Heterogeneous system integration and fault diagnosis operation and maintenance system based on big data analysis

Abstract

The invention discloses a heterogeneous system integration and fault diagnosis operation and maintenance system based on big data analysis, and the system is characterized in that the system comprises an acquisition cleaning module which is used for collecting structured data, semi-structured data and non-structured data, and carrying out the data cleaning; the mapping calculation module is used for dynamically mapping the cleaned data, and storing the data into a database after federal calculation; the feature extraction module is used for performing multi-modal extraction on the data in the database, constructing a knowledge graph and generating features for fault diagnosis; the model training module is used for constructing a fault diagnosis model and performing fault prediction and root cause analysis by using fault diagnosis features; the collaborative decision-making module is used for carrying out collaborative decision-making on the edge and the cloud according to the analysis result; and the feedback optimization module is used for feeding back the response processing result to the data center and carrying out updating iteration on the diagnosis model.

Description

Technical Field

[0001] This invention relates to the field of integration and operation and maintenance system technology, specifically a heterogeneous system integration and fault diagnosis operation and maintenance system based on big data analysis. Background Technology

[0002] Traditional solutions typically employ a single data source access model, relying on manually configured ETL tools or customized interfaces to achieve structured data collection. This approach suffers from severely insufficient support for semi-structured logs, unstructured text, and multimedia data. The data cleaning process heavily depends on preset rule templates and lacks adaptive anomaly detection capabilities, resulting in inefficient processing of noisy data. In industrial equipment maintenance scenarios, traditional systems can only collect sensor numerical data, ignoring crucial unstructured information such as equipment vibration waveforms and operator voice recordings, leading to insufficient completeness in fault feature extraction. Furthermore, data storage often utilizes centralized databases, which suffer from write bottlenecks when dealing with massive amounts of heterogeneous data and lack privacy protection mechanisms such as federated computing, hindering cross-domain collaborative data analysis.

[0003] Traditional heterogeneous system integration primarily relies on static middleware technologies, such as enterprise service buses or point-to-point API gateways, which use hard-coded methods to achieve protocol conversion and data format mapping between systems. This approach requires frequent modifications to the integration code and service restarts when dealing with system version upgrades or the addition of new data sources, resulting in high maintenance costs. Furthermore, it lacks unified knowledge graph construction capabilities, and the relationships between systems exist in the form of static configuration tables, making it difficult to support intelligent fault propagation path analysis.

[0004] Traditional fault diagnosis relies on rule-based expert systems or shallow machine learning models, such as support vector machines or decision trees. Their feature engineering depends on human experience and cannot automatically uncover deep correlations in multimodal data. In CNC machine tool fault diagnosis, traditional systems can only determine bearing condition based on vibration frequency thresholds, ignoring the multi-parameter coupling characteristics of temperature and current, leading to frequent missed early faults. The diagnostic results are presented in a limited way, providing only fault codes rather than root cause explanations, requiring maintenance personnel to conduct secondary analysis by consulting knowledge bases. Model updates rely on offline training, failing to incorporate new fault cases in real time, resulting in a significant decline in model accuracy over time.

[0005] Traditional operation and maintenance decision-making adopts a passive model of detection, alarm, and manual handling, with each link in the decision-making chain being isolated. Edge computing nodes only perform data collection and simple threshold judgments; complex analysis requires uploading to the cloud for processing, leading to decision delays in scenarios with high real-time requirements. Optimization strategy formulation relies on historical data backtracking, lacks online reinforcement learning mechanisms, and suffers from a missing feedback loop. Diagnostic results are disconnected from model iteration, and fault handling experience cannot be transformed into knowledge assets, creating a vicious cycle of repetitive work. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a heterogeneous system integration and fault diagnosis and maintenance system based on big data analysis, characterized in that it includes:

[0007] The data acquisition and cleaning module is used to acquire structured data, semi-structured data, and unstructured data, and to perform data cleaning.

[0008] The mapping calculation module is used to dynamically map the cleaned data and store it in the database after federated calculation.

[0009] The feature extraction module is used to extract multimodal data from the database, while constructing a knowledge graph to generate features for fault diagnosis.

[0010] The model training module is used to build fault diagnosis models and use fault diagnosis features to perform fault prediction and root cause analysis.

[0011] The collaborative decision-making module is used to make collaborative decisions between the edge and the cloud based on the analysis results;

[0012] The feedback optimization module is used to feed the response processing results back to the data center to update and iterate the diagnostic model.

[0013] Preferably, the collected data includes: structured data such as tables and database tables; semi-structured data such as tags, configuration files, responses, web pages, and logs; and unstructured data such as text, images, music, and videos.

[0014] The specific method of dynamic mapping is as follows: for structured data, the source table fields are matched with the target table fields through preset rules and matching patterns, the numerical precision and date are converted into a unified format, and a globally unique identifier is generated; for semi-structured data, nested data is extracted, the structure is automatically inferred through data tags, and fields are matched through semantic similarity to perform dynamic pattern conversion; for unstructured data, it is transformed into structured and semi-structured forms through feature extraction and semantic analysis.

[0015] Preferably, the method for extracting fault diagnosis features is as follows: extracting fault diagnosis-related features from multimodal data; extracting statistical features and time-series features from structured data, including the mean of continuous equipment operation time and temperature variance, and the time-series features including vibration signal frequency; extracting key fields and text semantics from semi-structured data; and extracting image features and document information from unstructured data.

[0016] The method for constructing the fault diagnosis model is as follows: an LSTM network prediction model is selected to capture the long-term dependence of vibration signals and predict the probability of faults in the future; CNN is used to process images and fully connected layers are used to process structured data to output the fault type; edge nodes are trained locally and then the parameters are aggregated in the cloud.

[0017] The root cause analysis specifically involves: locating the critical path by constructing a logic tree, confirming the root cause based on the output fault type and fault tree; and using a knowledge graph for reasoning, associating historical cases, and matching the latest maintenance solutions from the knowledge base.

[0018] The edge and cloud collaborative decision-making method is as follows: the edge layer performs real-time detection and marking of faults, and the cloud layer verifies the faults using a CNN model; the edge layer executes a pre-set emergency plan and generates a work order based on the fault type, and the cloud layer analyzes the root cause and issues new monitoring rules; the new model trained in the cloud is encrypted and then sent to the edge, and the edge node model is switched via OTA; the cloud periodically lowers the alarm threshold based on the device aging curve.

[0019] The model update and iteration method is as follows: the fault response results are structured and key information, including fault location, treatment measures, and causal relationships, is extracted, and the knowledge graph nodes are dynamically updated based on the extracted key information; federated incremental learning is adopted to collect new data from edge nodes to optimize the fault diagnosis model; and active learning is adopted to initiate manual annotation requests for samples with low confidence.

[0020] Combining all the above technical solutions, the positive effects of this invention are as follows: 1. This invention breaks through the data type limitations of traditional solutions by supporting the unified collection and cleaning of structured, semi-structured, and unstructured data. The dynamic mapping module adopts semantic similarity matching and federated computing technology to achieve intelligent conversion of cross-system data formats and privacy protection, solving the problems of rigid traditional static mapping rules and difficulties in cross-domain collaboration. For example, unstructured text in industrial equipment logs can be automatically parsed into structured fault features, while federated computing ensures that sensitive data does not leave the domain, which can improve data utilization and shorten the system integration cycle.

[0021] 2. The feature extraction module of this invention integrates statistical features, image features, and semantic features, and combines them with a knowledge graph to construct a fault propagation path model, improving the accuracy of root cause analysis and effectively avoiding the false negative rate caused by traditional solutions relying on a single data source. Through dual verification using logic tree and knowledge graph reasoning, fault location time is shortened, and dynamic association and matching with historical case libraries are supported.

[0022] 3. The collaborative decision-making module of this invention achieves a closed loop between real-time edge response and deep cloud analysis. Edge nodes perform initial fault probability screening using a lightweight LSTM model, while a cloud-based CNN model performs secondary verification on complex image features, reducing false alarm rates. The cloud dynamically adjusts alarm thresholds based on device aging curves, and edge nodes autonomously execute emergency plans while generating standardized work orders, improving decision-making efficiency and reducing network bandwidth consumption.

[0023] 4. The feedback optimization module shortens the model iteration cycle through federated incremental learning and active learning mechanisms. New data added to edge nodes is encrypted before participating in global model training. The system automatically extracts causal chains from the processing results to update the knowledge graph, reducing the need for manual intervention. Active learning initiates annotation requests for low-confidence samples, reducing model training costs and achieving a fully automated closed loop of diagnosis, optimization, and deployment. Attached Figure Description

[0024] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0025] Figure 1 This is a system framework diagram of the system of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See Figure 1 As shown, this invention proposes a heterogeneous system integration and fault diagnosis and maintenance system based on big data analysis, including a data acquisition and cleaning module, a mapping calculation module, a feature extraction module, a model training module, a collaborative decision-making module, and a feedback optimization module.

[0028] In a more specific application of this invention, in the data acquisition and cleaning module, when acquiring structured data, database connection technology is mainly used to obtain data for tables and database tables. Connections to relational databases can be established via JDBC or ODBC standard interfaces. SQL query scripts are used to extract specific fields and records from the database tables as needed. After acquisition, data cleaning tools are used to improve data quality through operations including setting data type validation rules, removing duplicate records, handling missing values, and correcting erroneous data formats.

[0029] For semi-structured data, tag and configuration file formats can be parsed using scripting languages. For example, Python's `xml.etree.ElementTree` library can be used to parse XML-formatted configuration files and tag data, and the `json` library can be used to parse JSON-formatted response data. Web page data can be collected using web crawling techniques, such as writing a crawler program using the Scrapy framework and extracting the required data from web pages according to set rules, using regular expressions or XPath expressions to precisely locate data nodes. Log data can be monitored in real time using the Flume tool, allowing newly generated log data to be quickly transferred to a specified storage location. During cleaning, tags and configuration files are formatted uniformly, noisy information such as advertisements and irrelevant HTML tags is removed from web page data, and outlier filtering and key information extraction are performed on log data.

[0030] When collecting unstructured data, text data can be extracted using text scraping tools, such as Python's BeautifulSoup library, from various documents and web pages. For image, music, and video data, data files are read in batches according to file paths through file system access interfaces, and relevant metadata such as filename, file size, and creation time are recorded. During data cleaning, text data undergoes preprocessing including word segmentation, stop word removal, and lemmatization; image data undergoes size normalization and removal of blurry or damaged images; music and video data are checked for file integrity, and files that cannot be played are removed. The cleaned structured, semi-structured, and unstructured data are then formatted uniformly and stored in a data warehouse or distributed storage system to provide a foundation for subsequent processing.

[0031] In the mapping calculation module, structured data is stored using a rule engine to build preset mapping rules, and then the field description information of the source and target tables is obtained through the metadata management system. After the field-level mapping configuration is completed using ETL tools, conversion scripts are written to unify the precision of numerical values ​​and convert date formats. To ensure data uniqueness, the system uses the UUID algorithm to generate globally unique identifiers. During implementation, data quality monitoring points need to be configured, and the mapping results are verified in real time using Apache NiFi, triggering an early warning mechanism for data with type mismatches or rule conflicts.

[0032] In the semi-structured data mapping process, nested data is extracted using JSONPath and XPath parsers, and the data structure can be automatically identified by combining pattern inference algorithms. Word vector models are used to calculate semantic similarity between fields, and cosine similarity algorithms are used for intelligent field matching. A rule template engine is used to dynamically generate transformation scripts, realizing the conversion from XML and JSON to relational models. During the mapping process, the system uses Apache Kafka to build a message queue to asynchronously process batch data transformation tasks, improving system throughput.

[0033] In the process of mapping unstructured data, a hierarchical parsing strategy can be adopted to transform unstructured data into a structured model. In text processing, a natural language processing (NLP) stack is deployed, sequentially performing Chinese word segmentation, named entity recognition, and sentiment analysis. Keyword feature vectors are generated based on a TF-IDF statistical model, and a structured encoding of text semantics is completed through the construction of an inverted index mechanism. For image data processing streams, a convolutional neural network model is built based on a deep learning framework. After multiple rounds of iterative training, visual feature representations are obtained, and finally, the image content is converted into high-dimensional feature vectors for persistent storage. For audio and video data streams, a hybrid processing architecture is adopted. OpenCV is used to process video frame sequences, while simultaneously extracting audio spectral features using the Librosa library. A speech recognition engine is then used to convert the temporal signal into structured text summaries. All multi-dimensional feature vectors generated in all processing steps are mapped and aligned with a pre-defined domain data model using feature engineering methods to form parsable structured data.

[0034] While mapping various types of data, a multi-party secure computing environment needs to be built using a horizontal federated learning framework, and data privacy protection should be achieved through homomorphic encryption technology. A distributed computing engine is deployed on the computing nodes to perform feature alignment and model training tasks. After training, the model parameter update process is recorded using blockchain technology to ensure data traceability. Finally, the computation results are written to a distributed database through a distributed transaction framework, and the data storage structure is optimized using database sharding and partitioning strategies, while query efficiency is improved through secondary indexes. The entire process is monitored using a system built with Prometheus and Grafana to track data flow and computational performance metrics in real time.

[0035] In the feature extraction module, the methods for extracting fault diagnosis-related features are as follows: for structured data, statistical features are calculated using database window functions, real-time statistics are achieved using the sliding window algorithm, and time-series features are extracted from vibration signals using Apache Flink combined with the Fast Fourier Transform algorithm; for semi-structured data, key fields in logs are extracted using regular expressions, and text semantic analysis is performed using the Word2Vec model and the TextRank algorithm; for unstructured data, image features are extracted using OpenCV and CNN models, and document information is processed using TesseractOCR and NLP technologies.

[0036] The knowledge graph is constructed as follows: BiLSTM-CRF model and dependency parsing are used to identify entities and relationships, and the association information between equipment hierarchy and fault causality is extracted from structured and semi-structured data; DS evidence theory is used to fuse multi-source entity information to resolve conflict issues and store knowledge in the Neo4j graph database; a rule-based reasoning engine is developed and fault diagnosis rules are defined to realize fault propagation path analysis and reasoning.

[0037] Through the above process, comprehensive and in-depth feature extraction from multimodal data is achieved, effectively mining fault-related information from structured, semi-structured, and unstructured data, significantly improving the accuracy and completeness of features. The construction of the knowledge graph connects multi-source data into an organic whole, forming a fault knowledge network with semantic reasoning capabilities. The combination of these two aspects provides high-quality feature input for fault diagnosis models, significantly enhancing the timeliness of fault prediction, the accuracy of root cause analysis, and improving the intelligence and efficiency of heterogeneous system operation and maintenance.

[0038] In the model training module, the fault diagnosis model is constructed as follows: For vibration signals, an LSTM network is used, generating a 3D tensor by slicing the data into windows at a certain time interval. An LSTM and bidirectional GRU structure is then constructed to handle temporal dependencies and sample imbalance. For equipment images, a CNN model is used, with transfer learning used to fine-tune fully connected layers and then unfreeze some convolutional blocks. Data augmentation is used to expand the sample size. For structured data, a multi-layer fully connected network is constructed, with input features processed through standardization and one-hot encoding. Batch Normalization and Leaky ReLU activation are added to each layer. All three models employ the Adam optimizer and an early stopping strategy. The LSTM outputs the fault probability, while the CNN and fully connected networks output the fault type. Finally, weighted voting is used to fuse the results of multiple models to improve diagnostic accuracy.

[0039] The root cause analysis specifically involves: constructing a logic tree with the fault as the top event based on the output fault type; calculating the minimum cut set through Boolean algebra simplification; calculating the probability importance of each basic event in combination with historical data to determine the most likely combination of root causes; then semantically matching the located root causes with entities in the knowledge graph; using graph embedding technology to retrieve similar historical cases; and generating customized maintenance solutions through template filling; ultimately forming a complete closed loop from fault location to cause tracing to solution generation, significantly improving the efficiency and accuracy of fault handling.

[0040] In the collaborative decision-making module, the edge and cloud collaborative decision-making process is as follows: edge nodes use lightweight models to detect vibration signals in real time. When the predicted fault probability exceeds a threshold, the nodes mark the fault and upload key features to the cloud. The cloud verifies the fault using a CNN model and then triggers a work order generation. Simultaneously, it issues new monitoring rules based on root cause analysis. Edge nodes execute pre-set emergency plans and generate work orders, and receive rule adjustments from the cloud. The cloud periodically encrypts and distributes the optimized model, and the edge nodes complete OTA upgrades after A / B testing. The cloud dynamically adjusts alarm thresholds based on the device aging curve, achieving full-process collaborative optimization from detection and decision-making to model iteration.

[0041] In the feedback optimization module, the response processing results are fed back to the data hub to update and iterate the diagnostic model. The model update and iteration method is as follows: the fault response results are structured, named entity recognition is performed using NLP technology stack, and key information about the fault location and handling measures is extracted; after mining causal relationships through dependency parsing, this information is transformed into triples of entity, relation, and entity, and the knowledge graph nodes are dynamically updated to ensure that the knowledge graph evolves with new cases.

[0042] In the federated incremental learning phase, edge nodes collect newly generated fault data using lightweight data acquisition tools. After local preprocessing, the data is incrementally trained using the federated learning framework. Homomorphic encryption is employed to protect data privacy during training. Edge nodes only upload model gradient update values ​​to the cloud, where the global model parameters are aggregated using the FedAvg algorithm. A resilient weight consolidation method is introduced to retain key weights from previous knowledge when updating parameters. After training, model performance evaluation determines whether to replace the online model.

[0043] An active learning mechanism is employed, specifically: when the model infers from new samples, it calculates the entropy value of the predicted probability distribution; the higher the entropy, the lower the confidence level. When the entropy exceeds a threshold, a manual annotation request is automatically triggered. The annotation system integrates the Labelbox API, encapsulating samples such as image slices and log fragments with contextual information such as device model and operating environment into annotation tasks, which are then assigned to the annotation team. After annotation is completed, active learning algorithms are used to filter high-value samples, which are then injected into the model through transfer learning. Simultaneously, the annotation status of the sample pool is updated, forming a closed-loop optimization process of model prediction, low-confidence filtering, manual annotation, and retraining.

[0044] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0045] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A heterogeneous system integration and fault diagnosis maintenance system based on big data analytics, characterized in that, include: The data acquisition and cleaning module is used to acquire structured data, semi-structured data, and unstructured data, and to perform data cleaning. The mapping calculation module is used to dynamically map the cleaned data and store it in the database after federated calculation. The feature extraction module is used to extract multimodal data from the database, while constructing a knowledge graph to generate features for fault diagnosis. The model training module is used to build fault diagnosis models and use fault diagnosis features to perform fault prediction and root cause analysis. The collaborative decision-making module is used to make collaborative decisions between the edge and the cloud based on the analysis results; The feedback optimization module is used to feed the response processing results back to the data center to update and iterate the diagnostic model.

2. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The collected data includes: structured data such as tables and database tables; semi-structured data such as tags, configuration files, responses, web pages, and logs; and unstructured data such as text, images, music, and videos.

3. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The specific method of dynamic mapping is as follows: for structured data, the source table fields are matched with the target table fields through preset rules and matching patterns, the numerical precision and date are converted into a unified format, and a globally unique identifier is generated; for semi-structured data, nested data is extracted, the structure is automatically inferred through data tags, and fields are matched through semantic similarity to perform dynamic pattern conversion; for unstructured data, it is transformed into structured and semi-structured forms through feature extraction and semantic analysis.

4. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The methods for extracting fault diagnosis features are as follows: extracting fault diagnosis-related features from multimodal data; extracting statistical features and time-series features from structured data; the statistical features include the mean of continuous equipment operation time and temperature variance; and the time-series features include vibration signal frequency; extracting key fields and text semantics from semi-structured data; and extracting image features and document information from unstructured data.

5. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The method for constructing the fault diagnosis model is as follows: an LSTM network prediction model is selected to capture the long-term dependence of vibration signals and predict the probability of faults in the future; CNN is used to process images and fully connected layers are used to process structured data to output the fault type; edge nodes are trained locally and then the parameters are aggregated in the cloud.

6. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The root cause analysis specifically involves: locating the critical path by constructing a logic tree, confirming the root cause based on the output fault type and fault tree; and using a knowledge graph for reasoning, associating historical cases, and matching the latest maintenance solutions from the knowledge base.

7. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The edge and cloud collaborative decision-making method is as follows: the edge layer performs real-time detection and marking of faults, and the cloud layer verifies the faults using a CNN model; the edge layer executes a pre-set emergency plan and generates a work order based on the fault type, and the cloud layer analyzes the root cause and issues new monitoring rules; the new model trained in the cloud is encrypted and then sent to the edge, and the edge node model is switched via OTA; the cloud periodically lowers the alarm threshold based on the device aging curve.

8. The heterogeneous system integration and fault diagnosis maintenance system based on big data analysis as described in claim 1, characterized in that: The model update and iteration method is as follows: the fault response results are structured and key information, including fault location, treatment measures, and causal relationships, is extracted, and the knowledge graph nodes are dynamically updated based on the extracted key information; federated incremental learning is adopted to collect new data from edge nodes to optimize the fault diagnosis model; and active learning is adopted to initiate manual annotation requests for samples with low confidence.

Images