A computer hardware troubleshooting method based on industrial big data fusion

Abstract

The application discloses a computer hardware troubleshooting method based on industrial big data fusion, and relates to the technical field of equipment fault diagnosis. The method comprises the following steps: collecting and preprocessing multi-modal data of industrial computer hardware; performing real-time anomaly detection at the edge, and dynamically scheduling computing tasks based on network bandwidth, edge computing resource utilization and fault urgency; extracting the spatiotemporal correlation features of hardware sensor data in the cloud, using an industrial pre-trained BERT model to extract the semantic and causal features of system logs, and performing three-dimensional fusion through a cross-attention mechanism; performing explainable fault diagnosis and root cause positioning based on the fused features; and finally collecting maintenance feedback information, combining an improved IA-EWC incremental learning algorithm to complete knowledge sedimentation and model iterative updating. The application solves the problems of weak data fusion capability, poor anti-interference capability and easy disaster forgetting in the existing industrial computer troubleshooting technology.

Description

Technical Field

[0001] This invention relates to the field of industrial big data and equipment fault diagnosis technology, and in particular to a method for troubleshooting computer hardware faults based on the fusion of industrial big data. Background Technology

[0002] Industrial computers, primarily comprising specialized computing devices such as industrial PCs, industrial servers, and edge computing nodes, serve as the core infrastructure for smart factories, automated production lines, and industrial data centers. They undertake core tasks including equipment control, data acquisition, computational analysis, and instruction issuance. Their operational stability directly determines the continuity and safety of industrial production. Unlike commercial computers, industrial computers operate in harsh industrial environments characterized by high electromagnetic interference, large temperature and humidity fluctuations, and high dust levels. Core hardware components such as CPUs, memory, hard drives, power supplies, and motherboards are prone to aging, wear and tear, and sudden failures. Hardware failure can easily lead to production line shutdowns, data loss, and significant direct economic losses for enterprises. In severe cases, it can even damage supporting automated equipment and trigger safety accidents. Therefore, accurate, rapid, and intelligent hardware fault diagnosis technology is crucial for ensuring the long-term stable operation of industrial computers.

[0003] Currently, the industry mainly uses two methods to troubleshoot industrial computer hardware failures: manual experience-based troubleshooting and intelligent automated diagnostics. The traditional manual troubleshooting method relies on the experience of maintenance engineers, using methods such as disassembly for testing, parameter comparison, and log screening to locate faults. This method is inefficient, labor-intensive, and the fault diagnosis results are easily influenced by subjective experience. It is unsuitable for large-scale, distributed industrial computer clusters and is only suitable for small-scale, simple fault troubleshooting, thus having very limited applicability.

[0004] Currently, most mainstream intelligent fault diagnosis solutions rely on big data and deep learning algorithms. They achieve fault early warning and classification diagnosis by collecting data such as hardware operating parameters and system logs. Compared with the traditional manual mode, they have the advantages of high efficiency and strong objectivity. However, in the actual industrial application, this type of technology still has several unresolved technical defects that restrict diagnostic performance and implementation effectiveness. First, the multimodal data fusion layer is too simple, resulting in low accuracy in fault root cause localization. Most existing intelligent solutions only perform shallow feature splicing or fixed-weight fusion on time-series data from hardware sensors, which can only analyze the abnormal patterns of parameters in a single dimension. They cannot simultaneously explore the spatiotemporal correlation characteristics of hardware parameters and the semantic causal characteristics of system log text, making it difficult to sort out the inherent transmission logic of "parameter abnormality - log error - hardware failure". They cannot distinguish between primary faults and induced secondary faults. Under normal scenarios, the accuracy of fault root cause localization is less than 85%, and the overall false alarm rate remains in the range of 8% to 15%, which is insufficient to meet the needs of high-precision troubleshooting.

[0005] Second, the rigid computing resource scheduling mechanism makes it impossible to balance real-time performance and accuracy. Existing technologies generally adopt a fixed computing architecture, either uploading all data to the cloud for full-process diagnosis, which is prone to latency due to long-distance data transmission and cannot quickly respond to sudden high-risk faults such as power short circuits and hardware overheating; or delegating all computing tasks to edge nodes, which is limited by the computing power limit of edge devices and cannot carry out complex feature fusion and deep root cause analysis. Neither of these modes can adapt to the differentiated fault diagnosis needs in complex industrial scenarios.

[0006] Third, deep learning models lack interpretability and have insufficient credibility in practical applications. Most current mainstream fault diagnosis models are end-to-end black-box models, only able to output fault type and probability, failing to demonstrate the model's decision-making basis, core judgment characteristics, and fault propagation paths to operations and maintenance personnel. Operations and maintenance personnel cannot intuitively understand the diagnostic logic and still need to conduct secondary manual verification, which not only negates the efficiency advantages of intelligent diagnosis but also fails to support engineers in completing fault debriefing and experience accumulation.

[0007] Fourth, the robustness of data preprocessing is insufficient, and its ability to adapt to harsh industrial environments is weak. External factors such as strong electromagnetic interference and voltage fluctuations in industrial settings can easily lead to problems such as impulse noise and missing data in the collected data. Existing preprocessing algorithms are mostly general-purpose algorithms that have not been adapted and optimized for the characteristics of industrial computer data. They cannot effectively filter interference noise, and the accuracy of missing data completion is poor. Inferior input data further reduces the stability of model diagnosis.

[0008] Fifth, the model iteration mechanism is imperfect, making it prone to catastrophic forgetting and performance drift. Existing technologies mostly use offline batch retraining to update models, which has a long update cycle and consumes a lot of resources. At the same time, it does not consider the performance aging and drift problem caused by long-term hardware operation, and conventional incremental learning algorithms are not adapted to industrial fault scenarios, and cannot take into account the learning needs of new and old fault modes. When learning new fault samples, it is easy to forget historical fault features, and the model's diagnostic accuracy will continue to decline after long-term operation.

[0009] In summary, at present, there is no integrated troubleshooting solution that can simultaneously address issues such as weak data fusion capabilities, rigid resource scheduling, black-box modeling, poor anti-interference capabilities, and model iteration defects. The industry urgently needs a troubleshooting method for industrial computer hardware that is adaptable to complex industrial environments, balances real-time performance and accuracy, has interpretability, and can be independently iteratively optimized. Summary of the Invention

[0010] To address the aforementioned issues, this application proposes a method for troubleshooting industrial computer hardware based on spatiotemporal-semantic 3D fusion. This method enables early warning, precise location, and root cause analysis of industrial computer hardware faults, while balancing real-time performance and accuracy. It also solves the problem of model black-box architecture and adapts to the aging characteristics of equipment performance through adaptive iterative updates. The method includes the following steps: A method for troubleshooting computer hardware faults based on industrial big data fusion includes the following steps: S1. Collect and perform adaptive robust preprocessing on multimodal data of industrial computer hardware, including hardware sensor data, system log data, performance index data and environmental data; S2. Deploy a lightweight anomaly detection model on edge computing nodes to perform real-time anomaly detection, and dynamically schedule subsequent computing tasks based on fault urgency, network bandwidth, and edge computing resource utilization. S3. In the cloud, perform spatiotemporal correlation feature extraction on the preprocessed sensor data, perform semantic causal feature extraction on the preprocessed system log data, and perform three-dimensional fusion of the two features through a spatiotemporal-semantic cross-attention mechanism. S4. Input the fused feature vector into the ensemble classifier for fault type identification, calculate the feature contribution using the SHS1P value, and generate an interpretable fault diagnosis report and root cause analysis through feature-rule positive mapping. S5. Collect maintenance feedback information to accumulate fault knowledge, and achieve incremental updates of the model through rule-feature back mapping and elastic weight consolidation algorithm.

[0011] Preferably, the adaptive robust preprocessing described in S1 specifically includes: S11. Each key parameter of the five core components—CPU, memory, hard drive, power supply, and motherboard—is redundantly collected using three independent sensors, and all data is uniformly stamped with a high-precision timestamp with an error of less than 1ms. S12. An adaptive noise filtering algorithm combining sliding window standard deviation and wavelet transform thresholding is adopted to automatically identify and filter impulse noise generated by industrial electromagnetic interference. S13. When the missing data length is less than or equal to 5 sampling points, linear interpolation is used to complete the data. When the missing data length is greater than 5 sampling points, conditional generative adversarial network CGS1N is used to generate complete data that conforms to the operating rules of the device. S14. Use regular expressions to extract timestamps, event types, error codes, and key parameters from system logs, and transform unstructured text into a structured event sequence.

[0012] Preferably, the lightweight anomaly detection model described in S2 is an isolated forest model that has undergone principal component analysis feature dimensionality reduction and 8-bit integer quantization.

[0013] Preferably, the dynamic task scheduling mechanism described in S2 is as follows: S21. When a serious fault is detected, regardless of the network condition, a local audible and visual alarm will be triggered immediately at the edge and non-critical loads will be cut off, with a response time of less than 500ms. S22. When the network bandwidth is less than 1MS2ps, all computing tasks are executed at the edge. S23. When the network bandwidth is greater than 10MS2ps and the edge CPU utilization is less than 80%, upload the in-depth analysis tasks for medium and minor faults to the cloud for execution. S24. When the edge CPU utilization rate is greater than 80%, upload all non-real-time computing tasks to the cloud for execution.

[0014] Preferably, the spatiotemporal correlation feature extraction described in S3 specifically includes: An industrial computer hardware topology graph is constructed with hardware components as nodes and physical connections and data interaction relationships as edges. Preprocessed sensor time-series data is input into the spatiotemporal graph convolutional network ST-GCN. The spatial dimension correlation is captured through graph convolutional layers, and the temporal dimension change trend is captured through gated recurrent units (GRUs). The output spatiotemporal feature vector has a dimension of 256.

[0015] Preferably, the semantic causal feature extraction described in S3 specifically includes: The structured system log event sequence is input into the S2ERT model, which is pre-trained using 1 million industrial computer log data for domain adaptation. A causal attention mechanism is introduced to automatically identify event pairs with causal relationships in the logs, and the output is a semantic feature vector with a dimension of 256.

[0016] Preferably, the computation process of the spatiotemporal-semantic cross-attention mechanism described in S3 is as follows: ; ; Among them, F st F is the spatiotemporal feature vector output by the spatiotemporal graph convolutional network. sem This refers to the semantic feature vector output by a pre-trained BERT model for industrial applications, where S4_k is the feature dimension, Q is the attention query vector, K is the attention key vector, V is the attention value vector, and K... T Let d be the transpose of the key vectors. k Let F be the dimension of the key vector, softmax(·) be the normalization exponential function, and F be the dimension of the key vector. fusion This is the global feature vector resulting from the cross-fusion of spatiotemporal and semantic features.

[0017] Preferably, the explainable fault diagnosis and root cause localization described in S4 specifically include: S41. Input the fused feature vector into the ensemble classifier consisting of random forest, XGS2oost and multilayer perceptron, and output the fault type and the corresponding fault probability. S42. Calculate the contribution of each input feature to the diagnostic result using the SHS1P value, and select the top 10 key features in terms of contribution. S43. Match key features with a rule base containing more than 500 verified industrial computer hardware fault rules to generate a fault explanation in natural language. S44. Construct a fault propagation path diagram based on the hardware topology diagram and feature contribution, and trace back to find the upstream component with the largest contribution as the root cause of the fault.

[0018] Preferably, the rule-feature inverse mapping described in S5 specifically refers to: The industrial S2ERT model is used to encode the natural language fault rules input by engineers into semantic vectors, and then maps them to the same dimensional space as the fused features through a fully connected layer to generate feature vectors that the model can recognize.

[0019] Preferably, the incremental model update described in S5 specifically refers to: Each week, newly added fault cases and new fault rules are automatically collected, and the IS1-EWC algorithm with adaptive elastic weight consolidation for industrial scenarios is used to incrementally train the ensemble classifier and fusion module. The IS1-EWC algorithm introduces a fault service importance weighting mechanism, sparse Fisher matrix calculation, and dynamic adjustment of data credibility. It calculates the importance of key weights in old tasks and applies penalties, while preserving the correlation between weight reorganizations to avoid catastrophic forgetting problems.

[0020] In summary, the computer hardware fault diagnosis method based on industrial big data fusion of the present invention has the following advantages compared with traditional technologies: 1. By integrating spatiotemporal and semantic dimensions, the spatiotemporal evolution of hardware parameters and the semantic causal relationship of system logs are simultaneously explored, overcoming the deficiency of insufficient information in single-modal data, and significantly improving diagnostic accuracy and root cause localization capabilities. 2. Relying on a dynamic task scheduling mechanism driven by multi-dimensional parameters, high-risk faults are pushed down to the edge to achieve rapid early warning, and complex faults are uploaded to the cloud to complete in-depth root cause analysis, balancing the timeliness of fault response and the accuracy of diagnosis, and adapting to the troubleshooting needs under different networks, computing power and fault levels. 3. Establish a feature-rule bidirectional mapping mechanism, combine SHAP value to quantify feature contribution, transform the black box decision logic of deep learning into standardized natural language fault explanation, clearly show the fault judgment basis and propagation path, eliminate the need for manual secondary verification, and shorten the average fault repair time. 4. It is equipped with an adaptive noise filtering and hierarchical data completion preprocessing scheme to address common problems such as electromagnetic interference and data loss in industrial settings. It effectively filters pulse noise and completes incomplete data with high precision, greatly improving the system's operational stability under complex and harsh working conditions. 5. An improved IA-EWC incremental learning algorithm is adopted, which combines fault business weights, data credibility and weight reorganization constraints, with a rule-based virtual sample training mechanism. This not only solves the catastrophic forgetting problem of traditional models, but also adapts to the aging and drift characteristics of hardware performance. At the same time, it reduces the computing power consumption of model updates, shortens the update cycle, takes into account the ability to identify new and old fault modes, and ensures the long-term diagnostic performance of the model.

[0021] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the steps of a computer hardware fault diagnosis method based on industrial big data integration. Figure 2 This is a schematic diagram illustrating the specific execution steps of the IA-EWC algorithm. Detailed Implementation

[0023] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0024] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0025] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, systems, and equipment should be considered part of the instruction manual.

[0026] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0028] Example 1 This embodiment mentions a method for troubleshooting computer hardware faults based on industrial big data fusion, such as... Figure 1 As shown, the specific steps are as follows: Step 1: Multimodal data acquisition and adaptive robust preprocessing: For the five core components of the industrial computer—CPU, memory, hard drive, power supply, and motherboard—three independent sensors are used for redundant data acquisition for each key parameter (such as CPU temperature, power supply voltage, memory utilization, and hard drive SMART parameters). Simultaneously, system logs, performance metrics (such as CPU utilization, disk IOPS, and network throughput), and environmental data (such as data center temperature, humidity, and power supply voltage fluctuations) are collected. All data is uniformly timestamped with a high precision of less than 1ms to ensure data synchronization.

[0029] Adaptive noise filtering is performed on the collected sensor data: a sliding window is used to calculate the standard deviation of the data in each window. When the standard deviation exceeds the dynamic threshold determined by the 3σ criterion of historical normal data, it is determined to be impulse noise and filtered using wavelet transform thresholding method; for non-impulse noise, weighted median filtering is used for smoothing.

[0030] Generative completion is performed on missing data: when the missing data length is less than or equal to 5 sampling points, linear interpolation is used for completion; when the missing data length is greater than 5 sampling points, a pre-trained Conditional Generative Adversarial Network (CGAN) is used to generate complete data that conforms to the device's operating rules. CGAN takes 20 sampling points before and after the missing data as conditional inputs and outputs the complete sequence after completion.

[0031] Perform structured preprocessing on system logs: Use regular expressions to extract timestamps, event types, error codes, and key parameters from the logs, transforming unstructured text logs into structured event sequences for easier subsequent semantic analysis.

[0032] Step 2: Dynamic task scheduling and real-time edge anomaly detection based on fault urgency: A lightweight isolated forest anomaly detection model is deployed on edge computing nodes. After principal component analysis (PCA) feature dimensionality reduction (preserving 95% of the variance) and 8-bit integer quantization, the model size is compressed to less than 1MB, and the inference time for a single data point is less than 1ms, enabling it to run efficiently on resource-constrained edge nodes.

[0033] The edge nodes run a lightweight model in real time to perform preliminary anomaly detection on the preprocessed data. When a serious fault is detected (such as CPU temperature > 90°C, power supply voltage fluctuations exceeding...), the model will take action. When a fault occurs (e.g., 15% error, double-bit memory error, etc.), a local audible and visual alarm is immediately triggered, and non-critical loads are cut off. The response time is less than 500ms, minimizing the losses caused by the fault.

[0034] A dynamic task scheduling mechanism is activated to dynamically determine the execution location of subsequent computation tasks based on the following three parameters: Network bandwidth: When the bandwidth is less than 1 Mbps, all computing tasks are executed at the edge; when the bandwidth is greater than 10 Mbps, complex computing tasks are uploaded to the cloud.

[0035] Edge CPU utilization: When the utilization rate is >80%, all non-real-time computing tasks are uploaded to the cloud.

[0036] Fault urgency: Critical faults must be initially handled at the edge, while medium and minor faults can be uploaded to the cloud for in-depth analysis.

[0037] For faults that require in-depth cloud analysis, the edge nodes will package and upload the pre-processed data from the most recent 72 hours to the cloud platform.

[0038] Step 3: Cloud-based spatiotemporal-semantic 3D feature fusion: Constructing an industrial computer hardware topology diagram: The CPU, memory, hard drive, power supply, and motherboard are considered nodes in the graph, and the physical connections and data interactions between components are considered edges. The weight of each edge is determined by the signal transmission delay and data flow between components. This topology diagram accurately reflects the structure and working principle of the industrial computer hardware.

[0039] Extracting spatiotemporal correlation features: The preprocessed sensor time-series data is input into a spatiotemporal graph convolutional network (ST-GCN). ST-GCN captures the correlation between different components in the spatial dimension through graph convolutional layers, effectively identifying the chain reaction caused by the failure of one component; it captures the changing trends of parameters in the temporal dimension through gated recurrent units (GRUs), enabling the detection of early signs of performance degradation. The final output is a spatiotemporal feature vector with a dimension of 256.

[0040] Semantic causal feature extraction: Structured system log event sequences are input into a pre-trained BERT model for the industrial domain. This model was pre-trained using domain-adaptive data from 1 million industrial computer logs, enabling it to better understand the technical terminology and expressions used in industrial computer logs. A causal attention mechanism is introduced to automatically identify event pairs with causal relationships in the logs (e.g., "driver loading failure - system blue screen - automatic restart"), providing direct evidence for root cause analysis. The final output is a semantic feature vector with a dimension of 256.

[0041] Spatiotemporal-semantic cross-fusion is performed: spatiotemporal feature vectors and semantic feature vectors are input into the cross-attention fusion module. Bidirectional feature interaction fusion is achieved using a scaled dot product attention algorithm, defining an attention query vector Q, a key vector K, and a value vector V, and using a scaling factor... To avoid the gradient vanishing problem caused by increased feature dimensionality, a bidirectional attention weight matrix is ​​obtained through normalization using the softmax activation function, and then superimposed with two sets of unidirectional attention features to obtain a globally fused feature; where d k F is the dimension of the feature vector. st F is the spatiotemporal feature vector output by the spatiotemporal graph convolutional network. sem Semantic feature vectors output by pre-trained BERT models for industrial applications.

[0042] Step 4: Interpretable Fault Diagnosis and Root Cause Localization Based on Fusion Features: The fused feature vectors are input into an ensemble classifier. This ensemble classifier consists of three base classifiers: Random Forest, XGBoost, and Multilayer Perceptron. It outputs the final fault type (CPU fault, memory fault, hard drive fault, power supply fault, motherboard fault) and its corresponding probability through a voting mechanism. The ensemble classifier exhibits higher accuracy and robustness compared to a single classifier.

[0043] Interpretability analysis was performed: the contribution of each input feature to the diagnostic results was calculated using the SHAP (SHapley Additive ex Planations) value. Based on game theory, the SHAP value fairly allocates the contribution of each feature to the prediction results. The top 10 key features with the highest contribution were selected; these features are the main basis for the fault diagnosis results.

[0044] Perform a positive mapping from features to rules: Match key features with a pre-built fault rule base. The fault rule base contains over 500 industrial computer hardware fault rules that have been validated in industrial fields, with each rule corresponding to a set of feature thresholds and fault phenomena. The matching process generates natural language descriptions of the faults, enabling engineers to intuitively understand the basis for their decisions.

[0045] Root cause analysis: Based on the industrial computer hardware topology and feature contribution, a fault propagation path graph is constructed. Starting from the component with the highest failure probability, the path is traced along the edges of the topology graph to find the upstream component with the largest contribution, which is the root cause of the fault. This effectively distinguishes between primary and secondary faults, avoiding misjudgments.

[0046] Generate a fault diagnosis report, including: fault type, fault probability, root cause analysis, key evidence (key features and corresponding log / sensor data), recommended repair plan, and estimated repair time.

[0047] Step 5: Fault knowledge accumulation and incremental model update: After the repair is completed, the engineer enters the actual cause of the fault, the repair method, and the information on the replaced parts into the system. The system stores the characteristic data of this fault, the diagnostic results, and the actual repair information into the fault case database, forming a valuable fault knowledge asset.

[0048] A reverse mapping from rules to features is performed: When engineers discover new fault rules, they can input them into the system in natural language. The system uses an industrial BERT model to encode the rule text into semantic vectors, and then maps them to the same dimensional space as the fused features through a fully connected layer, generating feature vectors that the model can recognize. This allows the system to quickly absorb the engineers' experiential knowledge.

[0049] Incremental model updates are performed: New fault cases and rules are automatically collected weekly, and the ensemble classifier and fusion module are incrementally trained using the Industrial Scenarios Adaptive Elastic Weight Consolidation (IA-EWC) algorithm. The IA-EWC algorithm is an incremental learning algorithm specifically improved for industrial computer hardware fault diagnosis scenarios, and its total loss function is: ; In the formula: L new The cross-entropy loss is for new faulty data; This is the global penalty coefficient, with a value of 0.01. F is the fault importance coefficient corresponding to the i-th weight; i The Fisher information value for the i-th weight; This represents the value of the current i-th weight; This represents the value of the i-th weight in the old model; The relevance penalty coefficient for weight reorganization; G is the set of model weight groups; W g This is the matrix of the current g-th weight group; This is the matrix of the g-th weight group in the old model; It is the Frobenius norm.

[0050] like Figure 2 As shown, the specific execution steps of the IA-EWC algorithm are as follows: 1. Quantification of Business Importance Due to Failures: Calculate the business importance coefficient for each failure mode based on the average downtime, average economic loss, and security risk level caused by the failure. The formula is: ; Among them, T m Let C be the average downtime (in hours) for the m-th type of fault. m S represents the average economic loss (in ten thousand yuan). mThe risk level is categorized into 1-5. The final business importance coefficient for each weight is the maximum value among all the failure modes it participates in identifying.

[0051] 2. Key Weight Selection and Sparsified Fisher Calculation: The average contribution of each weight to all fault mode diagnosis results is calculated using SHAP values. The top 20% of weights by contribution are selected as key weights. Fisher information matrices are calculated only for key weights; the F-values ​​for non-key weights are not considered. i Set to 0. This method reduces computational load by more than 80% while retaining over 95% of the model performance.

[0052] 3. Weighting and Relevance Constraints: The model parameters are divided into three weighting groups: spatiotemporal feature extraction, semantic feature extraction, and cross-attention fusion. A higher relevance penalty coefficient is applied to the cross-attention fusion group. =0.1) to preserve the coupling relationship between spatiotemporal features and semantic features; apply a low correlation penalty coefficient to the spatiotemporal feature extraction group and the semantic feature extraction group ( =0.05), to ensure their respective feature extraction capabilities.

[0053] 4. Dynamic Adjustment of Data Reliability: A reliability score is calculated for each data point based on the engineer's level, record completeness, and fault verification results. The formula is: ; Among them, E j For engineer level scores (Junior = 0.4, Intermediate = 0.7, Senior = 1.0), I j To record completeness scores (incomplete = 0.3, mostly complete = 0.7, complete = 1.0), V j The fault validation score is calculated as follows (unvalidated = 0.2, partially validated = 0.6, fully validated = 1.0). The learning rate for each data point is... ,in The base learning rate is set to 0.001.

[0054] 5. Virtual Sample Generation and Fusion Training: The natural language fault rules input by engineers are transformed into feature vectors through rule-feature back-mapping, and corresponding fault labels are assigned to generate virtual samples. The credibility score of the virtual samples is fixed at 1.0, and the learning rate is 1.5 times the base learning rate. The virtual samples are mixed with real fault data and used for incremental training to achieve collaborative updates of experiential knowledge and data knowledge.

[0055] The IA-EWC algorithm effectively avoids the "catastrophic forgetting" problem by calculating the importance of key weights in old tasks and applying differentiated penalties, while retaining the ability to identify old failure modes during the learning of new failure modes.

[0056] Perform model performance evaluation: Use an independent test set to evaluate the accuracy, recall, and false positive rate of the updated model. If the performance metrics meet the preset requirements (accuracy > 97%, false positive rate < 1%), the new model is deployed online; otherwise, a full retraining process is triggered.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for troubleshooting computer hardware faults based on industrial big data fusion, characterized in that, Includes the following steps: S1. Collect and perform adaptive robust preprocessing on multimodal data of industrial computer hardware, including hardware sensor data, system log data, performance index data and environmental data; S2. Deploy a lightweight anomaly detection model on edge computing nodes to perform real-time anomaly detection, and dynamically schedule subsequent computing tasks based on fault urgency, network bandwidth, and edge computing resource utilization. S3. In the cloud, perform spatiotemporal correlation feature extraction on the preprocessed sensor data, perform semantic causal feature extraction on the preprocessed system log data, and perform three-dimensional fusion of the two features through a spatiotemporal-semantic cross-attention mechanism. S4. Input the fused feature vector into the ensemble classifier for fault type identification, calculate the feature contribution using the SHS1P value, and generate an interpretable fault diagnosis report and root cause analysis through feature-rule positive mapping. S5. Collect maintenance feedback information to accumulate fault knowledge, and achieve incremental updates of the model through rule-feature back mapping and elastic weight consolidation algorithm.

2. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 1, characterized in that, The adaptive robust preprocessing described in S1 specifically includes: S11. Each key parameter of the five core components—CPU, memory, hard drive, power supply, and motherboard—is redundantly collected using three independent sensors, and all data is uniformly stamped with a high-precision timestamp with an error of less than 1ms. S12. An adaptive noise filtering algorithm combining sliding window standard deviation and wavelet transform thresholding is adopted to automatically identify and filter impulse noise generated by industrial electromagnetic interference. S13. When the missing data length is less than or equal to 5 sampling points, linear interpolation is used to complete the data. When the missing data length is greater than 5 sampling points, conditional generative adversarial network CGS1N is used to generate complete data that conforms to the operating rules of the device. S14. Use regular expressions to extract timestamps, event types, error codes, and key parameters from system logs, and transform unstructured text into a structured event sequence.

3. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 2, characterized in that, The lightweight anomaly detection model described in S2 is an isolated forest model that has undergone principal component analysis feature dimensionality reduction and 8-bit integer quantization.

4. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 3, characterized in that, The dynamic task scheduling mechanism described in S2 is as follows: S21. When a serious fault is detected, regardless of the network condition, a local audible and visual alarm will be triggered immediately at the edge and non-critical loads will be cut off, with a response time of less than 500ms. S22. When the network bandwidth is less than 1MS2ps, all computing tasks are executed at the edge. S23. When the network bandwidth is greater than 10MS2ps and the edge CPU utilization is less than 80%, upload the in-depth analysis tasks for medium and minor faults to the cloud for execution. S24. When the edge CPU utilization rate is greater than 80%, upload all non-real-time computing tasks to the cloud for execution.

5. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 4, characterized in that, The spatiotemporal correlation feature extraction described in S3 specifically refers to: An industrial computer hardware topology graph is constructed with hardware components as nodes and physical connections and data interaction relationships as edges. Preprocessed sensor time-series data is input into the spatiotemporal graph convolutional network ST-GCN. The spatial dimension correlation is captured through graph convolutional layers, and the temporal dimension change trend is captured through gated recurrent units (GRUs). The output spatiotemporal feature vector has a dimension of 256.

6. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 5, characterized in that, The semantic causal feature extraction described in S3 specifically refers to: The structured system log event sequence is input into the S2ERT model, which is pre-trained using 1 million industrial computer log data for domain adaptation. A causal attention mechanism is introduced to automatically identify event pairs with causal relationships in the logs, and the output is a semantic feature vector with a dimension of 256.

7. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 6, characterized in that, The computation process of the spatiotemporal-semantic cross-attention mechanism described in S3 is as follows: ； ； Among them, F st F is the spatiotemporal feature vector output by the spatiotemporal graph convolutional network. sem For industrial applications, the semantic feature vector output by the pre-trained BERT model is given, where Q is the attention query vector, K is the attention key vector, and V is the attention value vector. T Let d be the transpose of the key vectors. k Let F be the dimension of the key vector, softmax(·) be the normalization exponential function, and F be the dimension of the key vector. fusion This is the global feature vector resulting from the cross-fusion of spatiotemporal and semantic features.

8. The method for troubleshooting computer hardware faults based on industrial big data fusion according to claim 7, characterized in that, The explainable fault diagnosis and root cause localization described in S4 specifically include: S41. Input the fused feature vector into the ensemble classifier consisting of random forest, XGS2oost and multilayer perceptron, and output the fault type and the corresponding fault probability. S42. Calculate the contribution of each input feature to the diagnostic result using the SHS1P value, and select the top 10 key features in terms of contribution. S43. Match key features with a rule base containing more than 500 verified industrial computer hardware fault rules to generate a fault explanation in natural language. S44. Construct a fault propagation path diagram based on the hardware topology diagram and feature contribution, and trace back to find the upstream component with the largest contribution as the root cause of the fault.

9. A method for troubleshooting computer hardware faults based on industrial big data fusion as described in claim 8, characterized in that, The rule-feature inverse mapping described in S5 is specifically as follows: The industrial S2ERT model is used to encode the natural language fault rules input by engineers into semantic vectors, and then maps them to the same dimensional space as the fused features through a fully connected layer to generate feature vectors that the model can recognize.

10. A method for troubleshooting computer hardware faults based on industrial big data fusion as described in claim 9, characterized in that, The incremental model update described in S5 specifically refers to: Each week, newly added fault cases and new fault rules are automatically collected, and the IS1-EWC algorithm with adaptive elastic weight consolidation for industrial scenarios is used to incrementally train the ensemble classifier and fusion module. The IS1-EWC algorithm introduces a fault service importance weighting mechanism, sparse Fisher matrix calculation, and dynamic adjustment of data credibility. It calculates the importance of key weights in old tasks and applies penalties, while preserving the correlation between weight reorganizations to avoid catastrophic forgetting problems.

Images