Intelligent equipment fault diagnosis and reasoning method and system based on unsupervised learning

Abstract

The invention provides an intelligent equipment fault diagnosis and reasoning method and system based on unsupervised learning. Belongs to the technical field of industrial equipment fault diagnosis and prediction. 2) unsupervised abnormal feature extraction and detection; (3) dynamic causal reasoning based on unsupervised learning; (4) a dynamic self-optimization inference engine; (5) intelligent fault reasoning and exploration; (6) enhancing the dynamic knowledge graph; and (7) fault prediction and decision support. Aiming at the problems that a large amount of annotated data is needed, causal analysis is lacked, and multi-source data fusion is difficult, non-annotated efficient detection and classification, fault mode diagnosis, prediction and decision support of equipment faults are realized through feature extraction of an unsupervised depth model, construction of a dynamic causal relationship, and fusion of a real-time self-optimization inference engine and a dynamic knowledge graph; therefore, the downtime and the maintenance cost are greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault diagnosis and prediction, and in particular to an intelligent equipment fault diagnosis and reasoning method and system based on unsupervised learning. Background Art

[0002] With the rapid development of industrial Internet and intelligent manufacturing, the complexity and automation of industrial equipment are constantly improving, which puts forward higher requirements for real-time monitoring of equipment operation status and fault prediction. Traditional equipment fault diagnosis methods mainly rely on expert experience and rule bases, or supervised learning-based models, which require a large amount of labeled data. However, in actual industrial scenarios, equipment fault data is scarce and unevenly distributed, and it is costly to obtain high-quality labeled data, which limits the application of supervised learning methods.

[0003] At present, anomaly detection methods based on unsupervised learning are gradually gaining attention. For example, they use autoencoders to reconstruct normal data and detect anomalies with large reconstruction errors. These methods can model the normal behavior of equipment in the absence of labeled data. However, existing unsupervised anomaly detection methods mainly focus on static anomaly detection, lack causal analysis of faults, cannot capture the dynamic correlation between equipment parameters, and have difficulty in deeply understanding the evolution mechanism of faults.

[0004] In addition, traditional fault diagnosis systems often lack the effective integration of multi-source heterogeneous data, such as sensor data, operation logs, and environmental information. There are complex correlations between these data, and a single data source cannot fully reflect the operating status of the equipment. At the same time, existing methods lack the ability to diagnose new unknown fault modes and cannot adapt to changes in equipment status, resulting in reduced diagnostic accuracy and robustness.

[0005] In terms of knowledge management, existing fault diagnosis systems usually lack a dynamic knowledge update mechanism and are unable to promptly incorporate newly discovered fault modes and causal relationships into the system, resulting in insufficient timeliness and completeness of the knowledge base. In addition, the lack of effective use of knowledge graphs makes it difficult to fully explore the potential patterns and laws in equipment operation data, limiting the ability to predict faults and provide decision support.

[0006] In summary, the prior art has deficiencies in the following aspects:

[0007] 1) Reliance on a large amount of labeled data: Traditional supervised learning methods require a large amount of high-quality fault labeled data, which is costly to obtain and difficult to cover all fault types.

[0008] 2) Lack of causal analysis: Existing unsupervised anomaly detection methods mainly focus on the detection of abnormal points, lack the modeling of causal relationships between equipment parameters, and cannot deeply understand the mechanism of fault generation.

[0009] 3) Difficulty in processing multi-source heterogeneous data: Lack of effective fusion of multimodal data such as sensor data, operation logs and environmental information, unable to fully reflect the operating status of the equipment.

[0010] 4) Delayed knowledge updating: The existing system lacks a dynamic knowledge updating and management mechanism and is unable to promptly incorporate newly discovered failure modes and relationships, resulting in insufficient timeliness of the knowledge base.

[0011] 5) Limited prediction and decision support capabilities: Lack of modeling and prediction of the dynamic evolution of faults, inability to provide effective early warning and decision support, affecting the preventive maintenance of equipment.

[0012] In order to solve the above problems, it is necessary to propose a new method that can use unsupervised learning and causal reasoning in the absence of labeled data to automatically diagnose and discover equipment failure modes, build a dynamic knowledge graph, and realize the prediction and decision support of equipment failures, thereby improving the accuracy and efficiency of fault diagnosis and reducing equipment downtime and maintenance costs. Summary of the invention

[0013] To solve the above problems, one of the purposes of the present invention is to provide a method for intelligent device fault diagnosis and reasoning based on unsupervised learning, and the specific steps are as follows:

[0014] (1) Data collection and preprocessing: Collect multi-dimensional sensor data, operation logs, and environmental information of the equipment to form a multi-source data set D = {X, L, E}, where X is the sensor data, L is the operation log, and E is the environmental information. The collected data is cleaned, missing values ​​are filled, and normalized.

[0015] (2) Unsupervised abnormal feature extraction and detection: using unsupervised learning models such as deep autoencoders, we extract low-dimensional feature representations Z from preprocessed data and build a normal behavior model for the device. By minimizing the reconstruction error, the model learns the feature distribution of the device under normal operating conditions. We monitor the reconstruction error of new data in real time, and when the reconstruction error exceeds the set threshold τ, we determine it as an abnormal data point.

[0016] (3) Based on unsupervised learning, dynamic causal reasoning is used to map multimodal data such as sensor data, operation logs, and environmental information into a unified feature space Z. The causal relationship network between device parameters is constructed using an unsupervised causal discovery algorithm. Dynamic Bayesian networks or temporal causal models are used to model the dependency of variables over time and capture the dynamic evolution of faults.

[0017] (4) Dynamic self-optimizing inference engine: for newly arrived data, the model parameters θ are updated in real time to adapt the model to changes in the device status; when an unknown abnormal pattern is detected, experts are actively requested to label it to obtain high-quality labeled data; the model is updated using the newly labeled data to improve the ability to detect similar anomalies; reinforcement learning methods are used to dynamically adjust the inference strategy and optimize the inference path based on the feedback of the inference results;

[0018] (5) Intelligent equipment fault reasoning and diagnosis, building a multi-level reasoning architecture including data preprocessing layer, anomaly detection layer, causal reasoning layer and decision support layer; using clustering algorithms to classify abnormal data and discover unknown fault modes; incorporating newly discovered fault modes into the fault mode library to continuously enrich the knowledge of the system; combining Bayesian reasoning and reinforcement learning to optimize the decision-making strategy of fault diagnosis and improve diagnostic efficiency;

[0019] (6) Dynamic knowledge graph enhancement: The equipment operation data, failure modes, maintenance records and other information are constructed into a dynamic knowledge graph. When the unsupervised learning model discovers a new failure mode or relationship, the knowledge graph is updated in a timely manner, and new nodes and edges are added. The knowledge graph is extracted and represented using a graph neural network, and the structural information of the knowledge graph is integrated into the reasoning model. The structural information of the knowledge graph assists fault diagnosis and prediction, and realizes two-way feedback between the model and the knowledge graph.

[0020] (7) Fault prediction and decision support: using a self-optimized reasoning engine and enhanced reasoning model to predict the future state of the equipment and calculate the probability of a fault. When the predicted probability of a fault exceeds the set threshold, the system triggers an early warning and notifies relevant personnel. Combined with the fault mode library and knowledge graph, it provides maintenance suggestions and decision support to help technicians take timely measures.

[0021] Preferably, in step 4, the updating of the model parameters θ includes adjusting the weights and biases of the deep autoencoder, and optimizing the parameters of the causal reasoning model and the reinforcement learning model.

[0022] Preferably, a clustering algorithm is used to classify abnormal data, and the clustering algorithm used is K-means, density clustering or hierarchical clustering algorithm.

[0023] Preferably, a graph convolutional network (GCN) or a graph attention network (GAT) in a graph neural network is used to extract and represent features of the knowledge graph.

[0024] Preferably, when the predicted failure probability exceeds a threshold, the system automatically generates warning information, including possible causes of the failure, scope of impact, and recommended treatment measures.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1) Efficient anomaly detection and fault diagnosis without the need for a large amount of labeled data: By using unsupervised learning models such as deep autoencoders (step 2), a normal behavior model of the equipment is constructed, and abnormal data can be detected without relying on a large amount of fault labeled data. Using clustering algorithms to classify abnormal data (step 5) can discover unknown fault modes and improve fault diagnosis capabilities.

[0027] 2) Dynamic causal relationship modeling to deeply understand the failure mechanism: The causal relationship network between device parameters is constructed through unsupervised causal discovery algorithms (step 3) to capture the causal relationship between device parameters. Dynamic Bayesian networks or temporal causal models are used to model the dependency of variables over time, capture the dynamic evolution of faults, and deeply understand the mechanism of fault generation.

[0028] 3) Fusion of multi-source heterogeneous data to fully reflect the operating status of the equipment: Multimodal data such as sensor data, operation logs and environmental information are mapped to a unified feature space (step 3) to achieve effective fusion of multi-source heterogeneous data. The unified feature space provides a basis for causal reasoning and anomaly detection, and fully reflects the operating status of the equipment and environmental impact.

[0029] 4) Dynamic self-optimization and reinforcement learning to improve model adaptability and diagnostic efficiency: Through the dynamic self-optimization reasoning engine (step 4), the model parameters are updated in real time to adapt the model to changes in equipment status and improve the ability to detect similar anomalies. The reinforcement learning method is used to dynamically adjust the reasoning strategy and optimize the reasoning path based on the feedback of the reasoning results to improve the efficiency and accuracy of fault diagnosis.

[0030] 5) Build and enhance dynamic knowledge graphs to achieve continuous accumulation and application of knowledge: Update newly discovered fault modes and causal relationships to the knowledge graph in a timely manner (step 6) to achieve dynamic updating and accumulation of knowledge. Use graph neural networks to extract and represent features of the knowledge graph, integrate the structural information of the knowledge graph into the reasoning model, assist in fault diagnosis and prediction, and achieve two-way feedback between the model and the knowledge graph.

[0031] 6) Improve fault prediction and decision support capabilities to promote preventive maintenance: Use the self-optimized reasoning engine and enhanced reasoning model to predict the future state of the equipment and calculate the probability of failure (step 7). When the predicted failure probability exceeds the set threshold, the system triggers an early warning, provides maintenance suggestions and decision support, and helps technicians take timely measures to reduce equipment downtime and maintenance costs.

[0032] A second object of the present invention is to provide an intelligent device fault diagnosis and reasoning system based on unsupervised learning, the system comprising:

[0033] Data collection and preprocessing module: multi-source data acquisition, collecting multi-dimensional sensor data, operation logs and environmental information of equipment, and integrating them into data sets; cleaning, missing value filling and normalization of data to ensure data consistency and availability;

[0034] Unsupervised abnormal feature extraction and detection module: feature dimensionality reduction and representation learning, using unsupervised learning models such as deep autoencoders to extract low-dimensional feature representation Z from preprocessed data and build a behavioral model for normal equipment operation; abnormality detection, by monitoring the reconstruction error of real-time data, when the error exceeds the threshold τ, it is determined to be abnormal, and abnormal data points are discovered in time;

[0035] Dynamic causal reasoning module based on unsupervised learning: causal structure construction, using unsupervised causal discovery algorithm to extract the causal relationship between equipment parameters from the multimodal feature space Z and generate a causal network; dynamic modeling, introducing dynamic Bayesian network (DBN) or temporal causal model to capture the evolution of equipment parameters over time and the fault formation process;

[0036] Dynamic self-optimization inference engine module: Model adaptive update, for new data, real-time update of model parameters θ, so that the model can continuously adapt to changes in device status; human-computer interaction and annotation feedback, request expert annotation when unknown abnormal patterns are found, and use new annotation data to improve anomaly detection and diagnosis performance; reinforcement learning optimization, dynamically adjust the inference strategy and path according to the inference results and feedback information, and optimize the overall inference efficiency;

[0037] Intelligent equipment fault reasoning and diagnosis module: fault pattern discovery and classification, classifying abnormal data through clustering algorithms, and incorporating newly emerged fault patterns into the pattern library; multi-level reasoning architecture, building a multi-level framework including data preprocessing layer, anomaly detection layer, causal reasoning layer and decision support layer, to achieve layer-by-layer reasoning from data to decision; fault diagnosis strategy optimization, combining Bayesian reasoning and reinforcement learning methods, continuously optimizing fault diagnosis decision strategies and improving diagnosis efficiency;

[0038] Dynamic knowledge graph enhancement module: knowledge graph construction and update, constructing equipment operation data, failure modes, maintenance records and other information into a dynamic knowledge graph; when new failure modes or causal relationships are discovered, timely update graph nodes and edges; (2) graph representation and fusion, using graph neural network (GNN) to extract features from the knowledge graph, integrating the graph structure into the reasoning model, and realizing two-way information interaction between the reasoning model and the knowledge graph;

[0039] Fault prediction and decision support module: Fault prediction, based on the self-optimizing reasoning engine and enhanced causal reasoning model, predicts the future state of the equipment and calculates the probability of failure; early warning and decision support, issues an early warning when the predicted failure probability exceeds the threshold, and combines the fault mode library and knowledge graph to provide diagnostic suggestions and maintenance plans, support technicians to proactively perform preventive maintenance, and reduce equipment downtime and costs.

[0040] Preferably, the unsupervised abnormal feature extraction and detection module adopts a deep autoencoder model to learn the normal feature distribution of the device by minimizing the reconstruction error between the input data and the reconstructed data.

[0041] Preferably, the dynamic causal reasoning module based on unsupervised learning utilizes a PC algorithm or a LiNGAM algorithm to construct a causal network to capture the causal relationship between device parameters.

[0042] Preferably, the dynamic self-optimizing reasoning engine module adopts a reinforcement learning method to dynamically adjust the reasoning strategy according to the feedback of the reasoning results to optimize the path and efficiency of fault diagnosis; the intelligent device fault reasoning and diagnosis module uses a clustering algorithm to classify abnormal data, discover unknown fault modes, and incorporate them into a fault mode library.

[0043] Preferably, the dynamic knowledge graph enhancement module uses a graph neural network to extract and represent features of the knowledge graph, and integrates the structural information of the knowledge graph into the reasoning model to assist in fault diagnosis and prediction; when the predicted fault probability exceeds a set threshold, the fault prediction and decision support module combines the fault mode library and the knowledge graph to provide targeted maintenance suggestions and decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the intelligent device fault diagnosis and reasoning method based on unsupervised learning of the present invention;

[0045] Figure 2 This is a framework diagram of the intelligent device fault diagnosis and reasoning system based on unsupervised learning of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0048] The present invention provides a method and system for intelligent device fault diagnosis and reasoning based on unsupervised learning, which innovatively combines advanced technologies such as unsupervised learning, causal reasoning, reinforcement learning, and knowledge graphs, and proposes a method and system for device fault diagnosis and reasoning without a large amount of labeled data. The method and system can dynamically adapt to changes in device status, continuously learn and accumulate new fault modes and knowledge, improve the accuracy of fault detection and diagnostic efficiency, and promote the implementation of preventive maintenance. The preferred implementation of the present invention is described in detail below in conjunction with the accompanying drawings.

[0049] Embodiment 1:

[0050] This example provides a smart device fault diagnosis and reasoning method based on unsupervised learning, including: Step 1: data collection and preprocessing.

[0051] 1) Multi-source data collection: Multi-source data collection includes sensor data (X), operation log (L) data and environmental information (E) data. Sensor data (X) collects real-time data from multiple sensors on the device, such as temperature, pressure, vibration, current, voltage and other parameters. Operation log (L) records the device's operating status, control instructions, operator information and timestamp, etc. Environmental information (E) obtains external conditions such as temperature, humidity, dust concentration and other external conditions of the environment where the device is located.

[0052] 2) Data preprocessing: Data preprocessing includes data cleaning, missing value filling and normalization. Data cleaning removes obviously erroneous data points, such as extreme value data caused by sensor failure. Missing value filling uses interpolation, mean filling or prediction methods based on time series models to fill missing data. Normalization standardizes data of different dimensions to the same scale range, such as using Min-Max normalization or Z-score normalization.

[0053] Step 2: Unsupervised anomaly feature extraction and detection.

[0054] 1) Build a deep autoencoder model: Design a multi-layer neural network, including encoder and decoder parts. The encoder compresses the high-dimensional input data into a low-dimensional feature space Z, and the decoder reconstructs the feature space Z back to the original data space. The autoencoder is trained using preprocessed normal operation data, with the goal of minimizing the reconstruction error between the input and the reconstructed output. Alternative implementations of the unsupervised anomaly detection method include:

[0055] Variational Autoencoder (VAE): An alternative to deep autoencoders, it uses a probabilistic generative model to learn the underlying distribution of data and detect abnormal data points by calculating reconstruction errors or probability densities.

[0056] Generative Adversarial Network (GAN): GAN is used to generate normal data distribution, and the discriminator is used to identify abnormal data points. The accuracy of anomaly detection is improved by training the adversarial process between the generator and the discriminator.

[0057] Isolation Forest: An unsupervised anomaly detection method based on random forests. It randomly divides data, detects the degree of isolation of data, and identifies anomalies.

[0058] Gaussian Mixture Model (GMM): uses probability density estimation to cluster and detect anomalies in data and identify outliers that are different from the main data group.

[0059] 2) Real-time anomaly detection: During the operation of the device, new data is input into the trained autoencoder in real time to calculate the reconstruction error of each data point. A reconstruction error threshold τ is set. When the reconstruction error exceeds τ, the data point is determined to be an abnormal data point, triggering subsequent processing.

[0060] Step 3: Dynamic Causal Inference Based on Unsupervised Learning

[0061] 1) Multimodal data fusion: Map sensor data X, operation log L, and environmental information E to a unified feature space Z to form a comprehensive feature representation.

[0062] 2) Causal relationship discovery: Use the PC algorithm or LiNGAM algorithm to infer the causal relationship of the variables in the feature space Z and build a causal relationship network G = (V, E). Use time series data to build a dynamic Bayesian network (DBN) or a temporal causal model to capture the dependencies between variables that change over time. Alternative causal relationship discovery implementation methods include:

[0063] Granger Causality Test: For time series data, determine whether one variable has Granger causality to another variable and establish a causal relationship between variables.

[0064] Transfer Entropy: Based on information theory, it quantifies the information transfer from one random process to another and discovers the causal relationship of nonlinear and non-Gaussian distributions.

[0065] Structural Equation Modeling (SEM): Utilizes statistical models to analyze the causal relationship between observed variables and latent variables and establish associations between device parameters.

[0066] Causal Effect Inference: Use the potential outcome framework and instrumental variable method to estimate causal effects and build causal models.

[0067] Step 4: Dynamic self-optimizing inference engine

[0068] 1) Real-time update of model parameters: Online learning strategy, for newly arrived data, uses online learning algorithm to update model parameters θ, including the weights and biases of the autoencoder, and the structure and parameters of the causal model. This enables the model to adapt to changes in device status and improve sensitivity to new patterns and anomalies. Alternative implementation methods for model self-optimization:

[0069] Adaptive learning rate optimizer: Uses optimization algorithms such as Adam and RMSprop to dynamically adjust the learning rate of the model to improve the efficiency and stability of model training.

[0070] Federated Learning: Distribute training models across multiple devices, merge model updates from each device, improve model generalization capabilities, and protect data privacy.

[0071] Transfer Learning: Transfer the knowledge of the pre-trained model to the new model to accelerate the training and optimization of the model and adapt it to new devices or new environments.

[0072] Meta-Learning: By learning the ability to learn, the model can quickly adapt to new tasks and data, improving the model's self-optimization ability.

[0073] 2) Expert labeling and model improvement: When an unknown abnormal pattern is detected, the system actively requests experts to label it and obtain accurate label information. Using the expert-labeled data, the model is retrained or fine-tuned to improve the model's ability to detect and classify similar anomalies.

[0074] 3) Reinforcement learning optimization: Using reinforcement learning algorithms, dynamically adjust the reasoning strategy based on the feedback of the reasoning results, optimize the reasoning path and resource allocation. Set reward functions based on diagnostic accuracy, response time, and resource consumption to guide the model to develop in the optimal direction.

[0075] Step 5: Intelligent device fault reasoning and diagnosis

[0076] 1) Multi-level reasoning architecture construction: The data preprocessing layer processes and normalizes the input data. The anomaly detection layer detects abnormal data points in real time. The causal reasoning layer analyzes the causal relationship of abnormal data and finds the source of the fault. The decision support layer provides fault diagnosis results and maintenance suggestions.

[0077] 2) Abnormal data clustering and new fault mode discovery: Cluster abnormal data and use algorithms such as K-means and density clustering to identify similar abnormal patterns. Add newly discovered fault modes to the fault mode library to improve system knowledge. Alternative implementation methods for abnormal data processing and fault mode discovery include:

[0078] Density clustering algorithms: such as DBSCAN, which clusters data based on density, can discover clusters of any shape and identify abnormal data.

[0079] Principal Component Analysis (PCA): Reduces the dimension of the data and captures the main features. Abnormal data appears as deviated points in the principal component space.

[0080] Support Vector Data Description (SVDD): Construct a minimal hypersphere to enclose normal data, and data outside the hypersphere are considered abnormal.

[0081] 3) Strategy optimization: Combine prior knowledge and current observations to calculate the posterior probability of different faults and prioritize high-risk faults. Based on historical diagnosis results, continuously optimize decision-making strategies to improve diagnosis efficiency.

[0082] Step 6: Dynamic Knowledge Graph Enhancement

[0083] 1) Knowledge graph construction: Nodes include equipment components, sensors, failure modes, maintenance records, etc. Edge relationships represent the relationships between components, between components and failures, and between failures and maintenance measures, such as causal relationships, association relationships, etc.

[0084] 2) Knowledge graph update: When new failure modes or causal relationships are discovered, corresponding nodes and edges are added to the knowledge graph in a timely manner to ensure consistency between the newly added information and existing knowledge to avoid conflicts.

[0085] 3) Graph Neural Network Application: Use GNN (such as GCN, GAT) to learn features of knowledge graphs and obtain high-dimensional representations of nodes. Incorporate the structural information of knowledge graphs into fault diagnosis models to improve the reasoning ability of the models.

[0086] 4) Bidirectional feedback between model and knowledge graph: The reasoning results of the model can be used to improve the knowledge graph. The information of the knowledge graph improves the model's ability to diagnose complex faults. Alternative knowledge representation and update methods:

[0087] Ontology: Build an ontology model in the device field, define concepts, attributes and relationships, and use semantic web technology to express, reason and update knowledge.

[0088] Knowledge base and rule engine: Establish a rule-based knowledge base, use the rule engine for reasoning, and support dynamic addition and modification of knowledge.

[0089] Case-Based Reasoning (CBR): Provides fault diagnosis and decision support by storing and retrieving past fault cases and performing similarity matching.

[0090] Experience replay in reinforcement learning: In the reinforcement learning model, the experience replay mechanism is used to accumulate historical experience, continuously update knowledge, and improve decision-making performance.

[0091] Step 7: Failure prediction and decision support

[0092] 1) Equipment status prediction: Use deep learning models such as LSTM and GRU to predict future trends of key sensor parameters. Calculate the probability of failure based on the prediction results and historical data. Alternative implementation methods for fault prediction and early warning methods:

[0093] Health index modeling quantifies the overall health status of the equipment.

[0094] Bayesian network prediction utilizes prior and observed data to derive the posterior failure probability.

[0095] Threshold-based monitoring: Set normal ranges or thresholds for key parameters and trigger alerts when the parameters exceed the range.

[0096] 2) Early warning mechanism: Set a fault probability threshold. When the predicted fault probability exceeds the threshold, the system automatically triggers an early warning. Relevant personnel are notified via SMS, email, alarm, etc.

[0097] 3) Repair suggestions and decision support: Combine knowledge graphs and fault mode libraries to analyze possible causes of failures. Provide specific repair and maintenance suggestions, including replacing parts, adjusting parameters, and inspecting specific parts. Model optimization and decision support alternative implementation methods:

[0098] Multi-objective optimization algorithms: such as genetic algorithms and particle swarm optimization algorithms, which optimize multiple performance indicators of the model, such as accuracy, response time, and resource consumption.

[0099] Fuzzy logic system: handles uncertainty and ambiguity, and uses fuzzy rules for fault diagnosis and decision support.

[0100] Expert system: Based on expert knowledge and rules, an inference engine is established to diagnose faults and provide treatment suggestions.

[0101] 4) Implementation of preventive maintenance: According to the prediction results, formulate a preventive maintenance plan and arrange maintenance time and personnel. Through advance maintenance, sudden failures can be avoided and downtime and maintenance costs can be reduced.

[0102] Example 2

[0103] This embodiment also provides an intelligent device fault diagnosis and reasoning system based on unsupervised learning, such as Figure 2 As shown, the system includes:

[0104] Data acquisition and preprocessing module:

[0105] Multi-source data acquisition: Collect multi-dimensional sensor data, operation logs, and environmental information of the equipment, and integrate them into a data set D = {X, L, E}.

[0106] Improve data quality: clean the data, fill in missing values, and normalize the data to ensure data consistency and availability.

[0107] Unsupervised anomaly feature extraction and detection module:

[0108] Feature dimensionality reduction and representation learning: Use unsupervised learning models such as deep autoencoders to extract low-dimensional feature representations Z from preprocessed data and build a behavioral model for normal equipment operation.

[0109] Anomaly detection: By monitoring the reconstruction error of real-time data, when the error exceeds the threshold τ, it is determined to be abnormal and abnormal data points are discovered in time.

[0110] Dynamic causal reasoning module based on unsupervised learning:

[0111] Causal structure construction: Use unsupervised causal discovery algorithms (such as PC and LiNGAM) to extract the causal relationship between device parameters from the multimodal feature space Z and generate a causal network G = (V, E).

[0112] Dynamic modeling: Dynamic Bayesian networks (DBNs) or temporal causal models are introduced to capture the evolution of equipment parameters over time and the fault formation process.

[0113] Dynamic self-optimizing inference engine module:

[0114] Model adaptive update: For new data, the model parameters θ are updated in real time so that the model can continuously adapt to changes in device status.

[0115] Human-computer interaction and annotation feedback: Request expert annotation when unknown abnormal patterns are found, and use new annotation data to improve anomaly detection and diagnosis performance.

[0116] Reinforcement learning optimization: Dynamically adjust the reasoning strategy and path based on the reasoning results and feedback information to optimize the overall reasoning efficiency.

[0117] Intelligent equipment fault reasoning and diagnosis module:

[0118] Fault mode discovery and classification: Abnormal data is classified through clustering algorithms, and new fault modes are included in the pattern library.

[0119] Multi-level reasoning architecture: Build a multi-level framework including data preprocessing layer, anomaly detection layer, causal reasoning layer and decision support layer to achieve layer-by-layer reasoning from data to decision.

[0120] Fault diagnosis strategy optimization: Combine Bayesian reasoning and reinforcement learning methods to continuously optimize fault diagnosis decision-making strategies and improve diagnostic efficiency.

[0121] Dynamic knowledge graph enhancement module:

[0122] Knowledge graph construction and update: Equipment operation data, failure modes, maintenance records and other information are constructed into a dynamic knowledge graph G = (V, E). When new failure modes or causal relationships are discovered, graph nodes and edges are updated in a timely manner.

[0123] Graph representation and fusion: Use graph neural network (GNN) to extract features from knowledge graphs, integrate graph structures into reasoning models, and realize two-way information interaction between reasoning models and knowledge graphs.

[0124] Fault prediction and decision support module:

[0125] Fault prediction: Based on the self-optimizing reasoning engine and enhanced causal reasoning model, the future state of the equipment is predicted and the probability of fault occurrence is calculated.

[0126] Early warning and decision support: When the predicted failure probability exceeds the threshold, an early warning is issued, and combined with the failure mode library and knowledge graph, diagnostic suggestions and maintenance plans are provided to support technicians to proactively perform preventive maintenance and reduce equipment downtime and costs.

[0127] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0128] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A method for intelligent device fault diagnosis and reasoning based on unsupervised learning, characterized in that: The following steps are involved: (1) Data collection and preprocessing: Collect multi-dimensional sensor data, operation logs, and environmental information of the equipment to form a multi-source data set D = {X, L, E}, where X is the sensor data, L is the operation log, and E is the environmental information. The collected data is cleaned, missing values ​​are filled, and normalized. (2) Unsupervised abnormal feature extraction and detection: using unsupervised learning models such as deep autoencoders, we extract low-dimensional feature representations Z from preprocessed data and build a normal behavior model for the device. By minimizing the reconstruction error, the model learns the feature distribution of the device under normal operating conditions. We monitor the reconstruction error of new data in real time, and when the reconstruction error exceeds the set threshold τ, we determine it as an abnormal data point. (3) Based on unsupervised learning, dynamic causal reasoning is used to map multimodal data such as sensor data, operation logs, and environmental information into a unified feature space Z, and an unsupervised causal discovery algorithm is used to construct a causal relationship network between device parameters. Use dynamic Bayesian networks or temporal causal models to model the dependencies of variables over time and capture the dynamic evolution of faults; (4) Dynamic self-optimizing inference engine: for newly arrived data, the model parameters θ are updated in real time to adapt the model to changes in the device status; when an unknown abnormal pattern is detected, experts are actively requested to label it to obtain high-quality labeled data; the model is updated using the newly labeled data to improve the ability to detect similar anomalies; reinforcement learning methods are used to dynamically adjust the inference strategy and optimize the inference path based on the feedback of the inference results; (5) Intelligent equipment fault reasoning and diagnosis, building a multi-level reasoning architecture including data preprocessing layer, anomaly detection layer, causal reasoning layer and decision support layer; Use clustering algorithms to classify abnormal data and discover unknown fault modes; incorporate newly discovered fault modes into the fault mode library to continuously enrich the knowledge of the system; combine Bayesian reasoning and reinforcement learning to optimize the decision-making strategy of fault diagnosis and improve diagnostic efficiency; (6) Dynamic knowledge graph enhancement: The equipment’s operating data, failure modes, maintenance records and other information are constructed into a dynamic knowledge graph. When the unsupervised learning model discovers new failure modes or relationships, the knowledge graph is updated in a timely manner, and new nodes and edges are added. Using graph neural networks, we extract and represent features of knowledge graphs, and integrate the structural information of knowledge graphs into reasoning models. The structural information of knowledge graphs assists fault diagnosis and prediction, and realizes two-way feedback between models and knowledge graphs. (7) Fault prediction and decision support: using a self-optimized reasoning engine and enhanced reasoning model to predict the future state of the equipment and calculate the probability of a fault. When the predicted probability of a fault exceeds the set threshold, the system triggers an early warning and notifies relevant personnel. Combined with the fault mode library and knowledge graph, it provides maintenance suggestions and decision support to help technicians take timely measures.

2. The method according to claim 1, characterized in that In step 4, the updating of the model parameters θ includes adjusting the weights and biases of the deep autoencoder, and optimizing the parameters of the causal reasoning model and the reinforcement learning model.

3. The method according to claim 1, characterized in that In step 5, a clustering algorithm is used to classify abnormal data, and the clustering algorithm used is K-means, density clustering or hierarchical clustering algorithm.

4. The method according to claim 1, characterized in that: In step 6, the graph convolutional network (GCN) or graph attention network (GAT) in the graph neural network is used to extract and represent features of the knowledge graph.

5. The method according to claim 1, characterized in that In step 7, when the predicted failure probability exceeds a threshold, the system automatically generates warning information, including possible causes of the failure, impact range, and recommended treatment measures.

6. An intelligent device fault diagnosis and reasoning system based on unsupervised learning, comprising: Data collection and preprocessing module: multi-source data acquisition, collecting multi-dimensional sensor data, operation logs and environmental information of equipment, and integrating them into data sets; Clean the data, fill in missing values, and normalize the data to ensure data consistency and availability; Unsupervised abnormal feature extraction and detection module: feature dimensionality reduction and representation learning, using unsupervised learning models such as deep autoencoders to extract low-dimensional feature representation Z from preprocessed data and build a behavioral model for normal equipment operation; abnormality detection, by monitoring the reconstruction error of real-time data, when the error exceeds the threshold τ, it is determined to be abnormal, and abnormal data points are discovered in time; Dynamic causal reasoning module based on unsupervised learning: causal structure construction, using unsupervised causal discovery algorithm to extract the causal relationship between equipment parameters from the multimodal feature space Z and generate a causal network; dynamic modeling, introducing dynamic Bayesian network (DBN) or temporal causal model to capture the evolution of equipment parameters over time and the fault formation process; Dynamic self-optimization inference engine module: Model adaptive update, for new data, real-time update of model parameters θ, so that the model can continuously adapt to changes in device status; human-computer interaction and annotation feedback, request expert annotation when unknown abnormal patterns are found, and use new annotation data to improve anomaly detection and diagnosis performance; reinforcement learning optimization, dynamically adjust the inference strategy and path according to the inference results and feedback information, and optimize the overall inference efficiency; Intelligent equipment fault reasoning and diagnosis module: fault mode discovery and classification, classifying abnormal data through clustering algorithm, and incorporating new fault modes into the pattern library; Multi-level reasoning architecture: construct a multi-level framework including data preprocessing layer, anomaly detection layer, causal reasoning layer and decision support layer to achieve layer-by-layer inference from data to decision; fault diagnosis strategy optimization: combine Bayesian reasoning and reinforcement learning methods to continuously optimize fault diagnosis decision strategies and improve diagnosis efficiency; Dynamic knowledge graph enhancement module: knowledge graph construction and update, which constructs equipment operation data, failure modes, maintenance records and other information into a dynamic knowledge graph; When new fault modes or causal relationships are discovered, graph nodes and edges are updated in a timely manner; (2) Graph representation and fusion: Graph neural networks (GNNs) are used to extract features from knowledge graphs, and graph structures are integrated into reasoning models to achieve two-way information interaction between reasoning models and knowledge graphs; Fault prediction and decision support module: Fault prediction, based on the self-optimizing reasoning engine and enhanced causal reasoning model, predicts the future state of the equipment and calculates the probability of fault occurrence; Early warning and decision support: When the predicted failure probability exceeds the threshold, an early warning is issued, and combined with the failure mode library and knowledge graph, diagnostic suggestions and maintenance plans are provided to support technicians to proactively perform preventive maintenance and reduce equipment downtime and costs.

7. The system according to claim 6, characterized in that The unsupervised abnormal feature extraction and detection module adopts a deep autoencoder model to learn the normal feature distribution of the device by minimizing the reconstruction error between the input data and the reconstructed data.

8. The system according to claim 6, characterized in that The dynamic causal reasoning module based on unsupervised learning uses the PC algorithm or the LiNGAM algorithm to construct a causal relationship network to capture the causal relationship between device parameters.

9. The system according to claim 6, characterized in that The dynamic self-optimizing reasoning engine module adopts a reinforcement learning method to dynamically adjust the reasoning strategy according to the feedback of the reasoning results to optimize the path and efficiency of fault diagnosis; the intelligent device fault reasoning and diagnosis module uses a clustering algorithm to classify abnormal data, discover unknown fault modes, and incorporate them into the fault mode library.

10. The system according to claim 6, characterized in that The dynamic knowledge graph enhancement module uses a graph neural network to extract and represent features of the knowledge graph, integrates the structural information of the knowledge graph into the reasoning model, and assists in fault diagnosis and prediction; when the predicted fault probability exceeds a set threshold, the fault prediction and decision support module combines the fault mode library and the knowledge graph to provide targeted maintenance suggestions and decision support.

Images