An intelligent fault prediction and self-repairing method for an internet of things terminal

Abstract

The application discloses a kind of Internet of Things terminal intelligent fault prediction and self-repairing method, specifically related to edge computing and equipment health management technical field;Collect the multi-source operation data of Internet of Things terminal, construct multimodal feature time series tensor;Characteristic drift rate is calculated using high-dimensional space perturbation evolution model, and failure symptom saliency probability atlas is generated;Based on significant abnormal point, autoregressive-attention fusion prediction network is activated, and evolution sensitive path is extracted as fault chain prediction vector;By topological mapping with equipment failure knowledge graph, suspected fault site and type are identified;If prediction probability exceeds threshold, call matching recovery strategy, and real-time monitor repair effect, complete state label update and strategy success rate evaluation, realize closed-loop repair.The method has high sensitivity, high adaptability and full-process autonomy characteristics.

Description

Technical Field

[0001] This invention relates to the field of edge computing and device health management technology, specifically to a method for intelligent fault prediction and self-repair of Internet of Things (IoT) terminals. Background Technology

[0002] With the widespread deployment of IoT technology and the rapid growth of terminal device types, intelligent sensing nodes are gradually being applied in key scenarios such as intelligent manufacturing, smart agriculture, environmental monitoring, and edge computing. Because these devices operate in complex, heterogeneous edge environments characterized by strong electromagnetic interference, limited energy consumption, high network latency, and drastic environmental changes, minor, latent faults such as hardware aging, firmware anomalies, data drift, and protocol conflicts frequently occur. These faults often manifest as unstable noise in their early stages, making them easily masked by system filtering algorithms or misinterpreted as random fluctuations.

[0003] Traditional IoT fault detection technologies mostly rely on rule base matching, finite state machine analysis, or shallow models trained through manual annotation, which have the following technical bottlenecks: It is impossible to accurately capture microscopic signal mutations caused by quantum fluctuations and metastable jumps in millisecond-level communication, making it difficult to detect early fault signs of "not completely failed but already behaving abnormally" in a timely manner; current methods mostly use sliding window trend extrapolation or exponential weighted moving average, lacking the ability to model the behavioral patterns of high-frequency nonlinear failure chains; even if a node fails, most solutions still rely on manual restart or network operation and maintenance scheduling, failing to achieve closed-loop control of "active diagnosis-self-organized repair" for edge autonomy.

[0004] Especially in the deep-ground communication scenario of smart agriculture, since the equipment is often deployed in the depths of the field without signal shielding and protection and without human maintenance, it is in a working state of "weak power supply + strong interference" for a long time. The unpredictability of hidden faults can easily lead to the break of critical data links, which in turn paralyzes the sensing system of the entire agricultural area and brings large-scale economic losses. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent fault prediction and self-repair method for Internet of Things (IoT) terminals to address the shortcomings in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent fault prediction and self-repair of IoT terminals, comprising: Acquire multi-source operational data of the target IoT terminal within a fixed operating cycle; The multi-source operational data is subjected to noise suppression and morphological reconstruction to construct a multimodal feature time series tensor. The tensor is used to extract a set of weak steady-state perturbation feature vectors based on a time sliding window. The feature vector set is used to calculate the feature drift rate through a high-dimensional space perturbation evolution model, and a probability map of the significance of fault symptoms is generated based on a preset micro-failure threshold. Based on the anomalous time points with the highest significance in the probability graph, a lightweight autoregressive-attention fusion prediction network is dynamically activated to construct a short-term behavior mapping of the running state before and after the time point, and the evolution-sensitive sub-path is extracted from the mapping as the failure chain prediction vector. The predicted vectors are mapped with the equipment failure knowledge graph with high similarity to output the predicted probability distribution of the suspected fault initiation location and its fault type. If the predicted probability exceeds the device's set safety threshold, a recovery strategy matching the predicted fault type will be invoked. After implementing the repair strategy, monitor in real time whether the fault symptoms decrease, and update the equipment operating status label and fault strategy success rate evaluation matrix within a continuous time window when the characteristics continue to return to below the safety threshold, thus forming a repair feedback loop.

[0007] Preferably, the multi-source operating data includes, but is not limited to, transient power supply voltage, current mutation sequence, communication delay fluctuation trajectory, memory occupancy time sequence, and abnormal log fragments.

[0008] Preferably, the construction of the multimodal feature temporal tensor includes: The running data from different data sources are normalized separately. The normalization process adopts the minimum-maximum normalization method to make each type of data uniformly distributed in the interval [0,1]. A multi-level wavelet threshold denoising algorithm is applied to the normalized data, in which the Daubechies fourth-order wavelet basis function is selected, and the high-frequency wavelet coefficients are nonlinearly compressed to filter out random interference signals and retain abrupt change edge information. Multimodal feature temporal tensors are constructed based on a sliding time window mechanism. Specifically, normalized and denoised multi-source data are spliced ​​together within a time window of length T to generate a three-dimensional tensor of the form [d×T×N], where d is the number of modal dimensions, T is the length of the time window, and N is the number of samples under the sliding window step. For each time window tensor, a temporal gradient variability rate detection algorithm is used to extract significant perturbation points. Feature vectors that satisfy the perturbation variability rate below a set threshold are marked as weak steady-state perturbation vectors and collected to form a feature vector set.

[0009] Preferably, the step of calculating the feature drift rate of the feature vector set using a high-dimensional space perturbation evolution model includes: The extracted weak steady-state perturbation feature vector set is input into the embedding space, which is constructed using the t-distributed random neighborhood embedding algorithm. In the embedding space, the rate of change of the Euclidean distance of the centroid of the feature vector within a continuous time window is calculated as the basic measure of the feature drift rate, where the centroid is defined as the mean coordinate point of all feature vectors in the current window in the embedding space. The drift rate time series is modeled using an exponential growth fitting function to form the drift trend curve, and the growth slope of the curve is analyzed by derivative to extract high growth segments as potential fault evolution paths. A micro-failure threshold ε is set. If the feature drift rate continuously exceeds the threshold ε within a certain time period, a probability map of the significance of fault symptoms is generated based on the multi-window statistical significance score.

[0010] Preferably, the step of dynamically activating a lightweight autoregressive-attention fusion prediction network based on the abnormal time point with the highest significance in the probability graph, and constructing a short-term behavior mapping of the running state before and after the time point, includes: In the probability map of fault symptoms, the time point T_max with the highest significance probability value is selected, and the multimodal feature tensors within a fixed window length before and after it are extracted as prediction input sequences, and short-term behavior input matrices are constructed from the sequences. A lightweight autoregressive-attention fusion prediction network is dynamically loaded. The network includes: a one-layer causal one-dimensional convolutional autoregressive module for modeling local temporal dependencies, and a one-layer channel-temporal dual attention module for modeling intermodal coupling relationships and evolutionary importance scores. The input matrix is ​​input into the fusion prediction network, and the feature state prediction results and attention weight distribution within the future window are generated simultaneously through parallel computation to form a behavior mapping map of the short-term running state. The behavior mapping map represents the dynamic trajectory of the state changing with time and modality in the form of a three-dimensional tensor. In the behavior map, based on the attention weight heat and feature mutation path extraction algorithm, evolution-sensitive path segments that continuously cross multiple modalities and whose attention concentration is greater than a preset threshold are identified and used as failure chain prediction vectors.

[0011] Preferably, the step of performing a high-similarity topological mapping between the predicted vector and the device failure knowledge graph includes: Construct a knowledge graph of equipment failure. The knowledge graph represents the nodes of each sub-module of the equipment, the functional connection relationship and the historical fault type in a graph structure. Each node contains a module attribute vector, a status label and an associated historical fault weight factor. The edges represent the structural dependencies or signal links between modules. The fault chain prediction vector is embedded and encoded using a structure-aware graph neural network. The graph neural network generates the topological embedding vector of each node by aggregating the state of neighboring nodes and edge weight information, while preserving local connectivity features and global semantic context. The cosine similarity between the predicted vector and the embedding vector of each node in the graph is calculated, and a weighted similarity scoring matrix is ​​constructed by combining the path length correction factor. Several nodes with the highest scores are selected as a candidate set of suspected fault initiation nodes. Based on the candidate node set, Bayesian probability fusion inference is performed according to the distribution of historical fault types and the current prediction vector pattern distribution of the nodes to output the predicted probability distribution of the suspected fault initiation location and its fault type.

[0012] Preferably, the real-time monitoring of whether the fault symptom characteristics decrease, and within a continuous time window during which the characteristics continuously return to below the safety threshold, includes: Within a specified observation period after the implementation of the repair strategy, fault symptom feature sequences of the same category before repair are collected in real time. The symptom features include, but are not limited to, communication delay variation rate, voltage fluctuation frequency and abnormal memory growth rate. A comparison vector set is constructed using the same feature extraction method as the prediction stage. The feature vector set is used to determine whether the feature values ​​continuously regress to below a preset safety threshold based on a time sliding window mechanism. If all monitored features remain within the safe threshold range for multiple consecutive time windows, the current device operating status label will be updated from "Under Repair" to "Normal", and the label switching time will be recorded in the device operating status table. Simultaneously, the effectiveness of the currently executed repair strategy on the device is evaluated, and the success rate evaluation matrix between fault type and repair strategy is updated. The matrix has fault type as the row and strategy type as the column, recording the effectiveness label and cumulative success rate after each repair.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs an intelligent fault prediction and self-repair method for IoT terminals that integrates fault symptom prediction, embedded modeling, saliency scoring, path analysis, knowledge graph matching, and policy self-loop feedback. This method achieves real-time perception of device operating status, accurate identification and classification of early fault trends, and, upon triggering a fault risk, intelligently invokes repair strategies based on historical knowledge base and prediction models. After system stability verification, the method completes policy evaluation and optimization updates. This approach effectively overcomes the limitations of traditional methods that rely on rule matching and manual maintenance.

[0014] 2. The high-dimensional perturbation modeling and embedding alignment method introduced in this invention can extract potential fault paths from extremely subtle changes in behavioral features; combined with the fault knowledge graph topology mapping technology of graph neural networks, it realizes high-confidence identification of the fault initiation location and type in complex structure equipment; coupled with the self-repair mechanism based on probability judgment and feature regression feedback, it enables the system to have sustainable evolution and learning capabilities, providing practical technical support for the intelligent health management of large-scale distributed IoT terminals. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a mind map of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 As shown in this embodiment, an intelligent fault prediction and self-repair method for IoT terminals includes: Acquire multi-source operational data of the target IoT terminal within a fixed operating cycle; The multi-source operational data is subjected to noise suppression and morphological reconstruction to construct a multimodal feature time series tensor. The tensor is used to extract a set of weak steady-state perturbation feature vectors based on a time sliding window. The feature vector set is used to calculate the feature drift rate through a high-dimensional space perturbation evolution model, and a probability map of the significance of fault symptoms is generated based on a preset micro-failure threshold. Based on the anomalous time points with the highest significance in the probability graph, a lightweight autoregressive-attention fusion prediction network is dynamically activated to construct a short-term behavior mapping of the running state before and after the time point, and the evolution-sensitive sub-path is extracted from the mapping as the failure chain prediction vector. The predicted vectors are mapped with the equipment failure knowledge graph with high similarity to output the predicted probability distribution of the suspected fault initiation location and its fault type. If the predicted probability exceeds the device's set safety threshold, a recovery strategy matching the predicted fault type will be invoked. After implementing the repair strategy, monitor in real time whether the fault symptoms decrease, and update the equipment operating status label and fault strategy success rate evaluation matrix within a continuous time window when the characteristics continue to return to below the safety threshold, thus forming a repair feedback loop.

[0019] In this embodiment, multi-source operational data refers to a collection of data reflecting the operational status of the target terminal, obtained through various sensors or system calls, encompassing physical, electrical, network, and software dimensions. This includes, but is not limited to, transient power supply voltage, current surge sequences, communication delay fluctuation trajectories, memory occupancy time series, and abnormal log fragments. The following outlines the data acquisition methods and implementation steps for each data dimension.

[0020] Transient power supply voltage acquisition: A low-power ADC (analog-to-digital converter) module deployed on the terminal power supply path samples the voltage of the terminal power supply pin at 1-millisecond intervals within a specified operating cycle. To improve the stability and anti-interference capability of the sampled data, an adaptive median filtering algorithm is used for preliminary data cleaning. Specifically, within each time window, the median value of the current sampling sequence is extracted as the representative voltage value at that time point, effectively eliminating voltage spikes introduced by high-frequency pulses or electromagnetic interference.

[0021] Current mutation sequence extraction: Current signals are synchronously sampled in the power supply loop to construct a current change sequence. Then, a differential operator is applied to this sequence to calculate the difference between two consecutive sampling points to identify potential "mutation points." When the difference at a certain time point exceeds a preset mutation threshold (e.g., the current change rate exceeds 0.3 A / ms), it is marked as a "current mutation event." Further analysis using a sliding window zero-crossing rate is used to eliminate occasional jitter caused by noise, generating a stable current mutation trajectory.

[0022] Construction of Communication Delay Fluctuation Trajectory: A point-to-point network probing mechanism is constructed using the lightweight UDP Ping protocol (suitable for resource-constrained environments). Delay measurement data packets are periodically sent to the edge gateway, and the round-trip time (RTT) is recorded. The continuous RTT value sequence is smoothed using a multi-scale exponentially weighted moving average (MS-EWMA) algorithm. This algorithm applies different exponential decay factors at each scale, thus taking into account both short-term jumps and long-term trends, providing a reliable modeling basis for communication layer stability fluctuations.

[0023] Memory Usage and Anomaly Log Extraction: By calling the diagnostic interfaces exposed by the IoT terminal operating system (such as / proc / meminfo and / var / log / syslog files in Linux systems), the total memory usage, used space, and cache ratio are read every second to construct a continuous time series of memory usage. Simultaneously, error stack traces, abnormal termination (such as segmentation faults), and process restart records are extracted from the log files to form a set of anomaly event fragments associated with timestamps, serving as an important reference indicator for abnormal system behavior.

[0024] To ensure the quality of the collected results and the reliability of subsequent model training, this invention sets up multiple "anomaly filtering rules" and "sampling fault tolerance mechanisms": when the number of consecutive sampling failures exceeds 5, or the noise intensity of a certain data source is higher than the preset signal-to-noise ratio threshold (such as less than 20 dB), the data segment will be automatically marked as "suspicious data" and trigger the backup data source or compensation model to fill it.

[0025] Because the physical quantities involved in multi-source operational data vary greatly in units and magnitudes (e.g., voltage in volts, current in amperes, memory usage as a percentage, communication latency in milliseconds), direct feature fusion or modeling can easily lead to problems such as model convergence difficulties and feature bias due to scale inconsistencies. Therefore, this invention first normalizes various types of data to ensure the comparability and equal weighting of feature dimensions.

[0026] Normalization uses the Min-Max Normalization method, which processes data as follows: For any type of time series data, let its minimum value within the current time window be A_min, its maximum value be A_max, its original value be A_i, and its normalized value A_norm be calculated by the following formula: A_norm = (A_i - A_min) ÷ (A_max - A_min); This method maps all input values ​​to the interval [0,1], adapts to subsequent nonlinear compression and tensor splicing operations, and has the characteristics of lightweight computation and no bias.

[0027] After normalization, noise suppression is required for each type of running data sequence to filter out high-frequency transient interference signals while preserving as much of the true weak disturbance characteristics as possible. Traditional filters such as mean filtering and median filtering can reduce noise, but they cannot identify variant edge features and are prone to missing the "mutation signs" of early fault behavior.

[0028] Therefore, this invention introduces a multi-level wavelet threshold denoising algorithm as the core technology for noise suppression. The specific steps are as follows: The wavelet basis function is selected as the fourth-order Daubechies function (db4), which has both compact support and edge preservation capabilities, and is suitable for preserving abrupt changes in the signal. Perform a three-level wavelet decomposition on each type of data sequence to obtain the corresponding set of wavelet coefficients; For high-frequency wavelet coefficients, a soft threshold compression method is used. That is, when the absolute value of a coefficient is less than a set threshold λ, it is set to zero; if it is greater than the threshold, it is linearly compressed according to the difference of distance λ. The threshold λ can be obtained through an empirical formula. Estimate the value, where σ is the standard deviation of the high-frequency coefficients and n is the number of samples. Finally, the denoised signal sequence is reconstructed using inverse wavelet transform, which serves as the basic morphological data for the reconstruction.

[0029] This wavelet denoising process can significantly suppress high-frequency pseudo-random fluctuations (such as transient electromagnetic interference, sudden communication anomalies, etc.), while preserving the slow deformation, sudden rise or fall trend caused by micro-failure mechanisms.

[0030] To achieve joint modeling of multi-source features in the time domain, this invention proposes a multimodal feature tensor construction method based on a sliding time window mechanism.

[0031] The specific construction method is as follows: Set the sliding time window length T (in seconds or the number of data frames, e.g., T=50) and the step size N (e.g., N=5) to represent the step size of each sliding window movement; Within each sliding window, the data sequence within that window is extracted for each type of data that has undergone noise reduction processing (such as voltage, current, delay, memory usage, etc.). All normalized and denoised multi-source sequences within the same time window are synchronously concatenated along the time axis to construct a three-dimensional feature tensor of the form [d×T×M], where: d represents the number of modalities (i.e., the number of data types); T represents the time step per window; M represents the number of times the sliding window is sampled (i.e., the number of samples); The resulting tensors can be directly used as input for subsequent deep models, supporting parallel processing of multi-dimensional features, such as temporal convolution (1D-CNN), attention mechanisms, or graph structure learning.

[0032] The construction method of this tensor has temporal structural integrity and can effectively capture the coupling relationship and dynamic change trend between multiple modalities within the same time period.

[0033] After forming the multimodal temporal tensor, this invention further proposes a perturbation detection algorithm based on the temporal gradient variability rate, which is used to extract the set of feature subvectors that are in a weak steady state but have generated abnormal perturbations.

[0034] The implementation method is as follows: In the time dimension of each tensor, the gradient difference between adjacent time points under each mode is calculated to form a gradient sequence; Set a perturbation variation rate threshold η (e.g., η=0.05) to measure the magnitude of gradient fluctuations in the sequence; If, within a certain sliding time period, the gradient variation rate under a certain mode is consistently lower than η, but its trend shows a continuous deviation from the steady state (i.e., non-zero mean drift), then the data segment is determined to be a "weak steady-state disturbance region". The corresponding tensor quantum vectors are extracted from the perturbation region to form a "weak steady-state perturbation feature vector set", which is used for subsequent failure chain identification and prediction modeling.

[0035] The weakly steady-state perturbation feature vector set extracted through the preceding steps is input into a high-dimensional embedding space to achieve structural reconstruction of the features in the nonlinear space and modeling of the visualized evolution trajectory. The construction of the embedding space is based on the t-distributed random neighborhood embedding algorithm.

[0036] t-SNE is an embedding method for dimensionality reduction and preserving local structure. Its main function is to map the original high-dimensional feature vectors to a two-dimensional or three-dimensional visualization space while preserving the nearest neighbor relationships between sample points to the greatest extent possible. The steps are as follows: For each high-dimensional feature vector, calculate its conditional probability distribution with other vectors, representing the probability that a vector is a "neighbor" of another vector; In a low-dimensional space, a similar probability distribution can be constructed; By minimizing the Kullback-Leibler divergence between the two, an optimal embedding is found that keeps the local adjacency relationships in the high-dimensional space consistent in the low-dimensional space. This invention maps all feature vectors to a two-dimensional embedding space using t-SNE, obtaining the coordinate point representation of each feature in that space, which facilitates subsequent spatial evolution modeling.

[0037] In the constructed embedding space, this invention calculates the geometric centroid of the feature point set within each sliding time window, and uses the rate of change of the Euclidean distance of the centroid within a continuous time window as the core indicator of the system feature evolution, which is defined as the "feature drift rate".

[0038] The specific implementation method is as follows: Set the sliding time window length to T and the window step size to S; Within each window, the mean coordinates of all feature vectors in the embedding space are calculated and used as the "centroid" of the current time window; For the centroid coordinates of two consecutive windows, calculate their Euclidean distance, denoted as D_i; The characteristic drift rate V_i is defined as the ratio of the centroid distance D_i between the i-th time window and the (i-1)-th time window to the time interval Δt, i.e.: The drift rate reflects the dynamic shift of the overall structure of the feature in the embedding space over time and is an important indicator for measuring the intensity of system state changes.

[0039] To further quantify the speed and direction of system state evolution, this invention inputs the drift rate time series calculated above into an exponential growth fitting function model to characterize the nonlinear growth trend in potential fault paths.

[0040] The exponential fitting function takes the following form: Where f(t) is the drift rate at time t, a is the initial drift rate, and b is the growth factor. By fitting the drift rate sequence using the least squares method, parameters a and b can be obtained, and a continuous drift trend curve can be constructed.

[0041] Next, the first derivative of the function is calculated at each time point, i.e.: The derivative represents the rate of change of the drift rate per unit time. If the derivative continues to rise and maintains a positive growth trend over multiple time windows, it indicates that the system state is deviating from its original stable trajectory at an accelerated pace, suggesting that fault symptoms are intensifying and evolving.

[0042] To extract key risk segments, this invention sets a derivative growth rate threshold δ (e.g., δ = 0.02) to determine "high-growth segments," i.e.: If f′(t) ≥ δ at n consecutive time points, then the time period is marked as a "potential fault evolution path".

[0043] After extracting the fault evolution path, this invention introduces a micro-failure threshold ε as a criterion for judging whether the system state crosses the safety boundary. Specifically, if the continuous drift rate V_i ≥ ε within a certain time period (e.g., setting ε = 0.1), it is considered that the characteristic of that segment has shown "significant deviation".

[0044] To avoid misjudgment caused by a single outlier, this invention employs a multi-window statistical significance scoring mechanism for map construction. The specific steps are as follows: Divide all time window drift rates into multiple overlapping statistical segments, each segment having a length of L; Calculate the percentage of samples P_i in each segment whose drift rate exceeds the threshold ε; Scoring is based on P_i and the set significance factor (e.g., the significance threshold is set to 50%). The scoring results of all sub-segments are normalized to generate a probability map on a one-dimensional time axis; The final output is a set of significance probability values ​​that change over time, forming a "fault symptom significance probability map". This map serves as an input guidance signal or attention mechanism weight reference for subsequent deep prediction models, and has the characteristics of being structured, quantifiable, and interpretable.

[0045] After completing the fault symptom significance probability map output by the preceding module, this invention first locates the time point with the highest significance probability value in the map, denoted as T_max. This time point is considered to be the potential critical failure point where the system exhibits the strongest evolution signal within the sliding time window.

[0046] Subsequently, taking T_max as the center, time segments of length L are selected forward and backward (e.g., L=10 time steps), and their corresponding multimodal feature tensor sequences are extracted to construct a unified prediction input matrix X with the dimension of: [number of modalities × time step length × feature dimension].

[0047] This input matrix serves as a key representation of the short-term behavior of the system before and after T_max, and is the direct input to the subsequent prediction network. To prevent inconsistent feature scales from affecting prediction accuracy, this invention performs normalization processing on all values ​​in the input matrix (using the min-max normalization method).

[0048] The prediction model described in this invention adopts a fusion design of "lightweight autoregressive network" and "channel-time dual attention module", which has the advantages of low computational complexity, strong feature modeling ability and high interpretability.

[0049] The model consists of the following two layers: The Autoregressive Causal Convolution Layer is used to capture the dependencies between the current moment and historical states in a time series. It uses a one-dimensional convolution method to perform autoregressive processing on the input data. The convolution kernel size is set to 3, the stride is 1, and a causal mask is used to prevent the leakage of future information and maintain the consistency of the time sequence.

[0050] The output is a predicted sequence within a local time window, with a lightweight structure that does not require complex parameter updates, making it suitable for resource-constrained terminals.

[0051] The Channel-Time Dual Attention module, to further enhance the model's ability to identify key features and key moments, introduces a dual attention mechanism, modeling feature weights from both the modal channel dimension and the time step dimension: Channel attention: By fusing the global average pooling and max pooling results of different modal channels, the results are input into a two-layer fully connected neural network, and the output channel weight vector is used to enhance the response of the fault-critical mode. Temporal attention: Perform a similar structure in the time dimension to identify the time window in which feature perturbations are most active and assign them high response weights.

[0052] The attention output matrix and the autoregressive convolution result are weighted and superimposed element-wise to form the final prediction result, which has dynamic adjustable characteristics.

[0053] The results of the fusion prediction network are concatenated with the original input matrix at the time axis level, and the predicted values ​​of each modality channel at each time step are visualized to form a three-dimensional tensor structure called the "short-time behavior map".

[0054] The mapping graph has the following dimensions: [number of modes × number of time steps × behavioral state feature dimension]. It can be regarded as a dynamic tensor sequence that describes the perturbation evolution trend of the system state in the multimodal space near the abnormal time point T_max.

[0055] The mapping diagram not only reflects the fluctuations of the state over time, but also reveals the synchronicity and evolutionary coupling paths between modes, providing a basic structure for the extraction of sensitive paths.

[0056] To identify the key behavioral paths most likely to trigger fault evolution, this invention introduces an evolution-sensitive sub-path extraction mechanism in the behavior map, based on attention response heat and feature variation intensity.

[0057] The specific steps are as follows: In the output of the attention fusion network, the attention weight values ​​of each modality channel at each time step are extracted to form a heatmap matrix; For the state features at each time step, the local variation rate is calculated, which is the Euclidean difference of the state features of adjacent time steps. Set a sensitivity threshold θ (e.g., θ = 0.7), and mark channel sequences with attention weights higher than the threshold and variation rates higher than the variation threshold ν (e.g., ν = 0.1) in a continuous time period as sensitive path segments; These path segments are sorted by time and concatenated into a complete evolution path, which is then used as the fault chain prediction vector and input into the subsequent fault knowledge graph matching module.

[0058] Equipment failure knowledge graphs are a data representation method based on structured graph models, designed to model the structural dependencies, functional interaction logic, and historical failure evolution paths among sub-modules within equipment. This knowledge graph adopts a directed graph structure G = (V, E, A), where: V represents the set of nodes, and each node v ∈ V corresponds to a specific sub-module or sensing unit of the device; E represents the set of edges, and each edge e ∈ E represents a signal link or dependency between modules, which has directionality; A represents the set of attributes, including node attributes and edge weight attributes.

[0059] Each node v contains the following information: Module attribute vector: It consists of various features such as module type encoding, function category, power consumption level, and I / O interface type; Status labels: Current or historically collected running status markers (such as normal, abnormal, degraded, etc.); Historical Fault Factor Weight Vector: Represents the fault types associated with this module in past samples and their frequency or weight, used for subsequent probability inference.

[0060] The properties of edge e include structural coupling strength (such as communication frequency and control command dependency) and time delay weight.

[0061] This map can be automatically constructed or manually labeled using multi-source data such as the device BOM structure, sensor topology configuration, and historical operation and maintenance records.

[0062] After obtaining the predicted vector of the fault chain to be identified, in order to realize the semantic space mapping between it and the nodes of each module in the equipment failure knowledge graph, this invention introduces a structure-aware graph neural network (GNN) for embedded encoding.

[0063] The core idea of ​​the GNN model is that the final representation vector of each node v is a weighted aggregation of its own features and the features of its neighboring nodes, which can form a node embedding that integrates structural information and contextual semantics after multiple iterations.

[0064] The specific implementation steps are as follows: Initialize the attribute vector of each node in the knowledge graph as a first-order embedding vector. ; In each round of propagation, the vector representation of node v The update is performed based on the representations of its neighboring nodes u ∈ N(v) and the weights w_uv of the connecting edges. The update formula is as follows: σ is the activation function. and Let Σ be a trainable weight matrix, where Σ represents a weighted summation over neighboring nodes.

[0065] After multiple rounds of propagation, the final node embedding representation is obtained. This is used for subsequent matching calculations with the predicted vector.

[0066] The prediction vector X_p will undergo a linear mapping transformation of the same dimension, corresponding to the nodes in the graph. Maintaining consistency across dimensions, proceed to the similarity matching stage.

[0067] III. Similarity Scoring Mechanism and Candidate Node Selection The similarity between the predicted vector X_p and the embedding vector h_v^T of each node is calculated. This invention uses cosine similarity as the basic similarity index, and its calculation formula is as follows: Where "·" represents the dot product operation, and "||·||" represents the vector magnitude. The closer this index is to 1, the higher the similarity.

[0068] Considering the influence of a node's structural position in the graph, its reachability and centrality determine its priority in the fault propagation path. Therefore, this invention further introduces a path length correction factor γ_v to penalize low-weight nodes that are far from the main control path.

[0069] The final similarity scoring formula is: Where α is the path penalty coefficient (e.g., α = 0.3), and γ_v represents the shortest path length from the core node of the graph to node v. Based on the final scoring results, the top K nodes (e.g., K = 3) are selected as the candidate set of suspected fault initiation nodes.

[0070] After obtaining the candidate node set, in order to determine its specific fault type, this invention introduces a Bayesian probabilistic fusion inference mechanism, combining the following two distributions: Node historical fault type distribution P(F|v): represents the probability that node v will experience various types of faults Fi in its history; The current pattern distribution P(X_p|Fi) of the predicted vector is calculated from the goodness of fit of the feature patterns of vector X_p to each class Fi in the training set.

[0071] According to Bayes' theorem, for each candidate node v and fault type Fi, the joint probability distribution is calculated as follows: Through normalization, the predicted probability distribution for each node-fault type pair is obtained. The final output includes: The most likely module node where the failure originated; Several fault types associated with it and their corresponding probability values.

[0072] This result not only supports fault identification at the system level, but can also be used to drive subsequent modules to perform operations such as fault cascade graph matching and autonomous repair strategy selection.

[0073] Each type of terminal device in the system has a predefined predictive response safety threshold, denoted as P_th. This threshold represents the system's maximum tolerance range for predicted fault probabilities. When the probability of a certain predicted fault type exceeds this threshold, it is considered that the fault is highly likely to have occurred or is about to occur, and immediate recovery actions must be taken.

[0074] The device safety threshold P_th can be set in two ways: Static configuration: set by the equipment manufacturer or during engineering deployment based on empirical values ​​(e.g., P_th = 0.7); Dynamic adjustment: The system automatically corrects itself based on historical prediction accuracy and false alarm rate during long-term operation, and uses a moving average update strategy for adjustment.

[0075] The setup method is as follows: If a certain type of fault Fi has a predicted probability greater than p and an actual occurrence probability exceeding x% in historical samples, then P_th = p × σ can be set, where σ is a risk mitigation factor (e.g., 0.9).

[0076] The set of failure probabilities output by the prediction module is denoted as The system checks the existence of each fault type sequentially: If the judgment result is true, then the recovery strategy matching and invocation process will begin.

[0077] To support automatic repair functionality, the system pre-defines a fault type-recovery strategy mapping table (or recovery strategy library) to store the association between different fault types and their corresponding recovery actions. The principles for building this library are as follows: Each mapping contains three elements: Fault type ID (e.g., F_COMM_LOSS indicates a communication failure); Policy categories (such as soft reboot, parameter rollback, etc.); Strategy execution template (including action instructions, delay time, dependent modules, etc.).

[0078] For example: ; Once the predicted probability exceeds the threshold, the system searches for a matching policy in the policy library based on the fault type Fi and loads its corresponding template.

[0079] Strategy execution includes the following steps: Policy loading: Retrieves the corresponding instruction set from the policy library based on the mapping relationship; Execution environment detection: Determine whether the current device status meets the policy execution conditions (such as whether it is under high temperature, heavy load, etc.); Strategy execution: Execute specific recovery actions through the system command interface or embedded control module; Status readback monitoring: Monitor changes in key parameters after policy execution, such as whether memory usage has decreased and whether communication latency has returned to normal. Successful marking and failure recovery: If key indicators recover, record as successful; if there is no improvement, switch to backup strategy or trigger upper-level manual early warning interface.

[0080] After the repair strategy is executed, the system enters a specified observation window (e.g., set to 60 to 300 seconds) to periodically collect key fault symptoms consistent before and after execution, in order to perform trend regression analysis. These symptoms include, but are not limited to, the following categories: Communication-related anomalies include: average network latency, jitter frequency, and packet loss rate. Electrical fluctuation characteristics: such as the frequency of short-cycle voltage fluctuations and the duration of current peaks; System resource characteristics: such as memory usage growth rate, number of thread restarts due to abnormalities; Sensor-related abnormal characteristics: such as sensor data drift rate and sensor response delay.

[0081] The sampling period for each feature is consistent with the previous modeling phase (e.g., once per second), and they are uniformly converted into a standardized feature vector V(t), which is the set of state vectors at time point t. Continuous sampling yields a sliding time series. .

[0082] To accurately determine whether a repair strategy has a positive impact on the system's operating status, this invention designs a multi-feature regression judgment algorithm based on a time sliding window. The core of this algorithm is that it not only requires a single feature to fall within the normal range, but also that multiple features remain stable within a safe range over a continuous time window to confirm the repair's effectiveness.

[0083] The specific judgment logic is as follows: For each type of feature, the system predefines a safety threshold range, which is determined through long-term operational data analysis and is typically set as follows: The safety threshold is the average value of the feature under steady-state conditions ± one standard deviation. For example, if the average memory growth rate is 0.01 and the standard deviation is 0.005 under steady-state conditions, then the safety threshold is [0.005, 0.015].

[0084] A sliding time window mechanism is introduced, with the continuous decision window length set to W (e.g., W = 5 seconds) and the step size to 1 second; In each sliding window, determine whether all monitored features fall within their respective safety threshold ranges; If the above conditions are met for K consecutive sliding windows (e.g., K = 3), the system characteristics are considered to have continuously returned to the normal state, and the repair behavior is considered successful.

[0085] Once the repair is deemed effective, the system immediately updates the device's current operating status label from "Repairing" to "Normal." This label is written to the device operating status table for subsequent module reference and overall system health assessment.

[0086] Tag updates also record the following metadata: Tag switching timestamp; Summary of the repair strategy type and executed commands; Total time taken for feature recovery (total time from policy execution to continuous normal operation). Whether this state transition was triggered by an automatic mechanism (as a learning marker).

[0087] Status label updates are not only used for system visualization, but also serve as feedback signals for the subsequent adaptive training of the prediction model.

[0088] To optimize strategy prioritization in the face of similar future failures, this invention introduces a strategy historical performance recording method based on a two-dimensional evaluation matrix of equipment failure type and repair strategy. The construction logic of this matrix is ​​as follows: The rows represent the historically predicted fault type numbers (such as F_MEM_LEAK, F_COMM_LOSS, etc.). The types of repair strategies supported by the system are listed in columns (such as soft reboot, parameter rollback, module hot replacement, etc.). Each cell records the following metrics: Total number of executions: N; Success count S (i.e., the number of times the system feature regression was successful); Strategy average recovery time T_avg; Success rate P = S÷N.

[0089] After each repair action is completed and confirmed to be effective, the system automatically updates the corresponding cells of the above matrix and uses this matrix as input for the next round of strategy recommendation ranking.

[0090] For example: ; This matrix is ​​not only retained at the device level, but can also be aggregated at the platform level for larger-scale scenario analysis such as group strategy optimization and cross-device model migration.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent fault prediction and self-repair of IoT terminals, characterized in that: include: Acquire multi-source operational data of the target IoT terminal within a fixed operating cycle; The multi-source operational data is subjected to noise suppression and morphological reconstruction to construct a multimodal feature time series tensor. The tensor is used to extract a set of weak steady-state perturbation feature vectors based on a time sliding window. The feature vector set is used to calculate the feature drift rate through a high-dimensional space perturbation evolution model, and a probability map of the significance of fault symptoms is generated based on a preset micro-failure threshold. Based on the anomalous time points with the highest significance in the probability graph, a lightweight autoregressive-attention fusion prediction network is dynamically activated to construct a short-term behavior mapping of the running state before and after the time point, and the evolution-sensitive sub-path is extracted from the mapping as the failure chain prediction vector. The predicted vectors are mapped with the equipment failure knowledge graph with high similarity to output the predicted probability distribution of the suspected fault initiation location and its fault type. If the predicted probability exceeds the device's set safety threshold, a recovery strategy matching the predicted fault type will be invoked. After implementing the repair strategy, monitor in real time whether the fault symptoms decrease, and update the equipment operating status label and fault strategy success rate evaluation matrix within a continuous time window when the characteristics continue to return to below the safety threshold, thus forming a repair feedback loop.

2. The method for intelligent fault prediction and self-repair of an IoT terminal according to claim 1, characterized in that: The multi-source operational data includes, but is not limited to, transient power supply voltage, current mutation sequence, communication delay fluctuation trajectory, memory occupancy time sequence, and abnormal log fragments.

3. The method for intelligent fault prediction and self-repair of an IoT terminal according to claim 2, characterized in that: The construction of the multimodal feature temporal tensor includes: The running data from different data sources are normalized separately. The normalization process adopts the minimum-maximum normalization method to make each type of data uniformly distributed in the interval [0,1]. A multi-level wavelet threshold denoising algorithm is applied to the normalized data, in which the Daubechies fourth-order wavelet basis function is selected, and the high-frequency wavelet coefficients are nonlinearly compressed to filter out random interference signals and retain abrupt change edge information. Multimodal feature temporal tensors are constructed based on a sliding time window mechanism. Specifically, normalized and denoised multi-source data are spliced ​​together within a time window of length T to generate a three-dimensional tensor of the form [d×T×N], where d is the number of modal dimensions, T is the length of the time window, and N is the number of samples under the sliding window step. For each time window tensor, a temporal gradient variability rate detection algorithm is used to extract significant perturbation points. Feature vectors that satisfy the perturbation variability rate below a set threshold are marked as weak steady-state perturbation vectors and collected to form a feature vector set.

4. The method for intelligent fault prediction and self-repair of an IoT terminal according to claim 3, characterized in that: The step of calculating the feature drift rate using a high-dimensional spatial perturbation evolution model for the feature vector set includes: The extracted weak steady-state perturbation feature vector set is input into the embedding space, which is constructed using the t-distributed random neighborhood embedding algorithm. In the embedding space, the rate of change of the Euclidean distance of the centroid of the feature vector within a continuous time window is calculated as the basic measure of the feature drift rate, where the centroid is defined as the mean coordinate point of all feature vectors in the current window in the embedding space. The drift rate time series is modeled using an exponential growth fitting function to form the drift trend curve, and the growth slope of the curve is analyzed by derivative to extract high growth segments as potential fault evolution paths. A micro-failure threshold ε is set. If the feature drift rate continuously exceeds the threshold ε within a certain time period, a probability map of the significance of fault symptoms is generated based on the multi-window statistical significance score.

5. The method for intelligent fault prediction and self-repair of an IoT terminal according to claim 4, characterized in that: The method involves dynamically activating a lightweight autoregressive-attention fusion prediction network based on the most significant anomalous time point in the probability graph, and constructing a short-term behavior mapping of the operating state before and after the time point, including: In the probability map of fault symptoms, the time point T_max with the highest significance probability value is selected, and the multimodal feature tensors within a fixed window length before and after it are extracted as prediction input sequences, and short-term behavior input matrices are constructed from the sequences. A lightweight autoregressive-attention fusion prediction network is dynamically loaded. The network includes: a one-layer causal one-dimensional convolutional autoregressive module for modeling local temporal dependencies, and a one-layer channel-temporal dual attention module for modeling intermodal coupling relationships and evolutionary importance scores. The input matrix is ​​input into the fusion prediction network, and the feature state prediction results and attention weight distribution within the future window are generated simultaneously through parallel computation to form a behavior mapping map of the short-term running state. The behavior mapping map represents the dynamic trajectory of the state changing with time and modality in the form of a three-dimensional tensor. In the behavior map, based on the attention weight heat and feature mutation path extraction algorithm, evolution-sensitive path segments that continuously cross multiple modalities and whose attention concentration is greater than a preset threshold are identified and used as failure chain prediction vectors.

6. The method for intelligent fault prediction and self-repair of an IoT terminal according to claim 5, characterized in that: The step of performing a high-similarity topological mapping between the predicted vector and the device failure knowledge graph includes: Construct a knowledge graph of equipment failure. The knowledge graph represents the nodes of each sub-module of the equipment, the functional connection relationship and the historical fault type in a graph structure. Each node contains a module attribute vector, a status label and an associated historical fault weight factor. The edges represent the structural dependencies or signal links between modules. The fault chain prediction vector is embedded and encoded using a structure-aware graph neural network. The graph neural network generates the topological embedding vector of each node by aggregating the state of neighboring nodes and edge weight information, while preserving local connectivity features and global semantic context. The cosine similarity between the predicted vector and the embedding vector of each node in the graph is calculated, and a weighted similarity scoring matrix is ​​constructed by combining the path length correction factor. Several nodes with the highest scores are selected as a candidate set of suspected fault initiation nodes. Based on the candidate node set, Bayesian probability fusion inference is performed according to the distribution of historical fault types and the current prediction vector pattern distribution of the nodes to output the predicted probability distribution of the suspected fault initiation location and its fault type.

7. The method for intelligent fault prediction and self-repair of an IoT terminal according to claim 1, characterized in that: The real-time monitoring of whether fault symptom characteristics decrease, and within a continuous time window during which the characteristics continuously return to below the safety threshold, includes: Within a specified observation period after the implementation of the repair strategy, fault symptom feature sequences of the same category before repair are collected in real time. The symptom features include, but are not limited to, communication delay variation rate, voltage fluctuation frequency and abnormal memory growth rate. A comparison vector set is constructed using the same feature extraction method as the prediction stage. The feature vector set is used to determine whether the feature values ​​continuously regress to below a preset safety threshold based on a time sliding window mechanism. If all monitored features remain within the safe threshold range for multiple consecutive time windows, the current device operating status label will be updated from "Under Repair" to "Normal", and the label switching time will be recorded in the device operating status table. Simultaneously, the effectiveness of the currently executed repair strategy on the device is evaluated, and the success rate evaluation matrix between fault type and repair strategy is updated. The matrix has fault type as the row and strategy type as the column, recording the effectiveness label and cumulative success rate after each repair.

Images