Cloud-group-border lake and warehouse data active prefetching method oriented to industrial small and medium-sized enterprise groups

Abstract

The invention belongs to the field of industrial internet and intelligent manufacturing, and particularly relates to a cloud-group-border lake and warehouse data active prefetching method oriented to industrial small and medium-sized enterprise groups. Comprising the following steps: step 1, constructing a cloud-group-edge three-level asymmetric storage architecture, and executing static physical settlement of data in a normal state; step 2, constructing an LSTM-based time sequence fault prediction model and a multi-dimensional equipment space association map; step 3, extracting sliding window features in real time at an enterprise edge end, and performing LSTM online inference; 4, executing smooth anti-shake logic, and judging and locking a source trigger node based on an early warning threshold value; step 5, traversing the multi-dimensional equipment space correlation graph to carry out space retrieval, and generating a to-be-prefetched equipment set; and step 6, sensing the current downlink bandwidth load, and executing multi-granularity self-adaptive prefetching scheduling based on joint scoring. According to the method, accurate reverse preheating of cold / temperature data is realized, and cross-domain collaboration and zero-delay troubleshooting are realized while the storage overhead of an edge end is reduced.

Description

Technical Field

[0001] This invention belongs to the field of industrial internet and intelligent manufacturing, and specifically relates to a proactive data pre-fetching method for cloud-cluster-edge lake warehouse for industrial SMEs. Background Technology

[0002] In existing equipment monitoring and intelligent operation and maintenance scenarios targeting industrial SMEs, various sensors within a single enterprise continuously generate high-frequency time-series data. Due to the limited storage resources of edge computing nodes (such as enterprise local servers and factory gateways) within a single SME, existing technologies generally adopt a traditional two-tier storage architecture with cloud-edge collaboration, combined with a data lifecycle management solution based on static rules.

[0003] Its main operating mechanism is as follows: A two-tier cloud-edge storage architecture based on static time thresholds: The system is physically divided into a "hot storage layer" at the enterprise edge and a "cold storage layer" in the industry cloud. Time-series data collected in real-time by sensors within SMEs is initially stored in the enterprise's local storage. The system typically sets a fixed static time threshold (e.g., retaining data from the most recent 7 days). When the data residency time exceeds this threshold, the system executes a scheduled task to package, compress, and upload this data to the industry cloud's central database for long-term archiving. Simultaneously, a local copy of this data is deleted at the enterprise edge to free up storage space.

[0004] The reactive data retrieval mechanism is triggered after the fact: In routine monitoring, the enterprise edge only processes and displays real-time sensor waveforms. When a specific sensor reading at an enterprise site exceeds a safety threshold and triggers a system alarm, the enterprise's operations engineer will intervene to investigate. At this time, if the engineer needs to access historical normal waveforms or historical fault slices from several months ago for comparative analysis, the system will, upon receiving a clear query request from the engineer, initiate a data download command to the industry cloud, retrieving the sensor's historical data from the cloud's cold storage to the enterprise edge for decompression and display.

[0005] Independent query logic based on a single device dimension: Existing data retrieval engines treat the data from each sensor as an independent time-series object. During data archiving and subsequent historical data retrieval, the system performs exact matching queries based solely on the single device ID and time range input by the user, independently scheduling data blocks from each sensor in a "question and answer" manner.

[0006] In the complex equipment operation and maintenance and troubleshooting scenarios of industrial SMEs, existing technologies have the following three obvious objective shortcomings, which severely restrict the system's coordination and response efficiency: (1) The "cloud-edge" two-level static storage architecture lacks flexibility: For high-frequency shared data located in industrial parks or the central hub of industrial chain collaboration, if it is directly and crudely dumped from individual enterprises to the industry cloud, it will result in extremely high call costs when collaborating across enterprises and production lines. The existing two-level architecture lacks a physical "temperature buffer" facing the park or small enterprise group.

[0007] (2) High latency in passive data retrieval: The existing system adopts a reactive response model of "fault occurrence -> manual request -> cloud download". The volume of cold data in the cloud is huge, and due to the limited public network bandwidth at the site of small and medium-sized enterprises, it usually takes a long time to temporarily retrieve massive amounts of historical data across networks. This high latency prevents engineers from comparing historical data immediately after arriving at the site, prolonging equipment downtime and causing significant production losses.

[0008] (3) Lack of spatial topology association leads to "data silos" and multiple fetching requests: There are strong physical and technological correlations between the sensors of industrial equipment. The isolated data query logic of the existing system means that when sensor A alarms, the system will only fetch data from A; if the engineer finds that strongly related sensors B and C also need to be investigated, multiple fetching requests must be initiated again. This data scheduling method that lacks "associative context" not only increases the overhead of network communication, but also greatly reduces the efficiency of complex fault root cause analysis. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention proposes a cloud-cluster-edge collaborative lakeware data adaptive hierarchical storage and proactive prefetching method for industrial SMEs, based on spatiotemporal correlation maps and LSTM prediction. This method leverages the predictability of fault trends in SME sensor data over time series and the strong spatial correlation across device / enterprise topologies to establish a joint scoring analytical relationship between sensor fault probabilities and multi-dimensional device maps. Combined with dynamic perception of real-time network bandwidth load at the edge, it executes multi-granular adaptive degradation prefetching scheduling. Through adaptive scoring decisions, it achieves precise reverse preheating of cold / warm data, reducing edge storage overhead for individual enterprises while enabling cross-domain collaboration and zero-latency fault diagnosis.

[0010] Technical solution A proactive data prefetching method for cloud-cluster-edge-lake warehouses targeting industrial SMEs includes the following steps: Step 1: Construct a three-tier asymmetric storage architecture of cloud-cluster-edge, and perform static physical settling of data under normal conditions; Step 2: Construct an LSTM-based time-series fault prediction model and a multi-dimensional device spatial correlation map; Step 3: Extract sliding window features in real time at the enterprise edge and perform online LSTM inference; Step 4: Execute the smoothing and debouncing logic, and determine and lock the source trigger node based on the warning threshold; Step 5: Traverse the multi-dimensional device spatial association map to perform spatial retrieval and generate a set of devices to be prefetched; Step 6: Sensing the current downlink bandwidth load, and executing multi-granularity adaptive prefetching scheduling based on joint scoring.

[0011] Beneficial effects This invention addresses the pain points of high latency and high overhead in existing industrial IoT data retrieval by integrating a cloud-cluster-edge three-level asymmetric architecture for SMEs with a spatiotemporal dual-driven proactive prefetching mechanism, resulting in significant synergistic efficiency improvement and retrieval acceleration.

[0012] In terms of time, this system utilizes LSTM networks to transform reactive, post-event queries into proactive, pre-event warm-up. By leveraging a predictive window before equipment degradation alarms, it silently distributes the necessary cold / warm data to the enterprise edge gateway, allowing engineers' queries to directly access local memory. This mechanism completely eliminates the cumbersome operations required for cross-network access in traditional reactive response modes—including engineers manually assembling multi-dimensional query commands, establishing and addressing long-distance connections across the public network, waking up and performing high-I / O retrieval of remote cold / warm databases, packaging and compressing massive amounts of historical data, and the lengthy backhaul and local decompression under limited and unstable industrial public network bandwidth. By overcoming these intermediate steps that are highly susceptible to network congestion and consume significant time, this system drastically reduces the originally lengthy cross-domain data acquisition process to near-zero latency. This zero-latency data readiness state, to some extent, eliminates the "data transmission black hole" time caused by fetching massive historical waveforms across networks during traditional post-event troubleshooting. Considering the exorbitant costs of unplanned downtime in industrial production lines, ranging from thousands to tens of thousands of yuan per minute, this method, with its extremely low deployment cost of lightweight algorithms, shortens the average repair time of equipment. The downtime losses saved for small and medium-sized enterprises far exceed the cost of the system itself, demonstrating extremely significant industrial economic value.

[0013] In terms of spatial dimension, the system breaks through the limitation of isolated addressing of single devices based on a multi-dimensional device association map. When the source node is alerted, it automatically radiates and "pulls" the context data of strongly associated devices in groups, eliminating the tedious operation of repeatedly issuing query commands and greatly improving the efficiency of root cause analysis of complex faults.

[0014] Meanwhile, to avoid blindly prefetching the entire data and causing overload on the enterprise's limited memory and bandwidth, this algorithm innovatively combines real-time network load with spatiotemporal joint scoring to adaptively perform multi-granular on-demand fetching of high-precision original waveforms or downsampled statistical features, taking into account the trade-off between diagnostic accuracy and system physical overhead.

[0015] In addition, the innovative "cluster-end" temperature buffer layer deployed in the industrial park effectively intercepts high-frequency shared data within the warranty period of the supply chain between enterprises, breaking down the data silo barrier in cross-domain and cross-organizational collaboration of SMEs in the park with a lower storage transformation cost. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a model diagram of a CNC machine tool monitoring scenario system for small and medium-sized enterprises, as described in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0018] This invention is applicable to multi-level time-series databases and edge computing systems in resource-constrained industrial IoT scenarios, as follows: A proactive data prefetching method for cloud-cluster-edge-lake warehouse for industrial SMEs includes the following steps: (e.g.) Figure 1 ) Step 1: Construct a three-tier asymmetric storage architecture of cloud-cluster-edge, and perform static physical settling of data under normal conditions; The cloud-cluster-edge three-tier asymmetric storage architecture comprises edge nodes (hot layer), cluster nodes (warm layer), and cloud nodes (cold layer), constructing an asymmetric lakehouse architecture with a three-layer buffer system of hot, warm, and cold layers. Details are as follows: Edge nodes (hot layer): Deployed at the production site of a specific SME (such as the enterprise's local data center or factory smart gateway), with a built-in in-memory database. Limited by the extremely small hardware storage space of a single enterprise, this layer is only responsible for storing two types of data within the enterprise: real-time streaming data generated by sensors (short residence time, such as within 7 days); and high-value historical slices temporarily pulled by the "active prefetching algorithm." This layer ensures microsecond-level response time for the enterprise's local large-screen monitoring and emergency shutdown control.

[0019] The sensors mentioned are temperature, vibration, and pressure timing sensors deployed on key industrial production equipment (such as CNC machine tools, servo motors, air compressors, etc.) for condition monitoring. Sensor failure refers to abnormal conditions, performance degradation, or mechanical failures (such as increased vibration due to spindle bearing wear, or abnormal temperature rise due to lubrication failure) occurring in the underlying key industrial equipment monitored by the sensor.

[0020] Cluster Nodes (Temperature Layer): Shared servers deployed in industrial park data centers or regional small business clusters serve as a new "collaborative buffer" in the system. They centrally store mid-term business flow data (such as data from the past three months) that has gradually cooled down from the edge of each individual enterprise within the park. This temperature layer not only breaks down the data silos of individual SMEs and provides a global physical hub for cross-factory spatial correlation graph retrieval, but also serves as the highly efficient first-choice hit library for the "active prefetching algorithm" of this invention. Since most degradation characteristics of industrial equipment occur within this time window, when the LSTM algorithm triggers the prefetching instruction, the system can preferentially retrieve the required contextual historical features from this temperature layer via the park's local area network, completely avoiding the bandwidth bottlenecks and long latency encountered when requesting from the industry's large cloud across the public network. When joint troubleshooting and quality traceability occur across enterprises and factories, the cluster nodes can provide millisecond-level data support.

[0021] Cloud Node (Cold Layer): An object storage service (such as MinIO / OSS) deployed on an industry public cloud. It is used to store all raw, non-destructive sensor data for extended periods beyond the thermocline dwell time (e.g., more than 3 months).

[0022] Initial static settlement and flow process: Under normal conditions without any abnormalities, the system executes a unidirectional static physical settlement and flow process based on a preset time window threshold. Specifically, let the current time be... The timestamp of the sensor data block is A first settlement threshold is set at the edge. (Typical value is 7 days to ensure microsecond-level response for monitoring and shutdown control), a second settlement threshold is set at the group terminal. (Typical value is 90 days, ensuring millisecond-level support for troubleshooting across factory areas). When At that time, data is migrated from the edge to the group; when At that time, the data is migrated from the client to the cloud for long-term archiving.

[0023] Step 2: Construct an LSTM-based time-series fault prediction model and a multi-dimensional device spatial correlation map; Offline computation and model training of the "spatiotemporal dual-drive engine" can be completed in advance on cloud / cluster servers with sufficient computing power.

[0024] Step 2.1: Training the LSTM prediction model; Use historically archived sensor stream data with a fixed sliding window. Extracting the input sequence The reverse time window extraction method is used to capture the time when the fault occurs. The length of the front is The samples within the window are labeled as The normal period sample label is .

[0025] Long-range degradation features are extracted using a multi-layer LSTM network and mapped to fault probabilities using a Sigmoid function in the output layer. .

[0026] To balance the extraction accuracy of the long-term evolution characteristics of industrial equipment from "sub-health to degradation to failure" with the limited memory computing power at the edge, the multilayer LSTM network adopts a vertical stacking architecture with a depth of 2 to 4 layers (preferably 3 layers). The number of neurons in each hidden layer is set between 32 and 256 (preferably 64) and the Tanh activation function is used. At the same time, a dropout regularization mechanism with a dropout rate of 0.2 is introduced between layers to prevent overfitting during small sample training.

[0027] During training, the weights are updated via backpropagation using the binary cross-entropy loss function. After the loss converges, the lightweight LSTM model is deployed to the enterprise edge.

[0028] Step 2.2: Construct a multi-dimensional spatial association map of devices. ; Offline mining of historical fault work orders and query logs to build a sensor-based node system. The association weight is the edge The topology diagram. Two nodes. and Association weight between It is derived by combining the co-occurrence probability and physical distance. The mathematical calculation method is as follows: in, To trigger a query within a set historical troubleshooting time window (typically within 2 hours before and after the fault). Simultaneous access The conditional probability; This represents the actual physical topology hop distance between the two devices; This is a confidence level adjustment factor (the empirical typical value is 0.7). This is the distance attenuation factor (typically 1.0).

[0029] After completing the weight calculation between each pair of nodes, the system targets all nodes involved in the monitoring. Each sensor node is mapped to its inter-node correlation weight values ​​according to its row and column indices to construct a... A dimensional adjacency matrix of the relational graph. In this matrix, the th... Line 1 The matrix elements of a column are the corresponding nodes. With nodes Association weight between Furthermore, it is stipulated that the elements on the main diagonal of the matrix (i.e., the degree of association between a node and itself) are... After the calculation and construction are completed, the adjacency matrix of the association graph is serialized to generate an initial graph file, which is then pre-saved in the edge memory for subsequent online triggering of spatial adjacency traversal.

[0030] Meanwhile, to adapt to the dynamic evolution of industrial field equipment topology (such as production line modifications, process adjustments, or sensor additions or removals), the system introduces a dynamic update mechanism for the association graph at the cluster / cloud level: including periodic time-triggered and event-triggered events. The periodic time-triggered events are triggered monthly; the event-triggered events are triggered by setting up a driver listener to monitor the registration / cancellation change events of the underlying equipment ledger in real time to trigger the dynamic update of the association graph.

[0031] The dynamic update process is as follows: Using the latest historical fault work orders and query logs, the correlation weights between each node are reassessed and updated according to the aforementioned calculation method. After generating and serializing the adjacency matrix of the association graph, it is sent to the edge end to overwrite the original graph file with hot updates.

[0032] Step 3: Extract sliding window features in real time at the enterprise edge and perform online LSTM inference; During system online operation, edge nodes collect sensor data in real time. The data, after data cleaning, yields a sequence. This provides the LSTM prediction model to the sensor. The model undergoes forward propagation, calculating and outputting the sensor's predictions in real time. In the future Instantaneous failure probability within the window .

[0033] Step 4: Execute the smoothing and debouncing logic, and determine and lock the source trigger node based on the warning threshold; To prevent false alarms caused by transient electromagnetic interference in industrial settings, the system incorporates a moving average anti-jitter algorithm into the instantaneous probability sequence: in, The number of sampling cycles is set for smoothing and anti-shake operation. This is to effectively cover the time window of transient interference and adapt to the different sampling rates of heterogeneous sensors in industrial environments. The value of is not a fixed constant, but depends on the target sensor. Actual physical sampling frequency With preset image stabilization physical time window The specific mathematical mapping relationship is dynamically determined and satisfies: In the formula, Indicates a round-up operation. Anti-shake physical time window. The selection must strictly cover the maximum duration of typical transient electromagnetic interference in the field (typical value setting). (Between 1.0 and 2.0 seconds).

[0034] Set security warning thresholds (Typical value is 0.85), when the following conditions are met At that time, it was determined that the sensor monitoring device had established a degradation trend, and it was marked as a "source trigger node". "and generate signaling to activate the space diffusion mechanism.

[0035] Step 5: Traverse the multi-dimensional device spatial association map to perform spatial retrieval and generate a set of devices to be prefetched; Set topology cutoff threshold (Typical value is 0.3), the edge is triggered by the source node. Using a circle as the center, traverse the pre-defined adjacency matrix of the association graph. Extract all circles that satisfy the association weight. neighboring nodes This constitutes the set of targets to be prefetched. .

[0036] Step 6: Sensing the current downlink bandwidth load, and executing multi-granularity adaptive proactive prefetching scheduling based on joint scoring; Source node and sets The nodes are merged to form a set of nodes to be evaluated, and for each target node in the set... The spatiotemporal joint score of its "prefetch urgency" is calculated using the evaluation function. ,as follows: Among them, for the source node Self, constant The weight parameters satisfy For example, if the emphasis is on lead time... Taking 0.6, focusing on the completeness of obstacle removal. Take 0.6.

[0037] Meanwhile, the edge device monitors the downlink bandwidth load rate from the current cold / warm layer to the hot layer in real time. Set network congestion threshold .

[0038] Set a high-value rating threshold Medium scoring threshold Traverse the set of nodes to be evaluated and dynamically issue prefetch instructions according to the following closed-loop rules: Full high-precision waveform prefetch: If And currently If the network is working properly, then a command is sent to the cold / warm layer to pull the node. Non-destructive, high-precision raw waveforms down to the edge thermal layer from historically similar time periods. The "historically similar time period" is defined as the time period used for fault root cause comparison analysis, including a "historical fault degradation window" and a "healthy baseline comparison window," with the specific time range as follows: (1) Historical Fault Deterioration Window: refers to the source trigger node in the system's historical work order database. The last time a deterministic fault or alarm occurred Previous time period ,in The length of the pre-fault prediction window is kept consistent with the length of the LSTM model during offline training in step 2. (2) Health baseline reference window: refers to the time since the current warning. A known stationary operating period with a fixed time offset is denoted as ,in The operating condition offset period (typically taken as the same day and time of the previous period, i.e., offset by 7 days).

[0039] Downsampling statistical feature prefetching (degradation strategy): If Or although the rating is extremely high ( But currently (Network congestion) To prevent network crashes from triggering degradation, nodes are only pulled from remote locations (including warm and cold layers). Downsampling statistical characteristic values ​​(such as extreme values ​​and mean envelope) in the time domain to the thermal layer.

[0040] Ignore pull operations: If The system determined it to be a weakly correlated interference and did not perform any cross-domain fetching action.

[0041] At this point, the system has completed a closed loop of precise reverse data scheduling from the remote collaborative storage pool to the enterprise edge memory.

[0042] Example To enable those skilled in the art to better understand the technical solution of this invention, the specific implementation process of this invention will be further described in detail below, taking the monitoring scenario of CNC machine tools in a small and medium-sized enterprise in an industrial park as an example. The system model diagram is shown below. Figure 2 As shown.

[0043] Implementation Scenario and Offline Initialization: A high-end CNC machine tool was deployed in the workshop of a small and medium-sized enterprise. Its enterprise edge was an industrial gateway with only 4GB of memory. The machine tool was equipped with three key sensors: spindle vibration (node ​​A), spindle temperature (node ​​B), and coolant pressure (node ​​C). During system initialization, the static settling time threshold for data from the edge to the industrial park's cluster server was set to 7 days, and the settling threshold from the cluster server to the industry cloud was set to 90 days. In the offline phase, the industrial park's cluster server used archived data from the past year to train lightweight LSTM models for vibration, temperature, and pressure prediction according to physical quantity type, and then distributed them to the enterprise edge gateway. Simultaneously, the system constructed a multi-dimensional spatial correlation map of equipment by mining historical work orders, concluding that spindle vibration and spindle temperature were strongly correlated (weighted). The correlation with coolant pressure is weak (weight). ).

[0044] Online Inference and Anti-jitter Trigger Mechanism: During real-time machine tool operation, streaming data from the spindle vibration sensor (node ​​A) is continuously input into the edge-end LSTM model. At a certain moment, the model outputs an instantaneous probability of future failure that spikes to 0.90. At this point, the system triggers anti-jitter logic, given the known sampling frequency of the vibration sensor. Set the stabilization time window to 10Hz. The time is 1.2 seconds, and the system calculates according to the formula. The number of smooth sampling periods is calculated. The edge gateway then calculates the moving average failure probability of the past 12 sampling points as 0.88. Since this smoothing value is greater than the preset safety warning threshold of 0.85, the system officially determines that the spindle is about to undergo mechanical degradation and locks node A as the "source trigger node".

[0045] Spatial collaborative retrieval and adaptive scheduling closed loop: After locking the source node A, the system traverses the association graph with A as the center. Under the set topological truncation threshold (0.3), node B (0.85) is added to the prefetch set, while node C (0.15) is removed. Subsequently, the system sets time and spatial weight parameters. The spatiotemporal joint pre-acquisition of target node B is calculated. Before executing the data retrieval, the system dynamically senses the downlink bandwidth load rate of the current edge gateway. If the current network is unobstructed (e.g., load rate of 0.3, below the congestion threshold of 0.7), and node B's score (0.865) is greater than the high-value threshold (0.8), the system immediately sends an instruction to the park cluster terminal to fully pre-fetch the high-precision historical waveforms of node B into the gateway memory; if the current network is congested (e.g., load rate of 0.85), the system triggers a degradation protection strategy, only retrieving downsampled statistical features such as "daily extreme values ​​and average values" of node B. A few minutes later, when the machine tool experiences a real alarm shutdown and engineers arrive on-site to investigate, the strongly correlated historical comparison data is already ready in the gateway memory, achieving "zero-latency" fault diagnosis under cross-domain collaboration.

[0046] When the LSTM model predicts the aforementioned equipment failure trend, or when a real alarm occurs, the following actions are typically required in industrial settings: the control system issues hardware and software interlock commands for emergency shutdown or reduced speed to avoid physical damage; simultaneously, on-site maintenance engineers or cross-enterprise collaborative troubleshooting experts need to intervene immediately, reviewing the historical normal waveforms, historical fault slices, or statistical characteristics of the warning equipment and its strongly correlated devices from several months ago, conducting in-depth comparisons and root cause analysis to determine whether to replace parts, adjust process parameters, or continue operation with the fault. The mechanism of this invention triggering "reverse active prefetching" of cold / warm data to the edge satisfies this need for rapid access to cross-domain historical context data during emergency troubleshooting, and is the direct driving force behind this invention's "reverse active prefetching" of cold / warm data to the edge.

[0047] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A proactive data prefetching method for cloud-cluster-edge-lake warehouse for industrial SMEs, characterized in that, Includes the following steps: Step 1: Construct a three-tier asymmetric storage architecture of cloud-cluster-edge, and perform static physical settling of data under normal conditions; Step 2: Construct an LSTM-based time-series fault prediction model and a multi-dimensional device spatial correlation map; Step 3: Extract sliding window features in real time at the enterprise edge and perform online LSTM inference; Step 4: Execute the smoothing and debouncing logic, and determine and lock the source trigger node based on the warning threshold; Step 5: Traverse the multi-dimensional device spatial association map to perform spatial retrieval and generate a set of devices to be prefetched; Step 6: Sensing the current downlink bandwidth load, and executing multi-granularity adaptive prefetching scheduling based on joint scoring.

2. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 1, characterized in that, In step 1, The cloud-cluster-edge three-tier asymmetric storage architecture includes edge nodes, cluster nodes, and cloud nodes, as detailed below: Edge nodes, also known as thermal layers, are deployed at the production site of a specific small or medium-sized enterprise and have a built-in in-memory database. They are only responsible for storing two types of data within the enterprise: real-time streaming data generated by sensors and high-value historical slices temporarily pulled by active prefetching algorithms. Cluster nodes, also known as temperature layers, are shared servers deployed in industrial park data centers or regional small business clusters, serving as a "collaborative buffer" for the system. They centrally store intermediate business flow data that gradually cools down and settles from the edge of each individual enterprise within the park. When joint troubleshooting and quality traceability occur across enterprises and factories, the cluster nodes provide millisecond-level data support, avoiding the extremely high latency caused by retrieving data from cloud nodes. Cloud nodes, also known as cold layers: Object storage services deployed on industry public clouds are used to store all original historical data of sensors that are non-destructive and extend beyond the temperature layer's dwell time. Initial static settlement and flow process: Under normal conditions without any abnormalities, the system executes a unidirectional static physical settlement and flow process based on a preset time window threshold; specifically, let the current time be... The timestamp of the sensor data block is Set the first settlement threshold at the edge. Set a second settlement threshold at the group end. ;when At that time, data is migrated from the edge to the group; when At that time, the data is migrated from the client to the cloud for long-term archiving.

3. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 1, characterized in that, Step 2 specifically involves: Offline computation and model training of the "spatiotemporal dual-drive engine" can be completed in advance on cloud / cluster terminals with sufficient computing power; Step 2.1: Training the LSTM prediction model; Use historically archived sensor stream data with a fixed sliding window. Extracting the input sequence ; Using the reverse time window extraction method, the time of fault occurrence is... The length of the front is The samples within the window are labeled as The normal period sample label is Long-range degradation features are extracted using a multi-layer LSTM network, and then mapped to fault probabilities in the output layer using a Sigmoid function. ; During training, the weights are updated via backpropagation using the binary cross-entropy loss function. After the loss converges, the lightweight LSTM model is deployed to the enterprise edge. Step 2.2: Construct a multi-dimensional spatial association map of devices. ; Offline mining of historical fault work orders and query logs to build a sensor-based node system. The association weight is the edge Topology graph; two nodes and Association weight between The results, derived from combining co-occurrence probability and physical distance, are as follows: in, To trigger a query within the set historical troubleshooting time window Simultaneous access The conditional probability; This represents the actual physical topology hop distance between the two devices; This is a confidence level adjustment factor; This is the distance attenuation coefficient; After calculating the weights between each pair of nodes, the system targets all nodes involved in the monitoring. Each sensor node is mapped to its inter-node correlation weight values ​​according to its row and column indices to construct a... A dimensional adjacency matrix of the correlation graph; in this matrix, the th... Line number The matrix elements of a column are the corresponding nodes. With nodes Association weight between And it is specified that the main diagonal elements of the matrix are After the calculation and construction are completed, the adjacency matrix of the association graph is serialized to generate a static graph file and pre-saved in the memory of the edge end for spatial adjacency traversal when triggered online later.

4. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 3, characterized in that, A dynamic update mechanism for the association graph is introduced on the group / cloud: including periodic time triggering and event triggering; the periodic time triggering is a monthly timed triggering; the event triggering is triggered by setting a driver listener to monitor the registration / cancellation change events of the underlying device ledger in real time to trigger the dynamic update of the association graph; The dynamic update process is as follows: using the latest historical fault tickets and query logs, the correlation weights between each node are reassessed and updated. After generating and serializing the adjacency matrix of the association graph, it is sent to the edge end to overwrite the original graph file with hot updates.

5. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 1, characterized in that, Step 3 specifically involves the edge nodes collecting sensor data in real time during system online operation. The data, after data cleaning, yields a sequence. The LSTM prediction model is provided to the sensor; the model undergoes forward propagation, and the sensor's predictions are calculated and output in real time. In the future Instantaneous failure probability within the window .

6. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 5, characterized in that, Step 4 specifically involves: The system introduces a moving average anti-shake algorithm for the instantaneous probability sequence: in, The number of sampling periods for image stabilization and smoothing; Set security warning thresholds When satisfied At that time, determine the sensor The degradation trend of the monitoring equipment was established, and it was marked as a "source trigger node". "and generate signaling to activate the space diffusion mechanism.

7. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 6, characterized in that, The value is determined by the sensor. Actual physical sampling frequency With preset image stabilization physical time window The specific mathematical mapping relationship is dynamically determined and satisfies: ， in, Indicates rounding up, debounce physical time window The selection covers the maximum duration of typical transient electromagnetic interference in the field.

8. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 1, characterized in that, Step 5 specifically involves: Set topology cutoff threshold Edge end with source trigger node Using a circle as the center, traverse the pre-defined adjacency matrix of the association graph; extract all circles that satisfy the association weight. neighboring nodes This constitutes the set of targets to be prefetched. .

9. The cloud-cluster-edge-lake warehouse data proactive prefetching method for industrial SMEs as described in claim 1, characterized in that, Step 6 specifically involves sensing the current downlink bandwidth load and executing multi-granularity adaptive proactive prefetching scheduling based on joint scoring. Source node and sets The nodes are merged to form a set of nodes to be evaluated, and for each target node in the set... The spatiotemporal joint score of its "prefetch urgency" is calculated using the evaluation function. ,as follows: Among them, for the source node Self, constant The weight parameters satisfy ; Meanwhile, the edge device monitors the downlink bandwidth load rate from the current cold / warm layer to the hot layer in real time. Set network congestion threshold ; Set a high-value rating threshold Medium scoring threshold Traverse the set of nodes to be evaluated and dynamically issue prefetch instructions according to the following closed-loop rules: Full high-precision waveform prefetch: If And the network is currently unobstructed. Then, a command is sent to the cold / warm layer to pull the node. Lossless, high-precision original waveforms from similar historical periods to the edge thermal layer; Downsampling statistical feature prefetching: if Or even though the rating is extremely high However, the current network is congested. To prevent network crashes from triggering degradation, only nodes are pulled from the remote end. Downsampling statistical eigenvalues ​​in the time domain to the thermal layer; Ignore pull operations: If The system determines this to be a weakly correlated interference and does not perform any cross-domain fetching action; At this point, the system has completed a closed loop of precise reverse data scheduling from the remote collaborative storage pool to the enterprise edge memory.

Images