Edge computing based device failure prediction system

Abstract

The application relates to the technical field of industrial intelligent operation and maintenance and edge computing, and particularly discloses a device fault prediction system based on edge computing. The system comprises an edge perception node, a local immune modeling unit, a distributed antibody sharing network and a central coordination server; the edge perception node collects and pre-processes device state data, the local immune modeling unit takes normal data as a'self' benchmark, dynamically generates 'antibodies' for identifying unknown faults through a clone selection mechanism, the distributed antibody sharing network lightweightly propagates effective antibodies among adjacent nodes, and the central coordination server only performs macroscopic strategy regulation and control and global antibody archiving. Through the technical scheme, autonomous identification and group collaborative immunity of the edge side to zero-day faults are realized, the dependence on fault samples and network bandwidth is reduced, and the self-healing capability and prediction timeliness of the system in a weak network environment are improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent operation and maintenance and edge computing technology, specifically involving an equipment fault prediction system based on edge computing. Background Technology

[0002] With the convergence of the Industrial Internet and edge computing technologies, real-time status monitoring and predictive maintenance of industrial equipment have become key technologies for improving manufacturing efficiency and ensuring production safety. By deploying computing resources close to the data source at the edge, the system can achieve real-time processing and decision-making of massive amounts of sensor data, reducing network latency and improving response speed to abnormal conditions. This distributed architecture provides a solid foundation for the intelligent operation and maintenance of large-scale industrial facilities, making it possible to complete complex equipment health assessments and trend analyses locally.

[0003] Edge computing-based fault prediction technology focuses on deeply analyzing multi-dimensional parameters during equipment operation by constructing intelligent discriminative models. This technology typically utilizes pre-trained mathematical models or classification algorithms to map real-time collected signals such as vibration, temperature, and pressure to a predefined fault space. The aim is to continuously monitor and identify signs of equipment performance degradation, thereby providing a scientific basis for maintenance decisions.

[0004] Existing predictive technologies still have significant limitations when facing highly complex and dynamically changing industrial environments. Traditional predictive models mostly employ static parameter structures, and because their reasoning logic heavily relies on prior knowledge from offline training sets, the systems lack the ability to autonomously identify and learn online when faced with unknown fault modes. Industrial environments exhibit extreme imbalanced sample characteristics; the massive amounts of normal operation data and the extremely scarce fault cases make it difficult for conventional supervised learning methods to extract effective anomaly features.

[0005] Existing systems generally need to send data back to the cloud and retrain globally after discovering new types of faults. This centralized evolution mode not only consumes a lot of bandwidth, but also fails to achieve rapid knowledge sharing and collaborative defense among edge nodes, resulting in low efficiency of the system's collective response to sudden faults. Therefore, a device fault prediction solution based on edge computing is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a device fault prediction system based on edge computing, which can effectively solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The edge computing-based device failure prediction system includes edge sensing nodes, local immune modeling units, a distributed antibody sharing network, and a central coordination server. The edge sensing node is configured to collect multi-dimensional status parameters of industrial equipment in real time during operation, and to preprocess and extract features from the raw data to form a data stream characterizing the current health status of the equipment. The local immune modeling unit is deployed inside each edge sensing node to build and maintain a dynamically evolving fault identification model. This fault identification model uses normal operation data as a "self" benchmark, identifies abnormal patterns that deviate from the "self" benchmark as potential "antigens", and activates a local clone selection mechanism when an unknown antigen is detected. By performing high-frequency mutation and affinity screening on existing model parameters, it quickly generates a special identification rule, i.e., an "antibody", for the new type of fault. The distributed antibody sharing network consists of lightweight communication links established between adjacent edge sensing nodes, which are used to broadcast verified and effective antibody parameters in a local area, enabling neighboring nodes to synchronously obtain the ability to identify novel faults without relying on a central server, thereby achieving collaborative immunization at the population level. The central coordination server is located in the cloud or a regional data center. It is used to monitor the operating status of all edge nodes in the network, receive antibody effectiveness feedback information reported by each node, and make macroscopic adjustments to the global immunization strategy when necessary, but does not participate in the daily fault identification and antibody generation process.

[0008] Preferably, the local immune modeling unit is further configured to construct its core logic using artificial immune network theory, define the historical normal operation dataset of the device as the "self" space, and continuously calculate the distance metric between the current input data and the "self" space. When the distance exceeds a preset threshold, it is determined that there is a non-self signal, and the antibody generation process is triggered.

[0009] Furthermore, during antibody generation, the local immune modeling unit performs multiple rounds of replication, mutation, and selection operations on candidate models based on a clonal selection algorithm. The mutation intensity is dynamically adjusted according to the degree of antigen abnormality, and the affinity assessment is based on the accuracy of the candidate model's backtracking matching of similar abnormal fragments in history. Finally, several antibodies with the highest affinity are retained as locally usable fault identifiers.

[0010] Furthermore, the distributed antibody sharing network adopts a propagation mechanism based on topological proximity, sending newly generated antibody parameters only to edge sensing nodes that are physically or logically adjacent, and the transmitted content is only a subset of the model's key weights or compressed feature mapping rules, ensuring that communication overhead remains within limits.

[0011] Preferably, the edge sensing node has a built-in adaptive data cleaning module, which is used to filter out false abnormal signals caused by environmental noise or sensor drift before feature extraction, so as to avoid misjudging non-fault fluctuations as antigens and improve the specificity of the immune system.

[0012] Furthermore, the central coordination server is configured with a global antibody library, which is used to archive the effective antibodies reported by each region and their applicable scenario descriptions. When a common fault mode across regions is detected, a standardized initial antibody template is pushed to the relevant edge clusters to accelerate the immune initialization process of newly deployed nodes.

[0013] Furthermore, the local immune modeling unit supports a multi-antibody coexistence mechanism, allowing multiple identification models for different fault types to be loaded simultaneously, and dynamically activating the most matching antibody based on the feature distribution of real-time input data for discrimination, thereby achieving accurate tracking of complex faults or multi-stage degradation processes.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The device fault prediction system based on edge computing provided by this invention introduces an artificial immune mechanism, enabling edge nodes to have the adaptive evolutionary ability of organisms. It can autonomously identify and quickly respond to unknown fault modes without cloud intervention, solving the technical problem that traditional static models are helpless against "zero-day failures".

[0015] 2. This invention focuses on "self" recognition as its core, and focuses on anomaly detection rather than fully supervised classification, which greatly reduces the dependence on scarce fault samples and effectively alleviates the modeling difficulties caused by the extreme imbalance of samples in industrial sites. Through a distributed antibody sharing network, newly generated fault recognition knowledge can be efficiently propagated among neighboring nodes to form a local immune barrier, thereby improving the speed and coverage of the entire edge cluster's collective defense against sudden new faults.

[0016] 3. Since antibody generation and sharing are both completed at the edge, this invention avoids frequent backhaul of original data or complete models to the cloud, saving network bandwidth resources and enhancing the system's continuous operation capability in weak or offline environments. The overall architecture achieves a deep integration of biologically inspired mechanisms and federated edge learning, which not only improves the accuracy and timeliness of predictions, but also endows the system with intelligent characteristics of continuous self-healing and collaborative evolution, providing a new technical path for highly reliable industrial operation and maintenance. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic evolution of the local fault identification model based on the clone selection mechanism in this invention; Figure 3 This is a flowchart illustrating the logical process of adaptive data cleaning and device health status extraction for edge sensing nodes in this invention. Figure 4 This is a schematic diagram of the topological proximity propagation and multi-node collaborative interaction of the distributed antibody sharing network in this invention; Figure 5 This is a flowchart illustrating the logical flow of the central coordination server monitoring the global immunization strategy and pushing common fault antibody templates in this invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] The edge computing-based equipment fault prediction system includes edge sensing nodes, local immune modeling units, a distributed antibody sharing network, and a central coordination server. The edge sensing nodes are connected to the sensor array of the industrial equipment to acquire and preprocess the raw state data of the industrial equipment.

[0020] The local immune modeling unit is electrically connected to the edge sensing node and is used to construct a fault identification model based on an artificial immune mechanism; the distributed antibody sharing network is connected between multiple local immune modeling units and is used to realize the interaction and synchronization of model parameters at the edge; the central coordination server is connected to the edge sensing node and the local immune modeling unit through a wide area network and is used to execute global-level policy distribution and status monitoring.

[0021] The edge sensing node is configured as a data access and cleaning center in the industrial field. In terms of physical construction, the edge sensing node typically adopts an industrial-grade embedded computing board with high electromagnetic interference resistance, which integrates multi-channel analog signal acquisition interfaces, digital signal processing circuits, and large-capacity non-volatile storage media.

[0022] During operation, the edge sensing node collects multi-dimensional state parameters of the industrial equipment in real time through its physical interface. These multi-dimensional state parameters include, but are not limited to, vibration acceleration signals on the equipment surface, temperature sequences at key bearings, pressure fluctuation data of the lubrication system, current load characteristics of the drive motor, and environmental variables such as ambient humidity.

[0023] The edge sensing node incorporates an adaptive data cleaning module, configured to perform a series of complex signal purification operations before feature extraction. This module employs an adaptive bandpass filter, dynamically adjusting the cutoff frequency based on the sampling frequency to filter out electromagnetic noise caused by the start-up and shutdown of large electrical equipment in industrial settings. Furthermore, the adaptive data cleaning module compensates for zero-point drift generated by the sensor using a moving average algorithm or an exponential smoothing algorithm, eliminating non-faulty low-frequency fluctuations.

[0024] During the data preprocessing stage, the edge sensing node also performs data standardization, mapping raw physical quantities of different dimensions and orders of magnitude to a unified numerical range to ensure the computational stability of subsequent modeling units. The cleaned data is then further fed into a feature extraction operator to extract key representation vectors, including root mean square value, peak factor, and skewness index in the time domain, and center frequency and power spectral density distribution in the frequency domain, ultimately forming a data stream representing the current health status of the device.

[0025] The local immune modeling unit is deployed within the internal computing environment of each edge sensing node, and its core logical architecture is based on the theory of artificial immune networks. The local immune modeling unit is configured to build and maintain a dynamically evolving fault identification model, which is designed to simulate the biological immune system's ability to distinguish between "self" and "non-self".

[0026] The local immune modeling unit defines the historical normal operation dataset collected by the device during its stable operation period as the "self" space. Mathematically, the "self" space is abstracted as a manifold composed of a large number of high-dimensional orthogonal basis vectors, or a set of multi-dimensional hyperspheres covering the normal operation points.

[0027] The local immune modeling unit continuously monitors the data stream input from the edge sensing nodes and continuously calculates the distance metric between the feature vector corresponding to the current input data and the "self" space. The distance metric can employ Euclidean distance, Mahalanobis distance, or cosine similarity, among other criteria. When the distance metric exceeds a preset anomaly detection threshold, the local immune modeling unit determines that the current device state deviates from the normal baseline, indicating a "non-self" signal, and identifies this abnormal pattern as a potential "antigen." The system recognizes the possible emergence of a new, unrecorded fault.

[0028] To address the identified "antigen," the local immune modeling unit immediately initiates a local clone selection mechanism. During this process, the local immune modeling unit does not seek instructions from the cloud controller but instead generates a set of candidate recognizers based on the existing model parameter library. The clone selection mechanism includes multiple rounds of replication, mutation, and selection operations. The local immune modeling unit first performs large-scale replication of the current candidate recognizer model, generating a large number of clone copies. The local immune modeling unit then performs high-frequency mutation operations on these clone copies, that is, randomly perturbs the model's weight parameters, network topology, or the threshold of the discriminant function.

[0029] The intensity of the mutation operation is dynamically positively correlated with the degree of abnormality of the antigen. When the distance metric between the current input data and the "self" space increases, indicating that the abnormality is extremely severe or novel, the local immune modeling unit will increase the mutation probability and perturbation step size to search for potential "antibodies" in a wider parameter space; conversely, if the degree of abnormality is mild, a low mutation intensity will be used for local fine-tuning.

[0030] After the mutation generates a series of new models, the local immune modeling unit performs affinity assessment. This affinity assessment is based on either the new model's accuracy in retrospectively matching similar historical anomalous fragments, or the new model's confidence in classifying current anomalous features. Through multiple rounds of selection, the local immune modeling unit ultimately retains several models with the highest affinity as locally usable dedicated identification rules, i.e., "antibodies."

[0031] The distributed antibody sharing network is a key architecture for achieving herd immunity. It consists of lightweight communication links established between adjacent edge sensing nodes, supporting various industrial communication protocols. The distributed antibody sharing network employs a propagation mechanism based on topological proximity. When the local immune modeling unit within one of the edge sensing nodes successfully generates and verifies an effective antibody, the distributed antibody sharing network broadcasts the key parameters of this distributed antibody to physically adjacent or logically similar neighboring nodes.

[0032] To reduce communication overhead, the transmitted content is not the complete deep learning model, but rather a compressed subset of key weights, feature mapping rules, or lightweight knowledge distillation instructions. This "infection" mechanism enables the entire edge cluster to synchronously acquire the ability to identify new types of faults without relying on a central server, forming a regional immune barrier.

[0033] The central coordination server in the system is responsible for macro-level optimization and global knowledge archiving. It is equipped with a global antibody library to receive verified antibodies and their corresponding operational descriptions reported by each edge node. Although the central coordination server does not participate in real-time antibody generation and fault diagnosis, it is responsible for monitoring the operational health status of all edge nodes in the network. When the central coordination server discovers through big data analysis that a certain type of antibody is being repeatedly generated in multiple unrelated regions, it determines that this novel fault mode has global prevalence.

[0034] The central coordination server utilizes its powerful computing resources to aggregate and optimize these dispersed antibodies, generating standardized initial antibody templates. When a new node joins the system or a node undergoes an initialization reset, the central coordination server proactively pushes these standardized templates, accelerating the immune initialization process of the newly deployed node.

[0035] Furthermore, the local immune modeling unit internally supports a multi-antibody coexistence mechanism. This mechanism allows each edge node to simultaneously load and run multiple identification models for different fault types. During actual operation, the local immune modeling unit dynamically activates one or more antibodies that best match the current features for discrimination based on the statistical distribution characteristics of the real-time input data. This concurrent processing mode enables the system to cope with complex compound faults and accurately track and predict the multi-stage process of equipment degradation from minor degradation to complete failure.

[0036] The edge sensing node employs a multi-core processor based on the ARM architecture, supplemented by an FPGA coprocessor for real-time fast Fourier transform processing of high-frequency sampled data. The clone selection algorithm of the local immune modeling unit runs in the real-time operating system environment of the ARM processor, utilizing multi-threading technology to perform mutation and affinity calculations of clone copies in parallel.

[0037] The distributed antibody sharing network is implemented at the physical layer based on industrial Ethernet or 5G microcellular networks, and dynamically maintains the adjacency list between nodes through software-defined networking technology. The central coordination server is deployed on the enterprise's private cloud platform, and uses distributed database technology to classify, index, and store massive amounts of fault characteristics and antibody parameters.

[0038] Example 2: Based on the edge computing-based equipment fault prediction system described in Example 1, this example provides a system architecture variant for large-scale discrete manufacturing scenarios, focusing on its adaptive immune mechanism and cross-level data governance logic in a heterogeneous hardware environment.

[0039] In an alternative system architecture, the edge sensing nodes are divided into a sensing sublayer and a computing-assisted sublayer. The sensing sublayer consists of miniature intelligent sensor units distributed throughout the production line, directly embedded within the equipment, responsible for the lowest-level signal conditioning. The computing-assisted sublayer comprises industrial edge computing gateways placed within workshop control cabinets. The sensing and computing sublayers are connected via Bluetooth Low Energy or ZigBee wireless sensor networks. This layered design allows the system to cover a wider physical area while shifting complex immune computing tasks from the power-constrained sensor side to the computationally powerful gateway side.

[0040] In this embodiment, the local immune modeling unit is configured as a logical entity with a hierarchical defense structure. The first level of defense is a preliminary filter based on statistical process control, used to identify anomalies that are clearly outside the tolerance range; the second level of defense is the core artificial immune network, responsible for analyzing subtle performance degradation. When constructing the "self" space, the local immune modeling unit introduces a dynamic time warping algorithm to address the non-stationary data characteristics of industrial equipment under different speeds and loads. The calculation process of the distance metric is transformed into a similarity search of non-stationary time series, ensuring that the system does not generate false alarms under variable speed conditions.

[0041] In this embodiment, the distributed antibody sharing network employs a publish-subscribe communication mechanism. Each edge-aware node acts as either a publisher or subscriber to a topic. When a node identifies an antigen and generates an antibody, it tags the antibody with a fault type and operational context label before publishing it. Only neighboring nodes with the same equipment model or in similar production processes will receive and load the antibody. This logically related propagation method further optimizes bandwidth utilization.

[0042] In this embodiment, the central coordination server incorporates a global immune diversity maintenance module. This module assesses the coverage of the antibody library across the entire network. If certain types of faults are found to have never been identified by any node historically, or if the existing antibody library's coverage of certain extreme conditions is insufficient, the central coordination server utilizes its internally integrated physical simulation model to generate synthetic fault antigen data and forcibly distributes simulated antigens to edge nodes, inducing them to pre-evolve targeted "vaccine" antibodies locally. This preventative immune enhancement improves the system's early warning capabilities in the face of rare faults.

[0043] The edge-aware node in this embodiment also includes a self-repairing diagnostic unit. When the local immune modeling unit detects redundancy or conflict in its own antibody library, the self-repairing diagnostic unit initiates a competitive inhibition mechanism. Based on the predicted performance scores of antibodies over a period of time, it performs apoptosis operations on antibodies with low affinity or high false alarm rates, i.e., deletes these invalid discrimination rules from local memory. This simulates the immune tolerance and negative selection process in the biological immune system, ensuring that the memory consumption and computational load of the edge node remain within a reasonable threshold during long-term operation.

[0044] To address common sensor failure issues in industrial settings, the edge sensing node is also equipped with a cross-modal data consistency verification module. This module is configured to determine the reliability of input data by analyzing the physical correlation between different sensors. If a signal generated by a faulty sensor is determined to be a false anomaly, the local immune modeling unit will temporarily block the corresponding antibody generation triggering logic and send a hardware repair request to the central coordination server.

[0045] Example 3: Based on the above examples, this example further describes in detail the implementation of the edge computing-based device fault prediction system under a software-defined architecture, and how to improve the global universality of the antibody while protecting data privacy through federated edge learning technology.

[0046] In this embodiment, the system employs a fully containerized deployment model. The functional components of the edge-aware nodes and the local immune modeling unit run as independent microservices within a lightweight container engine. This architecture enables the system to be rapidly migrated and expanded across industrial PCs or embedded control boards on different hardware platforms.

[0047] The local immune modeling unit incorporates an enhanced clone selection algorithm during antibody generation. This clone selection algorithm employs a parameter optimization strategy based on generative adversarial networks during the mutation phase. The local immune modeling unit internally includes a discriminator and a generator. The generator is responsible for generating candidate mutation schemes for the model parameters, while the discriminator uses local historical anomaly logs to evaluate the effectiveness of the mutation schemes. Through this adversarial internal evolution, the generated antibodies exhibit significantly improved specificity, enabling more precise identification of the root cause of the failure.

[0048] In this embodiment, the distributed antibody sharing network integrates a secure aggregation protocol for federated learning. When nodes exchange antibody parameters, the distributed antibody sharing network first performs homomorphic encryption on the antibody parameters. Neighboring nodes receive the encrypted parameter gradients, and through weighted aggregation within the encrypted space, each node can jointly optimize a more powerful antibody model without exposing its own local raw production data. This approach is significant in multi-enterprise collaborative supply chain scenarios, achieving both herd immunity and protecting the core process data of each participant.

[0049] In this embodiment, the central coordination server acts as the coordinator of the federated learning process. It is responsible for defining the global model architecture standards and calculating the contribution weight of each node in the global immune network based on the antibody operation statistics reported by each node. For nodes with large contributions, the central coordination server will prioritize allocating more computing resource scheduling quotas to them.

[0050] The system also includes an augmented reality maintenance interface. When the local immune modeling unit generates a high-confidence fault antibody and triggers an alert, this augmented reality maintenance interface displays the fault characteristics associated with the fault antibody, the affected physical parts, and the suggested repair plan to the on-site operators in real time via a visual terminal. This closed-loop feedback from prediction to decision execution further enhances the system's practical value.

[0051] To verify the stability of the edge computing-based device fault prediction system, a continuous integration and testing module was also designed in this embodiment. This module runs on the central coordination server and periodically sends obfuscated simulated data streams to the edge nodes to simulate sensor failures, network congestion, and sudden hardware reset events under various extreme environments. By observing the self-healing time of the edge nodes, the central coordination server can dynamically evaluate the robustness of the entire system and adjust the self-healing strategy parameters on the edge side as necessary.

[0052] In terms of data storage logic, the edge sensing node adopts a circular buffer structure based on a time-series database. The size of the circular buffer is dynamically adjusted according to the device's acquisition frequency to ensure that high-frequency raw sampling data from the most recent 24 hours is always retained. Once an antigen is detected and the antibody generation process is triggered, the circular buffer permanently labels the context data related to the anomaly and dumps it to a long-term storage area, providing valuable labeled samples for subsequent deep root cause analysis and model backtracking. This local data governance strategy avoids meaningless massive data backhauling while retaining core fault evidence.

[0053] Example 4: This example focuses on describing the application details of the edge computing-based equipment fault prediction system in a cross-regional, ultra-large-scale industrial internet environment, especially the technical implementation in handling the non-stationary fault evolution process with time-varying characteristics.

[0054] In a large-scale deployment scenario, the system comprises tens of thousands of edge-aware nodes distributed across different geographical regions and process units. The distributed antibody-sharing network employs a hierarchical hash ring topology. Nodes within each region form small autonomous clusters, with the leader node of each cluster connecting to the central coordination server. In this hierarchical communication model, the propagation of antibody parameters is confined to a set of strongly correlated nodes. The topological proximity considers not only physical Euclidean distance but also logical similarities such as equipment model, production date, and load characteristics.

[0055] The local immune modeling unit introduces a memory B-cell module based on a recurrent neural network to address common performance degradation processes in industrial equipment. In biological immunology, memory cells are responsible for maintaining a long-term memory of previously encountered antigens. The memory B-cell module is configured to store antibody parameter sequences that have been proven to have high predictive value. When the equipment exhibits slow, trending degradation signals, the local immune modeling unit recalls historical antibody templates from the memory module and uses a gated recurrent unit to model the degradation trend. This mechanism enables the system not only to identify sudden failures but also to predict the specific time window in which the failure will occur, i.e., remaining lifespan prediction.

[0056] In the antibody affinity assessment stage, this embodiment employs a multi-criteria decision analysis method. The affinity is no longer a single numerical value, but a multi-dimensional vector encompassing accuracy, computational cost, real-time performance, and generalization ability. When selecting the optimal antibody, the local immune modeling unit dynamically adjusts the weights of each criterion based on the current system resource status of the edge nodes.

[0057] When the processor utilization of the edge sensing node is high, the system will prioritize lightweight antibodies with low computational overhead; while when the system is in a resource idle period, it will evolve and select deep antibodies with more complex structures and higher accuracy.

[0058] The central coordination server also runs an antibody evolution competition module at the global level. This module periodically collects different antibody schemes targeting the same type of fault from the entire network and puts them to a "survival competition" on a standardized test dataset. High-performing antibody models are assigned a higher ranking and used as seed models to be pushed to all relevant edge nodes for secondary mutation. This artificially intervened collaborative evolution mechanism simulates the expansion process of dominant species in nature, accelerating the system's ability to identify emerging global faults.

[0059] To address the issue of unstable network connectivity in industrial environments, the distributed antibody sharing network features retransmission after disconnection and latency compensation. When an edge node temporarily loses contact with its neighbors, it stores newly generated antibodies in its local queue and uses timestamp technology to evaluate the validity of antibody parameters. Once the network is restored, the system automatically performs differentiated synchronization to ensure that the immune knowledge base of each node is up-to-date.

[0060] To further enhance the system's security capabilities, the edge-aware node also integrates a security audit module based on anomaly detection. This security audit module utilizes an immune mechanism similar to fault prediction to identify network attacks targeting the edge. If the generation behavior of an antibody does not conform to the preset evolution logic, or if the antibody parameters contain illegal characters, the security audit module will immediately isolate the local immune modeling unit and trigger a global-level warning to prevent malicious "viruses" from spreading in the distributed network.

[0061] The system in this embodiment also supports a human-machine collaborative antibody labeling mechanism. Through a mobile app, on-site maintenance engineers can provide feedback on the "neoantigens" identified by the system. If the engineer confirms that the anomaly does indeed correspond to a specific mechanical fault, the feedback information is sent to the local immune modeling unit, which converts it into a "mature antibody" with a clear semantic label. This real-time injection of expert knowledge allows the system's artificial immune network to continuously learn from human experience, achieving a deep integration of machine intelligence and human wisdom.

[0062] Example 5: This example provides a specific implementation scheme based on hardware acceleration and low-latency communication optimization, which aims to solve the problem of capturing transient faults in industrial equipment at ultra-high sampling frequencies.

[0063] In certain specialized industrial scenarios, such as high-speed machine tools or aero-engine test benches, sensor sampling frequencies often reach hundreds of thousands of hertz. In such cases, traditional general-purpose processor-based computing models struggle to meet real-time requirements. In this embodiment, the edge sensing node is configured with a parallel processing architecture based on hardware logic gates. The feature extraction operator is embedded in the FPGA's logic circuitry, enabling zero-copy real-time analysis of streaming data.

[0064] The local immune modeling unit employs an artificial immune model based on a spiking neural network. Compared to traditional deep neural networks, spiking neural networks offer higher energy efficiency and temporal modeling capabilities, enabling direct processing of pulse-coded signals generated by sensors. In this architecture, antigen recognition is transformed into matching neuronal spiking patterns. When the spiking pattern of the input signal deviates from a preset pattern in the "self" space, the local clone selection logic is triggered.

[0065] To achieve ultra-fast sharing of antibody parameters at the microsecond level, the distributed antibody sharing network in this embodiment employs a low-latency communication protocol based on remote direct memory access technology. Adjacent edge sensing nodes are connected via dedicated fiber optic links, allowing model parameters generated by a node to be directly mapped into the memory address space of its neighbors, eliminating latency caused by protocol stack processing. This deep physical coupling makes the entire edge cluster logically behave like a massive, real-time self-healing distributed brain network.

[0066] The central coordination server, in this high-performance configuration, undertakes the maintenance tasks of the digital twin entity. It establishes a high-precision digital simulation model for each field device. Whenever an edge node reports a new fault, the central coordination server replays the anomaly process in the digital twin model to verify the physical mechanism of the fault. This dual verification method based on physical models and data-driven approaches reduces the false alarm rate of fault prediction and provides maintenance personnel with in-depth failure mechanism analysis reports.

[0067] To address the fusion problem of multi-source heterogeneous data, the edge sensing node is equipped with an adaptive weight allocator. Before performing immune modeling, this adaptive weight allocator dynamically adjusts the weights of each sensor in the feature vector based on the signal-to-noise ratio of each sensor. For example, if the vibration sensor is found to be saturated due to a strong impact, the system automatically increases the weights of current and temperature features to ensure that the immune system can maintain high recognition accuracy even when some sensing capabilities are impaired.

[0068] In terms of system lifecycle management, the central coordination server incorporates an antibody decommissioning management module. As industrial equipment is upgraded or ages, certain early failure modes may no longer be applicable. This module periodically analyzes the usage frequency and predicted gain of the antibody library, archiving "stale antibodies" that have not been activated for a long time and do not conform to current equipment health trends. This strategy ensures that edge computing resources are always focused on resolving the most pressing failure threats, improving the overall system operating efficiency.

[0069] The system in this embodiment also boasts excellent compatibility. The edge sensing node provides a standardized API interface, allowing third-party developers to write custom immune rules or mutation operators based on specific industrial application scenarios. Through this open ecosystem, the system can continuously absorb the latest machine learning achievements and fault diagnosis algorithms, maintaining its technological leadership.

[0070] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Without departing from the spirit and scope of the present invention, those skilled in the art can make various modifications and variations, all of which fall within the scope of the claims of the present invention.

[0071] It should be noted that all embodiments mentioned in this specification can be combined with each other unless there is a clear logical conflict between these features. For example, the layered defense structure in Embodiment 2 can be fully applied to the high-performance hardware platform of Embodiment 5 to realize an industrial prediction system with both real-time performance and defense-in-depth capabilities.

[0072] In the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that technical feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0073] Those skilled in the art will understand that, in one or more embodiments of the present invention, unless the context explicitly specifies otherwise, the connections may include direct connections or indirect connections implemented through an intermediate medium. The modules, units, or servers may be physically integrated into the same physical device or distributed across different physical devices. The calculation and judgment processes may be implemented by specific hardware logic or by a general-purpose processor executing computer program instructions stored in non-volatile storage media.

[0074] In summary, the edge computing-based equipment fault prediction system provided by this invention, through the deep integration of biological immune mechanisms and a distributed edge computing architecture, constructs an intelligent operation and maintenance system with adaptive evolution, group collaboration, and privacy protection capabilities. This system not only solves the pain points of difficult identification of unknown faults, imbalanced samples, and high response latency in industrial settings, but also lays a solid technical foundation for the future development of the Industrial Internet towards self-organization and self-healing. In long-term industrial practice, this system can continuously self-optimize as the equipment operating environment changes, reducing enterprise maintenance costs and improving production safety, demonstrating broad application prospects and extremely high economic value.

Claims

1. A device fault prediction system based on edge computing, characterized in that, This includes edge-aware nodes, local immune modeling units, a distributed antibody sharing network, and a central coordination server; The edge sensing node is configured to collect multi-dimensional state parameters during the operation of industrial equipment in real time, and to preprocess and extract features from the raw data to form a data stream characterizing the current health status of the equipment. The local immune modeling unit is deployed inside each of the edge sensing nodes and is used to build and maintain a dynamically evolving fault identification model. The fault identification model uses historical normal operation data of the equipment during the stable operation period as its self-benchmark, identifies abnormal patterns that deviate from the self-benchmark as potential antigens, and activates a local clone selection mechanism when an unknown antigen is detected. By performing mutation and affinity screening on the existing model parameters, antibodies against novel faults are generated. The distributed antibody sharing network consists of communication links established between adjacent edge sensing nodes, used to broadcast verified antibody parameters in a local area, enabling nearby edge sensing nodes to synchronously acquire the ability to identify novel faults. The central coordination server is connected to each of the edge sensing nodes, and is used to monitor the operating status of all edge nodes in the network, receive antibody feedback information reported by the edge sensing nodes, and make macroscopic adjustments to the global immunization strategy.

2. The device fault prediction system based on edge computing according to claim 1, characterized in that: The edge sensing node has a built-in adaptive data cleaning module; The adaptive data cleaning module is equipped with an adaptive bandpass filter, which dynamically adjusts the cutoff frequency according to the sampling frequency to filter out electromagnetic noise caused by external electrical equipment. The adaptive data cleaning module is also equipped with a compensation operator, which is used to compensate for the zero-point drift generated by the sensor through a moving average algorithm or an exponential smoothing algorithm, so as to eliminate non-faulty low-frequency fluctuations. The edge sensing node is also equipped with a data standardization unit, which is used to map raw physical quantities with different dimensions and orders of magnitude to a unified numerical range. The edge sensing node extracts time-domain and frequency-domain indicators from the preprocessed data using feature extraction operators. The time-domain indicators include root mean square value, peak factor, and skewness index, while the frequency-domain indicators include center frequency and power spectral density distribution, forming the data stream of the health status.

3. The device fault prediction system based on edge computing according to claim 1, characterized in that: The local immune modeling unit is constructed based on the theory of artificial immune networks, and the historical normal operation dataset is defined as the self-space; The self-space is logically abstracted as a manifold composed of high-dimensional orthogonal basis vectors, or as a set of multi-dimensional hyperspheres covering normal operating conditions. The local immune modeling unit is equipped with a monitoring module for continuously calculating the distance metric between the feature vector corresponding to the data stream of the health status and the self-space; The distance metric is characterized by Euclidean distance, Mahalanobis distance, or cosine similarity; When the distance metric exceeds a preset anomaly detection threshold, the local immune modeling unit determines that the current device state contains a non-self signal and identifies the anomaly pattern as a potential antigen.

4. The device fault prediction system based on edge computing according to claim 3, characterized in that: The local immune modeling unit is configured to perform multiple rounds of replication, mutation, and selection operations during the operation of the local clone selection mechanism. The local immune modeling unit first replicates the current candidate recognizer model to generate a clone copy. The clone copy performs a mutation operation to generate a new model by applying random perturbations to the model's weight parameters, network topology, or the threshold of the discriminant function. The mutation intensity of the mutation operation is dynamically positively correlated with the degree of abnormality of the antigen, that is, the mutation intensity increases synchronously with the increase of the distance metric, so as to search in the parameter space by increasing the mutation probability and the perturbation step size.

5. The device fault prediction system based on edge computing according to claim 4, characterized in that: The local immune modeling unit is also equipped with an affinity assessment module for evaluation after a series of new models are generated by mutations; The affinity assessment module calculates the affinity value based on the accuracy of the new model in retrospectively matching similar abnormal fragments in the past, or based on the classification confidence of the new model for the current abnormal features. The local immune modeling unit performs multiple rounds of screening based on the affinity values, retains several models with the highest affinity values ​​as locally usable recognition rules, and defines the recognition rules as antibodies.

6. The device fault prediction system based on edge computing according to claim 1, characterized in that: The distributed antibody sharing network employs a propagation mechanism based on topological proximity. Once an edge sensing node successfully generates and verifies an antibody, the distributed antibody sharing network broadcasts the key parameters of the antibody to nearby nodes that are physically adjacent or have similar logical functions. To optimize communication bandwidth, the distributed antibody sharing network only sends compressed subsets of key weights, feature mapping rules, or knowledge distillation instructions during transmission. The distributed antibody sharing network is implemented at the physical layer based on industrial Ethernet or mobile communication networks, and uses software-defined networking technology to dynamically maintain the adjacency list between each edge sensing node.

7. The device fault prediction system based on edge computing according to claim 1, characterized in that: The central coordination server is configured with a global antibody library, which is used to archive antibodies reported by each region and their corresponding applicable scenario descriptions. The central coordination server uses big data analysis to identify the repeated generation of antibodies of different types in different regions. When it determines that the failure mode is globally universal, it aggregates and optimizes the scattered antibodies to generate standardized initial antibody templates. When a new node is added to the system or a node undergoes an initialization reset, the central coordination server pushes the standardized initial antibody template to the edge sensing node to accelerate the immune initialization process of the newly deployed node.

8. The device fault prediction system based on edge computing according to claim 1, characterized in that: The local immune modeling unit is configured with a multi-antibody coexistence module, which allows each edge sensing node to load and run multiple identification models for different fault types simultaneously. During operation, the local immune modeling unit dynamically activates one or more antibodies that best match the current characteristics based on the statistical distribution characteristics of the health status data stream to make a judgment, so as to track complex faults or multi-stage degradation processes of equipment. The local immune modeling unit also integrates a memory module based on a recurrent neural network to store antibody parameter sequences with predictive value and to model the long-term degradation trend of the device using a gated recurrent unit.

9. The device fault prediction system based on edge computing according to claim 1, characterized in that: The edge sensing node adopts a hierarchical architecture, including a sensing sub-layer and a computing-aided sub-layer; The sensing sublayer consists of miniature intelligent sensor units distributed throughout the production line, which are responsible for signal conditioning; The computing assistance sublayer consists of an industrial edge computing gateway, which is used to carry the computing tasks of the local immune modeling unit; The perception sublayer and the computing-aided sublayer are connected via a wireless sensor network; The local immune modeling unit is also equipped with a self-repair diagnostic unit, which is used to activate a competitive inhibition mechanism when redundancy or conflict is detected in the local antibody library, and to perform apoptosis operation on antibodies with low affinity or high false alarm rate according to the predicted performance of the antibody.

10. The device fault prediction system based on edge computing according to claim 1, characterized in that: The system adopts a containerized deployment mode, and the functional components of the edge sensing node and the local immune modeling unit run as independent microservices in the container engine. The distributed antibody sharing network integrates a secure aggregation protocol for federated learning. Before exchanging antibody parameters between nodes, the parameters are homomorphically encrypted. Nearest nodes receive the encrypted parameter gradients and perform weighted aggregation within the encrypted space. The edge-aware node also integrates a security audit module, which is used to identify network attack behavior using the immune mechanism, and isolate the corresponding local immune modeling unit when the antibody generation behavior does not conform to the preset evolution logic or the antibody parameters contain illegal characters.

Images