A control method for weak network high concurrency industrial edge cloud collaborative adaptive routing

Abstract

This invention relates to the field of Industrial Internet of Things (IIoT) technology and discloses a control method for collaborative adaptive routing between industrial edge cloud and digital network in low-network, high-concurrency environments. The method includes: comparing geospatial coordinates with geofences; determining whether to trigger a context engine based on the comparison results or changes in the status of digital work orders; if the context engine is triggered, clustering the work order semantics corresponding to the changed digital work order status in the cloud vector space based on the geospatial coordinates and / or the work order semantics, identifying and synchronizing a subset of hot knowledge points; performing emergency quantization, bandwidth probe evaluation, and complexity evaluation in parallel on industrial query requests to determine the maximum tolerable latency, uplink and downlink bandwidth, and total cloud latency; if the edge device is offline, encapsulating the data generated during the offline state into an offline cognitive log and performing local retrieval based on a knowledge base. This invention balances low-latency response in low-network environments with routing stability for high-concurrency requests.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IIoT) technology, and more specifically, to a control method for collaborative adaptive routing of industrial edge cloud in low-concurrency environments with weak networks. Background Technology

[0002] With the rapid development of industrial IoT and edge computing technologies, cloud-edge collaborative intelligent analysis systems have been gradually introduced into industrial scenarios such as mines, deep-sea drilling platforms, and closed discrete manufacturing workshops to achieve industrial equipment status monitoring, process knowledge retrieval, and on-site decision support.

[0003] Existing patent CN114268640A discloses a cloud-edge collaborative industrial IoT intelligent routing system. This system collects, cleans, transforms, and compresses industrial equipment data through edge computing nodes, then transmits the pre-processed data to cloud services. The cloud then monitors, processes, and analyzes the data based on pre-configured business processes, as well as monitoring the status of the industrial equipment. However, under conditions of weak network, network outages, or high concurrency, the routing mechanism relies on interconnected transmission of the complete context, leading to system link collapse or timeouts. Furthermore, processing nodes not only consumes significant bandwidth but also causes data inconsistencies and communication congestion during gateway synchronization, making it difficult to simultaneously achieve low-latency response in weak network environments and routing stability for high-concurrency requests.

[0004] Therefore, it is necessary to design a control method for collaborative adaptive routing of industrial edge cloud in weak network with high concurrency to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency conditions. It aims to solve the problems that the routing mechanism relies on the interconnection transmission of the complete context under weak network, network outage or high concurrency conditions, which leads to system link avalanche or timeout, and when processing nodes, it is difficult to take into account both the low latency response in weak network environment and the routing stability of high concurrency requests.

[0006] This invention proposes a control method for collaborative adaptive routing in industrial edge cloud under weak network and high concurrency conditions, comprising: Obtain the geospatial coordinates of the edge device and the status of the digital work order bound to the edge device. Compare the geospatial coordinates with the geofence and determine whether to trigger the context engine based on the comparison result or the change in the status of the digital work order. If the context engine is triggered, then based on the geospatial coordinates and / or the semantics of the work order corresponding to the changed digital work order status, clustering is performed in the cloud vector space to determine a subset of hot knowledge and synchronize it. During synchronization, semantic buckets with changed hash values ​​are pushed to determine the knowledge base. If the edge device receives an industrial query request, it performs emergency quantization, bandwidth probe evaluation, and complexity evaluation in parallel on the industrial query request to determine the maximum tolerable latency, uplink and downlink bandwidth, and total cloud latency. If the uplink and downlink bandwidth is less than the safe value, or the total cloud latency is greater than the maximum tolerable latency, the cloud path of the industrial query request is cut off and the knowledge base is invoked. Otherwise, the industrial query request is encrypted and sent to the cloud for task offloading. If the edge device is offline, the data generated during the offline state is encapsulated into an offline cognitive log, and a local retrieval is performed based on the knowledge base. If the confidence level of the retrieved knowledge is less than the confidence level threshold, a conservative disclaimer instruction is attached. When the offline state is restored, the offline cognitive log is pushed to the cloud.

[0007] Furthermore, when determining whether to trigger the context engine, the following steps are taken: The geofence includes equipment areas, work areas, and hazardous areas; The digital work order status includes work order dispatch status and work order execution status, and the work order execution status includes work order semantics of equipment object, job type and job area; If the geospatial coordinates cross from one region to another, or the digital work order status changes from the work order dispatch status to the work order execution status, then the context engine is determined to be triggered; otherwise, the context engine is determined not to be triggered.

[0008] Furthermore, when clustering the semantics of the work order corresponding to the changed digital work order status in the cloud vector space based on the aforementioned geospatial coordinates and / or, the clustering includes: Based on the geospatial coordinates, the other area crossed is used as a spatial retrieval condition, and the work order semantics are used as a semantic retrieval condition; Based on the density clustering algorithm and using the spatial retrieval conditions and / or semantic retrieval conditions, clustering is performed in the cloud vector space to determine candidate knowledge clusters, and the candidate knowledge clusters are determined as the hot knowledge subset.

[0009] Furthermore, the semantic buckets whose hash values ​​have changed are pushed to determine the knowledge base; When distributing the aforementioned hot knowledge subset, synchronization is performed based on a conflict-free replicated data type and a Merkle tree. The gateway serializes the semantic buckets whose hash values ​​have changed and pushes them based on a lightweight protocol to determine the knowledge base.

[0010] Furthermore, when performing emergency metricization, bandwidth probe evaluation, and complexity evaluation in parallel on the industrial query request, the following steps are included: Based on the industrial ontology, the system identifies the equipment objects, risk words, and operational intentions in the industrial query request. The system then inputs these information into an adaptive fuzzy inference system to determine an urgency score. Based on the urgency score, the system maps the maximum tolerable delay, with the urgency score and the maximum tolerable delay being inversely proportional.

[0011] Furthermore, when performing emergency metricization, bandwidth probe evaluation, and complexity evaluation in parallel on the industrial query request, the following steps are also included: The heartbeat packets of the edge device are obtained, and the uplink and downlink bandwidths are determined based on the round-trip time, packet loss status, and response interval of the heartbeat packets. Identify whether the industrial query request contains multi-hop logical reasoning involving cross-device association and cross-process segment association, and determine the total cloud latency based on the routing cost function.

[0012] Furthermore, the data generated during the offline state includes interactive data, sensor data, and local decision-making.

[0013] Furthermore, when performing a local search based on the knowledge base, it includes: If the confidence level of the retrieved knowledge is greater than or equal to the confidence threshold, then the conservative disclaimer instruction will not be attached. If the confidence level of the retrieved knowledge is less than the confidence level threshold, the conservative disclaimer instruction is attached based on the prompt word project. The conservative disclaimer instruction is used to indicate the network outage status.

[0014] Furthermore, when pushing the offline cognitive logs to the cloud, the following steps are included: When the network outage is restored, the offline cognitive log is asynchronously pushed to the cloud as incremental data based on the conflict-free replication data type, and the cloud performs a retrospective evaluation to identify whether the edge device caused a chain reaction of potential risks to the remote device during the network outage.

[0015] Furthermore, when pushing the offline cognitive logs to the cloud, the process also includes: If the cloud identifies that the edge device causes a chain reaction of potential problems to the remote device when the network is down, it generates a correction and compensation instruction and pushes it back to the edge device to remind manual intervention.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By judging the geospatial coordinates of edge devices and the status of digital work orders, a context engine can be triggered when a device enters a specific work area or when the work order status changes. This allows the cloud to determine a subset of hot knowledge based on the current spatial location and the current work order semantics, avoiding the distribution of the entire knowledge base to the edge and reducing the data synchronization pressure and storage burden on edge devices in weak network environments. Incremental synchronization of semantic buckets based on conflict-free replication data types and Merkle trees reduces redundant transmissions and alleviates communication congestion in high-concurrency scenarios. By performing emergency quantification, bandwidth probe evaluation, and complexity evaluation in parallel on industrial query requests, the security urgency of the request, the current network availability, and the task processing complexity are considered simultaneously, avoiding misjudgments caused by routing based solely on computational complexity or fixed network conditions. By cutting off the cloud path and calling the local knowledge base for processing, low-latency response can be guaranteed in weak network, network outage, or high-risk task scenarios, reducing security risks caused by cloud timeouts. By encapsulating offline cognitive logs during network outages and pushing them to the cloud for retrospective evaluation and compensation after network recovery, the reliability, security, and adaptability of industrial edge cloud collaborative systems are improved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a control method for collaborative adaptive routing between industrial edge cloud and weak network conditions with high concurrency, provided in an embodiment of the present invention; Figure 2 The flowchart illustrates the routing decision logic based on maximum tolerable latency, uplink and downlink bandwidth, and total cloud latency, as provided in this embodiment of the invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] See Figure 1-2 As shown in some embodiments of this application, a control method for collaborative adaptive routing in a weak network with high concurrency in an industrial edge cloud includes: S100: Obtain the geospatial coordinates of the edge device and the status of the digital work order bound to the edge device. Compare the geospatial coordinates with the geofence and determine whether to trigger the context engine based on the comparison result or the change in the status of the digital work order. S200: If the context engine is triggered, the semantics of the work order corresponding to the changed digital work order status will be clustered in the cloud vector space based on the geospatial coordinates and / or the changes. Hot knowledge subsets will be determined and synchronized. During synchronization, semantic buckets with changed hash values ​​will be pushed to determine the knowledge base. S300: If the edge device receives an industrial query request, it performs emergency quantization, bandwidth probe evaluation and complexity evaluation in parallel on the industrial query request to determine the maximum tolerable latency, uplink and downlink bandwidth and total cloud latency. If the uplink and downlink bandwidth is less than the safe value, or the total cloud latency is greater than the maximum tolerable latency, the cloud path of the industrial query request is cut off and the knowledge base is called. Otherwise, the industrial query request is encrypted and sent to the cloud for task offloading. S400: If the edge device is offline, the data generated during the offline state will be encapsulated into an offline cognitive log and a local retrieval will be performed based on the knowledge base. If the confidence level of the retrieved knowledge is less than the confidence level threshold, a conservative liability disclaimer instruction will be attached. When the offline state is restored, the offline cognitive log will be pushed to the cloud.

[0022] Specifically, the control method provided in this embodiment is applicable to industrial IoT environments with weak networks, high concurrency, or intermittent network outages, such as underground mines and deep-sea drilling platforms. Edge devices do not only refer to fixed edge servers, but can also be field computing terminals carried by personnel or installed on mobile devices. For example, edge devices can be industrial PDAs, inspection robots, or vehicle-mounted edge computing terminals. Edge devices typically include a miniature vector library (local storage space), a quantized small language model (local execution computing power), and multimodal sensors (for acquiring field input).

[0023] In this embodiment, the cloud can be an enterprise private cloud, a factory data center, or a cloud control platform within an industrial private network. The cloud typically includes an industrial knowledge graph, a vector index library, a context engine, a cloud-based large-scale model inference cluster, and a gateway. The industrial knowledge graph stores plant-wide equipment ledgers, equipment operating procedures, maintenance records, fault cases, process constraints, alarm rules, work order information, spatial area information, and safety procedures. The vector index library is used to vectorize and store this structured and unstructured knowledge. The context engine determines the required knowledge scope for the current edge device based on its spatial location and digital work order status. The gateway performs knowledge synchronization, log uploading, and correction command distribution between the cloud and edge devices. The cloud and edge devices can be connected via wireless or satellite communication links.

[0024] In the S100, different location sensing methods can be selected based on the application scenario when acquiring the geospatial coordinates of edge devices. For open-pit mines or field inspection scenarios, geospatial coordinates can be obtained through the Global Navigation Satellite System (GNSS) or BeiDou positioning. For scenarios where satellite signals are unavailable, such as underground mines, the geospatial coordinates of edge devices can be obtained through methods such as UWB positioning. If the edge device is installed on a vehicle, robot, or mobile terminal carried by personnel, the geospatial coordinates represent the current location of that mobile vehicle.

[0025] Geofencing is a pre-defined spatial area on an industrial site, comprising equipment areas, work areas, and hazardous areas. Equipment areas include, for example, high-voltage power distribution rooms, compressor rooms, pump stations, and wellhead areas of drilling platforms. Work areas include safe working areas around equipment, inspection routes, or production line workstations. Hazardous areas include flammable and explosive areas, areas with toxic gases, and confined spaces. If different areas (equipment areas, work areas, and hazardous areas) represent the same area, they are treated as a single unit.

[0026] Digital work order status is provided by Enterprise Resource Planning (ERP) or Manufacturing Execution System (MES). The digital work order status includes work order dispatch status and work order execution status. The work order execution status contains work order semantics such as equipment object, job type, and work area. The equipment object can be the name, number, system, or critical component of the equipment to be repaired. The job type can be inspection, maintenance, switchover, isolation, leak sealing, parameter verification, fault handling, or safety confirmation. The work area can be the actual plant location, equipment room, or workstation where the work order is executed. If the geospatial coordinates cross from one region to another, or the digital work order status changes from work order dispatch status to work order execution status, the context engine is triggered; otherwise, it is not triggered. By triggering the context engine through both location changes and work order changes, rather than simply synchronizing data to edge devices at fixed intervals, synchronization is tied to on-site operational needs, reducing transmissions unrelated to the current task.

[0027] In S200, if the context engine is triggered, clustering is performed in the cloud vector space based on geospatial coordinates and / or the semantics of the work order corresponding to the changed digital work order status. Specifically, when the trigger source is a geofence crossing, the area entered after the crossing is used as the spatial retrieval condition. The spatial retrieval condition may include the area boundary, equipment within the area, area safety procedures, and historical fault records of the area. When the trigger source is a change in the digital work order status, the work order semantics are used as the semantic retrieval condition, i.e., equipment object, job type, and job area. If both geospatial coordinate changes and digital work order status changes occur simultaneously, a joint retrieval is performed using both spatial and semantic retrieval conditions.

[0028] The cloud-based vector space is an index space, or index library, formed by converting knowledge entries, operation manual paragraphs, historical case records, alarm handling procedures, equipment parameter descriptions, process constraint texts, and sensor anomaly explanation texts from the industrial knowledge graph into vectors. Each knowledge entry includes metadata such as version number, security level, confidentiality level, and update time, in addition to vector data. The context engine does not distribute all data; instead, it uses a density clustering algorithm in the cloud-based vector space to delineate candidate knowledge clusters that are similar to the current spatial location and work order semantics. For example, it filters out the vector clusters of relevant operation manuals for the high-voltage substation and historical accident records required for the current maintenance work order. During clustering, it filters in the cloud-based vector space based on spatial retrieval conditions and / or semantic retrieval conditions. For example, if the work order points to "partial discharge detection of switchgear No. 1 in the high-voltage substation," the clustering result is determined based on the density of the sample distribution. Vectors related to switchgear No. 1, partial discharge, insulation detection, and high-voltage operation procedures are retained to form candidate knowledge clusters, which are then identified as hotspot knowledge subsets. The hotspot knowledge subset is not a simple cut-off of all knowledge data in the industrial knowledge graph, but rather a local set of knowledge that is most likely to be invoked by the current edge device in the current region and under the current task. The hotspot knowledge subset may include relevant chapters in the equipment operation manual, historical maintenance records of the equipment involved in the current work order, similar fault cases, alarm handling procedures, safety precautions for the current region, etc.

[0029] To adapt to the relatively weak bandwidth of industrial networks, this embodiment abandons full download and performs underlying protocol synchronization based on Conflict-free Replicated Data Types (CRDTs) and Merkle trees when downloading hot knowledge subsets. Specifically, the cloud divides the hot knowledge subset into multiple semantic buckets. A semantic bucket is the smallest synchronization unit divided according to region, device object, work order type, or knowledge topic. Each semantic bucket includes vector entries, metadata, and hash values. The local micro vector library of the edge device also stores synchronized knowledge according to the same semantic bucket structure. During synchronization, the gateway identifies semantic buckets in the cloud and edge devices where hash values ​​have changed and serializes them. The serialized semantic bucket includes a data header, semantic bucket identifier, vector data, and metadata. After serialization, it is pushed through a lightweight protocol, which allows the edge device's micro vector library to silently and in milliseconds update its surrounding environmental knowledge even in extremely weak network conditions. Incremental synchronization using semantic buckets and Merkle trees, instead of downloading the entire knowledge base on every trigger, allows edge devices to obtain the most relevant knowledge to the current space and work order, even under conditions of limited memory and computing power. Especially in bandwidth-constrained scenarios such as underground, offshore platforms, or satellite links, the cloud only pushes semantic buckets with changing hash values, reducing bandwidth consumption and synchronization waiting time.

[0030] In S300, if an edge device receives an industrial query request, it performs emergency quantification, bandwidth probe assessment, and complexity assessment in parallel. Industrial query requests can be triggered by operators via voice, text, image, or video. For example, an operator can input "Desulfurization tower differential pressure alarm accompanied by severe vibration, please provide handling steps" through an AR headset, or an alarm query "Sudden pressure drop at pump station outlet" can be automatically generated by field sensors. After receiving the industrial query request, the edge device performs emergency quantification, bandwidth probe assessment, and complexity assessment in parallel. Emergency quantification is used to determine the security priority corresponding to the industrial query request. Specifically, the edge device identifies the equipment object, risk words, and operational intent in the industrial query request based on the industrial ontology.

[0031] Industrial ontology can be constructed through industry standard knowledge, field data, and expert verification. This involves extracting basic concepts such as equipment objects, component names, fault phenomena, risk terms, operational actions, and safety consequences from industry standards or expert verification. Then, based on operational constraints between equipment, semantic relationships are established between equipment, components, faults, and handling actions, forming an ontology library (industrial ontology) suitable for industrial sites. For example, "sudden pressure rise," "sudden pressure drop," "leakage," "abnormal noise," "temperature runaway," "electric arc," "tripping," and "toxic gas exceeding limits" can be identified as risk terms; "immediate shutdown," "valve shut-off," "evacuation," "reset," "bypass," and "venting" can be identified as operational intentions; and equipment objects include high-voltage switchgear, desulfurization towers, wellhead blowout preventers, main control pump stations, and gas pipelines.

[0032] After identifying the equipment object, risk words, and operational intent, these are input into an adaptive fuzzy inference system to determine the urgency score. First, historical alarm samples, digital work order samples, accident handling samples, sensor anomaly samples, and manual inspection question-and-answer samples are collected. Each sample contains at least the equipment object, risk words, operational intent, alarm frequency, and handling time limit. Process, safety, and maintenance personnel label these samples with urgency levels, classifying them into low, medium, high, and extremely high urgency levels. For example, ordinary parameter queries are considered low-urgency samples, minor equipment anomalies are considered medium-urgency samples, and sudden pressure increases, leaks, abnormal noises, and temperature runaways are considered high-urgency. Then, based on these labeled samples, the correspondence between each input feature and the urgency level is statistically analyzed to determine the membership interval and rule weights of each fuzzy variable, thereby establishing the adaptive fuzzy inference system. Finally, the urgency score is output, ranging from 0 to 1. The process of establishing the adaptive fuzzy inference system is lengthy and mature, and will not be described in detail here.

[0033] The closer the urgency score is to 1, the more likely the query is to be related to on-site personnel safety, equipment interlock protection, or the escalation of process accidents. The maximum tolerable delay is mapped based on the urgency score, with an inverse relationship between the urgency score and the maximum tolerable delay. For example, when the urgency score is close to 1, the maximum tolerable delay can be set to 300ms to 500ms; when the urgency score is 0.5, the maximum tolerable delay can be set to 1000ms to 3000ms; and when the urgency score is 0.3, a longer cloud analysis time is allowed. The maximum tolerable delay can be configured according to the safety procedures and control response requirements of different industrial sites. For example, in the scenario of a deep-sea oil and gas drilling platform, if an industrial query request involves events such as a sudden increase in wellhead pressure, natural gas leakage, desulfurization tower differential pressure alarm, or accompanying severe vibration, then according to the platform's safety operating procedures for emergency shutdown and personnel evacuation response requirements, the maximum tolerable delay for such requests is configured to 300ms to 500ms. If the industrial query request involves equipment inspection prompts, general valve status queries, or historical maintenance record queries, it does not fall under the category of events that immediately affect personal or equipment safety, and the maximum tolerable delay can be configured to be 1000ms to 3000ms.

[0034] Bandwidth probe assessment is used to determine the current communication conditions between edge devices and the cloud. When the edge device sends a heartbeat packet to the cloud, it captures the round-trip time, packet loss status, and response interval of the heartbeat packet. The heartbeat packet includes the edge device's identifier, current timestamp, current region identifier, queue length, and link status. After the cloud returns a heartbeat response, the edge device calculates the round-trip time and calculates the packet loss rate and response jitter over multiple consecutive heartbeat cycles. Based on the round-trip time, packet loss status, and response interval, the current uplink and downlink bandwidth can be determined, which is the available uplink and downlink bandwidth of the current channel. The theoretical communication delay that would occur if the current data were uploaded to the cloud can also be estimated. Obtaining the current communication status through real-time heartbeat packets avoids making cloud upload decisions based on historical bandwidth during sudden fluctuations in the network, thereby reducing the risk of link avalanche and timeout lock-up.

[0035] In this embodiment, the edge device determines the round-trip delay based on the transmission time and the reception response time, determines the packet loss status based on whether consecutive sequence numbers are missing, and determines the response interval stability based on the fluctuation of the arrival time of adjacent response packets. To estimate the available uplink and downlink bandwidth, the edge device sends probe packets of different lengths over multiple consecutive probe cycles and counts the number of valid bytes confirmed by the cloud per unit time. Furthermore, the estimation does not use single heartbeat results but instead forms a sliding window over the most recent few probe cycles to eliminate abnormal instantaneous peaks. The bandwidth is then reduced based on the packet loss rate. When the packet loss rate increases or the response interval fluctuation becomes larger, a safety margin is added to the available bandwidth to make it closer to the available transmission capacity under weak network conditions. After determining the available uplink and downlink bandwidth, the upload and download times are estimated based on the available uplink and downlink bandwidth, respectively. Then, the theoretical communication delay required to upload the current data to the cloud is obtained based on the round-trip latency. It should be noted that the bandwidth probe evaluation in this application can preferably be implemented using existing technologies in the field. It is not the focus of the improvement claimed in this application. The focus of the improvement in this application is the three-dimensional parallel quantitative evaluation. Those skilled in the art can combine existing public technologies or conventional engineering methods to complete the adaptation, replacement, or equivalent implementation without affecting the implementation of the technical solution of this application.

[0036] Complexity assessment is used to determine whether industrial query requests require cross-device association, cross-process segment association, or multi-hop logical reasoning. Specifically, edge devices identify whether industrial query requests contain multiple device objects, multiple process systems, historical trend analysis, cause inference, batch comparison, interlocking impact assessment, or multi-step plan generation. For example, "querying the current valve's operating steps" is a low-complexity request, while "judging the cause of compressor bearing abnormalities based on the vibration trend of the last three months" and "comparing the coating thickness of the current batch and analyzing the cause" are high-complexity requests. The complexity assessment results can be used to estimate local inference time and cloud inference time. For low-complexity industrial query requests, the edge device's small quantized language model can complete the inference, while for high-complexity macro-analysis requests, the large model at the cloud level has higher inference capabilities.

[0037] After obtaining the emergency quantification results, bandwidth probe evaluation results, and complexity evaluation results, the total cloud latency is determined based on the routing cost function. The routing cost function comprehensively considers theoretical communication latency, cloud queuing latency, cloud inference time, edge local inference time, communication energy consumption, current concurrent queue length, and the data out-of-domain privacy leakage risk penalty weight. The data out-of-domain privacy leakage risk penalty weight indicates whether the industrial query request contains core process formulas, top-secret operating parameters, production batch data, or other data that should not be sent out of the domain. If the request contains the aforementioned sensitive data, even with good network bandwidth and stronger cloud inference capabilities, it is not advisable to send the data to the cloud. It should be noted that the routing cost function in this application can be implemented using existing technologies in the field, and it is not the focus of the improvement claimed in this application. The focus of the improvement in this application is on routing decision-making. Those skilled in the art can combine existing publicly available technologies or conventional engineering methods to complete adaptation, replacement, or equivalent implementations without affecting the implementation of the technical solution of this application.

[0038] During routing decisions, if the uplink and downlink bandwidth is less than the safety value, or the total cloud latency exceeds the maximum tolerable latency, the cloud path for the industrial query request is cut off, and the knowledge base is invoked. The safety value can be determined based on the industrial site network standard, service type, and minimum response requirements. For example, for security alarm requests, a higher safety bandwidth value can be set to ensure that the cloud request does not time out due to insufficient bandwidth; for inspection explanation requests, a lower safety bandwidth value can be set. The total cloud latency is the cumulative result of theoretical communication latency, cloud queuing latency, cloud inference time, and other stages. If the uplink and downlink bandwidth is less than the safety value, or the total cloud latency exceeds the maximum tolerable latency, it means that even if the cloud can provide more complete analysis results, it cannot meet the security response time requirements of the request. In this case, the edge device directly calls the local micro-vector library and quantized small language model to perform low-latency local closed-loop inference. For other situations involving uplink / downlink bandwidth or total cloud latency, industrial query requests are encrypted. The encryption method can be a symmetric encryption algorithm such as AES or SM4. After encryption, the data is sent to the cloud and the task is offloaded to maximize service throughput. In this embodiment, the routing decision is not simply based on computational complexity to select local or cloud, but also considers urgency, network bandwidth, and computational complexity, ensuring the stability of routing decisions in weak network and high-concurrency environments.

[0039] In S400, if an edge device is offline, the data generated during this offline state is encapsulated as an offline cognitive log. Offline states can be determined through methods such as continuous heartbeat timeouts, gateway unreachability, DNS or service registration unreachability, satellite link interruption, or loss of signal from a mine base station. When an edge device enters an offline state, all attempts to upload industrial query requests to the cloud are immediately cut off to prevent requests from being blocked in the transmission queue for extended periods. Simultaneously, the edge device enters a degraded service mode and performs local searches based on the knowledge base. The data generated during the offline state includes interaction data, sensor data, and local decision-making data. Interaction data includes user questions, model responses, manual confirmation information, voice or text input, and interface click records. Sensor data includes field data such as pressure, temperature, and vibration collected during the offline period. Local decision-making data includes local search results and local inference paths.

[0040] When performing local searches based on a knowledge base, edge devices calculate the knowledge confidence score of the search. The knowledge confidence score is determined comprehensively based on search similarity, the number of search results, the degree of work order matching, and the degree of matching in the current region. For example, if the local micro-vector library contains operating procedures that match the current device object, the current fault type, and the current work area, and the knowledge version is relatively new, the knowledge confidence score is relatively high. If only general safety prompts are retrieved locally, without finding specific procedures for the current device, or if the local knowledge version is outdated, the knowledge confidence score is relatively low. If the knowledge confidence score is greater than or equal to the confidence threshold, no conservative disclaimer is applied; if the knowledge confidence score is less than the confidence threshold, a conservative disclaimer is applied based on Prompt Engineering. The conservative disclaimer is used to limit the output degrees of freedom of the quantized small language model in offline mode. Specifically, if the edge device is offline, to prevent the small computing power model at the edge from experiencing illusions, a conservative disclaimer is automatically applied based on Prompt Engineering. In other words, the forced model can only output the operating baseline according to regulations and standards, and will prompt "In offline mode, please refer to the operating standards of a certain device" in the response. By attaching a conservative disclaimer, the output of the small model during network outages is limited to a controllable range, reducing the risk of model illusions caused by lack of knowledge and misleading operators.

[0041] Once generated, offline cognitive logs cannot be tampered with. When the network is restored after an outage, i.e., when the edge device reconnects and regains network connectivity, the edge device asynchronously pushes the offline cognitive logs as incremental data to the cloud based on conflict-free replication data type (CRDT). The cloud's massive computing power model will perform a retrospective evaluation of the local operations during this offline period. If the handling suggestions output by the edge device during the outage only affect the local device, it is determined that there is no cascading risk. If the cloud identifies that a local decision made by the edge device during the outage will cause a cascading risk to the remote device, such as a local leak plugging operation affecting the liquid level of the main control pump station, a local valve adjustment changing the upstream pipeline pressure, or a local shutdown affecting the safety of downstream processes, it is determined that there is a cascading risk. Once the cloud discovers that an operation by the edge device during the outage will cause a cascading risk to the remote device, it immediately generates a correction and compensation instruction and pushes it back to the edge device to request manual intervention for correction, thereby constructing a system-level asynchronous self-healing closed loop.

[0042] In one specific embodiment, the deep-sea oil and gas drilling platform connects to the cloud via a satellite link. This link typically suffers from high latency and low bandwidth, and is prone to packet loss in severe weather. The edge device is an AR headset worn by the operator, which is equipped with a lightweight routing proxy, a local micro-vector library, and a quantized small language model. The cloud pre-establishes geofencing based on the spatial area of ​​the drilling platform, with the desulfurization equipment area, wellhead area, and high-pressure pump area designated as hazardous areas. When the operator enters the desulfurization equipment area, the cloud triggers a context engine based on the geographic coordinates of the AR headset. Combining this with the equipment object "desulfurization tower" and the work type "pressure anomaly handling" in the current maintenance work order, the cloud clusters hotspot knowledge subsets in the cloud vector space, including the desulfurization tower operation manual, differential pressure alarm handling procedures, historical pressure surge cases, and emergency shut-off valve sequences. Subsequently, the cloud uses Merkle tree comparison to find that the edge device lacks the latest differential pressure alarm handling procedures, and only pushes the corresponding semantic bucket incrementally to the AR headset.

[0043] During operation, the platform's sensors detected a sudden pressure surge in the desulfurization tower. The operator, using an AR headset, issued a text command: "Desulfurization tower differential pressure alarm accompanied by severe vibration; please provide immediate handling procedures." The edge device, based on industrial ontology, identified "desulfurization tower" as the equipment object, "differential pressure alarm" and "severe vibration" as risk terms, and "provide immediate handling procedures" as the operational intent. The adaptive fuzzy inference system output an urgency score of 0.95 and mapped a maximum tolerable latency of 300ms. Simultaneously, bandwidth probe assessment revealed severe packet loss on the current satellite link, with an estimated total cloud upload latency exceeding 2000ms. Since the total cloud upload latency exceeded the maximum tolerable latency, the routing agent immediately cut off the cloud upload path and invoked a local micro-vector library and quantized small language model to perform local closed-loop inference. The AR headset displayed the emergency shut-off valve sequence and on-site confirmation items within approximately 200ms. This avoided the long wait times for centralized cloud upload requests in weak network environments, enabling emergency handling to be completed within the safe tolerance time.

[0044] In summary, by determining the geospatial coordinates of edge devices and the status of digital work orders, a context engine can be triggered when a device enters a specific work area or when the work order status changes. This allows the cloud to determine a subset of hot knowledge based on the current spatial location and the semantics of the current work order, avoiding the need to distribute the entire knowledge base to the edge and reducing data synchronization pressure and storage burden on edge devices in weak network environments. Incremental synchronization of semantic buckets based on conflict-free replication data types and Merkle trees reduces redundant transmissions and alleviates communication congestion in high-concurrency scenarios. By performing emergency quantification, bandwidth probe assessment, and complexity assessment in parallel on industrial query requests, considering the security urgency of the request, the current network availability, and the task processing complexity, misjudgments caused by routing based solely on computational complexity or fixed network conditions are avoided. By cutting off the cloud path and calling the local knowledge base for processing, low-latency response can be guaranteed in weak network, network outage, or high-risk task scenarios, reducing security risks caused by cloud timeouts. By encapsulating offline cognitive logs during network outages and pushing them to the cloud for retrospective evaluation and compensation after network recovery, the reliability, security, and adaptability of the industrial edge cloud collaborative system are improved.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A control method for collaborative adaptive routing of industrial edge cloud in low-concurrency networks, characterized in that, include: Obtain the geospatial coordinates of the edge device and the status of the digital work order bound to the edge device. Compare the geospatial coordinates with the geofence and determine whether to trigger the context engine based on the comparison result or the change in the status of the digital work order. If the context engine is triggered, then based on the geospatial coordinates and / or the semantics of the work order corresponding to the changed digital work order status, clustering is performed in the cloud vector space to determine a subset of hot knowledge and synchronize it. During synchronization, semantic buckets with changed hash values ​​are pushed to determine the knowledge base. If the edge device receives an industrial query request, it performs emergency quantization, bandwidth probe evaluation, and complexity evaluation in parallel on the industrial query request to determine the maximum tolerable latency, uplink and downlink bandwidth, and total cloud latency. If the uplink and downlink bandwidth is less than the safe value, or the total cloud latency is greater than the maximum tolerable latency, the cloud path of the industrial query request is cut off and the knowledge base is invoked. Otherwise, the industrial query request is encrypted and sent to the cloud for task offloading. If the edge device is offline, the data generated during the offline state is encapsulated into an offline cognitive log, and a local retrieval is performed based on the knowledge base. If the confidence level of the retrieved knowledge is less than the confidence level threshold, a conservative disclaimer instruction is attached. When the offline state is restored, the offline cognitive log is pushed to the cloud.

2. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 1, characterized in that, When determining whether to trigger the context engine, the following should be included: The geofence includes equipment areas, work areas, and hazardous areas; The digital work order status includes work order dispatch status and work order execution status, and the work order execution status includes work order semantics of equipment object, job type and job area; If the geospatial coordinates cross from one region to another, or the digital work order status changes from the work order dispatch status to the work order execution status, then the context engine is determined to be triggered; otherwise, the context engine is determined not to be triggered.

3. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 2, characterized in that, When clustering the semantics of the work order corresponding to the changed digital work order status in the cloud vector space based on the aforementioned geospatial coordinates and / or, the clustering includes: Based on the geospatial coordinates, the other area crossed is used as a spatial retrieval condition, and the work order semantics are used as a semantic retrieval condition; Based on the density clustering algorithm and using the spatial retrieval conditions and / or semantic retrieval conditions, clustering is performed in the cloud vector space to determine candidate knowledge clusters, and the candidate knowledge clusters are determined as the hot knowledge subset.

4. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 3, characterized in that, Push semantic buckets whose hash values ​​have changed to determine the knowledge base; When distributing the aforementioned hot knowledge subset, synchronization is performed based on a conflict-free replicated data type and a Merkle tree. The gateway serializes the semantic buckets whose hash values ​​have changed and pushes them based on a lightweight protocol to determine the knowledge base.

5. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 4, characterized in that, When performing emergency metricization, bandwidth probe evaluation, and complexity evaluation in parallel on the industrial query request, the following are included: Based on the industrial ontology, the system identifies the equipment objects, risk words, and operational intentions in the industrial query request. The system then inputs these information into an adaptive fuzzy inference system to determine an urgency score. Based on the urgency score, the system maps the maximum tolerable delay, with the urgency score and the maximum tolerable delay being inversely proportional.

6. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 5, characterized in that, When performing emergency metricization, bandwidth probe evaluation, and complexity evaluation in parallel on the industrial query request, the following is also included: The heartbeat packets of the edge device are obtained, and the uplink and downlink bandwidths are determined based on the round-trip time, packet loss status, and response interval of the heartbeat packets. Identify whether the industrial query request contains multi-hop logical reasoning involving cross-device association and cross-process segment association, and determine the total cloud latency based on the routing cost function.

7. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 6, characterized in that, Data generated during network outages includes interactive data, sensor data, and local decision-making.

8. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 7, characterized in that, When performing a local search based on the knowledge base, the following are included: If the confidence level of the retrieved knowledge is greater than or equal to the confidence threshold, then the conservative disclaimer instruction will not be attached. If the confidence level of the retrieved knowledge is less than the confidence level threshold, the conservative disclaimer instruction is attached based on the prompt word project. The conservative disclaimer instruction is used to indicate the network outage status.

9. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 8, characterized in that, When pushing the offline cognitive logs to the cloud, the following are included: When the network outage is restored, the offline cognitive log is asynchronously pushed to the cloud as incremental data based on the conflict-free replication data type, and the cloud performs a retrospective evaluation to identify whether the edge device caused a chain reaction of potential risks to the remote device during the network outage.

10. The control method for collaborative adaptive routing of industrial edge cloud under weak network and high concurrency as described in claim 9, characterized in that, When pushing the offline cognitive logs to the cloud, the following is also included: If the cloud identifies that the edge device causes a chain reaction of potential problems to the remote device when the network is down, it generates a correction and compensation instruction and pushes it back to the edge device to remind manual intervention.

Images