all fresh produce in each frame image.

[0047] When performing spatiotemporal alignment of the surface state features and location distribution features of the fresh produce based on the path node information and time information in the transportation route optimization scheme, the geographical coordinates and estimated arrival time of each path node are first extracted from the transportation route optimization scheme to form a spatiotemporal reference frame. Then, the surface state features and location distribution features obtained through an image feature extraction network are organized according to the acquisition timestamp, and the acquisition time of each feature point is matched with the time information of the path node to determine the location of the transportation path node corresponding to each feature point. A linear interpolation method is used to calculate the relative position coordinates of feature points acquired between two path nodes on the path, achieving a precise association between the surface state features and location distribution features and the spatial location of the path. All feature points are arranged in chronological order, while retaining the path node identifier and relative time offset associated with each feature point, ultimately forming a sequence of flow state features associated with the transportation path.

[0048] When analyzing the anomaly level of each feature point in the flow state feature sequence using preset anomaly discrimination rules, it is first necessary to preset the anomaly score calculation method and the specific value of the preset anomaly threshold in the anomaly discrimination rules. The anomaly score calculation method adopts the Z-score model based on statistics. The Z-score calculation formula defined in the preset anomaly discrimination rules is the difference between the observed value of the feature point and the historical normal state mean divided by the historical normal state standard deviation. The preset anomaly threshold is set to 3.0. This preset anomaly threshold is determined based on the three-standard-deviation principle in statistics. When the Z-score of a feature point exceeds 3.0, it indicates that the deviation of the observed value from the normal mean has reached a statistically significant level. For each feature point in the flow state feature sequence, the anomaly degree scoring module first extracts the surface state feature vector and position distribution feature value of the feature point, compares them with the preset normal state distribution model, calculates the comprehensive deviation of the feature point in each dimension, and generates an anomaly degree score between 0 and 1. When the anomaly degree score exceeds the preset anomaly threshold of 3.0, the anomaly discrimination system automatically marks the feature point as a flow anomaly monitoring result, triggering subsequent alarm and recording processes.

[0049] When the abnormal monitoring results meet the preset abnormal conditions, abnormal handling information containing abnormal type identifier and abnormal location information is generated, and the abnormal handling information is sent to the corresponding mobile terminal, including: The abnormal flow monitoring results are matched against a preset abnormal condition rule base. The abnormal type identifier in the monitoring results is compared with the abnormal type field in the preset abnormal condition rule base. When the abnormal type identifier matches any abnormal type in the preset abnormal condition rule base, the handling priority and handling measure type corresponding to the abnormality are determined according to the matched rule. Based on the handling priority and handling measure type, and combined with the abnormality occurrence location information in the monitoring results, the handling task allocation weight for each candidate mobile terminal is calculated using an abnormality handling decision function. The mathematical expression of the abnormality handling decision function is: In the formula, Indicates the index of the candidate mobile device. Indicates the first The anomaly handling decision function value for each candidate mobile device. Indicates the location where the anomaly occurred. Indicates the first The current location of each candidate mobile device and the location where the anomaly occurred. Geographical distance between them This represents a very small constant that avoids division by zero. This represents the weighting coefficient of the distance factor (e.g., 0.5). Indicates the first Current task load of each candidate mobile device , Indicates the first The deviation value of the match between the skill level of the personnel handling the incident corresponding to each candidate mobile device and the current anomaly handling measure type. , This represents the weighting factor of the load factor (e.g., 0.2). The weighting coefficient (e.g., 0.3) represents the skill matching factor, calculated by comparing the corresponding factors for each candidate mobile terminal. Value, select The candidate mobile terminal with the largest value is selected as the target mobile terminal. Based on the selected target mobile terminal and the anomaly type identifier and anomaly location information in the abnormal flow monitoring results, the preset anomaly handling information template is filled with fields to generate anomaly handling information containing anomaly type identifier, anomaly location information and handling guidance content, and the anomaly handling information is sent to the corresponding target mobile terminal.

[0050] In this embodiment, the determination of the preset anomaly condition rule base is based on the analysis of historical anomaly data and business experience. Various anomalies are classified according to anomaly type identifiers. The value retrieval logic of the preset anomaly condition rule base is to compare the anomaly type identifier in the real-time monitoring results with the anomaly type field in the base. If a match is found, the predefined value corresponding to that anomaly type is directly retrieved from the rule base: handling priority (e.g., high, medium, low) and handling measure type (e.g., on-site repair, remote guidance).

[0051] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A whole-process digital management system for a fresh food supply chain, characterized in that: include: The data acquisition module is used to acquire real-time status data of fresh products throughout the entire fresh food supply chain through IoT devices, and to preprocess the status data to obtain a fused dataset containing timestamps and location identifiers. The fused dataset includes real-time sales data, real-time inventory data, temperature data, humidity data, and environmental image data. The model building module is used to construct a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network. It iteratively trains the fusion prediction network framework based on a historical fresh food supply chain dataset containing sales volume labeling information and inventory labeling information. When the prediction accuracy output by the trained fusion prediction network framework exceeds a preset prediction accuracy threshold, training stops, and a fresh food product demand prediction model is output. The historical fresh food supply chain dataset contains historical sales volume data, historical inventory data, historical temperature data, historical humidity data, and historical environmental image data of the same type as the fusion dataset. The sales volume labeling information is used to identify the actual sales volume data in the historical fresh food supply chain dataset, and the inventory labeling information is used to identify the actual inventory data in the historical fresh food supply chain dataset. The inventory control scheme generation module is used to input the fused dataset into the fresh product demand forecasting model for predictive analysis, obtain demand forecast results including the predicted demand amount and the fresh product inventory adjustment amount, and dynamically adjust the inventory turnover parameters of fresh products based on the demand forecast results and temperature data to obtain an inventory control scheme including replenishment time points and replenishment quantities. The route optimization scheme generation module is used to analyze the storage environment parameters based on the inventory control scheme and humidity data to obtain the storage environment control instructions, and to perform collaborative analysis on the transportation route planning parameters based on the storage environment control instructions and the inventory control scheme to obtain the transportation route optimization scheme. The anomaly monitoring module is used to monitor and analyze the circulation status of fresh products through the environmental image data and transportation route optimization scheme, obtain circulation anomaly monitoring results, and when the circulation anomaly monitoring results meet the preset anomaly conditions, generate anomaly handling information containing anomaly type identifier and anomaly location information, and send the anomaly handling information to the corresponding mobile terminal. 2.The system according to claim 1, characterized in that: The model building module constructs a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network, comprising: A temporal convolutional network component is constructed based on a preset inflation factor sequence and a preset convolutional kernel size. The temporal data in the historical fresh food supply chain dataset is processed by the temporal convolutional network component, and local temporal features of the temporal data under different inflation rates are extracted to obtain a local temporal feature set containing multi-scale time dependencies. A multi-head self-attention mechanism network component is constructed based on a preset number of attention heads and a preset feature dimension. The multi-head self-attention mechanism network component is then used to perform a linear transformation on each feature vector in the local temporal feature set to generate a query matrix, a key matrix, and a value matrix. The query matrix and key matrix are sequentially subjected to dot product, scaling, and softmax normalization to obtain the attention weight matrix. The value matrix is ​​then weighted and summed according to the attention weight matrix to obtain the context feature vector of each attention head. The context feature vector of each attention head is then sequentially concatenated and linearly transformed to obtain the enhanced temporal features containing global context information. A feature fusion layer is constructed based on the local temporal feature set and the enhanced temporal features. The local temporal feature set and the enhanced temporal features are concatenated through the feature fusion layer to obtain a concatenated feature vector. The concatenated feature vector is then input into a fully connected network for nonlinear transformation to obtain a fusion prediction network framework that includes a temporal convolutional network and an attention mechanism network. 3.The system according to claim 1, characterized in that: In the model building module, the fusion prediction network framework is iteratively trained based on a historical fresh food supply chain dataset containing sales volume and inventory labeling information. When the prediction accuracy output by the trained fusion prediction network framework is greater than a preset prediction accuracy threshold, training stops, and a fresh food product demand prediction model is output, including the following steps: The historical fresh food supply chain dataset is divided according to a preset dataset division ratio to obtain a data training set for model parameter updates and a data validation set for model performance evaluation. The training data set is input into the fusion prediction network framework. The input features in the training data set are forward propagated according to the preset loss function to obtain the sales forecast and inventory forecast of fresh products. The parameters of the temporal convolutional network and the attention mechanism network in the fusion prediction network framework are updated by gradient based on the back propagation algorithm. When the number of iterations reaches the preset maximum number of iterations, the parameter update is stopped. The prediction accuracy of the current fusion prediction network framework is verified based on the data validation set. When the verified prediction accuracy is greater than the preset prediction accuracy threshold, the fusion prediction network framework corresponding to the current iteration round is determined as the fresh product demand prediction model.

4. The digital management system for the entire fresh food supply chain as described in claim 3, characterized in that: A mathematical expression of the preset loss function is: , wherein, represents a loss function value, represents a total number of training samples, represents sales label information of the i-th training sample, represents a sales prediction value of the i-th training sample output by the fusion prediction network framework, represents a weight coefficient of the sales prediction error, represents inventory label information of the i-th training sample, represents an inventory prediction value of the i-th training sample output by the fusion prediction network framework, represents a weight coefficient of the inventory prediction error, represents a set of trainable parameters in the fusion prediction network framework, represents a square of an L2 norm of the set of trainable parameters, represents a regularization coefficient.

5. The digital management system for the entire fresh food supply chain as described in claim 1, characterized in that: In the inventory control plan generation module, the fused dataset is input into the fresh produce demand forecasting model for predictive analysis, yielding demand forecast results that include the predicted demand for fresh produce and the adjusted inventory levels for fresh produce. The real-time sales data and real-time inventory data in the fused dataset, as well as at least one of the temperature data, humidity data, and environmental image data in the fused dataset, are used as input data and fed into the fresh produce demand prediction model. Temporal features are extracted through a temporal convolutional network to obtain local temporal features. The local temporal features are then enhanced with global contextual features through a multi-head self-attention network to obtain enhanced fused features. The enhanced fused features are then fed into a fully connected output layer for linear mapping and nonlinear activation to obtain the predicted demand for fresh produce. The initial inventory adjustment is calculated based on the difference between the predicted demand for fresh produce and the current inventory, combined with the preset inventory volatility coefficient and safety stock level. The initial inventory adjustment amount is compared with the current available storage capacity upper limit threshold in the warehouse management system. When the initial inventory adjustment amount is greater than the current available storage capacity upper limit threshold, the current available storage capacity upper limit threshold is used as the fresh product inventory adjustment amount. When the initial inventory adjustment amount is less than or equal to the current available storage capacity upper limit threshold, the initial inventory adjustment amount is used as the fresh product inventory adjustment amount. The fresh product demand forecast amount and the fresh product inventory adjustment amount are combined and packaged into a demand forecast result.

6. The digital management system for the entire fresh food supply chain as described in claim 5, characterized in that: The initial inventory adjustment amount The mathematical expression is: In the formula, Represents the demand difference response coefficient. This indicates the forecast for demand for fresh produce. Indicates the current inventory level. Indicates the sensitivity coefficient to demand fluctuations. This represents the total number of historical prediction cycles. This represents the demand forecast for the t-th historical period. This represents the average of historical demand forecasts.

7. The digital management system for the entire fresh food supply chain as described in claim 1, characterized in that: In the inventory control plan generation module, the inventory turnover parameters of fresh produce are dynamically adjusted based on the demand forecast results and temperature data to obtain an inventory control plan that includes replenishment timing and replenishment quantity, including: Based on the fresh produce inventory adjustment amount in the demand forecast results, the net demand is calculated by combining the current inventory and the inventory in transit. The net demand is then matched with the preset order quantity model to obtain the initial replenishment quantity. Based on the initial replenishment quantity and the inventory turnover target, the theoretical replenishment time point is calculated. The theoretical replenishment time point is the moment when the inventory drops to the safety stock level. The temperature data is compared with the optimal storage temperature range corresponding to the fresh product category to obtain the temperature offset. The dynamic remaining shelf life under the current temperature condition is calculated based on the temperature offset and the preset shelf life decay function. The theoretical replenishment time point is forward or backward corrected based on the dynamic remaining shelf life to obtain the replenishment time point. The replenishment time point is matched with the supplier's delivery cycle. When the replenishment time point conflicts with the delivery cycle, the initial replenishment quantity is adjusted to obtain the replenishment quantity. An initial inventory control plan is generated based on the replenishment time point and replenishment quantity. The initial inventory control plan is then compared with the workload data in the warehouse operation scheduling system for collision detection. When it is detected that the warehouse operation load corresponding to the replenishment time point exceeds the preset operation capacity threshold, the replenishment time point is shifted forward or backward to an available time period when the operation load is lower than the preset operation capacity threshold, thus obtaining an inventory control plan that includes the replenishment time point and replenishment quantity.

8. The digital management system for the entire fresh food supply chain as described in claim 1, characterized in that: In the path optimization scheme generation module, the warehouse environment parameters are analyzed based on the inventory control scheme and humidity data to obtain warehouse environment control instructions, including: Based on the replenishment time and quantity in the inventory control plan, and combined with historical inventory data, the product density and space occupancy rate of each storage zone in the future preset period are predicted to obtain the storage load prediction data. The humidity data is then analyzed to determine the deviation between the humidity data and the optimal humidity range corresponding to the fresh product category to obtain the current humidity deviation value. The warehouse load prediction data and the current humidity deviation value are input into a preset fuzzy logic controller to reason about the priority of warehouse environment regulation, generate regulation weight coefficients for different temperature and humidity control areas, and perform joint optimization analysis on the operating parameters of refrigeration units, humidification equipment and ventilation equipment based on the regulation weight coefficients to obtain a comprehensive environmental regulation parameter set including target temperature value, target humidity value, equipment operating power and operating time. The comprehensive environmental control parameter set is compared with the real-time operating status parameters of each environmental control device to generate a sequence of device control instructions containing device identifiers, control mode switching instructions and parameter setting values. The sequence of device control instructions is then output to the centralized controller of the warehouse environment as a warehouse environment control instruction.

9. The digital management system for the entire fresh food supply chain as described in claim 1, characterized in that: In the anomaly monitoring module, the circulation status of fresh produce is monitored and analyzed using the environmental image data and transportation route optimization scheme to obtain anomaly monitoring results, including: The environmental image data is analyzed by frame sequence and target detection using an image feature extraction network to obtain the surface condition features and location distribution features of fresh products. Based on the path node information and time information in the transportation route optimization scheme, the surface state characteristics and location distribution characteristics of the fresh products are spatiotemporally aligned to obtain a flow state characteristic sequence associated with the transportation route. An anomaly degree analysis is performed on each feature point in the flow status feature sequence by using preset anomaly discrimination rules to obtain an anomaly degree score. When the anomaly degree score exceeds a preset anomaly threshold, the feature point corresponding to the preset anomaly threshold is marked as a flow anomaly monitoring result.

Images