Brake fluid filling device cooperative perception method and system based on swarm intelligence

Abstract

This invention discloses a collaborative perception method and system for brake fluid filling equipment based on swarm intelligence. The method includes: fusing multi-source information and using knowledge graph reasoning to pre-guess a protocol template; lightweight detection and extraction of feature fingerprints by the first device; edge nodes initiating collaborative perception, with each device performing federated learning to optimize the global model and generate an accurate template; storing experience in local contextual memory and cloud-based semantic memory; dynamically instantiating communication frames during process execution; if there is a conflict in the dual-mode memory predictions, a resolution engine driven by physical parameters selects the optimal template and updates the knowledge; and incremental evolution of experience is aggregated in the cloud. This invention breaks down information silos, achieves swarm intelligence, and significantly improves the speed and success rate of protocol adaptation for new vehicle models.

Description

Technical Field

[0001] This invention belongs to the field of automotive intelligent manufacturing technology, specifically relating to a collaborative sensing method and system for brake fluid filling equipment based on swarm intelligence, which enables multiple filling devices to collaboratively detect, adaptively generate, and continuously evolve the communication protocol of unknown vehicle models. Background Technology

[0002] In the automotive assembly line, brake fluid filling is a core process. The filling equipment needs to communicate in real time with the vehicle's braking control system (such as ABS / ESP) via an OBD interface to collaboratively complete operations such as vacuuming, pressure holding, and filling. Existing technologies are mainly divided into three stages: In the hard-coding and parameterized configuration stage: the device pre-stores communication protocol instruction sets for specific vehicle models. Importing new models requires the device manufacturer to redevelop or manually configure the instruction parameters. This method relies on manual input, is time-consuming and costly, and cannot adapt to the rapidly iterating needs of different vehicle models.

[0003] In the single-machine adaptation stage: multi-dimensional protocol feature fingerprints of vehicles are obtained by sending protocol probe sequences, and a pre-trained feature association network is used to deconstruct and generate protocol fingerprint templates. This method achieves autonomous protocol learning for a single device, but it suffers from the problem of information silos: when multiple devices on the same production line face the same unknown vehicle model, each device needs to independently perform a complete detection and learning process, resulting in wasted computing resources and adaptation efficiency bottlenecks; moreover, it cannot leverage the successful experience of other devices to accelerate its own adaptation and lacks collaborative evolution capabilities.

[0004] Summary of existing technological shortcomings: Information silos: Individual learning experiences cannot be shared, leading to duplication of work and delays.

[0005] Weak adaptability to long tails: For rare or modified vehicle models, a single device may fail due to insufficient detection, and there is a lack of a mechanism to aggregate the experience of multiple devices to improve the overall recognition rate.

[0006] Protocol generation is passive: it relies solely on detection responses and does not fully utilize multi-source information from the production line (such as MES data and visual recognition results from preceding workstations) to predict protocols.

[0007] No mechanism to handle conflicting experiences: When multiple sources of experience (such as successful cases in different environments) give conflicting predictions about the same vehicle model, existing technologies cannot intelligently resolve them, leading to decision-making errors.

[0008] Therefore, how to achieve collaborative perception and experience sharing among multiple refueling devices, leverage swarm intelligence to enhance the ability to quickly adapt to unknown protocols, and intelligently handle conflicts of multi-source experience has become a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0009] This invention aims to overcome the shortcomings of existing technologies and provide a collaborative perception method and system for brake fluid filling equipment based on swarm intelligence. By constructing an end-edge-cloud collaborative swarm intelligence architecture, introducing a dual-mode memory network and a conflict resolution mechanism driven by physical parameters, it achieves multi-device collaborative detection, federated learning experience sharing, and continuous evolution of protocol knowledge graph, thereby significantly improving the speed and success rate of new vehicle model adaptation and intelligently handling multi-source experience conflicts.

[0010] In a first aspect, this application provides a collaborative sensing method for brake fluid filling equipment based on swarm intelligence, comprising: S1. Multi-source information fusion and protocol pre-conjecture: Edge nodes acquire multi-source information of the target vehicle, input it into the protocol knowledge graph reasoning, and generate candidate protocol initial templates; By adopting the above technical solution, the core principle of this step lies in leveraging the associative reasoning capabilities of knowledge graphs to pre-guess protocols, rather than passively waiting for detection as in traditional solutions. Specifically, the protocol knowledge graph maps discrete multi-source information (VIN codes, part numbers, suppliers, etc.) to nodes in a graph structure. Through the link prediction algorithm of graph neural networks (GNNs), it infers the protocol features that unknown vehicle models may adopt from known vehicle model-protocol associations. The technical advantages are: before physical connections are established, the system already possesses prior knowledge of the new vehicle model's protocol, significantly reducing the search space for subsequent detection; simultaneously, by integrating preceding visual detection results, it can detect anomalies such as inconsistencies between part numbers and MES information in advance, providing a high-quality initial template for subsequent collaborative perception.

[0011] S2. Single-machine detection and feature extraction: The first refueling device sends a protocol detection sequence based on the candidate template, captures the response signal and extracts multi-dimensional protocol feature fingerprints, and uploads them to the edge node; By adopting the above technical solution: This step employs a lightweight detection strategy, that is, targeted detection sequences are sent based on the candidate templates generated by S1, rather than traversing all possible protocol combinations. The principle is that the candidate templates provide a general framework for the protocol (such as the diagnostic session ID range and service ID format), and the detection sequences only need to be fine-tuned and verified within this framework. The extraction of multi-dimensional protocol feature fingerprints covers multiple dimensions such as timing, padding bits, and content entropy. The technical benefits are: the number of detections is reduced by more than 60%, avoiding unnecessary interference with the vehicle control system; at the same time, multi-dimensional features provide rich discriminative basis for subsequent swarm intelligence collaboration, enabling different devices to perceive the protocol characteristics of the same vehicle from different perspectives.

[0012] S3. Group Collaborative Perception and Federated Learning: When the edge node determines that the vehicle model appears for the first time, it broadcasts the collaborative perception task to other devices; each participating device uses local historical data to calculate the model gradient and uploads it; the edge node aggregates the gradient to optimize the global protocol generation model and generates an accurate protocol fingerprint template. By adopting the above technical solution, the core innovation of this step lies in introducing federated learning into the industrial communication protocol adaptation scenario. The principle is as follows: each device's locally stored historical success / failure cases constitute a distributed training dataset. However, for data privacy and bandwidth considerations, the original data is not uploaded; instead, model gradients (i.e., how to adjust model parameters to adapt to the direction of current features) are uploaded. Edge nodes act as federated aggregators, using a federated averaging algorithm (FedAvg) to fuse the experiences of each device, generating a global model that integrates the experience of multiple devices. The technical effects are: Firstly, it achieves workshop-level collective memory. The protocol learned by the first device is indirectly transmitted to other devices through gradient sharing, enabling zero-latency adaptation for subsequent vehicles entering any workstation. Simultaneously, the federated learning mechanism naturally protects the data privacy of each device, preventing the leakage of core process data.

[0013] S4. Dual-mode memory storage: The experience tuple of this interaction is stored simultaneously in the local context memory network of the device and the semantic memory network of the cloud evolution center; By adopting the above technical solution: this step mimics the memory mechanism of the biological brain, constructing a complementary dual-mode memory architecture. The episodic memory network (local on the device) uses a fast write and retrieval mechanism to store original experience details for handling recent or rare situations; the semantic memory network (cloud-based) uses a slow update and abstract generalization mechanism to extract general rules from a large amount of experience for handling routine situations. Its technical effect is that it resolves the contradiction between a single model's inability to simultaneously achieve rapid adaptation to new situations and long-term stable memory. Episodic memory ensures the system can remember rare protocols learned yesterday, while semantic memory ensures the system will not waver in its core rules due to individual abnormal data; the two work together to give the system human-like learning capabilities.

[0014] S5. Protocol Instantiation and Process Execution: The refueling equipment calls the precise protocol fingerprint template at each process node, combines it with real-time physical parameters to dynamically instantiate and generate communication fingerprint frames, interact with the vehicle and control the physical actuators; By adopting the above technical solution, this step achieves dynamic binding between process logic and communication protocol. The principle is as follows: the protocol fingerprint template uses an abstract form with wildcards, and wildcard positions are reserved for filling real-time process parameters (such as vacuum level and filling volume). During runtime, these parameters are dynamically calculated and filled based on the current process node and sensor data, while checksums are calculated in real-time according to protocol requirements. The technical benefits are: it enables real-time correlation between communication content and process status; for example, the filling volume parameter is directly encoded into the communication frame, ensuring complete synchronization between the instructions received by the vehicle and physical execution; simultaneously, the templated design allows the same set of protocol logic to adapt to different process parameter changes without needing to generate independent instructions for each parameter combination.

[0015] S6. Memory Conflict Perception and Resolution: If the prediction results of contextual memory and semantic memory conflict during subsequent calls, the conflict resolution engine is triggered. It extracts the current real-time physical parameters and the physical parameters in the historical experience tuple for correlation analysis, selects the best matching result as the final protocol template, and updates the semantic memory network and knowledge graph as new knowledge associations. By adopting the above technical solution, this step is one of the core inventive features of this invention, solving the multi-source experience conflict problem that existing technologies cannot handle. Its principle is based on an important discovery: the same vehicle model may use different communication protocols under different environmental conditions (e.g., different verification algorithms are used due to battery voltage changes in low-temperature environments). The conflict resolution engine introduces physical world parameters (temperature, voltage, etc.) into the decision-making process of the information world, determining which historical experience the current environment is closer to through similarity calculations, thereby selecting the most suitable protocol template. Its technical effects are: improved protocol recognition accuracy, especially for special protocol variants in extreme environments (such as high cold or high temperature), giving the system an intelligent decision-making capability of "environmental perception"; simultaneously, the resolution results are stored in the knowledge graph in the form of part number + physical parameter threshold = protocol feature, enabling the system to continuously accumulate environment-protocol association rules, forming a physically parameter-sensitive protocol knowledge base.

[0016] S7. Experience Replay and Evolution: The cloud-based evolution center gathers experience tuples, updates the protocol knowledge graph and protocol generation model, and distributes them to edge nodes and devices.

[0017] By adopting the above technical solution, this step enables the system to continuously evolve. The principle is to use incremental learning algorithms (such as Elastic Weight Consolidation (EWC)) to learn new experiences while preserving old knowledge. A cloud-based evolution center aggregates interactive experiences from factories worldwide, and uses an experience replay buffer to sample recent important cases and long-term general rules for joint training. The technical effects are: the system becomes smarter with use; the knowledge base is enriched with each new vehicle processed; and the cross-regional experience aggregation allows the system to discover regional protocol patterns (such as protocol variations due to differences in fuel quality in a particular region), forming a globally unified protocol knowledge graph that any factory can directly access in similar situations in the future.

[0018] Furthermore, the multi-source information mentioned in step S1 includes the VIN code and optional package configuration information obtained from the manufacturing execution system, as well as the brake system part number image recognition result obtained from the preceding visual inspection station.

[0019] By adopting the above technical solution, this step achieves dual verification of vehicle identity by integrating MES system data with visually perceived physical data. The principle is as follows: VIN code and optional package information provide the vehicle's design configuration, while part number image recognition provides the actual assembled hardware information. Combining the two can detect anomalies in design-assembly inconsistencies (such as temporarily replacing controllers from different suppliers due to missing parts). The technical benefits are: early warning of potential protocol incompatibility risks, avoiding communication failures caused by hardware changes; simultaneously, the visually recognized part numbers can be directly used as key retrieval items in the knowledge graph, improving the accuracy of protocol prediction.

[0020] Furthermore, the protocol knowledge graph mentioned in step S1 is stored in the form of a graph database. Nodes include vehicle model, part number, supplier, protocol features, instruction template and physical parameters, and edges represent the relationships between entities. Inference uses a link prediction algorithm based on graph neural networks.

[0021] By adopting the above technical solution, this step constructs a structured protocol knowledge representation system. The principle is to transform the originally isolated protocol data into a graph structure. Vehicle models, part numbers, and protocol features are no longer independent table records, but rather interconnected nodes, with relationships represented by edges. The graph neural network performs inference by aggregating neighbor node information. For example, given part number A, it can infer its possible associated supplier B and protocol feature C. The technical effect is that it enables the system to have the ability to extrapolate from specific examples: even if a part number has never been seen before, as long as its supplier, vehicle model, and other information are associated with known nodes, a reasonable protocol template can be inferred; physical parameters are introduced as nodes, providing a structured knowledge foundation for subsequent conflict resolution.

[0022] Furthermore, the protocol generation model described in step S3 adopts a variational autoencoder architecture, where the encoder maps the multidimensional protocol feature fingerprint into a latent space vector, and the decoder reconstructs the protocol fingerprint template; in federated learning, each device only uploads the gradient of the decoder parameters, and the edge nodes use a federated averaging algorithm to update the global model.

[0023] By adopting the above technical solution, this step uses a generative model instead of a discriminative model to generate protocol templates. The principle is that the VAE encodes feature fingerprints as probability distributions in the latent space, and the decoder samples from this distribution to reconstruct the complete protocol template. The unique advantage of this generative architecture is that the latent space is continuous, and similar features are close together in the latent space. Therefore, for unseen features, reasonable templates can be generated through interpolation. Only the decoder gradients are exchanged during federated learning because the encoder is responsible for processing local private data features, which should not be shared. Its technical effects are: the model is creative and can generate never-before-seen protocol templates, rather than being limited to matching known patterns; the federated learning mechanism ensures that the core data assets of each device (such as complete interaction logs for a specific vehicle model) never leave the local machine.

[0024] Furthermore, the real-time physical parameters mentioned in step S6 include one or more of the following: ambient temperature, target vacuum level, pressure holding time, filling amount, and vehicle battery voltage; the correlation analysis uses similarity calculation and introduces a time decay factor; the new knowledge association is stored in the semantic memory network in the form of part number + physical parameter threshold = protocol feature.

[0025] By adopting the above technical solution, this step establishes a new paradigm for knowledge representation that links physical and information domains. The principle is as follows: similarity calculation (such as cosine similarity) quantifies the degree of matching between the current environment and historical experience; a time decay factor ensures that recent experience has higher weight, conforming to the objective law that similar recent working conditions in industrial production are more valuable for reference. The resolution results are stored in the knowledge graph in the form of rules, transforming one-time conflict resolution into reusable explicit knowledge. Its technical effects are: the system not only resolves current conflicts but also provides direct decision-making basis for similar situations in the future, forming a closed loop of practice-summary-application; the automatic mining of physical parameter thresholds enables the system to discover patterns that are difficult for humans to perceive (such as a protocol automatically switching communication modes when the voltage is below 11.5V).

[0026] A second aspect of this application provides a collaborative sensing system for brake fluid filling equipment based on swarm intelligence, used to implement the above-mentioned method, including: Multiple dispensing devices, each equipped with an OBD communication module, a local processor and storage unit, and a locally deployed scene memory network; Edge computing nodes communicate with each refueling device and are equipped with a conflict resolution engine. The cloud-based evolution center communicates with edge nodes and is equipped with a semantic memory network. This network aggregates experience tuples, updates the protocol knowledge graph and protocol generation model, and distributes the updated model to edge nodes and devices.

[0027] By adopting the above technical solution, this system architecture employs a three-tiered collaborative system of end-edge-cloud to optimize the allocation of computing resources. The principle is as follows: the end device is responsible for tasks with high real-time requirements (probe, execution) and rapid local memorization (contextual memory); edge nodes are responsible for regional collaborative tasks (federated aggregation, conflict resolution), achieving workshop-level intelligent collaboration; and the cloud is responsible for global tasks (knowledge accumulation, model evolution), aggregating the wisdom of multiple factories. The technical effect is that it forms an organic whole of rapid end-side response, intelligent edge-side collaboration, and continuous cloud-side evolution, ensuring both the real-time requirements of production rhythm and enabling cross-temporal and spatial knowledge transfer; the three-tiered architecture inherently possesses scalability, allowing new devices to share collective intelligence simply by connecting to the network.

[0028] Furthermore, the refueling device is locally deployed with a lightweight federated learning agent, which is used to calculate model gradient update information based on local historical data and upload the encrypted gradients; the gradient update information includes metadata tags associated with physical parameters.

[0029] By adopting the above technical solutions, this step ensures the security and interpretability of federated learning. The principle is as follows: gradient encryption uses homomorphic encryption or differential privacy technology, so even if the gradient is intercepted during transmission, the original data cannot be deduced; metadata tags (such as associating gradients with a 25°C environment) provide contextual information for the aggregation of edge nodes, enabling the federated averaging algorithm to perform weighted aggregation (e.g., assigning higher weights to experiences with similar environments). The technical effects are: it achieves a secure sharing model where data is usable but not visible, alleviating concerns about OEMs sharing core process data; metadata tags make gradient aggregation more intelligent, avoiding the mixing of experiences from completely unrelated scenarios.

[0030] Furthermore, the protocol knowledge graph is deployed in a cloud evolution center and periodically synchronized to edge nodes to enable fast local reasoning; the knowledge graph contains physical parameter nodes and their associated edges with protocol feature nodes.

[0031] By adopting the above technical solution, this step achieves hierarchical deployment and synchronization of knowledge. The principle is as follows: the cloud maintains a complete global knowledge graph, while edge nodes cache subgraphs related to local production (such as nodes related to commonly produced models in this workshop), enabling local processing of most inference requests; the introduction of physical parameter nodes allows the knowledge graph to support "environment-aware" queries (such as querying the protocol characteristics of part number X when the temperature is below 10℃). The technical benefits are: a balance between inference speed and knowledge integrity; edge nodes can complete protocol pre-guessing in milliseconds without needing to access the cloud for every query; and physical parameter-related edges give the knowledge graph spatiotemporal awareness, supporting complex environment-related queries.

[0032] Furthermore, the cloud-based evolution center adopts a distributed storage and computing architecture, supports multi-factory data aggregation and cross-regional federated learning, and performs pattern mining on the conflict cases reported by each factory through a global conflict resolution rule base to form cross-regional general protocol rules.

[0033] By adopting the above technical solution, this step achieves cross-regional swarm intelligence evolution. The principle is as follows: the distributed architecture supports the storage and processing of petabyte-scale experience data; cross-regional federated learning allows factories to collaboratively optimize the global model without sharing raw data; the global conflict resolution rule base discovers universal patterns from a large number of conflict cases through association rule mining (such as the FP-Growth algorithm). The technical effect is that the discovery of universal protocol rules may even indirectly assist automakers in optimizing their own protocol designs, forming standardized communication specifications.

[0034] Furthermore, the conflict resolution engine includes: The conflict detection unit monitors the predictive differences between episodic memory and semantic memory. The physical parameter extraction unit acquires the real-time physical parameters of the current process. The similarity calculation unit calculates the similarity between the current physical parameters and the physical parameters in the historical experience tuples; The decision output unit selects the optimal prediction result based on similarity ranking; The knowledge update unit transforms the resolution results into new knowledge associations and feeds them back into the memory network.

[0035] By adopting the above technical solution, this step constructs a complete conflict resolution closed loop. The principle of collaborative work among the units is as follows: the conflict detection unit continuously monitors the outputs of the two memory networks, triggering the resolution process upon detecting differences; the physical parameter extraction unit reads data from the PLC or sensors in real time; the similarity calculation unit uses a multi-dimensional vector matching algorithm (such as normalizing and weighting parameters of different dimensions, such as temperature and voltage); the decision output unit not only outputs the results but also records the decision basis (such as selecting semantic memory due to a temperature matching degree of 0.95); and the knowledge update unit transforms this resolution into new triplet knowledge. Its technical effects are: making the conflict resolution process transparent and traceable, providing analytical samples for subsequent model optimization; and the closed-loop feedback mechanism ensures that each conflict resolution enriches the knowledge base, making the system's conflict handling capability stronger with use.

[0036] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a collaborative swarm intelligence architecture involving the edge, cloud, and device, and introduces a dual-mode memory network and a conflict resolution mechanism driven by physical parameters. This enables the brake fluid filling device to collaboratively perceive, adaptively generate, and continuously evolve its communication protocol for unknown vehicle models. Compared with existing technologies, this invention offers the following significant advantages: I. Breaking down information silos to achieve workshop-level collective intelligence In existing technologies, each refueling device independently performs protocol detection and learning. When multiple devices on the same production line encounter the same unknown vehicle model, they need to repeat the entire adaptation process, resulting in wasted computing resources and efficiency bottlenecks. This invention, through a federated learning mechanism, allows each device to share learning results by uploading only model gradients (rather than raw data). After the first device successfully adapts to a new vehicle model, its learning experience is aggregated as gradients through edge nodes, instantly transforming into the optimization direction of the global model. Other devices can directly call the optimized and accurate protocol template without repeated detection. Its beneficial effects are: it achieves workshop-level learning once, benefiting all personnel; subsequent vehicles can achieve zero-latency adaptation at any workstation, shortening the adaptation cycle from weeks in traditional technology to minutes, thus improving production efficiency; simultaneously, federated learning only exchanges gradients, not raw data, effectively protecting the security of vehicle model protocol data from various automakers and eliminating concerns about data sharing.

[0037] II. Intelligent handling of multi-source experience conflicts significantly improves protocol recognition robustness. Existing technologies cannot handle conflicts arising from multiple sources of experience: when experiences accumulated from different devices or at different times provide contradictory predictions for the same vehicle model, the system cannot make intelligent decisions, often leading to communication failures. This invention innovatively introduces a dual-mode memory network (contextual memory + semantic memory) and a conflict resolution engine driven by physical parameters. Its beneficial effects are: for the first time, it introduces environmental parameters from the physical world into protocol decision-making in the information world, enabling the system to possess intelligent judgment capabilities based on environmental perception. For example, when contextual memory (recent experience) and semantic memory (general rules) conflict in the selection of verification algorithms, the conflict resolution engine analyzes the matching degree between current environmental parameters such as temperature and battery voltage and historical experience, selecting the protocol template most suitable for the current environment. This mechanism improves the accuracy of protocol recognition, especially for special protocol variants in extreme environments (such as -30℃ in frigid regions and 50℃ in hot regions). The system can accurately identify and adapt to these variants, completely solving the problems of amnesia and misjudgment in traditional technologies when the environment changes.

[0038] III. Dual-mode memory architecture, balancing rapid adaptation and long-term stability. Existing single-model architectures cannot simultaneously meet the contradictory needs of quickly memorizing new information and retaining old knowledge over a long period. This invention draws inspiration from the memory mechanisms of the biological brain, constructing a complementary dual-mode memory network: an episodic memory network (local on the device) employs a high-speed write and fast retrieval mechanism, specifically storing recent (e.g., within the past 24 hours) or rare (e.g., specially modified vehicles) experience details, supporting millisecond-level similarity retrieval; and a semantic memory network (cloud-based) employs a slow update and abstract generalization mechanism, extracting general protocol rules from hundreds of thousands of experience tuples to form a stable knowledge core. Its beneficial effects are: the system can quickly adapt to newly emerging rare vehicle models (e.g., test vehicles, modified vehicles) without being shaken by individual abnormal data points that undermine the core rules accumulated over a long period; episodic memory ensures that protocols learned yesterday are remembered, while semantic memory ensures that experiences accumulated three years ago are not forgotten; the synergy of these two aspects enables the system to possess human-like progressive learning capabilities.

[0039] IV. Protocol prediction and lightweight detection significantly reduce communication overhead. Existing technologies rely on manual configuration or traversal probing, resulting in numerous probes, long processing times, and unnecessary interference with the vehicle control system. This invention achieves protocol pre-guessing through multi-source information fusion and knowledge graph reasoning—before the physical connection is established, the system infers high-confidence candidate protocol templates based on prior information such as VIN codes, optional packages, and visual recognition results. Its beneficial effects include: reducing the probe search space by more than 80%, decreasing the number of probes from hundreds in traditional technologies to less than 20, and shortening the probe time from minutes to seconds; the lightweight probe strategy significantly reduces invalid access to the vehicle's braking control system, avoiding ECU crashes or fault code storage caused by excessive probing, thus improving production cycle time and equipment lifespan; simultaneously, the fusion of visual recognition results can detect design-assembly inconsistencies in advance (such as temporarily replacing controllers from different suppliers due to missing parts), enabling early risk warning.

[0040] V. Dynamic instantiation and physical parameter association enable deep process-communication collaboration. Existing technologies rely on relatively independent communication commands and process parameters, making dynamic adjustments impossible based on real-time operating conditions. This invention employs an abstract fingerprint frame template with wildcards, dynamically instantiating complete communication frames at each process node in conjunction with real-time physical parameters (such as target vacuum level, holding time, and filler volume), and calculating checksums in real time. Its advantages include: achieving deep coupling between process logic and communication protocol, ensuring complete synchronization between communication content and physical execution; for example, filler volume parameters are directly encoded into the communication frame, allowing the vehicle to confirm the physical execution target upon receiving the command, avoiding execution deviations caused by the separation of process parameters and communication commands in traditional solutions; dynamic checksum calculation ensures that each frame of data conforms to the current vehicle's checksum specifications, improving communication success rate; and the templated design allows the same protocol logic to adapt to a wide variety of process parameter combinations, eliminating the need to generate independent commands for each parameter and reducing protocol maintenance workload.

[0041] VI. Continuous evolution and cross-regional swarm intelligence: the system becomes smarter with use. Once deployed, existing technologies have largely fixed protocol recognition capabilities, unable to continuously learn from subsequent interactions. This invention, through an incremental learning mechanism (Elastic Weight Consolidation (EWC)) in a cloud-based evolutionary center and an experience replay buffer, enables the system to possess lifelong learning capabilities—each successful or failed interaction is transformed into an experience tuple, used to periodically update the knowledge graph and generative model. Its beneficial effects include: the system exhibits an evolutionary characteristic of becoming smarter with use; for every doubling of the number of vehicles processed, the protocol recognition accuracy improves by 2-3 percentage points; the cross-regional federated learning architecture aggregates the interaction experience of factories worldwide, instantly turning rare protocol variants encountered by one factory into public assets of the global knowledge base, which any factory can directly access in similar situations in the future; and the global conflict resolution rule base automatically discovers universal patterns across regions through association rule mining, which can even reverse-engineer automakers to optimize protocol design and promote industry standardization.

[0042] VII. Three-level collaboration between endpoint, edge, and cloud for optimized resource allocation and elastic scaling The three-tiered system architecture of this invention—edge-cloud—optimizes the allocation of computing resources: the receiving device (edge) handles real-time detection, execution, and local memory; edge nodes (edges) handle regional collaborative tasks (federated aggregation, conflict resolution) to achieve workshop-level intelligence; and the cloud-based evolution center (cloud) handles global knowledge accumulation and model evolution. Its advantages include: forming an organic whole with millisecond-level response on the edge, intelligent collaboration on the edge, and continuous evolution on the cloud, ensuring both the stringent real-time requirements of production cycle time (detection response <50ms) and enabling cross-temporal and spatial knowledge transfer; the three-tiered architecture inherently possesses elastic scalability, allowing new production lines or factories to share global collective intelligence simply by deploying edge nodes and connecting to the cloud, reducing system deployment costs by 60% and shortening the expansion cycle from several months to one week.

[0043] VIII. Data Privacy Protection and Compliance Design This invention prioritizes data privacy protection while achieving swarm intelligence: federated learning only exchanges model gradients, not the original data, and gradient encryption uses differential privacy technology to ensure that the original information cannot be reversed; local data of each device (such as complete interaction logs of a specific vehicle model) never leaves the local machine, and only anonymized metadata tags are uploaded. Its beneficial effects are: it completely eliminates the security concerns of automakers regarding the sharing of core process data, making cross-enterprise and cross-regional collaborative learning possible; the system design complies with GDPR and other data privacy regulations, providing a feasible paradigm for compliant data sharing in the field of intelligent manufacturing.

[0044] In summary, this invention, through the organic combination of a series of innovative mechanisms such as swarm intelligence architecture, dual-mode memory, conflict resolution, and dynamic instantiation, systematically solves the core problems of existing technologies, such as information silos, experience conflicts, poor environmental adaptability, and inability to continuously evolve. It achieves zero-latency adaptation of brake fluid filling equipment to unknown vehicle protocols, environmental perception decision-making, and lifelong continuous evolution. It has made breakthrough progress in adaptation speed, recognition accuracy, resource efficiency, and data security, and has significant industrial practical value and broad application prospects. Attached Figure Description

[0045] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is a diagram of the collaborative perception system architecture for brake fluid filling equipment based on swarm intelligence, as described in an embodiment of the present invention. It shows the three-level architecture of end-edge-cloud and its main modules.

[0047] Figure 2 The flowchart below shows the overall steps of the method of the present invention, illustrating the complete process from S1 to S7.

[0048] Figure 3 This is a schematic diagram of the pre-conjecture of the fusion of protocol knowledge graph and multi-source information, showing how MES information and visual recognition information are input into the knowledge graph and inferred to generate candidate templates.

[0049] Figure 4 This diagram illustrates the working principle of the dual-mode memory network and the conflict resolution engine, showing the complementary relationship between episodic memory and semantic memory, as well as the internal units and processing flow of the conflict resolution engine. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] I. System Architecture like Figure 1 As shown, the collaborative sensing system for brake fluid filling equipment based on swarm intelligence of the present invention adopts a three-level collaborative architecture of end-edge-cloud, specifically including: Multiple dispensing devices (endpoints): Taking devices A, B, C, and D as examples, each dispensing device is equipped with an OBD communication module, a local processor, and a storage unit for performing protocol detection, feature extraction, protocol instantiation, and process control. Each device locally deploys a context memory network (using a lightweight encoder based on Transformer) and a lightweight federated learning agent. The context memory network is responsible for quickly storing the original details of recent or rare experiences, supporting content-based similarity retrieval; the federated learning agent calculates the gradient of the protocol generation model based on local historical data, and uploads the gradient to the edge node after homomorphic encryption.

[0052] Edge computing nodes (edges): Connected to each refueling device via the workshop LAN, they are equipped with a conflict resolution engine, a local copy of the protocol knowledge graph (which is periodically synchronized from the cloud), and federated aggregation services. Edge nodes are responsible for multi-source information fusion, protocol pre-guessing, collaborative task scheduling, federated learning aggregation, and real-time processing of prediction conflicts between contextual and semantic memories.

[0053] The Cloud Evolution Center (Cloud) connects to various edge nodes via a wide area network and deploys a semantic memory network (using a graph neural network GNN architecture), a global protocol knowledge graph, an incremental learning engine, and a model distribution service. The Cloud Evolution Center aggregates experience tuples uploaded by all edge nodes, updates the global protocol generation model and knowledge graph through incremental learning algorithms (such as Elastic Weight Consolidation (EWC), and distributes the updated model to various edge nodes and devices.

[0054] II. Method Examples This embodiment takes a car assembly workshop as an example. The workshop has two parallel filling lines on the left and right, with a total of 4 filling devices (devices A, B, C, and D). A brand new test vehicle (VIN code LFVXXXXXXXXXXXXXX) enters the filling station and needs to communicate with the filling device to complete the brake fluid filling process.

[0055] Step S1: Multi-source information fusion and protocol prediction Before the target vehicle enters the filling station, the edge node obtains the vehicle's VIN code and optional package configuration information from the Manufacturing Execution System (MES), learning that the vehicle is equipped with a new generation integrated braking system, IPB-R. Simultaneously, a high-definition camera at the preceding visual inspection station captures the part number label on the vehicle's brake control unit, which is identified by OCR as IPB-R_2025. The edge node inputs this multi-source information (VIN code segment, optional configuration, part number) into a pre-built protocol knowledge graph for inference.

[0056] The protocol knowledge graph is stored in the Neo4j graph database format. Node types include: vehicle model nodes (attributes: VIN prefix, brand, year), part number nodes (attributes: part number string, supplier, hardware version), protocol feature nodes (attributes: diagnostic session ID, service ID, sub-function code, verification algorithm, instruction template), and physical parameter nodes (attributes: parameter type, threshold range). Edge types include: vehicle model and part number, part number and protocol feature, protocol feature and physical parameter, etc.

[0057] The inference process employs a link prediction algorithm based on graph neural networks, such as a relational graph convolutional network. Given part number and vehicle model information, the graph neural network predicts missing protocol feature nodes by aggregating the features of neighboring nodes. Ultimately, two initial candidate protocol templates and their confidence levels are generated: Template A: 85% confidence, Template B: 65% confidence. Based on protocol feature corrections, the edge nodes distribute these two candidate templates to device A.

[0058] Step S2: Single-machine detection and feature extraction Device A generates a set of lightweight protocol probe sequences based on template A (85%) with the highest confidence level. This sequence consists of different diagnostic session identifiers (such as 0x01 default session, 0x03 extended session), service identifiers (such as 0x10 diagnostic session control, 0x22 read data identifier) ​​and sub-function codes, and is sent to the vehicle braking control system one by one through the OBD interface according to a preset timing sequence (10ms interval).

[0059] Device A captures the raw timing response signal returned by the vehicle for each detection sequence, including the response timestamp, response data frame length, data link layer padding bit sequence, and application layer data content. After feature engineering, a multi-dimensional protocol feature fingerprint is extracted, including: Timing response characteristics: response delay sequence [5ms, 12ms, 8ms, ...]; Padding bit encoding characteristics: A fixed-length vector (e.g., 32 bits) formed by hashing the data link layer padding bit sequence. Data content entropy characteristics: Entropy values ​​and byte distribution statistics of application layer data; Session interaction pattern characteristics: number of interactions and timeouts in different diagnostic sessions.

[0060] Device A uploads the aforementioned multidimensional protocol feature fingerprint and vehicle unique identifier (VIN segment) to the edge node.

[0061] Step S3: Group Collaborative Perception and Federated Learning The edge node determines that the vehicle model is appearing for the first time based on the vehicle's VIN segment (after querying the local database, no matching record is found), and immediately broadcasts a collaborative perception task to devices B, C, and D, requesting each device to use the historical interaction data stored locally to perform similarity matching and protocol verification suggestions on the received feature fingerprint, and to associate it with historical physical parameter environmental labels (such as temperature, voltage, etc.).

[0062] Each participating device calculates gradient update information for the protocol generation model based on local historical data through a locally deployed lightweight federated learning agent. In this embodiment, the protocol generation model adopts a variational autoencoder (VAE) architecture. The encoder maps the multidimensional protocol feature fingerprint (128-dimensional) to a latent space vector (32-dimensional), and the decoder reconstructs the protocol fingerprint template (variable-length instruction sequence) from the latent vector. Each device only uploads the gradient of the decoder parameters, homomorphically encrypts the gradient, and attaches metadata tags (thus associating the gradient with a 25°C environment).

[0063] After collecting gradients uploaded by each device, the edge node aggregates them using a federated averaging algorithm, updates the global protocol generation model, and generates a precise protocol fingerprint template that integrates the intelligence of multiple devices. This template contains abstract fingerprint frames for each process node, as well as corresponding metadata such as timing, retry logic, and verification parameters. The edge node then distributes the precise template to device A.

[0064] In this embodiment, the specific implementation process of federated learning is as follows: Suppose there are K refueling devices participating in the collaborative sensing task, denoted as device 1, 2, ..., K. Each device k locally stores a historical interaction dataset. It contains several experience tuples representing successes and failures. The global protocol generation model adopts a variational autoencoder (VAE) architecture, and its parameters are denoted as... The goal of the model is to reconstruct the protocol fingerprint template by maximizing the log-likelihood estimate.

[0065] (a) Local gradient calculation Edge nodes first set the current global model parameters Broadcast to all participating devices. Each device In local dataset Calculate the loss function The loss function consists of three parts: reconstruction loss. , Divergence loss And auxiliary loss used for protocol verification : in: Reconstruction loss Cross-entropy loss is used to measure the difference between the protocol fingerprint frame reconstructed by the decoder and the real frame; Divergence loss The distance between the latent space distribution of the encoder output and the standard normal distribution is used to constrain the distance between them. Auxiliary loss This is an optional verification matching loss used to improve the prediction accuracy of protocol verification parameters; and This is a hyperparameter used to balance various losses.

[0066] Device k performs gradient descent in local training rounds (E is the number of local training rounds) to calculate the gradients of the model parameters. To protect data privacy, the device gradient Perform homomorphic encryption or differential privacy processing to obtain the encryption gradient. and attach metadata tags (e.g., the ambient temperature range of the current equipment, vehicle model distribution, etc.), forming tuples ( , Upload to the edge node.

[0067] (ii) Gradient aggregation of edge nodes Edge nodes collect encryption gradients uploaded by all participating devices. and its metadata tags The aggregation process uses a weighted federated average algorithm, with weights determined by the amount of data from each device. And the similarity of metadata is dynamically adjusted.

[0068] Aggregated global gradient The calculation formula is as follows: Among them, weight It consists of two parts: In the formula: The total amount of data from all participating devices; For equipment The similarity between the metadata tags of the current query task and the metadata of the current query task (generated by the edge node based on the current vehicle features) is calculated using cosine similarity. This is a balancing coefficient used to control the relative importance of data volume weight and metadata similarity weight.

[0069] After receiving gradients from all devices, the edge nodes perform aggregate calculations to obtain the global gradient update. And update the global model parameters accordingly: in, The global learning rate, This is an index for the current federated learning round.

[0070] (III) Node Processing Flow 1. Task Broadcasting and Device Filtering: Edge nodes generate query metadata based on current vehicle characteristics (VIN segment, part number). The system broadcasts a collaborative sensing task to all refueling equipment in the workshop. Upon receiving the task, each device first determines its local data volume... Similarity with metadata Determine whether to participate in this round of aggregation: If the similarity is lower than a preset threshold (e.g., 0.3), the device can selectively exit to avoid introducing irrelevant experience that may affect the model's accuracy.

[0071] 2. Local training and gradient upload: Participating devices execute the training locally. Round of training, calculating gradients To reduce communication overhead, the device employs gradient compression techniques (such as Top-K sparsity) during training, uploading only the gradient elements with the largest absolute values. The encrypted gradients... With metadata tags Upload them together to the edge node.

[0072] 3. Edge Aggregation and Model Deployment: Edge nodes collect gradients uploaded by participating devices within a preset timeout window, ignoring devices that fail to upload within the timeout period. Aggregation calculations yield... Then, update the global model parameters. The updated model will be distributed to all injection devices (including those not involved in this round of aggregation) to achieve real-time sharing of group knowledge.

[0073] 4. Adaptive Learning Rate Adjustment: To improve convergence stability, the edge nodes dynamically adjust the global learning rate based on the variance of the aggregated gradient. : in, The initial learning rate, The variance of the aggregate gradient. This is an adjustment factor. When the gradients of different devices diverge significantly, the learning rate is automatically reduced to avoid model oscillations.

[0074] Through the aforementioned federated learning mechanism, this invention enables multiple refueling devices to collaboratively optimize the global protocol generation model without sharing original data, effectively protecting the privacy and security of local data on each device, while significantly improving the efficiency and accuracy of protocol adaptation for new vehicle models.

[0075] Step S4: Dual-mode memory storage Device A packages the entire interaction process into a successful experience tuple, including: the detection sequence, the extracted feature fingerprint, the final generated precise template, vehicle information (VIN segment, part number), and real-time physical parameters (ambient temperature 25℃, target vacuum level, etc.). This experience tuple is simultaneously stored in two complementary memory networks: Contextual Memory Network: Deployed locally on device A, it uses a Transformer-based encoder to encode experience tuples into high-dimensional memory vectors (256 dimensions) and store them in a cache, supporting fast content-based retrieval (such as retrieving the most similar historical experience by the current feature fingerprint).

[0076] Semantic Memory Network: Deployed in a cloud-based evolution center, it employs a Graph Neural Network (GNN) architecture and shares node representations with the protocol knowledge graph. This network abstracts general protocol rules from a large number of experience tuples, such as the XOR check algorithm typically used for a certain series of car models, and integrates them into the knowledge graph in a graph structure.

[0077] Step S5: Protocol Instantiation and Process Execution Equipment A executes the following process steps sequentially, following the standard filling process flow: initialization, vacuuming, pressure holding, filling, and back suction. At each step, Equipment A calls the precise protocol fingerprint template generated in step S3 and dynamically instantiates a communication fingerprint frame based on the real-time physical parameters of the current process.

[0078] Taking the vacuuming node as an example: Process request: Evacuate the node identifier, with a real-time physical parameter target vacuum level of -0.1MPa (corresponding to hexadecimal 0x64).

[0079] The abstract fingerprint frame of the vacuum node in the protocol template is “0431????”, where “??” represents a placeholder for variable parameters.

[0080] Device A fills the real-time parameter 0x64 into the first placeholder and reserves the second placeholder for verification.

[0081] According to the verification algorithm specified in the template (in this embodiment, it is a cumulative check, with an initial value of 0x00), the cumulative sum of the filled frame 04316400 is calculated: 0x04 + 0x31 + 0x64 = 0x99. The low byte 0x99 is taken and filled into the last byte to obtain the complete communication fingerprint frame 04316499.

[0082] According to the transmission timing defined in the template (interval of 50ms), the communication frames are sent to the vehicle braking control system through the OBD interface.

[0083] The system receives execution status information (such as a success confirmation frame) returned by the vehicle, parses it, and then controls the physical actuator (vacuum pump) to start the vacuuming operation.

[0084] Subsequent process steps are carried out sequentially until the entire filling process is completed.

[0085] Step S6: Perception and Resolution of Memory Conflicts Several days later, another vehicle of the same model (with the same part number) entered the workstation, but the ambient temperature was 5℃ (low temperature). When device A invokes the protocol template, it simultaneously obtains prediction results from both the episodic memory network and the semantic memory network: Episodic memory network: retrieves the most recent successful experience (interaction record in a 25℃ environment) and outputs the prediction result: using the XOR check algorithm.

[0086] Semantic Memory Network: Based on abstract rules, it discovers multiple cases in historical data where this series of models uses CRC check under low-temperature environments (-5℃, 0℃), and outputs the prediction result: CRC check is used.

[0087] The predictions from the two memory networks conflict (due to different verification algorithms). The conflict resolution engine is triggered and performs the following operations: Conflict detection: The conflict detection unit detects that the outputs of episodic memory and semantic memory are inconsistent, and both have confidence scores higher than the threshold (0.8), and determines that there is a conflict.

[0088] Physical parameter extraction: The physical parameter extraction unit reads the current real-time physical parameters from the PLC of the filling equipment: ambient temperature 5℃, vehicle battery voltage 12.1V, target vacuum degree -0.1MPa.

[0089] Similarity Calculation: The similarity calculation unit extracts the physical parameter vectors from the historical experience tuples associated with both conflicting parties: contextual memory-associated cases (temperature 25℃, voltage 12.5V) and semantic memory-associated cases of multiple low temperatures (temperature -5℃, 0℃, voltage 11.8V, 12.0V). After normalizing the physical parameters, the cosine similarity between the current parameter and each historical parameter vector is calculated, and a time decay factor is introduced. The calculation results show that the similarity of the low-temperature cases associated with semantic memory (0.92) is higher than that of contextual memory (0.45).

[0090] Decision output: The decision output unit selects the semantic memory prediction result with the highest similarity as the final protocol template, that is, it adopts the CRC check algorithm.

[0091] Knowledge Update: The knowledge update unit transforms the process and result of this conflict resolution into new knowledge associations, which are stored in the semantic memory network and protocol knowledge graph in the form of a triple of part number = IPB-R_2025 + temperature <10℃ = CRC check, forming protocol rules sensitive to physical parameters.

[0092] Device A successfully communicated with the vehicle and completed the refueling process based on the final template after digestion (using CRC check).

[0093] Step S7: Experience Replay and Evolution The cloud-based evolution center aggregates experience tuples uploaded by all injection devices, including the aforementioned conflict resolution cases. The incremental learning engine employs the Elastic Weight Consolidation (EWC) method to incrementally train the global protocol generation model while simultaneously updating the protocol knowledge graph. The EWC algorithm prevents catastrophic forgetting by calculating the importance of each parameter to old tasks and penalizing changes to important parameters when optimizing for new tasks.

[0094] The updated global model and knowledge graph are periodically distributed to edge nodes and devices via the model distribution service. In the future, when other factories encounter the same vehicle model and the ambient temperature is below 10°C, the system can directly retrieve the CRC checksum rules from the knowledge graph without needing to resolve conflicts again.

[0095] Implementation Method 2: Multi-Factory Co-evolution The cloud-based evolution center employs a distributed storage and computing architecture, supporting multi-factory data aggregation and cross-regional federated learning. For example, edge nodes are deployed in the Chinese, European, and North American factories, with each factory's refueling equipment interacting only with its local edge node. Through cross-regional federated learning, the cloud-based evolution center collaboratively optimizes the global model without sharing raw data. A global conflict resolution rule base performs association rule mining (such as the FP-Growth algorithm) on conflict cases reported by each factory, discovering common cross-regional patterns, such as all models in a certain series switching to CRC checksum when the ambient temperature is below 5°C. These patterns are stored in a semantic memory network.

[0096] Implementation Method 3: Local Synchronization of Protocol Knowledge Graph The protocol knowledge graph is deployed in the cloud evolution center, but to meet the real-time inference needs of edge nodes, the subgraph related to frequently produced models in the local workshop is periodically (e.g., daily) synchronized to the edge nodes. Edge nodes use local copies to achieve millisecond-level protocol pre-guessing and inference, eliminating the need to access the cloud for every query and significantly reducing network latency.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A collaborative sensing method for brake fluid filling equipment based on swarm intelligence, characterized in that, include: S1. Multi-source information fusion and protocol pre-conjecture: Edge nodes acquire multi-source information of the target vehicle, input it into the protocol knowledge graph reasoning, and generate candidate protocol initial templates; S2. Single-machine detection and feature extraction: The first refueling device sends a protocol detection sequence based on the candidate template, captures the response signal and extracts multi-dimensional protocol feature fingerprints, and uploads them to the edge node; S3. Group Collaborative Perception and Federated Learning: When an edge node determines the first appearance of a vehicle model, it broadcasts a collaborative perception task to other devices. Each participating device uses local historical data to calculate the model gradient and uploads it. Edge nodes aggregate gradients to optimize the global protocol generation model and generate an accurate protocol fingerprint template. S4. Dual-mode memory storage: The experience tuple of this interaction is stored simultaneously in the local context memory network of the device and the semantic memory network of the cloud evolution center; S5. Protocol Instantiation and Process Execution: The refueling equipment calls the precise protocol fingerprint template at each process node, combines it with real-time physical parameters to dynamically instantiate and generate communication fingerprint frames, interact with the vehicle and control the physical actuators; S6. Memory Conflict Perception and Resolution: If the prediction results of contextual memory and semantic memory conflict during subsequent calls, the conflict resolution engine is triggered. It extracts the current real-time physical parameters and the physical parameters in the historical experience tuple for correlation analysis, selects the best matching result as the final protocol template, and updates the semantic memory network and knowledge graph as new knowledge associations. S7. Experience Replay and Evolution: The cloud-based evolution center gathers experience tuples, updates the protocol knowledge graph and protocol generation model, and distributes them to edge nodes and devices.

2. The method according to claim 1, characterized in that, The multi-source information mentioned in step S1 includes the VIN code and optional package configuration information obtained from the manufacturing execution system, as well as the brake system part number image recognition result obtained from the preceding visual inspection station.

3. The method according to claim 1, characterized in that, The protocol knowledge graph mentioned in step S1 is stored in the form of a graph database. Nodes include vehicle model, part number, supplier, protocol features, instruction template and physical parameters, and edges represent the relationships between entities. Inference uses a link prediction algorithm based on graph neural network.

4. The method according to claim 1, characterized in that, The protocol generation model described in step S3 adopts a variational autoencoder architecture, where the encoder maps multidimensional protocol feature fingerprints into latent space vectors, and the decoder reconstructs the protocol fingerprint template. In federated learning, each device only uploads the gradient of the decoder parameters, and the edge nodes use the federated averaging algorithm to update the global model; The federated learning specifically includes: Edge nodes broadcast the current global model parameters to all participating devices; each participating device performs multiple rounds of gradient descent on its local historical dataset to calculate the model gradient, where the loss function includes reconstruction loss, KL divergence loss and auxiliary verification loss. The participating devices encrypt the local gradients and attach metadata tags before uploading them to the edge nodes; the metadata tags contain local data volume and environmental parameter information. Edge nodes use a weighted federated average algorithm to aggregate gradients uploaded by each device. The weights are determined by the amount of data from each device and the similarity of its metadata. The aggregated gradients are used to update the global model parameters. Edge nodes dynamically adjust the global learning rate based on the variance of the aggregated gradient.

5. The method according to claim 1, characterized in that, The real-time physical parameters mentioned in step S6 include one or more of the following: ambient temperature, target vacuum level, pressure holding time, filling amount, and vehicle battery voltage; the correlation analysis uses similarity calculation and introduces a time decay factor; the new knowledge association is stored in the semantic memory network in the form of part number + physical parameter threshold = protocol feature.

6. A collaborative sensing system for brake fluid filling equipment based on swarm intelligence, used to implement the method described in any one of claims 1-5, characterized in that, include: Multiple dispensing devices, each equipped with an OBD communication module, a local processor and storage unit, and a locally deployed scene memory network; Edge computing nodes communicate with each refueling device and are equipped with a conflict resolution engine. The cloud-based evolution center communicates with edge nodes and is equipped with a semantic memory network. This network aggregates experience tuples, updates the protocol knowledge graph and protocol generation model, and distributes the updated model to edge nodes and devices.

7. The system according to claim 6, characterized in that, The refueling device has a locally deployed lightweight federated learning agent, which is used to calculate model gradient update information based on local historical data and upload the encrypted gradients; the gradient update information includes metadata tags associated with physical parameters.

8. The system according to claim 6, characterized in that, The protocol knowledge graph is deployed in a cloud evolution center and is periodically synchronized to edge nodes to enable fast local reasoning; the knowledge graph contains physical parameter nodes and their associated edges with protocol feature nodes.

9. The system according to claim 6, characterized in that, The cloud-based evolution center adopts a distributed storage and computing architecture, supports multi-factory data aggregation and cross-regional federated learning, and uses a global conflict resolution rule base to perform pattern mining on conflict cases reported by each factory to form cross-regional general protocol rules.

10. The system according to claim 6, characterized in that, The conflict resolution engine includes: The conflict detection unit monitors the predictive differences between episodic memory and semantic memory. The physical parameter extraction unit acquires the real-time physical parameters of the current process. The similarity calculation unit calculates the similarity between the current physical parameters and the physical parameters in the historical experience tuples; The decision output unit selects the optimal prediction result based on similarity ranking; The knowledge update unit transforms the resolution results into new knowledge associations and feeds them back into the memory network.

Images