Multi-stage early warning method and system for central kitchen based on industrial big data

Abstract

The application discloses a kind of multi-level early warning method and system of central kitchen based on industrial big data, and the application relates to the technical field of industrial big data, determine multiple work areas based on the division of current work route, and cross matching is carried out in combination with the real-time image of central kitchen and corresponding dish sequence, to output the industrial big data combination of each work area, determine the working state of corresponding work area according to the iteration of each industrial big data combination, improve the accuracy of the working state of work area. The working plan preset by the central kitchen is obtained, and the working state of each work area and the current work progress of the central kitchen are input into the digital twin space, determine a plurality of delay factors of different dimensions according to the trace of the progress delay content, further closed loop deduction is carried out in combination with the work history of the central kitchen and the corresponding industrial constraint relationship, to output multi-level early warning path, determine the accuracy of the early warning optimization sequence of the central kitchen.

Description

Technical Field

[0001] This invention relates to the technical field of industrial big data, and in particular to a multi-level early warning method and system for a central kitchen based on industrial big data. Background Technology

[0002] As the core hub of the modern catering supply chain, central kitchens are equipped with a large number of automated woks, conveyor belts, rapid cooling units, and other industrial equipment, exhibiting a high degree of continuity and automation. With the development of Industrial Internet of Things (IIoT) technology, current monitoring methods for central kitchens have largely evolved from manual inspections to online monitoring based on sensor data.

[0003] Existing monitoring methods typically divide central kitchens into fixed areas based on their static physical layout and rely solely on threshold values ​​from sensors in each area to determine their status. In actual operation, central kitchens have intertwined water, electricity, gas, and material pipelines, and multiple production lines may operate concurrently or switch flexibly. Existing technologies fail to achieve deep spatiotemporal alignment between massive amounts of equipment operating data and complex pipeline distribution maps, making it impossible to perceive "how materials / energy flow in the pipelines." Consequently, they cannot construct a dynamic topology map that reflects the actual physical connections, affecting the accuracy of the operating status of each work area in the central kitchen and resulting in low accuracy of the early warning optimization sequence for the central kitchen. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a multi-level early warning method and system for central kitchens based on industrial big data.

[0005] This invention provides a multi-level early warning method for central kitchens based on industrial big data, including: The system marks multiple industrial devices in the central kitchen, performs spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and constructs a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process; and determines the current working route of the central kitchen based on the identification of the dynamic topology map. Based on the division of the current work route, multiple work areas are determined, and cross-matching is performed with real-time images of the central kitchen and the corresponding food preparation sequence to output the industrial big data combination of each work area. The working status of the corresponding work area is determined according to the iteration of each industrial big data combination. The system obtains the pre-set work plan of the central kitchen, and inputs it into the digital twin space along with the work status of each work area and the current work progress of the central kitchen. A causal mechanism is introduced into the digital twin space to determine the corresponding progress delay content. Based on the tracing of the delay content, multiple delay factors in different dimensions are identified. Furthermore, a closed-loop deduction is carried out by combining the central kitchen's work process and corresponding industrial constraints, thereby outputting multi-level early warning paths and integrating multiple work areas to determine the early warning optimization sequence for the central kitchen.

[0006] This invention provides a multi-level early warning system for a central kitchen based on industrial big data. This system is applied to the aforementioned multi-level early warning method for a central kitchen based on industrial big data. The multi-level early warning system for a central kitchen based on industrial big data includes: The dynamic topology module is used to label multiple industrial devices in the central kitchen, perform spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and construct a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process; and determine the current working route of the central kitchen based on the identification of the dynamic topology map. The industrial big data module is used to determine multiple work areas based on the current work route division, and to perform cross-matching with the real-time images of the central kitchen and the corresponding dish production sequence to output the industrial big data combination of each work area. The working status of the corresponding work area is determined based on the iteration of each industrial big data combination. The digital twin module is used to obtain the work plan preset by the central kitchen, and input it into the digital twin space in combination with the work status of each work area and the current work progress of the central kitchen. A causal mechanism is introduced into the digital twin space to determine the corresponding progress delay content. The multi-level early warning module is used to identify multiple delay factors in different dimensions based on the tracing of the delay content. It further combines the central kitchen's work history and corresponding industrial constraints to conduct closed-loop deduction, thereby outputting multi-level early warning paths and integrating multiple work areas to determine the central kitchen's early warning optimization sequence.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Mark multiple industrial equipment in the central kitchen, perform spatiotemporal alignment of the working data of each industrial equipment with the pipeline distribution map of the central kitchen, and construct a dynamic topology map of the central kitchen along the time-series sliding window during the alignment process; determine the current working route of the central kitchen based on the identification of the dynamic topology map; determine multiple working areas based on the division of the current working route, and perform cross-matching with the real-time image of the central kitchen and the corresponding dish output sequence to output the industrial big data combination of each working area; determine the working status of the corresponding working area based on the iteration of each industrial big data combination; introduce the current working route of the central kitchen, further control the industrial big data combination of each working area, and improve the accuracy of the working status of the working area.

[0008] (2) Obtain the work plan preset by the central kitchen, and input the work status of each work area and the current work progress of the central kitchen into the digital twin space. In the digital twin space, a causal mechanism is introduced to determine the corresponding progress delay content. Based on the tracing of the progress delay content, multiple delay factors in different dimensions are determined. Furthermore, the work process of the central kitchen and the corresponding industrial constraints are combined to conduct closed-loop deduction, thereby outputting multi-level early warning paths. Multiple work areas are integrated to determine the early warning optimization sequence of the central kitchen, further control the progress delay content, fully consider the multi-level early warning paths and multiple work areas, and determine the accuracy of the early warning optimization sequence of the central kitchen. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the multi-level early warning method for a central kitchen based on industrial big data in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 in the multi-level early warning method for a central kitchen based on industrial big data in this embodiment of the invention. Figure 3 This is a flowchart illustrating step S12 in the multi-level early warning method for a central kitchen based on industrial big data in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the multi-level early warning method for a central kitchen based on industrial big data in this embodiment of the invention. Figure 5 This is a flowchart illustrating step S14 of the multi-level early warning method for a central kitchen based on industrial big data in an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of a multi-level early warning system for a central kitchen based on industrial big data in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 6 A multi-level early warning method for central kitchens based on industrial big data is proposed and applied to industrial big data scenarios. The multi-level early warning method for central kitchens based on industrial big data includes: Step S11: Mark multiple industrial devices in the central kitchen, perform spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and construct a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process; determine the current working route of the central kitchen based on the identification of the dynamic topology map. Step S12: Based on the division of the current work route, multiple work areas are determined, and cross-matching is performed with the real-time images of the central kitchen and the corresponding dish production sequence to output the industrial big data combination of each work area. The working status of the corresponding work area is determined according to the iteration of each industrial big data combination. Step S13: Obtain the work plan preset by the central kitchen, and input it into the digital twin space in combination with the work status of each work area and the current work progress of the central kitchen. In the digital twin space, a causal mechanism is introduced to determine the corresponding progress delay content. Step S14: Based on the tracing of the delay content, identify multiple delay factors in different dimensions, and further combine the central kitchen's work process and corresponding industrial constraints to conduct closed-loop deduction, thereby outputting multi-level early warning paths and integrating multiple work areas to determine the central kitchen's early warning optimization sequence.

[0012] refer to Figure 2 In step S11, the specific steps are as follows: S111: Based on the identification of the distribution map of the central kitchen, identify multiple industrial devices and mark the locations of the multiple industrial devices. Associate the multiple industrial devices along the communication channels of the central kitchen, obtain the working data of each industrial device, and perform spatiotemporal alignment along the cross-modal alignment mechanism in conjunction with the pipeline distribution map of the central kitchen. S112: In the alignment process, a time-series sliding window is introduced, using the abrupt change points of energy flow and material flow of industrial equipment as window segmentation anchors to dynamically generate a dynamic topology map of the central kitchen; a graph neural network is used to learn the feature manifold of the dynamic topology map, and a topology recognition mechanism is combined to remove background noise paths, thereby mapping the current working route of the central kitchen under complex working conditions.

[0013] In the embodiments of this application, multiple industrial devices are identified based on the distribution map of the central kitchen, and the locations of the multiple industrial devices are marked. The multiple industrial devices are associated along the communication channels of the central kitchen, the working data of each industrial device is obtained, and spatiotemporal alignment is performed along the cross-modal alignment mechanism in conjunction with the pipeline distribution map of the central kitchen. This is compatible with the overall consideration of the distribution map of the central kitchen and ensures the accuracy of the multiple industrial devices.

[0014] At this point, the system accesses the 2D CAD layout drawing or 3D BIM model of the central kitchen, and uses a deep learning-based instance segmentation network to perform pixel-level analysis of the equipment legends or 3D bounding boxes in the drawings, identifying the outlines and categories of various industrial equipment; the system extracts the absolute physical coordinates of the geometric center point of each piece of equipment and maps them to a pre-constructed global Cartesian coordinate system, completing the 3D vector (x,y,z) labeling of the spatial location of the equipment; based on this, the system assigns a unique global identifier (UID) to each identified physical equipment, thereby establishing a set of digital twin nodes in the digital space that correspond one-to-one with the physical entity.

[0015] The system reads the topology configuration table of the central kitchen's underlying industrial control network and performs logical addressing along the cascaded communication channels from switches and gateways to the end-point I / O modules. During this process, the system associates the device UID determined in the first step with its corresponding communication node. The system uses industrial protocols such as OPCUA or MQTT to pull and parse the register addresses of each communication node in real time according to the set sampling frequency. This allows the system to accurately mount the working data stream containing physical quantities such as current, voltage, temperature, and frequency to the corresponding device digital node, forming a real-time mapping relationship of "node-address-data stream".

[0016] The system first performs vectorized parsing on the pipeline distribution map, abstracting the complex pipeline network into a set of directed edges connecting various equipment nodes, and assigning edge attributes such as pipe material, pipe diameter, and medium type. Using the coordinates of the equipment connected at both ends of the pipeline as constraints, the system employs the nearest neighbor feature matching method to spatially interpolate and project the acquired equipment operating data, such as the temperature difference and pressure difference before and after the steam pipeline, along the spatial path of the directed edges. Through this mechanism, the data that originally existed in isolation on each equipment node is given a spatial topological meaning at the pipeline level, achieving a close fit between the data flow and the physical medium flow direction in the spatial dimension.

[0017] The system introduces a global high-precision clock synchronization mechanism as the absolute time reference. On this basis, the system abandons the traditional hard downsampling or upsampling and instead adopts an elastic time alignment method based on dynamic time warping (DTW) or causal inference. Using the timestamp of high-frequency equipment data as the reference anchor point, forward or backward edge interpolation is performed according to the physical change rate of low-frequency data. This ensures that under the same time slice Ti, all equipment parameters and pipeline status on the pipeline distribution map are in the same physical time section, completing the closed loop of cross-modal spatiotemporal alignment.

[0018] Specifically, the central kitchen is characterized by a multi-layered, three-dimensional layout, cross-control of raw and cooked food, and the concurrent operation of dozens of automated production lines. The system analyzes the three-dimensional BIM model of the central kitchen, accurately identifies and marks the "No. 1 fully automatic wok machine" (coordinates: X15.2, Y8.4, Z4.5) located in the second-floor cooking area and the "No. 1 spiral rapid cooling machine" (coordinates: X18.6, Y8.4, Z4.5) located slightly behind on the same floor, generating UID-001 and UID-002 for them.

[0019] The system searches along the gigabit industrial Ethernet channel of the central kitchen and finds that UID-001 is bound to PLC slave station Station-12 (IP: 192.168.1.12) and UID-002 is bound to Station-15. The system obtains the stirring motor current and pot temperature of UID-001 in real time, as well as the compressor frequency and refrigerant return pressure of UID-002.

[0020] The system retrieves the pipeline distribution map of the central kitchen and extracts a 3.4-meter-long "stainless steel material conveying trough" pipeline connecting the "wok machine outlet" to the "rapid cooler inlet," along Edge-A. The system projects and aligns the discharge signal of UID-001 and the feed signal of UID-002 along the spatial path of Edge-A. At this point, the originally isolated wok current data and rapid cooler pressure data are forcibly connected in the digital space by the physical pipeline "material conveying trough," forming a spatial correlation feature of "wok discharge - pipeline transmission - rapid cooler feed."

[0021] Suppose a minor material congestion suddenly occurs in the central kitchen; the photoelectric sensor installed at the inlet of the conveyor chute captures the material accumulation at timestamp T0; while the temperature probe installed inside the blast cooler only detects the temperature drop caused by the reduced feed at timestamp T1; the system, through a flexible time alignment method, uses T0 as the anchor point and combines a thermodynamic inertial model to perform reverse interpolation and deduction, accurately "backtracking" the temperature drop trend to time T0; the system clearly concludes on the spatiotemporally aligned digital base that at time T0, in the material pipeline space from the wok machine to the blast cooler on the second floor of the central kitchen, a cross-modal correlation event of "continuous front-end discharge: high current" and "back-end material obstruction: abnormal temperature trend" occurred simultaneously, laying an absolutely accurate spatiotemporal data foundation for the subsequent generation and early warning deduction of the S112 dynamic topology.

[0022] Furthermore, a temporal sliding window is introduced during the alignment process, using the abrupt change points of energy flow and material flow in industrial equipment as window segmentation anchors to dynamically generate a dynamic topology map of the central kitchen. A graph neural network is used to learn the feature manifold of this dynamic topology map, and a topology recognition mechanism is combined to remove background noise paths, thereby mapping the current working route of the central kitchen under complex working conditions.

[0023] At this point, the system continuously monitors the differential rate of change of two types of core physical variables on the data stream after S111 alignment: one is the energy flow characterizing the working state of the equipment, such as current step, sudden change in instantaneous gas flow, and steam pressure jump; the other is the material flow characterizing the material flow state, such as the weight change of the dynamic weighing module, the on / off pulse of the photoelectric switch, and the flow rate step of the flow meter. The system adopts a change detection method based on the sliding T test. When the rate of change of the above variables exceeds the adaptive dynamic threshold, an "event anchor point" is triggered. The system uses this anchor point as the starting boundary of the timing sliding window and extends it to the next steady-state anchor point that conforms to the process delay logic as the termination boundary, thereby constructing a flexible timing window that is strictly driven by physical events, ensuring that a complete physical operation process is encapsulated in each window.

[0024] The system uses the industrial equipment undergoing state transitions within the current window as the node set and calls the pipeline distribution map containing spatial topological relationships in S111 as a priori structural constraint. The system calculates the energy flow or material flow gradient of the nodes at both ends of the pipeline within the current window. If the gradient direction is consistent with the preset medium flow direction of the pipeline and the absolute value is greater than zero, a directed edge is activated between the two nodes. If the gradient is zero or in the opposite direction, the edge weight is reset to zero or cut off. Through this dynamic edge weight allocation mechanism based on physical flow direction and gradient, the system evolves and generates a dynamic directed topological graph reflecting the current transient physical connection relationship in real time within each sliding window.

[0025] Because a large number of concurrently operating devices exist in the central kitchen, the system inputs this dynamic topology graph into a graph neural network (such as a spatiotemporal graph convolutional network ST-GCN). In this graph neural network, such as ST-GCN, message passing occurs along first- or second-order neighbor nodes, aggregating the multidimensional temporal features of each node, such as waveform fragments of temperature, pressure, and rotational speed. During this process, the system introduces the manifold learning hypothesis: that the temporal fluctuations of the physical parameters of device nodes on the same real working path are essentially driven by the processing of the same batch of materials. The high-dimensional features of these nodes must be tightly clustered on the same low-dimensional submanifold surface in the manifold space. Through nonlinear mapping of the graph convolutional layer and dimensionality reduction using Laplacian feature mapping, the system projects the chaotic high-dimensional topological features onto the low-dimensional manifold space, causing the real collaborative node group to form a high-density manifold cluster, while background interference nodes are scattered as isolated outliers in the manifold space.

[0026] The system sets a local density reachability threshold in the low-dimensional manifold space and uses density-based clustering to accurately extract low-density, long-distance background noise paths, such as false association edges caused by constantly running exhaust systems and ambient temperature compensation equipment. The system extracts the core backbone subgraph with the strongest connectivity and the highest feature manifold density. The system uses graph isomorphic networks (GIN) or graph kernel functions to perform topological similarity matching between the cleaned core subgraph and the central kitchen's pre-set standard process route map library, thereby translating the abstract graph structure into concrete business logic and accurately mapping the main work routes and their specific numbers that the central kitchen is actually executing under the current complex concurrent conditions.

[0027] Specifically, in the cooking area on the second floor of the central kitchen, although the main "Kung Pao Chicken" line is being stir-fried at 180°C in a fully automatic wok (UID-001), at the same time, the gas steam generator (UID-003) on the adjacent "Vegetable Blanching" sub-line just completed an ignition start-up, causing a sudden change in energy flow: the gas flow rate instantly jumped from 0 to 2.5 m³ / h; the system immediately used the gas flow change point of UID-003 as the "starting anchor point" to open a new time-series sliding window Wt.

[0028] Within window Wt, the system traverses the pipeline distribution diagram as a constraint. At this time, due to the dense pipelines in the central kitchen, the gas steam generator (UID-003) is not only connected to the blanching tank (UID-004) of the auxiliary line, but the steam it generates also passes through the main steam distribution pipe row shared with the wok machine (UID-001). If only simple on / off judgment is used, the system will generate a huge and intricate dynamic topology diagram containing "steam generator > main steam pipe row > wok machine > material conveyor belt > quick cooler". On the surface, all devices are in an "actively connected" state.

[0029] During feature aggregation and manifold dimensionality reduction, waveform features were compared in depth: UID-003 triggered a waveform of slow water temperature rise, while UID-001 contained high-frequency, violent temperature fluctuations and stirring current waveforms; in the low-dimensional manifold space, the features of UID-003 and UID-004 clustered into a "low-frequency gradual change" manifold cluster; while UID-001, the conveyor belt photoelectric sensor, and UID-002, having all experienced the "material congestion anomaly event" previously identified by S111, exhibited highly consistent transient change characteristics in current and temperature, tightly clustering together to form another "high-frequency sudden change" high-density manifold cluster; the two were far apart in manifold distance, and their physical "shared steam pipe row" was determined to be a weak correlation without causal consistency.

[0030] Based on the manifold density threshold, the system decisively removed and cut off the background noise path "steam generator > main steam pipe > wok machine" from the topology graph. The core subgraph that was ultimately retained was only the strongly connected path "wok machine outlet > conveyor belt > quick-cooler inlet". The system performed graph matching on this subgraph with the central kitchen process library and accurately mapped the current underlying real working route as "Kung Pao Chicken main line - material transport and quick-cooling process route". Moreover, there were abnormal topological features of material flow obstruction on this route, which directly determined the unique physical target for the subsequent digital twin delay inference of S13.

[0031] refer to Figure 3 In step S12, the specific steps are as follows: S121: Dynamically segment the current work route along the graph segmentation mechanism and decouple the physical space of the central kitchen into multiple work areas with coupling relationships; synchronously acquire real-time images of the central kitchen, and determine the food output sequence of the central kitchen based on the identification of the food output data stream of the central kitchen; further combine multiple work areas and real-time images of the central kitchen for cross-matching, thereby aggregating equipment data, visual data and order sequence data into a corresponding industrial big data combination during the matching process; S122: Performs multi-level iterations on various industrial big data combinations. During the iteration process, a long short-term memory mechanism is introduced to simultaneously trigger nonlinear iteration and latent variable evolution analysis, and simultaneously filter out transient interference to output the working status of each working area at different granularities.

[0032] In the embodiments of this application, the current work route is dynamically segmented along a graph segmentation mechanism, and the physical space of the central kitchen is decoupled into multiple work areas with coupling relationships. Real-time images of the central kitchen are acquired synchronously, and the food output sequence of the central kitchen is determined based on the identification of the food output data stream of the central kitchen. Furthermore, the real-time images of multiple work areas and the central kitchen are cross-matched, thereby aggregating equipment data, visual data and order sequence data into a corresponding industrial big data combination during the matching process. This approach is compatible with the overall consideration of the food output data stream of the central kitchen and ensures the accuracy of the food output sequence of the central kitchen.

[0033] At this point, the system takes the dynamic topology graph of the current work route as input and adopts a graph partitioning method based on modularity, using the edge weights between device nodes as the partitioning basis. The system searches for community structures in the graph structure that are "sparsely connected but densely internal" and forcibly cuts the continuous physical topology lines into several relatively independent node clusters. While cutting off the physical connections, the system retains and marks the upstream and downstream time-series dependencies and material handover relationships between these clusters, thereby decoupling the rigid physical space of the central kitchen into multiple dynamic work areas with strong coupling relationships in business logic, such as raw material pretreatment area, heat processing area, and rapid cooling area, to achieve flexible reconstruction of the spatial dimension.

[0034] The system synchronously retrieves the industrial camera network deployed at key nodes in the central kitchen to acquire high-definition real-time image streams covering the aforementioned work areas. Simultaneously, the system directly connects to the central kitchen's Manufacturing Execution System (MES) or Order Management System (OMS) to perform real-time parsing and topological sorting of the issued dish output data streams. Based on the dish identifier (SKU) in the order, the planned issuance time, and process constraints, such as the exclusive constraint of serving meat dishes before vegetable dishes, the system constructs a dish output sequence queue with strict temporal logic. This dish output sequence not only includes the macroscopic attributes of the dishes but also carries the theoretical process node status of each dish.

[0035] The system uses the "multiple working areas" output from the first two stages as spatial anchor boxes and the "dish production sequence" as a temporal anchor point, employing a cross-modal attention mechanism for deep cross-matching. Specifically, the system uses computer vision methods, such as YOLO object detection or semantic segmentation, to extract appearance features from real-time images, such as the color and shape of ingredients in the tray and personnel gestures, and compares them with the underlying equipment sensor data in the corresponding working area, such as temperature curves and motor speeds. Cosine similarity is calculated at the feature layer. When the visual features and equipment features reach a high degree of consistency in attention weights, the system determines that the match is successful and immediately binds the current dish production sequence label. The system forcibly aligns and encapsulates the "underlying equipment physical quantities, intermediate visual spatial representations, and top-level order business logic" on the spatiotemporal coordinate axis, aggregating them into an "industrial big data combination" with full-dimensional mutual verification relationships, serving as the smallest complete set unit for subsequent state evaluation.

[0036] Specifically, for the "Kung Pao Chicken Main Line - Material Transfer and Rapid Cooling Process Route" determined by S112, the system performs graph partitioning logic for the topology route "Wok Machine > Conveyor Belt > Rapid Cooler". It finds that there is a significant difference in edge weights between the "Wok Machine" node and the subsequent "Rapid Cooler" node. The wok area involves high-frequency and intense energy exchange, while the rapid cooling area involves stable and continuous cold exchange. Therefore, the system "cuts off" the topology at the weak connection point of the "Conveyor Belt Discharge Port", dynamically decoupling the physical space of the second floor of the central kitchen into two coupled regions: "R-α Hot Processing and Discharge Area: including the wok machine and the front section of the conveyor belt" and "R-β Rapid Cooling Area: including the rapid cooler and the rear section of the conveyor belt". The two are physically separated, but logically they are strongly coupled by the constraint that "materials must be transferred on time".

[0037] The system synchronously acquires real-time images from industrial cameras covering the R-α and R-β zones; at the same time, it analyzes the current food delivery data stream of the central kitchen. Due to the lunch peak, the system sorts out the current food delivery sequence queue from the complex order pool: [Task-01: Kung Pao Chicken (estimated to be ready in T+2min) > Task-02: Shredded Pork with Garlic Sauce (T+5min) > Task-03: Stir-fried Vegetables (T+8min)].

[0038] The system performs cross-matching at the time node T+2min; in the R-α region, the visual method identifies "brown, peanut-containing lumps" in the image of the wok machine's outlet, while the equipment data in this region shows "a decrease in stirring current and a sudden drop in the pot temperature curve"; the system calculates through an attention mechanism and finds that the "visual brown lumps" highly match the "sudden drop in equipment temperature" and perfectly correspond to the business tag of Task-01: Kung Pao Chicken in the dish output sequence.

[0039] The system immediately triggered the aggregation mechanism, tightly packaging "R-α zone equipment current / temperature data + image features of brown ingredients at the wok outlet + Task-01 Kung Pao Chicken order tag" to generate the industrial big data combination Combo-Alpha; similarly, for the R-β zone, the system packaged "rapid chiller compressor frequency data + image features of the rear section of the conveyor belt + Task-01 expected feed tag" to generate Combo-Beta; in the extremely complex multi-layer concurrent production line of the central kitchen, the system successfully transformed the cold, impersonal pipeline anomalies into an industrial big data combination with clear business attributes, such as "Kung Pao Chicken has been cooked in the R-α zone but has not entered the R-β zone", completely eliminating data ambiguity in the scenario of cross-control of raw and cooked food.

[0040] Furthermore, the various industrial big data combinations are iterated at multiple levels. During the iteration process, a long short-term memory mechanism is introduced to simultaneously trigger nonlinear iteration and latent variable evolution analysis, and to simultaneously filter out transient interference, so as to output the working status of each work area at different granularities. The working status of each work area at different granularities is introduced, and at the same time, the current work route of the central kitchen is introduced to further control the industrial big data combinations of each work area and improve the accuracy of the working status of the work area.

[0041] At this point, the system transforms the industrial big data output by S121 into a multivariate time series vector and inputs it into the Long Short-Term Memory (LSTM) network in segments. During the multi-level iteration process, the LSTM selectively retains and forgets the waveforms of equipment parameters, the evolution trajectory of visual features, and the movement of order nodes in historical time steps through its internal forget gate, input gate, and output gate. This multi-level iteration is not a single forward propagation, but rather it continuously expands on the time axis by using the temporal state vector implicit in the previous iteration cycle as the initial input of the next iteration cycle. This allows the system to capture the macro-cyclical patterns of the working area from "material preparation-processing-discharge" over a long span, while determining the transient changes at the millisecond level over a short span.

[0042] In the iterative loop of LSTM, the system does not use a simple linear mapping, but introduces a nonlinear activation function and a latent variable space mapping mechanism. The system assumes that the surface data of the industrial big data combination is generated by a few unobservable "latent variables". In each nonlinear iteration, the system uses a variational autoencoder (VAE) or a hidden Markov nonlinear state-space model to forcibly project the high-dimensional surface data to a low-dimensional latent variable space. By tracking the evolution trajectory of the probability distribution of these latent variables in the iteration process, such as the contraction or diffusion of the Gaussian distribution, the system can peel off the differences in physical dimensions of the surface data and directly access the essential nonlinear evolution state of the physical processes inside the working area.

[0043] The central kitchen site is subject to numerous transient disturbances not inherent to the process itself, such as personnel obstructing cameras, minor voltage spikes in the power grid, and sudden packet loss from sensors. The system sets confidence intervals and evolution slope thresholds based on historical stationary distributions in the latent variable space. When a latent variable calculated in a particular iteration undergoes a drastic change, but this change quickly returns to its original evolution trajectory within a very short subsequent time window without triggering a coordinated response from the latent variables of adjacent nodes, the system classifies it as a transient disturbance lacking physical continuity. The system uses Kalman filtering or smoothing noise reduction to forcibly smooth and filter out these isolated pulse disturbances at the latent variable level, ensuring that the output evolution trajectory purely reflects the actual process progress.

[0044] The system reverse-maps the filtered and purified latent variable trajectories back to the physical business space. By setting decoding thresholds of different granularities, the system outputs two or more levels of operating status: at the "fine granularity" level, the system decodes the minute fluctuations of latent variables and outputs equipment-level or action-level status, such as "slight increase in agitator resistance"; at the "coarse granularity" level, the system combines the macroscopic period captured by LSTM long-range memory to make overall judgments on latent variable clusters and outputs regional-level macroscopic status, such as "regional full-load stable operation" or "risk of material backlog," thus providing a clear-cut status basis for subsequent early warnings.

[0045] Specifically, the system inputs a Combo-Alpha, which includes "wok current, image brown features, and Task-01 label," into the LSTM network. The network then begins multi-level iterations, recalling the standard current decay curve of the wok machine when it completed the "Fish-flavored Shredded Pork" task in the past 10 minutes. The current cycle is compared and iterated with the historical cycle in the hidden state layer to establish a time-series baseline for evaluating the current "Kung Pao Chicken" output process.

[0046] During the in-depth iteration, an unexpected incident occurred in the R-α zone of the central kitchen: a quality inspector briefly blocked the industrial camera monitoring the wok's discharge port while inspecting the food, resulting in a "black occlusion block" feature appearing in the visual data stream for 2 seconds. At this point, the system entered nonlinear iteration and latent variable space analysis. Surface data showed that the equipment current was normal (no change), but the visual feature changed drastically (black appeared). The system deduced this phenomenon in the latent variable space: the extracted latent variable Zvisual shifted, but the latent variable Zprocess, representing task progress, and the latent variable Zequipment, representing equipment health, still evolved smoothly along the expected trajectory of "discharge completed".

[0047] The system detected an isolated spike in the Zvisual offset in the latent variable space. After the quality inspector left 2 seconds later, the Zvisual instantly dropped back without causing any linkage offset in the Zprocess. Based on this, the system determined that this was a typical "non-process transient interference," such as personnel obstruction, rather than "outlet blockage or equipment failure." The system completely smoothed out this 2-second abnormal fluctuation in the latent variable layer to prevent it from contaminating the final state assessment.

[0048] Based on the latent variable trajectory, the system decodes and outputs the R-α and R-β regions: For the R-α region (fine granularity): it decodes the latent variable fluctuations and outputs the status as: "The wok machine has entered the discharge tail section, the stirring current is at a normal decay trend, and there is no mechanical jamming"; For the R-β region (coarse granularity): the system comprehensively evaluates the latent variable cluster of Combo-Beta and finds that the macro latent variable representing "logistics flow" is consistently lower than the expected baseline evolution curve. The system directly skips the details and outputs the coarse granular status as: "The working status of the R-β rapid cooling zone is abnormal - it is in the 'inefficient idling waiting state caused by logistics input blockage'". Through the deep iteration and interference filtering of S122, the central kitchen system successfully eliminated the visual interference false alarms caused by personnel movement. In the extremely complex concurrent pipeline noise, it confirmed the real regional working status of "the hot processing zone is normal, but the cooling zone is experiencing substantial business blockage" with a very high confidence level, providing boundary conditions for S13 to enter the digital twin space to extrapolate the delay.

[0049] refer to Figure 4 In step S13, the specific steps are as follows: S131: Collect the work database of the central kitchen, and determine the preset work plan of the central kitchen based on the detection of the work database. The work plan is used as a priori constraint and is synchronously loaded into the digital twin space. Combine the work status of each work area with the current work progress of the central kitchen to perform dynamic simulation of multiple physical fields in the digital twin space, and output the corresponding simulation results. S132: Load the causal mechanism corresponding to the central kitchen into the simulation results and trigger the causal tracing of the simulation results. In this way, the interference of non-critical variables can be eliminated through multi-level reasoning, and the corresponding causal chain can be further determined. The schedule deviation can be further transformed into schedule delay content with clear causal orientation. This causal chain covers equipment performance degradation, material flow obstruction or personnel operation fluctuations.

[0050] In the embodiments of this application, the working database of the central kitchen is collected, and the preset work plan of the central kitchen is determined based on the detection of the working database. The work plan is used as a priori constraint and is synchronously loaded into the digital twin space. The working status of each work area and the current work progress of the central kitchen are combined to perform dynamic simulation of multiple physical fields in the digital twin space, and the corresponding simulation results are output. This approach takes into account the overall consideration of the detection of the working database and ensures the accuracy of the preset work plan of the central kitchen.

[0051] At this point, the system directly connects to the central kitchen's Enterprise Resource Planning (ERP) and Advanced Planning and Scheduling (APS) databases to perform structured parsing of unstructured production instructions. The system does not simply extract a list of orders, but deeply examines and extracts a "work plan network" containing strict temporal logic, resource dependencies, and process standard parameters. The system transforms these plan networks into inequality constraint matrices and directed acyclic graph (DAG) time window constraints in mathematical programming, defining them as "prior constraint boundaries" of physical evolution in the digital twin space, ensuring that any subsequent virtual deductions cannot deviate from the logical framework of the established production outline.

[0052] The system awakens a digital twin space corresponding to the central kitchen at a 1:1 scale in the 3D rendering engine, and mounts the prior constraints extracted in the first stage as the rule skeleton of the underlying physical engine. The system performs high-frequency state synchronization, and forcibly superimposes the real working status of each work area output by S122 and the current work progress directly collected by the underlying PLC, such as the percentage of completed processes and the current cycle time, onto the corresponding virtual equipment model and area mesh in the twin space through the spatiotemporal mapping operator. At this time, the virtual objects in the twin space are given a "current physical condition" and "real-time progress label" that are completely consistent with the physical site.

[0053] The system activates a multiphysics coupled solver in the twin space. Considering the characteristics of a central kitchen, it typically constructs thermodynamic, fluid dynamic, and kinetic fields in parallel. Starting from the current real-time loading state, the system combines standard process curves from the prior work plan, such as standard heating rate curves, and performs forward integration on the time axis. During this process, the physical fields do not operate in isolation but interact and transfer energy and matter through cross-coupling terms. For example, changes in the fluid field alter the convective heat transfer coefficient of the thermodynamic field in real time, thereby simulating the continuous dynamic evolution of the working state over a virtual time period. Optionally, the thermodynamic field covers temperature distribution and heat exchange rate; the fluid dynamic field covers steam flow direction and material conveyor velocity; and the kinetic field covers robotic arm kinematics and motor torque.

[0054] At each virtual time step of the multiphysics forward extrapolation, the system continuously performs collision detection between the extrapolated virtual progress, virtual energy consumption, and virtual field distribution and the prior constraint boundaries set in the first stage. Once it is found that the extrapolation trajectory exceeds the tolerance threshold of the prior plan at some point in the future, such as the predicted temperature exceeding the food safety red line or the predicted discharge time being later than the planned cycle time, the system immediately captures the deviation, extracts the timestamp, spatial coordinates, business nodes involved, and degree of deviation of the deviation, and encapsulates it into a structured extrapolation result vector output, providing an accurate "target" for subsequent causal attribution.

[0055] Specifically, the system deeply examines the APS working database of the central kitchen and extracts stringent prior constraints for the "Kung Pao Chicken" task: Constraint-1: The material must enter the rapid cooling machine within 15 minutes after leaving the pot; Constraint-2: The core temperature must drop from 85℃ to below 10℃ within 45 minutes; Constraint-3: The first material handover to the packaging area (R-γ zone) must be completed at T+10 minutes. These constraints are written into the bottom layer of the solver in the twin space.

[0056] On the digital twin 3D screen in the central kitchen, the system forcibly assigns the "R-β zone logistics blockage status" output by S122 to the virtual rapid cooling machine model; at this time, the R-β zone conveyor belt in the virtual space shows a standstill, while the R-α zone virtual wok model continues to pour out high-temperature Kung Pao Chicken ingredients according to the real rhythm; the virtual time is anchored at "3 minutes after leaving the wok, i.e. T+3".

[0057] The system initiates a forward simulation in virtual space, assuming the current blocking state remains unchanged, and simulates the situation over the next 12 minutes. The thermodynamic field solver begins calculation: due to the stagnation of the conveyor belt in the R-β region, the high-temperature Kung Pao Chicken pieces piled on the conveyor belt lose the exchange of forced air cooling and refrigerant from the rapid cooling machine, and rely solely on natural convection for heat dissipation. The fluid dynamics field shows that the cold airflow in this region is heated by the high-temperature material to form a thermal resistance layer. The multiphysics coupling simulation shows that the material temperature drop curve is extremely flat.

[0058] When the virtual clock extrapolated to "14 minutes after cooking", the system performed constraint collision detection and found that Constraint-1 (not entering the rapid cooling machine within 15 minutes) was triggered. The system immediately stopped the invalid extrapolation of this timeline, extracted key deviation data, and output structured extrapolation results: [Extrapolation results: Under the current R-β zone blocking state, it is expected that at T+14 minutes, Task-01 (Kung Pao Chicken) will seriously violate the prior constraint of "entering the cold machine within 15 minutes" for cross-control of raw and cooked food, causing a high risk of food safety, and the subsequent T+10 packaging and handover progress will have an absolute delay of at least 15 minutes]; The central kitchen system no longer just stayed at the superficial cognition of "the conveyor belt has stopped", but through the physical law calculation of the digital twin space, it accurately predicted the certain consequences of "the imminent triggering of the food safety red line and serious progress breach" 1 minute in advance.

[0059] Furthermore, the causal mechanism corresponding to the central kitchen is loaded into the simulation results, and the causal tracing of the simulation results is triggered. In this way, the interference of non-critical variables is eliminated through multi-level reasoning, the corresponding causal chain is further determined, and the schedule deviation is further transformed into schedule delay content with clear causal orientation. This causal chain covers equipment performance degradation, material flow obstruction or personnel operation fluctuations, and introduces equipment performance degradation, material flow obstruction or personnel operation fluctuations.

[0060] At this point, the system retrieves a pre-built structural causal model from the historical fault database and process mechanism database of the central kitchen. This model is usually represented as a directed acyclic graph (DAG) containing nodes and directed edges. The system uses the deduction results output by S131 as "target effect nodes" and forces them to be mapped to the end of this causal graph. At the same time, the system extracts all the related variables generated during the deduction process, such as sudden temperature drop, current fluctuation, and visual occlusion, and attaches them as "candidate cause nodes" to the graph network, thereby establishing an inverse search space with the deviation to be explained as the root node. The deduction results include the progress deviation amount at a certain moment and the abnormal distribution of the physical field.

[0061] In complex industrial environments, numerous variables change synchronously with deviations, but these are not necessarily the causes of the deviations. The system employs a counterfactual reasoning framework, proposing counterfactual hypotheses for each candidate causal node in the causal graph: "If, at the current moment, all other conditions remain unchanged, and only the state of this candidate node is restored to the normal baseline, will the progress deviation in the deduction result still occur?" Through multi-level logical calculations and probability integration in the causal graph, the system calculates the counterfactual causal effect value of each candidate node one by one. For candidate nodes whose counterfactual effect values ​​approach zero, the system classifies them as accompanying non-critical interference variables and resolutely removes them, thereby significantly shrinking the solution space of the causal search.

[0062] For nodes with high causal effect values ​​that are retained, the system introduces Pearl's Do calculus mechanism. By performing intervention operations on the graph, all incoming edges pointing to the core node are forcibly cut off, that is, upstream interference is eliminated, and only the path probability of the node as a pure cause transmitting the effect downstream is calculated. The system performs a depth-first traversal along the directed edges from the core node to the target effect node, and calculates the product of the path coefficients of all connected paths. The system selects the path with the largest product of path coefficients that exceeds the preset confidence threshold and establishes it as the unique or dominant deterministic causal chain, completing the adjustment from "network divergent correlation" to "chain unidirectional causality".

[0063] The system extracts the sequence of underlying physical nodes traversed by the causal chain established in the third stage and maps it to three predefined business semantic category pools in the central kitchen: equipment performance degradation, material flow obstruction, and personnel operation fluctuations. Based on the category of the first node in the causal chain, the system semantically reconstructs the general "schedule deviation" output by S131, generating a structured "schedule delay content" containing "qualitative description of the responsible party + description of the mechanism + quantified consequences of delay," completely eliminating the ambiguity of the warning information. Optionally, equipment performance degradation focuses on energy flow and mechanical characteristics; material flow obstruction focuses on material flow and spatial logistics characteristics; and personnel operation fluctuations focus on visual behavior and interaction timing characteristics. Specifically, S131 deduced that "Kung Pao Chicken materials are expected to trigger the food safety red line and schedule breach at T+14 minutes due to blockage in the R-β zone". In the causal graph of the central kitchen, the system set "T+14 minute schedule delay and over-temperature risk" as the target effect node. At this time, the graph is loaded with a large number of candidate cause nodes, such as "small fluctuations in the compressor current of the R-β zone quick-cooler", "large visual area of ​​the wok discharge in the R-α zone", and "continuous obstruction of the photoelectric sensor of the R-β zone conveyor belt".

[0064] Regarding the "fluctuation of current in the blister cooler" node, the system calculated: "Assuming the current in the blister cooler is completely stable, but the conveyor belt does not deliver the material, will the delay be eliminated?" The calculation result was "no," and this node was determined to be a non-critical interference variable and removed. Regarding the "large visual area of ​​the wok discharge," the system calculated: "Assuming the discharge volume is standardized, but the material still cannot enter the blister cooler due to the conveyor belt being stopped, will the delay be eliminated?" The calculation result was still "no," and it was also removed. The only node that passed the counterfactual test was "the photoelectric sensor of the conveyor belt in the R-β area is continuously blocked."

[0065] The system performs Do-calculus intervention on the event "conveyor belt photoelectric sensor is blocked", cutting off other upstream influences and purely tracking the downstream effects of the event. The system extracts a high-weight unidirectional path in reverse from the graph: "conveyor belt photoelectric sensor blocked → conveyor belt PLC safety interlock mechanism triggered → conveyor belt drive motor forced to shut down → high temperature material accumulates in the absence of strong cooling → material cooling slope violates process curve → 15-minute cooling prior constraint failure triggered → schedule delay occurs". This logically rigorous link is established as the dominant causal chain.

[0066] The system examines the initial trigger source of this causal chain, "the photoelectric sensor is blocked"; combined with the central kitchen's cross-control regulations for raw and cooked food, this non-equipment malfunction blocking must originate from physical intervention; the system accurately maps it to the semantic pool of "personnel operation fluctuations", and finally transforms the cold physical deduction result of S131 into a clearly targeted structured progress delay content: [Progress delay content: due to "personnel operation fluctuations" (specifically: illegal cross-detention operation in the photoelectric sensor area of ​​the R-β zone conveyor belt, triggering the safety interlock shutdown of the conveyor belt, directly causing "material flow obstruction: high temperature accumulation of Kung Pao Chicken", which in turn leads to equipment performance degradation: the rapid cooling machine runs idle, ultimately causing the current work route to face an absolute food safety red line breach risk and a systemic progress delay of more than 15 minutes].

[0067] refer to Figure 5 In step S14, the specific steps are as follows: S141: Trace the delay content and decouple multiple delay factors in different dimensions along the timeline. These delay factors cover equipment-level, process-level, and system-level dimensions. Simultaneously acquire the central kitchen's work history and corresponding industrial constraints, and input them into the closed-loop simulation engine in combination with multiple delay factors. This closed-loop simulation engine performs forward simulation in the digital twin space through reinforcement learning and simulates multiple intervention measures to further generate a multi-level early warning path. This multi-level early warning path includes response strategies for different stages. S142: Dynamically integrate multi-level early warning paths with multiple work areas, perform global optimization of multi-level early warning paths in each work area, eliminate early warning conflicts between work areas, and output an optimized early warning sequence for the central kitchen with spatiotemporal optimality.

[0068] In the embodiments of this application, the delay content is traced back, and multiple delay factors of different dimensions are decoupled in reverse along the timeline. These multiple delay factors cover equipment-level, process-level, and system-level dimensions. The work process of the central kitchen and the corresponding industrial constraints are obtained simultaneously, and the multiple delay factors are input into the closed-loop simulation engine. The closed-loop simulation engine performs forward simulation in the digital twin space through a reinforcement learning mechanism and simulates multiple intervention measures to further generate a multi-level early warning path. This multi-level early warning path includes response strategies for different stages, thus introducing a multi-level early warning path that includes response strategies for different stages.

[0069] At this point, the system uses the delay occurrence time determined by S132 as the zero point and performs reverse slicing tracing in the historical database along the reverse time pointer. The system not only focuses on the direct triggering source, but also adopts the root cause decomposition method to decompose the underlying logic that caused this delay into three orthogonal dimensions: in the "equipment-level dimension", the parameter drift or interlock response characteristics of the equipment before the anomaly occurs are decoupled; in the "process-level dimension", the cycle time mismatch of the preceding and following processes and the accumulation of work-in-process (WIP) due to the current node blockage are decoupled; in the "system-level dimension", the topology disruption and capacity loss characteristics of this local delay on the global production schedule are decoupled, thereby constructing a three-dimensional multi-dimensional set of delay factors.

[0070] The system synchronously extracts "work history" from the central kitchen's MES and knowledge management base, namely the handling trajectory and final evolution result under similar past working conditions. More importantly, the system extracts "industrial constraint relationships" and transforms them into a mixed integer programming constraint matrix containing hard and soft constraints. The system uses these process characteristics and constraint matrices as prior knowledge and injects them into the state space and reward function design of the closed-loop deduction engine to ensure that any strategy generated by the engine is feasible in industrial logic. Hard constraints include the rapid cooling time and maximum operating power limit of equipment stipulated by food safety laws; soft constraints include energy consumption priority strategies and product output order rules.

[0071] The system initializes a deep reinforcement learning agent in a digital twin space. Using the current set of multidimensional delay factors as the initial environmental state, the agent outputs a series of actions in each virtual time step, such as adjusting equipment parameters, switching valves, and modifying order priorities. These actions act on the multiphysics solver in the twin space, triggering a transition in the environmental state, and feeding back the new state to the agent. During this process, a reward function is constructed by closely integrating the industrial constraints injected in the second stage: if several pre-actions violate hard constraints, a large penalty is imposed; if several pre-actions can quickly restore the process cycle time and reduce system-level delays, a positive reward is given. Through hundreds of thousands of closed-loop trial-and-error interactions, the agent learns the optimal strategy cluster for dealing with the current specific delay under complex constraints.

[0072] The system performs cluster analysis and temporal arrangement on the optimal strategy clusters explored by the agent. Based on "intervention time margin" and "resource mobilization amount", it decomposes them into multi-level early warning paths with a progressive relationship. The path includes response strategies at different stages: the primary stage strategy focuses on fine-tuning at the equipment and process levels, such as parameter limit correction and local buffering, which requires execution within a very short time window; the intermediate stage strategy focuses on resource reallocation between regions, such as redundant equipment intervention and process bypassing, which requires execution within a medium time window; the advanced stage strategy focuses on global reconstruction at the system level, such as plan rescheduling and order circuit breaking, as the ultimate fallback solution, and finally outputs a structured early warning path with clear triggering conditions and execution sequence.

[0073] Specifically, the system uses the downtime of zone R-β as a baseline to trace back and output a multi-dimensional set of factors: [Equipment-level factors]: The conveyor belt drive motor is interlocked and cut off by the PLC, and the quick-cooling machine enters a low-frequency idle standby state due to no material input; [Process-level factors]: Kung Pao Chicken physically accumulates on the suspended conveyor belt between the discharge port of the hot processing zone (R-α) and the inlet of the quick-cooling zone (R-β), cutting off the logistics channel from zone R-α to zone R-γ (packaging zone); [System-level factors]: According to the current APS production schedule, if zone R-β is shut down for more than 8 minutes, it will cause the "vegetarian blanching line" of the subsequent 3 orders to cascade down due to the packaging table being occupied.

[0074] The system urgently extracted the industrial constraints of the central kitchen and injected them into the simulation engine: Hard constraint 1: Kung Pao Chicken must be kept in the pot for no more than 15 minutes from cooking to cooling, with only 12 minutes remaining; Hard constraint 2: Manual handling of high-temperature and high-risk materials is strictly prohibited in the cross-control area for raw and cooked food; Soft constraint: Prioritize the continuity of large group meal orders.

[0075] The system initiates a reinforcement learning agent in the digital twin space to conduct trial and error. The agent tries various actions, such as "forcibly unlocking the conveyor belt" and "manually removing the backlog of vegetables". After thousands of evolutions, the agent finally converges on a set of legal and efficient action sequences: Action 1: remotely issue a bypass interlock command to reverse the conveyor belt and return the material to the R-α zone insulation temporary storage pool; Action 2: dispatch AGV carts to transfer the material from the backup channel to the adjacent backup small rapid cooling machine (UID-005); Action 3: adjust the feeding rhythm of the subsequent vegetable line to make time difference for the packaging area.

[0076] The system encapsulates the above action sequence into a tiered, multi-level early warning path and outputs it to the central kitchen control room: [Level 1 Warning Path - Immediate Execution (Equipment / Process Level)]: The strategy is "reverse risk elimination and physical isolation"; immediately trigger the reverse jog return command of the R-β zone conveyor belt to return the Kung Pao Chicken to the R-α zone heat preservation tank, eliminate the threat to the food safety red line, and remove the personnel from the violation status; [Level 2 Early Warning Path - Synchronous Follow-up (Regional Level)]: The strategy is "flexible entry of redundant paths"; while executing the Level 1 action, the backup small rapid cooling machine (UID-005) on the second level is scheduled to go online, and a temporary logistics bridging path is established through AGV to absorb the backlog of materials in the R-α zone insulation pool and restore the flow of local processes; [Level 3 Early Warning Path - Global Monitoring (System Level)]: The strategy is "adaptive rescheduling of cycle time"; the processing rate of the backup rapid cooling machine is monitored in real time. If it is found that the Level 2 path still cannot digest the backlog within 8 minutes, the APS system is automatically triggered to modify the planned issuance time of the subsequent "vegetable line", and actively give up the resources of the packaging area to prevent local delays from evolving into a systemic disaster of global line stoppage.

[0077] Furthermore, the multi-level early warning paths are dynamically integrated with multiple work areas, and the multi-level early warning paths of each work area are globally optimized to eliminate early warning conflicts between different work areas. This results in the output of an optimized early warning sequence for the central kitchen with spatiotemporal optimality. At the same time, the progress delays are further controlled, and the accuracy of the optimized early warning sequence for the central kitchen is determined by fully considering the multi-level early warning paths and multiple work areas.

[0078] At this point, the system constructs a multi-dimensional spatiotemporal resource tensor network based on the global physical space and future time axis of the central kitchen. The system analyzes the multi-level early warning paths generated by S141 for each work area and extracts the "spatial nodes: such as a certain piece of equipment or a certain section of conveyor belt", "time windows: such as the 2nd to 5th minute in the future", and "resource consumption: such as AGV capacity, cooling capacity, and power load" on each path. The system uses these extracted spatiotemporal resource requests as new edge or node weights and forcibly overlays and maps them onto the global spatiotemporal resource tensor network to realize the digital dynamic fusion of local early warning paths of multiple work areas on a unified base.

[0079] In the fused spatiotemporal network, the system employs multi-agent game theory or conflict graph coloring to traverse all activated early warning path nodes. The system focuses on detecting three types of fatal conflicts: space occupancy conflicts, resource competition conflicts, and logical mutual exclusion conflicts. For detected conflicts, the system introduces a preset industrial priority constraint matrix, such as food safety red line intervention being absolutely higher than energy consumption optimization intervention. Through local backtracking and constraint propagation, the system postpones or reroutes low-priority paths in time, forcibly eliminating all hard conflict nodes in the network and generating a set of feasible solutions. Optional conflicts include space occupancy conflicts such as intervention actions in two areas requiring the use of the same conveyor belt at the same time; resource competition conflicts such as two areas simultaneously calling the only remaining spare AGV; and logical mutual exclusion conflicts such as area A requiring increased exhaust ventilation while area B's early warning requires shutting down fresh air for insulation.

[0080] The system constructs a multi-objective optimization function, which typically includes: minimizing the global schedule delay penalty cost, minimizing the additional equipment wear and energy consumption costs caused by intervention measures, and maximizing the on-time delivery rate of the overall food production sequence. The system adopts a non-dominated sorting genetic method (NSGA-II) or a particle swarm optimization method (PSO), using the set of conflict-free early warning paths generated in the second stage as the initial population, and performs multi-dimensional iterative optimization in the solution space of time forward shift and spatial topological constraints. By evaluating the Pareto front of each set of path arrangements, a globally optimal intervention topology sequence that is seamlessly connected on the time axis and has no redundant backtracking in spatial scheduling is finally converged.

[0081] The system performs fine-tuning and smoothing on the time axis of the globally optimal topology output in the third stage to eliminate the tiny time gaps caused by discretization. The system strings together specific intervention actions across different working areas into a single "early warning optimization sequence" with absolute time sequence correctness, according to the absolute time sequence, such as T+0, T+1, and T+3. This sequence not only specifies "what to do" but also precisely specifies "which node in which area, in which second in the future, and by which carrier to execute it", thus completing the final closed loop from complex system deduction to precise machine instructions.

[0082] Specifically, S141 generates a local multi-level path for the "Kung Pao Chicken line (R-β zone blockage)" which is "return material - activate backup chiller - adjust subsequent vegetable line". At this time, it is assumed that the "clean vegetable cutting area (R-0 zone)" on the first floor of the central kitchen also triggers a local path due to blade wear: "call backup AGV to transport new blades, estimated time 3 minutes"; the system projects the path of the second-floor R-β zone and the path of the first-floor R-0 zone onto the three-dimensional spatiotemporal resource network of the central kitchen at the same time; at this time, the two originally parallel local plans intersect in the digital space.

[0083] The system accurately identified two conflicts using a conflict graph coloring method: Space / Resource Conflict: The "backup AGV" required by R-0 area is currently parked near R-β area on the second floor, and the secondary warning path of R-β area implies the need to use this AGV to transfer the returned Kung Pao Chicken to the backup refrigeration unit; one AGV cannot perform the transfer from the first floor to the second floor at the same time. Logical mutual exclusion conflict: The level 3 warning in the R-β area requires "adjusting the subsequent vegetable production line rhythm", but the source of the vegetable production line is the R-0 area. If the R-0 area stops to change the blades for 3 minutes, the vegetable production line will be cut off. At this time, forcibly adjusting the vegetable production line rhythm will cause the computing power to idle.

[0084] Faced with the conflict, the system, based on the priority matrix of "food safety and core hot food supply line priority over clean vegetable cutting", forces the R-0 area path to be downgraded and rerouted in a time sequence: the application to immediately call the AGV in the R-0 area is rejected, and instead the "manual backup tool replacement" path inside the R-0 area is triggered, which takes 5 minutes, but does not require the AGV.

[0085] The system performs a multi-objective evaluation of the rerouting solution; it finds that after eliminating AGV conflicts, although the manual knife change in the R-0 zone causes a 2-minute delay in the supply of clean vegetables, this coincides perfectly with the time window of "adjusting the pace of the subsequent vegetable line in the R-β zone and delaying it by 3 minutes". Not only does it not increase the additional global delay, but it also enables the AGV to fully guarantee the emergency transfer of the high-temperature Kung Pao Chicken on the second floor, achieving Pareto optimality.

[0086] The system abandoned the original scattered instructions divided by region and generated a precise "early warning optimization sequence" for global scheduling of the central kitchen, which is then issued to the actuators in each region: [T+0 minutes 00 seconds] Instruction issued to R-β region: triggers the conveyor belt to reverse jog, and the Kung Pao Chicken returns to the R-α warming pool; [T+0 minutes 05 seconds] Instruction issued to R-0 region: cancels AGV scheduling and triggers the manual replacement process for spare knives in R-0 region; [T+1 minutes 02 seconds] Instruction issued to the global AGV scheduling pool: identifies the only spare AGV and plans... The optimal path is used to travel to the R-α heat preservation tank to stand by; [T+2 minutes 30 seconds] the instruction is issued to the R-β equipment layer: the standby small rapid cooling machine (UID-005) has completed preheating and opened the feeding valve; [T+3 minutes 0 seconds] the instruction is issued to the AGV and R-β area: the AGV picks up the Kung Pao Chicken from the R-α area and feeds it to UID-005 through the temporary channel; [T+5 minutes 0 seconds] the instruction is issued to the R-0 area and the vegetable line: the manual knife change in the R-0 area is completed and the feeding is resumed, and the vegetable line resumes operation according to the delayed new rhythm.

[0087] Please see Figure 6The multi-level early warning system for central kitchens based on industrial big data is applied to the aforementioned multi-level early warning method for central kitchens based on industrial big data; the multi-level early warning system for central kitchens based on industrial big data includes: The dynamic topology module 21 is used to mark multiple industrial devices in the central kitchen, perform spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and construct a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process; and determine the current working route of the central kitchen based on the identification of the dynamic topology map. The industrial big data module 22 is used to determine multiple work areas based on the division of the current work route, and to perform cross-matching with the real-time images of the central kitchen and the corresponding dish production sequence, thereby outputting the industrial big data combination of each work area, and determining the working status of the corresponding work area based on the iteration of each industrial big data combination. The digital twin module 23 is used to obtain the work plan preset by the central kitchen, and input it into the digital twin space in combination with the work status of each work area and the current work progress of the central kitchen. A causal mechanism is introduced in the digital twin space to determine the corresponding progress delay content. The multi-level early warning module 24 is used to determine multiple delay factors in different dimensions based on the tracing of the progress delay content, and further combine the central kitchen's work history and corresponding industrial constraints to conduct closed-loop deduction, thereby outputting a multi-level early warning path and integrating multiple work areas to determine the central kitchen's early warning optimization sequence.

[0088] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0089] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-level early warning method for central kitchens based on industrial big data, characterized in that, include: The system marks multiple industrial devices in the central kitchen, performs spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and constructs a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process. The current work route of the central kitchen is determined based on the identification of this dynamic topology map; Based on the division of the current work route, multiple work areas are determined, and cross-matching is performed with real-time images of the central kitchen and the corresponding food preparation sequence to output the industrial big data combination of each work area. The working status of the corresponding work area is determined according to the iteration of each industrial big data combination. The system obtains the pre-set work plan of the central kitchen, and inputs it into the digital twin space along with the work status of each work area and the current work progress of the central kitchen. A causal mechanism is introduced into the digital twin space to determine the corresponding progress delay content. Based on the tracing of the delay content, multiple delay factors in different dimensions are identified. Furthermore, a closed-loop deduction is carried out by combining the central kitchen's work process and corresponding industrial constraints, thereby outputting multi-level early warning paths and integrating multiple work areas to determine the early warning optimization sequence for the central kitchen.

2. The multi-level early warning method for central kitchens based on industrial big data according to claim 1, characterized in that, The system marks multiple industrial devices in the central kitchen, performs spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and constructs a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process. The current work routes of the central kitchen are determined based on the identification of this dynamic topology map, including: Based on the identification of the central kitchen's distribution map, multiple industrial devices are identified and their locations are marked. The communication channels of the central kitchen are used to associate the multiple industrial devices, obtain the working data of each industrial device, and perform spatiotemporal alignment along the cross-modal alignment mechanism in conjunction with the central kitchen's pipeline distribution map.

3. The multi-level early warning method for central kitchens based on industrial big data according to claim 2, characterized in that, The system marks multiple industrial devices in the central kitchen, performs spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and constructs a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process. Determining the current workflow of the central kitchen based on the identification of this dynamic topology map also includes: A temporal sliding window is introduced during the alignment process, using the abrupt changes in energy flow and material flow of industrial equipment as window segmentation anchors to dynamically generate a dynamic topology map of the central kitchen. A graph neural network is used to learn the feature manifold of this dynamic topology map, and a topology recognition mechanism is combined to remove background noise paths, thereby mapping the current working route of the central kitchen under complex working conditions.

4. The multi-level early warning method for central kitchens based on industrial big data according to claim 1, characterized in that, The process involves defining multiple work areas based on the current work route, cross-matching real-time images from the central kitchen with corresponding food preparation sequences, and outputting a combination of industrial big data for each work area. The working status of the corresponding work area is determined based on the iteration of these industrial big data combinations, including: The current work route is dynamically segmented along a graph segmentation mechanism, and the physical space of the central kitchen is decoupled into multiple work areas with coupling relationships. Real-time images of the central kitchen are acquired simultaneously, and the food output sequence of the central kitchen is determined based on the identification of the food output data stream. Furthermore, the real-time images of multiple work areas and the central kitchen are cross-matched, thereby aggregating equipment data, visual data, and order sequence data into a corresponding industrial big data combination during the matching process.

5. The multi-level early warning method for central kitchens based on industrial big data according to claim 4, characterized in that, The process of determining multiple work areas based on the current work route division, and cross-matching the real-time images of the central kitchen with the corresponding dish-producing sequences to output industrial big data combinations for each work area, and determining the working status of the corresponding work area based on the iteration of each industrial big data combination, also includes: The system performs multi-level iterations on various industrial big data combinations. During the iteration process, a long short-term memory mechanism is introduced to simultaneously trigger nonlinear iteration and latent variable evolution analysis, and simultaneously filter out transient interference, so as to output the working status of each working area at different granularities.

6. The multi-level early warning method for central kitchens based on industrial big data according to claim 1, characterized in that, The process involves acquiring the pre-set work plan of the central kitchen, combining it with the work status of each work area and the current work progress of the central kitchen, and inputting this data into a digital twin space. A causal mechanism is then introduced into this digital twin space to determine the corresponding progress delays, including: The system collects the work database of the central kitchen and determines the preset work plan of the central kitchen based on the detection of the work database. The work plan is then used as a priori constraint and synchronously loaded into the digital twin space. The system combines the work status of each work area with the current work progress of the central kitchen to perform dynamic simulation of multiple physical fields in the digital twin space and outputs the corresponding simulation results.

7. The multi-level early warning method for central kitchens based on industrial big data according to claim 6, characterized in that, The process of obtaining the pre-set work plan of the central kitchen, combining the work status of each work area and the current work progress of the central kitchen, and inputting it into the digital twin space, and introducing a causal mechanism in this digital twin space to determine the corresponding progress delays, also includes: The causal mechanism corresponding to the central kitchen is loaded into the simulation results, and the causal tracing of the simulation results is triggered. In this way, the interference of non-critical variables is eliminated through multi-level reasoning, the corresponding causal chain is further determined, and the schedule deviation is further transformed into schedule delay content with clear causal orientation. This causal chain covers equipment performance degradation, material flow obstruction, or personnel operation fluctuations.

8. The multi-level early warning method for central kitchens based on industrial big data according to claim 1, characterized in that, The process involves identifying multiple delay factors across different dimensions based on the tracing of the delay content, further combining the central kitchen's work history and corresponding industrial constraints to perform closed-loop deduction, thereby outputting multi-level early warning paths, and integrating multiple work areas to determine the central kitchen's early warning optimization sequence, including: The delay was traced back, and multiple delay factors of different dimensions were decoupled in reverse along the timeline. These delay factors covered equipment-level, process-level, and system-level dimensions. The work process of the central kitchen and the corresponding industrial constraints were obtained simultaneously, and the multiple delay factors were input into the closed-loop simulation engine. The closed-loop simulation engine performed forward simulation in the digital twin space through reinforcement learning mechanism and simulated multiple intervention measures to further generate a multi-level early warning path. This multi-level early warning path includes response strategies for different stages.

9. The multi-level early warning method for central kitchens based on industrial big data according to claim 8, characterized in that, The process of identifying multiple delay factors across different dimensions based on the tracing of the delay content, further combining the central kitchen's work history and corresponding industrial constraints for closed-loop deduction, thereby outputting multi-level early warning paths, and integrating multiple work areas to determine the central kitchen's early warning optimization sequence, also includes: By dynamically integrating multi-level early warning paths with multiple work areas, and globally optimizing the multi-level early warning paths of each work area, early warning conflicts between different work areas are eliminated, thereby outputting an optimized early warning sequence for the central kitchen with spatiotemporal optimality.

10. A multi-level early warning system for a central kitchen based on industrial big data, characterized in that: The multi-level early warning system for a central kitchen based on industrial big data is applied to the multi-level early warning method for a central kitchen based on industrial big data as described in any one of claims 1-9. The multi-level early warning system for the central kitchen based on industrial big data includes: The dynamic topology module is used to label multiple industrial devices in the central kitchen, perform spatiotemporal alignment of the working data of each industrial device with the pipeline distribution map of the central kitchen, and construct a dynamic topology map of the central kitchen along a time-series sliding window during the alignment process; and determine the current working route of the central kitchen based on the identification of the dynamic topology map. The industrial big data module is used to determine multiple work areas based on the current work route division, and to perform cross-matching with the real-time images of the central kitchen and the corresponding dish production sequence to output the industrial big data combination of each work area. The working status of the corresponding work area is determined based on the iteration of each industrial big data combination. The digital twin module is used to obtain the work plan preset by the central kitchen, and input it into the digital twin space in combination with the work status of each work area and the current work progress of the central kitchen. A causal mechanism is introduced into the digital twin space to determine the corresponding progress delay content. The multi-level early warning module is used to identify multiple delay factors in different dimensions based on the tracing of the delay content. It further combines the central kitchen's work history and corresponding industrial constraints to conduct closed-loop deduction, thereby outputting multi-level early warning paths and integrating multiple work areas to determine the central kitchen's early warning optimization sequence.

Images