An intelligent canteen meal supply amount intelligent regulation and control system based on end-cloud cooperation

Abstract

This invention discloses an intelligent supply control system for smart canteens based on edge-cloud collaboration, comprising: a data normalization module for acquiring and preprocessing multi-source supply data; a state graphing module for aggregating states to form a set of state nodes and constructing a supply state graph by combining continuous succession relationships and conditional transition probabilities; a graph recursion module for determining the state propagation neighborhood and aggregating messages between adjacent state nodes; a deviation discrimination module for performing causal convolution inference and imposing interval constraints on the inference results; a control solution module for performing rolling optimization and introducing a state deviation penalty term during the solution process to generate supply control instructions; and a feedback correction module for collecting execution feedback states and correcting the supply state graph. This invention enables dynamic prediction and intelligent control of the supply of smart canteen meals, improving the timeliness of supply adjustments.

Description

Technical Field

[0001] This invention relates to the field of edge-cloud collaboration technology, and in particular to a smart canteen food supply intelligent control system based on edge-cloud collaboration. Background Technology

[0002] With the advancement of smart canteen construction, edge data acquisition terminals, cloud management platforms, and data analysis models are gradually being introduced into the canteen supply process to collect, analyze, and support decision-making based on on-site operational data. Existing technologies typically estimate subsequent supply volumes based on historical supply records, fixed meal service time rules, or simple prediction models, and then adjust the on-site supply process in conjunction with human experience.

[0003] However, most existing smart canteen supply control methods still rely on data statistics and rule matching, lacking a structured expression of state changes during continuous operation, making it difficult to accurately depict the relationship and evolution of different supply states. When dining demand experiences local fluctuations, short-term concentrations, or periodic changes, simply relying on historical averages, fixed thresholds, or single-point predictions can easily lead to lagging supply control.

[0004] Meanwhile, existing cloud-based analytics methods often directly predict the collected data, lacking collaborative modeling between real-time edge status and long-term cloud patterns. Insufficient correlation updates between on-site feedback and cloud models result in the system struggling to make timely adaptive corrections when there are discrepancies between the predicted results and the actual execution status.

[0005] In addition, existing control methods usually record the execution results after the fact, lacking a closed-loop mechanism that uses the execution feedback to update the state relationship and correct the control parameters. This results in insufficient dynamic adaptability of the system during the continuous meal supply cycle, making it difficult to stably support the refined control of the food supply in smart canteens.

[0006] Therefore, how to provide a smart canteen food supply intelligent control system based on edge-cloud collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent control system for the food supply of a smart canteen based on edge-cloud collaboration. This invention fully utilizes edge-cloud collaboration, DBSCAN density clustering, graph neural networks, causal temporal convolution, quantile regression, and model predictive control technologies. It describes in detail the intelligent implementation process of aggregating and constructing graphs of the supply status of the smart canteen, recursively updating graphs, identifying deviations, solving rolling control, and providing feedback correction. It has the advantages of accurate status perception, timely supply control, and strong online adaptability.

[0008] According to an embodiment of the present invention, a smart canteen meal supply intelligent control system based on edge-cloud collaboration includes: The data standardization module is used to acquire and preprocess multi-source supply data from the smart canteen to form a standard supply dataset, while establishing a collaborative control space in the cloud. The state graph module is used to aggregate states based on a standard supply dataset using the DBSCAN clustering algorithm to form a set of state nodes. It then constructs a supply state graph by combining the continuous connection relationships and conditional transition probabilities between the state nodes. The graph recursion module is used to determine the state propagation neighborhood based on the node connectivity relationship within the collaborative control space. It uses a graph neural network to aggregate messages between adjacent state nodes and recursively updates the current node state through a gated loop unit. The deviation discrimination module is used to perform causal convolutional deduction on the state evolution trajectory within a continuous control cycle based on the recursively updated state nodes and a temporal convolutional network. It then uses quantile regression to constrain the deduction results to determine the supply deviation state. The regulation and solution module is used to establish a set of candidate actions based on the supply deviation state, perform rolling optimization solution using model predictive control method, and introduce a state offset penalty term in the solution process to generate supply regulation instructions; The feedback correction module is used to collect the execution feedback status of supply control instructions and correct the supply status diagram using an incremental gradient update method.

[0009] Optionally, the data straightening module includes: The multi-source supply data of the smart canteen is rearranged in time sequence according to the collection timestamp, and the rearranged multi-source supply data is window merged according to a fixed sampling window. The data fields within the same sampling window are unified in field name, converted in field type and calibrated in value scale to obtain regularized supply data. Among them, the multi-source supply data includes the supply object identifier, sampling time, supply status value, usage status value, settlement status value and historical control status value collected at the edge; Missing positions and abnormal positions are marked in the regular supply data. The missing positions are filled in by interpolation results of adjacent sampling windows, and the abnormal positions are replaced by median filtering results to obtain valid supply data. The valid supply data is organized into samples according to the sampling window order. The valid supply data corresponding to each sampling window is used as a status sample, and a window index is configured for the status sample to form a standard supply dataset. Establish a collaborative control space in the cloud, register the standard supply dataset according to the window index, and establish a data synchronization relationship between the edge and the collaborative control space.

[0010] Optionally, the state mapping module includes: Based on the window index corresponding to each state sample in the standard supply dataset, the state samples are read in the order of the sampling window, and the scale is made consistent for each state sample. The DBSCAN clustering algorithm is used to perform density aggregation on the scale-consistent state samples. Core samples are determined based on the sample distribution within the neighborhood radius, and state samples that satisfy the density reachability relationship with the core samples are grouped into the same cluster. Each cluster is treated as a state node, and state samples belonging to the same cluster are mapped to the same state node to form a set of state nodes. According to the window index, record the order in which each state node appears in the continuous sampling window. When two state nodes correspond to state samples in adjacent sampling windows, establish a connection edge between the two state nodes. The node transition distribution is statistically analyzed based on the connection relationship of state nodes in the continuous sampling window, and the connection edges are weighted using conditional transition probabilities. A supply state graph is constructed using the set of state nodes as graph nodes, the connecting edges as graph edges, and the weights of the connecting edges as edge attributes.

[0011] Optionally, the formation of the clusters includes: According to the order of the state samples in the standard supply dataset, the state samples to be judged are selected one by one, and the neighboring state samples are retrieved within the preset neighborhood radius according to the Euclidean distance, with the state sample to be judged as the center, to form a neighborhood sample set. The number of state samples in the neighborhood sample set is counted. State samples whose number of state samples reaches the minimum sample number parameter are marked as core samples. State samples that do not reach the minimum sample number parameter but fall within the neighborhood of core samples are marked as boundary samples. Using the core sample that has not been assigned to a cluster as the starting point for aggregation, the boundary samples that are in the same neighborhood as the core sample are merged into the current cluster, and the neighborhood sample set of the newly added core sample in the current cluster is retrieved. Density reachability expansion is performed on the neighborhood sample set of the newly added core sample, and state samples that satisfy the density reachability relationship are incorporated into the current cluster until there are no more core samples in the current cluster that can be expanded. For state samples that are not assigned to the current cluster, continue to perform core sample determination and density reachability expansion to obtain multiple clusters; For state samples that are not assigned to a cluster, the nearest core sample is determined. State samples that meet the neighborhood radius condition are merged into the corresponding cluster, and state samples that do not meet the neighborhood radius condition are treated as isolated clusters.

[0012] Optionally, the graph recursion module includes: Using the current state node in the supply state diagram as the recursive object, the adjacent state nodes of the current state node are determined based on the node connectivity relationship; Based on the connection weights, the propagation strength of adjacent state nodes to the current state node is determined, thus obtaining the state propagation neighborhood; Within the state propagation neighborhood, a graph neural network is used to modulate the node states of adjacent state nodes according to the propagation strength, and the modulated adjacent node states are aggregated into the neighborhood propagation message of the current state node. Based on the neighborhood propagation message and the node state of the current state node, the gated loop unit is used to calculate the update gate state and the reset gate state. Based on the reset gate state, the retention ratio of the current state node participating in the candidate state calculation is adjusted to obtain the candidate update state. Based on the update gate state, the node state of the current state node and the candidate update states are recursively fused to update the node state of the current state node. Based on the node connectivity in the supply state graph, the state propagation neighborhood determination, message modulation aggregation, and gated recursive fusion are repeatedly performed on each state node, and the recursion stops when the preset recursion round is reached, resulting in the recursively updated state node.

[0013] Optionally, the formation of the candidate update state includes: Based on the neighborhood propagation message and the node state of the current state node, a gating association transformation is performed on the neighborhood propagation message and the node state of the current state node to obtain the reset gate state. Based on the reset gate state, the node state of the current state node is subjected to component-level gating screening. The low-response historical state component corresponding to the reset gate state is suppressed, while the high-response historical state component is retained, thus obtaining the controlled node state. The controlled node state and the neighborhood propagation message are jointly transformed into a state, so that the neighborhood propagation message participates in the candidate state reconstruction of the current state node, and the intermediate candidate state is obtained. The intermediate quantities of the candidate states are nonlinearly activated and compressed to obtain the candidate updated states; The candidate update state is used as the state to be merged in the current state node in this round of graph recursion, and is used for subsequent recursive fusion with the node state of the current state node.

[0014] Optionally, the deviation detection module includes: Based on the recursive updated state nodes, the node states are arranged into a node state sequence according to the window index, and the node state sequence corresponding to the continuous control cycle is extracted as the sequence to be deduced. A causal temporal convolutional network is used to perform state inference on the sequence to be inferred. The causal temporal convolutional network only associates the node states before the current convolution position in the convolution operation, and extracts the state changes at different intervals within a continuous control period through dilated convolution to obtain the state inference result. Based on the state deduction results, quantile regression is performed to determine the lower bound quantile state and upper bound quantile state corresponding to each deduction position. The lower bound quantile state and upper bound quantile state constitute the state constraint interval of the corresponding deduction position. Based on the state projection results of each projection location, the state constraint interval of the corresponding projection location is compared. The supply deviation state is determined according to the projection location, deviation direction and deviation magnitude that exceed the state constraint interval.

[0015] Optionally, the formation of the state deduction result includes: According to the arrangement order of the node states in the sequence to be deduced, for each deduction position, extract the node states within the continuous sampling window before that deduction position to obtain the historical node state fragment; According to the preset dilation interval, select the node states with interval distribution in the historical node state segment, and feed the selected node states into the corresponding dilation convolution kernel for convolution operation to obtain the local state change amount under the corresponding time span. The local state changes over different time spans are accumulated and fused according to the same projection location. The residual connection is then used to superimpose the accumulated and fused results with the node state corresponding to the projection location to obtain the position fusion state of the projection location. The positional fusion state of each simulation position is nonlinearly compressed and combined according to the arrangement order of each simulation position within the continuous control cycle to form the state simulation result.

[0016] Optionally, the control and solution module includes: The action direction is set based on the deviation direction in the supply deviation state, and the action level is divided according to the deviation magnitude to form a set of candidate actions within the continuous control cycle; Using the current node state that generates the supply deviation state as the initial state for rolling optimization, the candidate action set is combined into a candidate action sequence within a continuous control cycle, and the candidate state trajectory corresponding to each candidate action sequence is recursively obtained. Based on the offset distance between the candidate state trajectory and the state constraint interval of the corresponding inferred position, a state offset penalty term is constructed and embedded into the optimization objective of the model predictive control. The model predictive control method is used to perform rolling optimization on the candidate action sequence and select the candidate action sequence that minimizes the state offset penalty term. In the candidate action sequence, the target control action corresponding to the current control cycle is encoded, control parameters are configured according to the action direction and action level, and the control parameters are encapsulated into supply control instructions that can be executed at the edge.

[0017] Optionally, the feedback correction module includes: The feedback status after the execution of the supply control command is collected at the edge, and the feedback status is incorporated into the corresponding control cycle according to the window index corresponding to the control command. Based on the set of state nodes formed by the state diagram module, the nearest state node is determined for the feedback state, the arrival state node corresponding to the feedback state is determined, and the state node corresponding to the supply deviation state is taken as the starting state node. Based on the starting state node and the arriving state node, determine the actual receiving edge in the supply state diagram. When the actual receiving edge does not exist, add a connection edge between the starting state node and the arriving state node. Based on the offset difference between the feedback status and the supply deviation status, the number of acceptances of the actual receiving edge is incrementally corrected, and the total number of outgoing acceptances corresponding to the same starting status node is updated synchronously. Based on the corrected number of state acceptances and the total number of outgoing acceptances, the conditional transition probability is redefined and used as the connection weight of the actual acceptance edge. The supply state diagram is then corrected based on the updated connection weights.

[0018] The beneficial effects of this invention are: First, this invention organizes and states the multi-source supply data of the smart canteen through edge-cloud collaboration, and uses the DBSCAN clustering algorithm to form a set of state nodes. It then constructs a supply state graph by combining continuous succession relationships and conditional transition probabilities, so that the changes in supply state during continuous operation can be expressed in a graph structure, reducing the problem of missing state relationships caused by simply relying on historical statistics or fixed rules.

[0019] Secondly, this invention utilizes graph neural networks for message aggregation within the collaborative control space and recursively updates state nodes through gated recurrent units. It combines causal temporal convolutional networks and quantile regression to deduce and constrain the state evolution within continuous control cycles, thereby making the judgment of supply deviation no longer dependent on single-point prediction results and enhancing the ability to identify short-term fluctuations and continuous changing trends.

[0020] Finally, this invention establishes a set of candidate actions based on the supply deviation state, uses a model predictive control method for rolling optimization, and incrementally corrects the supply state diagram through execution feedback, so that the supply quantity control command can be continuously adjusted according to the feedback state, thereby improving the timeliness, stability and online adaptability of the smart canteen's food supply quantity control. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a module structure diagram of an intelligent control system for the supply of food in a smart canteen based on edge-cloud collaboration, as proposed in this invention. Figure 2 This invention presents a flowchart illustrating the generation of supply control instructions for a smart canteen food supply intelligent control system based on edge-cloud collaboration. Figure 3 This is a flowchart illustrating the supply status diagram correction process of a smart canteen food supply intelligent control system based on edge-cloud collaboration proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 A smart canteen food supply intelligent control system based on edge-cloud collaboration includes: The data standardization module is used to acquire and preprocess multi-source supply data from the smart canteen to form a standard supply dataset, while establishing a collaborative control space in the cloud. The state graph module is used to aggregate states based on a standard supply dataset using the DBSCAN clustering algorithm to form a set of state nodes. It then constructs a supply state graph by combining the continuous connection relationships and conditional transition probabilities between the state nodes. The graph recursion module is used to determine the state propagation neighborhood based on the node connectivity relationship within the collaborative control space. It uses a graph neural network to aggregate messages between adjacent state nodes and recursively updates the current node state through a gated loop unit. The deviation discrimination module is used to perform causal convolutional deduction on the state evolution trajectory within a continuous control cycle based on the recursively updated state nodes and a temporal convolutional network. It then uses quantile regression to constrain the deduction results to determine the supply deviation state. The regulation and solution module is used to establish a set of candidate actions based on the supply deviation state, perform rolling optimization solution using model predictive control method, and introduce a state offset penalty term in the solution process to generate supply regulation instructions; The feedback correction module is used to collect the execution feedback status of supply control instructions and correct the supply status diagram using an incremental gradient update method.

[0024] In this embodiment, the data straightening module includes: The multi-source supply data of the smart canteen is rearranged in time sequence according to the collection timestamp, and the rearranged multi-source supply data is window merged according to a fixed sampling window. The data fields within the same sampling window are unified in field name, converted in field type and calibrated in value scale to obtain regularized supply data. Among them, the multi-source supply data includes the supply object identifier, sampling time, supply status value, retrieval status value, settlement status value and historical control status value collected at the edge. The edge is a computing node set up on-site in the smart canteen and used for data collection, short-term caching and local preprocessing. Missing positions and abnormal positions are marked in the regular supply data. The missing positions are filled in by interpolation results of adjacent sampling windows, and the abnormal positions are replaced by median filtering results to obtain valid supply data. The valid supply data is organized into samples according to the sampling window order. The valid supply data corresponding to each sampling window is used as a status sample, and a window index is configured for the status sample to form a standard supply dataset. Establish a collaborative control space in the cloud, register the standard supply dataset according to the window index, and establish a data synchronization relationship between the edge and the collaborative control space.

[0025] In this embodiment, the state mapping module includes: Based on the window index corresponding to each state sample in the standard supply dataset, the state samples are read in the order of the sampling window, and the scale is made consistent for each state sample. The DBSCAN clustering algorithm is used to perform density aggregation on the scale-consistent state samples. Core samples are determined based on the sample distribution within the neighborhood radius, and state samples that satisfy the density reachability relationship with the core samples are grouped into the same cluster. Each cluster is treated as a state node, and state samples belonging to the same cluster are mapped to the same state node to form a set of state nodes. According to the window index, record the order in which each state node appears in the continuous sampling window. When two state nodes correspond to state samples in adjacent sampling windows, establish a connection edge between the two state nodes. Based on the connection relationships of state nodes within a continuous sampling window, the node transition distribution is statistically analyzed, and conditional transition probabilities are used to assign weights to connecting edges. Specifically, this includes: Traverse the state node numbers corresponding to the state samples according to the window index, take the previous state node in the adjacent sampling window as the starting state node, take the next state node as the arriving state node, and record the state succession relationship from the starting state node to the arriving state node. The state continuity relationships between nodes with the same starting state and nodes with the same arriving state are accumulated and statistically analyzed to obtain the number of state continuity for the corresponding connecting edges. The total number of state acceptance relationships issued by the same starting state node is obtained by summing and counting all such relationships. The conditional transition probability of the connecting edge is determined based on the ratio between the number of times the connecting edge accepts a state and the total number of times the corresponding initial state node accepts a state. Write the conditional transition probability into the weight attribute of the corresponding connection edge, and construct the supply state graph with the state node set, connection edge, and the weight attribute of the connection edge; A supply state graph is constructed using the set of state nodes as graph nodes, the connecting edges as graph edges, and the weights of the connecting edges as edge attributes.

[0026] In this embodiment, the formation of clusters includes: According to the order of the state samples in the standard supply dataset, the state samples to be judged are selected one by one, and the neighboring state samples are retrieved within the preset neighborhood radius according to the Euclidean distance, with the state sample to be judged as the center, to form a neighborhood sample set. The number of state samples in the neighborhood sample set is counted. State samples whose number of state samples reaches the minimum sample number parameter are marked as core samples. State samples that do not reach the minimum sample number parameter but fall within the neighborhood of core samples are marked as boundary samples. Using the core sample that has not been assigned to a cluster as the starting point for aggregation, the boundary samples that are in the same neighborhood as the core sample are merged into the current cluster, and the neighborhood sample set of the newly added core sample in the current cluster is retrieved. Density reachability expansion is performed on the neighborhood sample set of the newly added core sample, and state samples that satisfy the density reachability relationship are incorporated into the current cluster until there are no more core samples in the current cluster that can be expanded. For state samples that are not assigned to the current cluster, continue to perform core sample determination and density reachability expansion to obtain multiple clusters; For state samples that are not assigned to a cluster, the nearest core sample is determined. State samples that meet the neighborhood radius condition are merged into the corresponding cluster, and state samples that do not meet the neighborhood radius condition are treated as isolated clusters.

[0027] In this embodiment, the graph recursion module includes: Using the current state node in the supply state diagram as the recursive object, the adjacent state nodes of the current state node are determined based on the node connectivity relationship; Based on the connection weights, the propagation strength of neighboring state nodes to the current state node is determined, resulting in the state propagation neighborhood, which specifically includes: Using the current state node as the recursion center, determine the adjacent state nodes along the connection edges in the supply state graph, and use the connection weight between the current state node and the adjacent state node as the initial propagation coefficient. The neighborhood proportion is calculated for multiple initial propagation coefficients corresponding to the same current state node to obtain the neighborhood competition coefficient of each adjacent state node. Based on the initial propagation coefficient and the neighborhood competition coefficient, the propagation contribution of each neighboring state node is corrected to obtain the propagation strength; The adjacent state nodes participating in message aggregation are determined based on the propagation strength, and the adjacent state nodes participating in message aggregation and their propagation strength constitute the state propagation neighborhood. Within the state propagation neighborhood, a graph neural network is used to modulate the node states of adjacent state nodes according to the propagation strength, and the modulated adjacent node states are aggregated into the neighborhood propagation message of the current state node. Based on the neighborhood propagation message and the node state of the current state node, the gated loop unit is used to calculate the update gate state and the reset gate state. Based on the reset gate state, the retention ratio of the current state node participating in the candidate state calculation is adjusted to obtain the candidate update state. Based on the update gate state, the node state of the current state node and the candidate update states are recursively fused to update the node state of the current state node. Based on the node connectivity in the supply state graph, the state propagation neighborhood determination, message modulation aggregation, and gated recursive fusion are repeatedly performed on each state node, and the recursion stops when the preset recursion round is reached, resulting in the recursively updated state node.

[0028] In this embodiment, the formation of the candidate update state includes: Based on the neighborhood propagation message and the node state of the current state node, a gating association transformation is performed on the neighborhood propagation message and the node state of the current state node to obtain the reset gate state. Based on the reset gate state, the node state of the current state node is subjected to component-level gating screening. The low-response historical state component corresponding to the reset gate state is suppressed, while the high-response historical state component is retained, thus obtaining the controlled node state. The controlled node state and the neighborhood propagation message are jointly transformed into a state, so that the neighborhood propagation message participates in the candidate state reconstruction of the current state node, and the intermediate candidate state is obtained. The intermediate quantities of the candidate states are nonlinearly activated and compressed to obtain the candidate updated states; The candidate update state is used as the state to be merged in the current state node in this round of graph recursion, and is used for subsequent recursive fusion with the node state of the current state node.

[0029] In this embodiment, the deviation detection module includes: Based on the recursive updated state nodes, the node states are arranged into a node state sequence according to the window index, and the node state sequence corresponding to the continuous control cycle is extracted as the sequence to be deduced. A causal temporal convolutional network is used to perform state inference on the sequence to be inferred. In the convolution operation, the causal temporal convolutional network only associates the state of the nodes before the current convolution position, and extracts the state changes at different intervals within the continuous control period through dilated convolution to obtain the state inference result. Based on the state extrapolation results, quantile regression is performed to determine the lower and upper bound quantile states corresponding to each extrapolation position. These lower and upper bound quantile states then constitute the state constraint interval for the corresponding extrapolation position, specifically including: According to the arrangement order of each predicted position within the continuous control cycle, the state prediction result is divided into the position prediction state corresponding to each prediction position, and the position prediction state of adjacent prediction positions is differentially calculated to obtain the position residual sequence. The position projection state and the corresponding position residual are used as the constraint input of the quantile regression network. Asymmetric quantile loss is calculated according to the lower bound quantile and the upper bound quantile respectively to obtain the lower bound quantile state and the upper bound quantile state corresponding to each projection position. Monotonic constraints are applied to the lower and upper bound quantile states corresponding to the same deduction position. When the lower bound quantile state is higher than the upper bound quantile state, the lower and upper bound quantile states are rearranged according to the boundary order. Based on the position residual sequence, the lower bound quantile states of adjacent inferred positions are smoothed, and the upper bound quantile states of adjacent inferred positions are smoothed to match the change amplitude of the boundary of adjacent inferred positions with the change amplitude of the position residual. The lower bound quantile state after boundary smoothing is used as the lower boundary of the state constraint interval, and the upper bound quantile state after boundary smoothing is used as the upper boundary of the state constraint interval, thus forming the state constraint interval for the corresponding deduction position. Based on the state projection results of each projection location, the state constraint interval of the corresponding projection location is compared. The supply deviation state is determined according to the projection location, deviation direction and deviation magnitude that exceed the state constraint interval.

[0030] In this embodiment, the formation of the state deduction result includes: According to the arrangement order of the node states in the sequence to be deduced, for each deduction position, extract the node states within the continuous sampling window before that deduction position to obtain the historical node state fragment; According to the preset dilation interval, select the node states with interval distribution in the historical node state segment, and feed the selected node states into the corresponding dilation convolution kernel for convolution operation to obtain the local state change amount under the corresponding time span. The local state changes over different time spans are accumulated and fused according to the same projection location. The residual connection is then used to superimpose the accumulated and fused results with the node state corresponding to the projection location to obtain the position fusion state of the projection location. The positional fusion state of each simulation position is nonlinearly compressed and combined according to the arrangement order of each simulation position within the continuous control cycle to form the state simulation result.

[0031] In this embodiment, the control and solution module includes: The action direction is set based on the deviation direction in the supply deviation state, and the action level is divided according to the deviation magnitude to form a set of candidate actions within the continuous control cycle; Using the current node state that generates the supply deviation state as the initial state for rolling optimization, the candidate action set is combined into a candidate action sequence within a continuous control cycle, and the candidate state trajectory corresponding to each candidate action sequence is recursively obtained. Based on the offset distance between the candidate state trajectory and the state constraint interval of the corresponding inferred position, a state offset penalty term is constructed and embedded into the optimization objective of the model predictive control. The model predictive control method is used to perform rolling optimization on the candidate action sequence and select the candidate action sequence that minimizes the state offset penalty term. In the candidate action sequence, the target control action corresponding to the current control cycle is encoded, control parameters are configured according to the action direction and action level, and the control parameters are encapsulated into supply control instructions that can be executed at the edge.

[0032] In this embodiment, the feedback correction module includes: The feedback status after the execution of the supply control command is collected at the edge, and the feedback status is incorporated into the corresponding control cycle according to the window index corresponding to the control command. Based on the set of state nodes formed by the state diagram module, the nearest state node is determined for the feedback state, the arrival state node corresponding to the feedback state is determined, and the state node corresponding to the supply deviation state is taken as the starting state node. Based on the starting state node and the arriving state node, determine the actual receiving edge in the supply state diagram. When the actual receiving edge does not exist, add a connection edge between the starting state node and the arriving state node. Based on the offset difference between the feedback status and the supply deviation status, the number of acceptances of the actual receiving edge is incrementally corrected, and the total number of outgoing acceptances corresponding to the same starting status node is updated synchronously. Based on the corrected number of state acceptances and the total number of outgoing acceptances, the conditional transition probability is redefined and used as the connection weight of the actual acceptance edge. The supply state diagram is then corrected based on the updated connection weights.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to the scenario of food supply control in a comprehensive smart canteen. The canteen has deployed edge data acquisition devices and a cloud management platform during the food service process. However, the original method mainly relied on historical statistics and manual experience to determine the supply. When concentrated food collection, status fluctuations, or local deviations occur during a continuous food service cycle, the system struggles to identify changes in supply status in a timely manner, easily leading to problems such as insufficient supply, supply redundancy, and delayed control commands.

[0034] In this scenario, the edge device first acquires multi-source supply data during the operation of the smart canteen, and performs timestamp calibration, field unification, outlier replacement, and sampling window merging on the data to form a standard supply dataset. The cloud-based collaborative control space is based on the standard supply dataset, and uses the DBSCAN clustering algorithm to perform density aggregation on the state samples to form a set of state nodes. Then, a supply state diagram is constructed based on the continuous connection relationship and conditional transition probability between the state nodes.

[0035] Subsequently, the system determines the state propagation neighborhood based on the node connectivity and connection weights in the supply state graph, uses a graph neural network to complete message modulation and aggregation, and recursively updates the state nodes through a gated loop unit.

[0036] During continuous regulation, the deviation discrimination module uses a causal temporal convolutional network to predict the evolution of subsequent states based on the recursively updated state nodes, and forms state constraint intervals for each prediction position through quantile regression. When the state prediction result exceeds the corresponding interval, the system determines the supply deviation state based on the deviation direction and magnitude.

[0037] The regulation and control solution module further establishes a set of candidate actions, employs model predictive control methods for rolling optimization, and encapsulates the target regulation action corresponding to the current regulation cycle into an executable supply regulation command at the edge. After the command is executed, the edge collects feedback status and adjusts the number of state acceptances and conditional transition probabilities based on the actual receiving edge corresponding to the feedback status, thereby continuously updating the supply status diagram.

[0038] In continuous operation verification, the accuracy rate of supply deviation identification of the original control method was 82.6%, which was improved to 93.8% after adopting the present invention; the average control response delay was reduced from 7.4 minutes to 2.6 minutes, of which the response delay during peak fluctuation periods was reduced from 9.1 minutes to 3.2 minutes; the supply shortage occurrence rate was reduced from 9.7% to 3.1%, and the supply redundancy rate was reduced from 12.4% to 5.2%.

[0039] In terms of state projection, the average state deviation of the original method was 0.186, which was reduced to 0.071 after adopting this invention. After continuous operation, the state deviation stabilized at a minimum of around 0.064. The command execution matching rate increased from 84.3% to 94.6%, indicating a significant enhancement in the consistency between control commands and on-site feedback states. After feedback correction, the edge weight stability of the supply state diagram increased from 0.781 to 0.923, and the fluctuation amplitude of the state node connection relationship decreased from 0.154 to 0.068, demonstrating that this invention can continuously correct the supply state diagram through feedback, thereby improving the stability of subsequent control.

[0040] Table 1. Comparison of the effects of intelligent regulation of food supply in smart canteens

[0041] As can be seen from the data in Table 1, after adopting the present invention, the intelligent canteen's food supply regulation has been significantly improved in terms of supply deviation identification, regulation response speed, supply stability, and feedback correction capability. It can identify changes in supply status more promptly and generate matching regulation instructions. At the same time, it continuously optimizes the supply status diagram through feedback correction, thereby improving the intelligent regulation effect in the continuous food supply process.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart canteen meal supply intelligent control system based on edge-cloud collaboration, characterized in that, include: The data standardization module is used to acquire and preprocess multi-source supply data from the smart canteen to form a standard supply dataset, while establishing a collaborative control space in the cloud. The state graph module is used to aggregate states based on a standard supply dataset using the DBSCAN clustering algorithm to form a set of state nodes. It then constructs a supply state graph by combining the continuous connection relationships and conditional transition probabilities between the state nodes. The graph recursion module is used to determine the state propagation neighborhood based on the node connectivity relationship within the collaborative control space. It uses a graph neural network to aggregate messages between adjacent state nodes and recursively updates the current node state through a gated loop unit. The deviation discrimination module is used to perform causal convolutional deduction on the state evolution trajectory within a continuous control cycle based on the recursively updated state nodes and a temporal convolutional network. It then uses quantile regression to constrain the deduction results to determine the supply deviation state. The regulation and solution module is used to establish a set of candidate actions based on the supply deviation state, perform rolling optimization solution using model predictive control method, and introduce a state offset penalty term in the solution process to generate supply regulation instructions; The feedback correction module is used to collect the execution feedback status of supply control instructions and correct the supply status diagram using an incremental gradient update method. 2.The intelligent meal supply regulation system based on end-cloud cooperation according to claim 1, wherein The data straightening module includes: The multi-source supply data of the smart canteen is rearranged in time sequence according to the collection timestamp, and the rearranged multi-source supply data is window merged according to a fixed sampling window. The data fields within the same sampling window are unified in field name, converted in field type and calibrated in value scale to obtain regularized supply data. Among them, the multi-source supply data includes the supply object identifier, sampling time, supply status value, usage status value, settlement status value and historical control status value collected at the edge; Missing positions and abnormal positions are marked in the regular supply data. The missing positions are filled in by interpolation results of adjacent sampling windows, and the abnormal positions are replaced by median filtering results to obtain valid supply data. The valid supply data is organized into samples according to the sampling window order. The valid supply data corresponding to each sampling window is used as a status sample, and a window index is configured for the status sample to form a standard supply dataset. Establish a collaborative control space in the cloud, register the standard supply dataset according to the window index, and establish a data synchronization relationship between the edge and the collaborative control space. 3.The intelligent meal supply regulation system based on end-cloud cooperation according to claim 1, characterized in that, The state mapping module includes: Based on the window index corresponding to each state sample in the standard supply dataset, the state samples are read in the order of the sampling window, and the scale is uniformized for each state sample. The DBSCAN clustering algorithm is used to perform density aggregation on the scale-consistent state samples. Core samples are determined based on the sample distribution within the neighborhood radius, and state samples that satisfy the density reachability relationship with the core samples are grouped into the same cluster. Each cluster is treated as a state node, and state samples belonging to the same cluster are mapped to the same state node to form a set of state nodes. According to the window index, record the order in which each state node appears in the continuous sampling window. When two state nodes correspond to state samples in adjacent sampling windows, establish a connection edge between the two state nodes. The node transition distribution is statistically analyzed based on the connection relationship of state nodes in the continuous sampling window, and the connection edges are weighted using conditional transition probabilities. A supply state graph is constructed using the set of state nodes as graph nodes, the connecting edges as graph edges, and the weights of the connecting edges as edge attributes.

4. The intelligent canteen meal supply amount intelligent regulation and control system based on end-cloud cooperation according to claim 3, characterized in that, The formation of the clusters includes: According to the order of the state samples in the standard supply dataset, the state samples to be judged are selected one by one, and the neighboring state samples are retrieved within the preset neighborhood radius according to the Euclidean distance, with the state sample to be judged as the center, to form a neighborhood sample set. The number of state samples in the neighborhood sample set is counted. State samples whose number of state samples reaches the minimum sample number parameter are marked as core samples. State samples that do not reach the minimum sample number parameter but fall within the neighborhood of core samples are marked as boundary samples. Using the core sample that has not been assigned to a cluster as the starting point for aggregation, the boundary samples that are in the same neighborhood as the core sample are merged into the current cluster, and the neighborhood sample set of the newly added core sample in the current cluster is retrieved. Density reachability expansion is performed on the neighborhood sample set of the newly added core sample, and state samples that satisfy the density reachability relationship are incorporated into the current cluster until there are no more core samples in the current cluster that can be expanded. For state samples that are not assigned to the current cluster, continue to perform core sample determination and density reachability expansion to obtain multiple clusters; For state samples that are not assigned to a cluster, the nearest core sample is determined. State samples that meet the neighborhood radius condition are merged into the corresponding cluster, and state samples that do not meet the neighborhood radius condition are treated as isolated clusters. 5.The intelligent meal supply regulation system based on end-cloud cooperation according to claim 1, wherein, The graph recursion module includes: Using the current state node in the supply state diagram as the recursive object, the adjacent state nodes of the current state node are determined based on the node connectivity relationship; Based on the connection weights, the propagation strength of adjacent state nodes to the current state node is determined, thus obtaining the state propagation neighborhood; Within the state propagation neighborhood, a graph neural network is used to modulate the node states of adjacent state nodes according to the propagation strength, and the modulated adjacent node states are aggregated into the neighborhood propagation message of the current state node. Based on the neighborhood propagation message and the node state of the current state node, the gated loop unit is used to calculate the update gate state and the reset gate state. Based on the reset gate state, the retention ratio of the current state node participating in the candidate state calculation is adjusted to obtain the candidate update state. Based on the update gate state, the node state of the current state node and the candidate update states are recursively fused to update the node state of the current state node. Based on the node connectivity in the supply state graph, the state propagation neighborhood determination, message modulation aggregation, and gated recursive fusion are repeatedly performed on each state node, and the recursion stops when the preset recursion round is reached, resulting in the recursively updated state node. 6.The intelligent meal supply regulation system based on end-cloud cooperation according to claim 5, wherein, The formation of the candidate update state includes: Based on the neighborhood propagation message and the node state of the current state node, a gating association transformation is performed on the neighborhood propagation message and the node state of the current state node to obtain the reset gate state. Based on the reset gate state, the node state of the current state node is subjected to component-level gating screening. The low-response historical state component corresponding to the reset gate state is suppressed, while the high-response historical state component is retained, thus obtaining the controlled node state. The controlled node state and the neighborhood propagation message are jointly transformed into a state, so that the neighborhood propagation message participates in the candidate state reconstruction of the current state node, and the intermediate candidate state is obtained. The intermediate quantities of the candidate states are nonlinearly activated and compressed to obtain the candidate updated states; The candidate update state is used as the state to be merged in the current state node in this round of graph recursion, and is used for subsequent recursive fusion with the node state of the current state node.

7. The intelligent canteen meal supply amount intelligent regulation and control system based on end-cloud cooperation according to claim 1, characterized in that, The deviation detection module includes: Based on the recursive updated state nodes, the node states are arranged into a node state sequence according to the window index, and the node state sequence corresponding to the continuous control cycle is extracted as the sequence to be deduced. A causal temporal convolutional network is used to perform state inference on the sequence to be inferred. The causal temporal convolutional network only associates the node states before the current convolution position in the convolution operation, and extracts the state changes at different intervals within a continuous control period through dilated convolution to obtain the state inference result. Based on the state deduction results, quantile regression is performed to determine the lower bound quantile state and upper bound quantile state corresponding to each deduction position. The lower bound quantile state and upper bound quantile state constitute the state constraint interval of the corresponding deduction position. Based on the state projection results of each projection location, the state constraint interval of the corresponding projection location is compared. The supply deviation state is determined according to the projection location, deviation direction and deviation magnitude that exceed the state constraint interval.

8. The intelligent canteen meal supply amount intelligent regulation and control system based on end-cloud cooperation according to claim 7, characterized in that, The formation of the state deduction result includes: According to the arrangement order of the node states in the sequence to be deduced, for each deduction position, extract the node states within the continuous sampling window before that deduction position to obtain the historical node state fragment; According to the preset dilation interval, select the node states with interval distribution in the historical node state segment, and feed the selected node states into the corresponding dilation convolution kernel for convolution operation to obtain the local state change amount under the corresponding time span. The local state changes over different time spans are accumulated and fused according to the same projection location. The residual connection is then used to superimpose the accumulated and fused results with the node state corresponding to the projection location to obtain the position fusion state of the projection location. The positional fusion state of each simulation position is nonlinearly compressed and combined according to the arrangement order of each simulation position within the continuous control cycle to form the state simulation result. 9.The intelligent meal supply regulation system based on end-cloud cooperation according to claim 1, wherein, The control and solution module includes: The action direction is set based on the deviation direction in the supply deviation state, and the action level is divided according to the deviation magnitude to form a set of candidate actions within the continuous control cycle. Using the current node state that generates the supply deviation state as the initial state for rolling optimization, the candidate action set is combined into a candidate action sequence within a continuous control cycle, and the candidate state trajectory corresponding to each candidate action sequence is recursively obtained. Based on the offset distance between the candidate state trajectory and the state constraint interval of the corresponding inferred position, a state offset penalty term is constructed and embedded into the optimization objective of the model predictive control. The model predictive control method is used to perform rolling optimization on the candidate action sequence and select the candidate action sequence that minimizes the state offset penalty term. In the candidate action sequence, the target control action corresponding to the current control cycle is encoded, control parameters are configured according to the action direction and action level, and the control parameters are encapsulated into supply control instructions that can be executed at the edge.

10. The intelligent canteen food supply control system based on edge-cloud collaboration according to claim 1, characterized in that, The feedback correction module includes: The feedback status after the execution of the supply control command is collected at the edge, and the feedback status is incorporated into the corresponding control cycle according to the window index corresponding to the control command. Based on the set of state nodes formed by the state diagram module, the nearest state node is determined for the feedback state, the arrival state node corresponding to the feedback state is determined, and the state node corresponding to the supply deviation state is taken as the starting state node. Based on the starting state node and the arriving state node, determine the actual receiving edge in the supply state diagram. When the actual receiving edge does not exist, add a connection edge between the starting state node and the arriving state node. Based on the offset difference between the feedback status and the supply deviation status, the number of acceptances of the actual receiving edge is incrementally corrected, and the total number of outgoing acceptances corresponding to the same starting status node is updated synchronously. Based on the corrected number of state acceptances and the total number of outgoing acceptances, the conditional transition probability is redefined and used as the connection weight of the actual acceptance edge. The supply state diagram is then corrected based on the updated connection weights.

Images