A Fault Diagnosis System and Method for Compressed Air Energy Storage Power Station
By constructing a fault diagnosis system based on semantic anchoring and physical information reasoning, the problems of multimodal data fusion and physical law constraints in the fault diagnosis of compressed air energy storage power stations were solved, achieving high diagnostic accuracy, physical compliance rate and precise spatial positioning, and improving the robustness and interpretability of the system.
Patent Information
- Application Number
- CN202611323716.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing fault diagnosis technologies for compressed air energy storage power stations lack high diagnostic accuracy, high physical compliance, precise spatial positioning capabilities, and multimodal data fusion capabilities.
A fault diagnosis system based on semantic anchoring and physical information reasoning is adopted. By constructing digital semantic models, multimodal coding, hybrid reasoning and knowledge reasoning modules, it realizes the fusion of multimodal data and physical law constraints. It combines BIM entities as anchor points to perform cross-modal feature alignment and physical constraints, and uses knowledge graphs for structured reasoning.
It achieves deep integration of multimodal heterogeneous data, organic unity of physical laws and data-driven approaches, precise fault spatial location capability, interpretable and traceable reasoning process, strong system robustness, knowledge accumulation and sustainable evolution, and friendly engineering deployment.
Smart Images

Figure CN123115897A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for compressed air energy storage power stations, and in particular to a fault diagnosis system and method for compressed air energy storage power stations. Background Technology
[0002] Compressed air energy storage (CAES) technology, as a large-scale, long-life, and environmentally friendly energy storage method, plays an important role in peak shaving and valley filling and promoting the consumption of renewable energy in new power systems. As CAES power plants develop towards larger capacity and higher parameters, their equipment composition is becoming increasingly complex, and the degree of system coupling is deepening, which places higher demands on fault diagnosis technology.
[0003] Currently, there are three main technical approaches for fault diagnosis in CAES power plants, all of which have significant drawbacks: (i) Rule-Based Systems. This type of method relies on domain experts to manually compile an "IF-THEN" diagnostic rule base, which makes logical judgments based on equipment operating parameters and preset thresholds. Its advantage lies in its high physical compliance (approximately 100%), meaning that all diagnostic conclusions strictly conform to physical laws. However, this method has the following significant shortcomings: (1) The rule base has a limited coverage, and the diagnostic accuracy drops sharply when faced with unseen composite fault modes, with an accuracy of only about 76.3% in typical application scenarios; (2) The rule maintenance cost is high, and a large number of rules need to be rewritten manually after the power plant equipment is upgraded; (3) It lacks spatial perception capabilities and cannot accurately locate faults.
[0004] (ii) Traditional Machine Learning. This type of method uses traditional machine learning models such as support vector machines, random forests, and gradient boosting trees to classify faults based on manually designed statistical features (such as time-domain features, frequency-domain features, wavelet packet energy, etc.). Compared with rule-based systems, this type of method has a certain generalization ability, but it has the following defects: (1) It is heavily dependent on the quality of feature engineering, requiring domain experts to manually extract features, which is cumbersome and prone to missing key information; (2) It cannot effectively integrate multimodal heterogeneous data (text, images, time series, acoustics, point clouds, etc.), resulting in low information utilization; (3) The model has poor interpretability and cannot provide a reasoning process that conforms to physical laws; (4) In scenarios where fault samples are scarce, the model's generalization performance is insufficient.
[0005] (III) General-Purpose LLM Fine-Tuning. In recent years, general-purpose LLM models, represented by GPT, BERT, and LLaMA, have demonstrated powerful capabilities in natural language processing tasks. Some studies have attempted to apply general-purpose LLM fine-tuning to industrial fault diagnosis. However, this type of method has the following fundamental defects: (1) General-purpose LLM models lack domain physics knowledge, and the physical compliance rate after fine-tuning is only 65.8%, meaning that more than one-third of the diagnostic conclusions violate physical laws; (2) They cannot understand the BIM three-dimensional spatial structure and equipment topology of CAES power plants; (3) They cannot process multimodal data such as sensor time-series data and images, and can only process text information; (4) The reasoning process is uncontrollable and may generate seemingly reasonable but actually incorrect diagnostic conclusions.
[0006] In summary, existing technologies lack a CAES power plant intelligent fault diagnosis solution that can simultaneously achieve high diagnostic accuracy, high physical compliance, precise spatial positioning capabilities, and multimodal data fusion capabilities. Summary of the Invention
[0007] To address the above problems, the present invention provides a fault diagnosis system and method for compressed air energy storage power stations, which solves the problems of poor generalization ability of rule systems, weak multimodal fusion ability of traditional machine learning methods, low physical compliance rate and lack of spatial awareness of general large language models in existing fault diagnosis technologies for compressed air energy storage power stations.
[0008] According to a first aspect of the present invention, an intelligent fault diagnosis system for energy storage power stations based on semantic anchoring and physical information reasoning is provided, comprising: The semantic backbone module is used to construct a digital semantic model of the energy storage power station based on preset data standards. A multimodal coding module, connected to the semantic backbone module, is used to receive sensing data from the energy storage power station, extract features from the sensing data, and perform cross-modal feature alignment using the device entities in the semantic backbone module as anchor points to generate multimodal fusion features. The hybrid reasoning module, connected to the multimodal encoding module, is used to perform fusion reasoning on the multimodal fusion features and apply physical law constraints to output deep features that conform to physical laws. The knowledge reasoning module, connected to the hybrid reasoning module, is used to perform structured reasoning on the deep features and output fault diagnosis conclusions.
[0009] In the above scheme, the digital semantic model of the energy storage power station is constructed according to a preset data standard, including: Based on the IFC4.3 standard, the special entity type of the energy storage power station is extended, and the equipment, sensors and pipeline connections of the power station are constructed into a topology map with spatial coordinates; Assign each entity a spatial anchor containing three-dimensional geometric coordinates, partition identifiers, and system identifiers; Four semantic layers are constructed, comprising a process system layer, an equipment layer, a sensor layer, and a space layer, with each layer interconnected through semantic relationships; The topology graph, the spatial anchor, and the four semantic layers together constitute the digital semantic model.
[0010] In the above scheme, the topology graph is equipped with a topology attention mask, which is used to determine the attention weights between different device nodes; Among them, directly connected device nodes receive the highest attention weight, device nodes with the shortest path hop count not exceeding a preset threshold receive attention weight related to the hop count, and device nodes with the shortest path hop count exceeding a preset threshold have their attention weight set to a minimum value.
[0011] In the above scheme, the multimodal coding module includes multiple dedicated encoders, which are used to process text data, image data, time series data, acoustic data and point cloud data respectively; the multimodal coding module uses the device entity identifier as the anchor point, and through comparative learning, it brings different modal features belonging to the same device entity closer in the feature space, and pushes features belonging to different device entities further apart.
[0012] In the above scheme, the image encoder generates a region of interest mask based on the semantic backbone module to guide the attention mechanism; the temporal encoder maps temporal data to the corresponding device entity based on the installation relationship between the sensor and the device; the acoustic encoder uses the spatial anchor to assist in sound source localization; and the output of the point cloud encoder is registered with the coordinate system of the semantic backbone module.
[0013] In the above scheme, the hybrid inference module includes alternately stacked data-driven blocks and physical constraint blocks; The data-driven block is used to learn the hidden patterns in the data, and the physical constraint block is used to introduce the control equations of the energy storage power station as soft constraints into the network training; the weights of the physical constraints gradually increase from the initial value to the final value according to the training process.
[0014] In the above scheme, the loss function of the hybrid inference module includes a data fitting loss term and a physical constraint loss term. The physical constraint loss term is calculated based on the residuals of the control equation, and the weight of the physical constraint is used to dynamically adjust the influence of the physical constraint loss term on the total loss.
[0015] In the above scheme, the knowledge reasoning module includes: A two-layer knowledge graph, comprising a hard knowledge graph and a soft knowledge graph, wherein the hard knowledge graph is used to store formalized physical equation knowledge, and the soft knowledge graph is used to store historical fault cases and operation and maintenance experience; The inference engine is used to execute the structured inference process, which decomposes the complex fault symptoms step by step and performs evidence retrieval and confidence calculation based on the two-layer knowledge graph. The verification unit is used to verify the reasoning result according to the physical constraints. If the verification fails, the reasoning is retried.
[0016] In the above scheme, the knowledge reasoning module further includes an adversarial training component, which is used to enhance the robustness of the system by generating adversarial examples. The adversarial training component includes an attack unit and a defense unit. The attack unit is used to generate adversarial examples that simulate sensor anomalies, and the defense unit is used to correct the adversarial examples based on historical failure cases in the soft knowledge graph.
[0017] According to a second aspect of the present invention, a method for intelligent fault diagnosis of an energy storage power station based on any of the above-described solutions is provided, comprising the following steps: Step S1: Construct a digital semantic model of the energy storage power station according to preset data standards; Step S2: Receive the sensing data of the energy storage power station, extract features from the sensing data, and perform cross-modal feature alignment using the device entities in the semantic backbone as anchor points to generate multimodal fusion features; Step S3: Perform fusion inference on the multimodal fusion features and apply physical law constraints to output deep features that conform to physical laws; Step S4: Perform structured reasoning on the deep features based on the knowledge graph and output the fault diagnosis conclusion.
[0018] The beneficial effects of this invention are: (i) Deep fusion capability of multimodal heterogeneous data. This invention achieves cross-modal unified representation of five heterogeneous sensing data, namely text, image, time series, acoustic, and point cloud, through the collaborative work of five dedicated encoders in the multimodal coding module and the cross-modal comparison and alignment strategy with BIM entities as anchor points, making full use of the existing multi-source information of the power station.
[0019] (II) Organic unity of physical laws and data-driven approach. This invention embeds the CAES control equations as soft constraints into the network training process through a PINN-Transformer hybrid backbone network, and adopts an annealing scheduling strategy to ensure that the model is primarily data-driven in the early stages of training and strictly follows physical laws in the later stages of training, thus balancing the flexibility of data fitting with the rigor of physical laws.
[0020] (III) Precise fault spatial location capability. This invention assigns spatial anchors and topological attention masks to each equipment entity through the BIM semantic backbone module, enabling the system to accurately locate the three-dimensional spatial coordinates of the faulty equipment and the corresponding process system partition, providing direct guidance for maintenance personnel to quickly locate the fault.
[0021] (iv) The reasoning process is explainable and traceable. This invention makes each step traceable and verifiable through a five-step reasoning process. Combined with the forward verification and backtracking mechanism of the physical equations in the verification module, it fundamentally overcomes the "illusion" problem common in general large language models, and significantly improves the credibility and engineering practicality of the system.
[0022] (v) The system is highly robust. Through a red-blue team-based continuous error correction mechanism, the alternating attack and defense training of this invention enables the system to maintain stable diagnostic performance when faced with actual engineering interferences such as sensor drift, communication interruption, and environmental noise, thus meeting the deployment requirements of complex power plant environments.
[0023] (vi) Knowledge can be accumulated and continuously evolved. This invention uses a two-layer knowledge graph architecture. The hard knowledge graph ensures the accuracy of physical knowledge, while the soft knowledge graph continuously absorbs new fault cases and operation and maintenance experience, enabling the system's diagnostic capabilities and coverage to have the potential for continuous evolution.
[0024] (vii) User-friendly engineering deployment. The four core modules of this invention have clear functional boundaries and well-defined interfaces, supporting flexible deployment in a centralized or distributed manner. They automatically ignore missing modes through a masking mechanism and have good fault tolerance for incomplete field data. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0026] Figure 1 This is a schematic diagram of the overall architecture of the intelligent fault diagnosis system in an embodiment of the present invention.
[0027] Figure 2 This is a schematic diagram of the structure of the BIM semantic backbone module in an embodiment of the present invention.
[0028] Figure 3 This is a schematic diagram of the structure of the multimodal sensing coding module in an embodiment of the present invention.
[0029] Figure 4This is a schematic diagram of the structure of the PINN-Transformer hybrid backbone network module in an embodiment of the present invention.
[0030] Figure 5 This is a flowchart illustrating the workflow of the knowledge-enhanced thinking chain reasoning engine module in an embodiment of the present invention.
[0031] Figure 6 This is an overall flowchart of the intelligent fault diagnosis method in an embodiment of the present invention.
[0032] Figure 7 This is a schematic diagram illustrating the generation of the topological attention mask in an embodiment of the present invention.
[0033] Figure 8 This is a schematic diagram of the weight change curves for the four-stage annealing scheduling in an embodiment of the present invention.
[0034] Figure 9 This is a comparison chart of fault diagnosis results in a 300MW CAES demonstration power plant, based on an embodiment of the present invention.
[0035] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0036] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.
[0037] The terms "first," "second," etc., used in this disclosure are for distinguishing similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein.
[0038] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.
[0039] Multiple, including two or more.
[0040] And / or, it should be understood that, for the purposes of this disclosure, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.
[0041] According to a first aspect of the present invention, an intelligent fault diagnosis system for energy storage power stations based on semantic anchoring and physical information reasoning is provided, such as... Figure 1 As shown, it includes: The semantic backbone module is used to construct a digital semantic model of the energy storage power station based on preset data standards. The multimodal coding module, connected to the semantic backbone module, is used to receive sensing data from the energy storage power station, extract features from the sensing data, and perform cross-modal feature alignment using the device entities in the semantic backbone module as anchor points to generate multimodal fusion features. The hybrid reasoning module, connected to the multimodal coding module, is used to perform fusion reasoning on multimodal fusion features and apply physical law constraints to output deep features that conform to physical laws. The knowledge reasoning module, connected to the hybrid reasoning module, is used to perform structured reasoning on deep features and output fault diagnosis conclusions.
[0042] Specifically, such as Figure 2 As shown, the semantic backbone module is extended based on the IFC4.3 standard to construct a complete digital semantic model of the CAES power plant. Building upon the existing general entities in the IFC4.3 standard, such as IfcBuildingElementProxy, IfcSensor, and IfcFlowTerminal, the module extends the model with CAES power plant-specific entity types through the IfcProxy mechanism. These include: IfcCavernStorage (salt cave / artificial chamber gas storage), IfcTurboExpander (turboexpander unit), IfcMultiStageCompressor (multi-stage compressor unit), IfcRegenerativeHeatExchanger (regenerative heat exchanger), and IfcUndergroundSeal (underground gas storage sealing structure). Each entity is associated with its BIM spatial coordinates, its assigned zone, and its associated process system. In this embodiment, the digital semantic model consists of a topology map, spatial anchors, and four semantic layers.
[0043] Among them, such as Figure 7 As shown, the construction of the topology graph includes: abstracting the CAES power station as a directed graph G=(V,E), where V is the set of device nodes and E is the set of connecting edges (including pipe connections and electrical connections). For any two device nodes u and v, their topology attention mask... is calculated as follows: when device node u and device node v are directly connected (the number of hops h of the shortest path is 1), =1; when the number of hops h of the shortest path between u and v satisfies 1<h≤3, =1 / h; when the number of hops h of the shortest path between u and v is greater than 3, =ε, where ε is a minimum value close to 0. This mechanism ensures that the attention of fault diagnosis is focused on local areas with close device associations.
[0044] Assigning a spatial anchor to each entity includes: each entity is assigned a spatial anchor . Where x, y, z are three-dimensional geometric coordinates, adopting the unified coordinate system of the power station; zone is the partition identifier; system is the identifier of the affiliated process system. The spatial anchor data is derived from the design BIM model, and is extracted and converted from the IfcLocalPlacement attribute of each entity in the IFC file.
[0045] The four semantic layers are organized in order from the high level to the low level: the process system layer includes a compression system, an expansion power generation system, a gas storage system, a heat exchange system, a control system, etc.; the device layer includes all CAES power station device entities and their attribute parameters; the sensor layer includes temperature sensors, pressure sensors, flow sensors, vibration sensors, acoustic sensors, etc., and each sensor is associated with the installed device through a "MountedOn" relationship; the spatial layer is organized according to the functional partitions and spatial grids of the power station.
[0046] In the present invention, a digital semantic model is jointly constructed by the topological graph, the spatial anchors and the four semantic layers, so as to provide structured semantic support for subsequent multimodal perception encoding and hybrid reasoning.
[0047] This embodiment takes a certain 300MW-class advanced CAES demonstration power station as an example to describe in detail the implementation manner of the BIM semantic backbone module. The CAES demonstration power station includes the following main process areas and devices: (1) Compression area: multiple multi-stage centrifugal compressors are configured, interstage intercoolers are arranged, and the designed outlet pressure and temperature are reached. In the BIM model, an entity is established for each compressor, assigned the IfcMultiStageCompressor type, which includes attributes: rated power, number of stages, designed pressure ratio, designed inlet / outlet temperature / pressure values, etc.
[0048] (2) Gas storage area: an underground salt cavern is used as a gas storage container, with a designed volume of about 3×10 5 m 3The operating pressure range is 4.5-10MPa. The BIM model uses IfcCavernStorage solid modeling and assigns attributes such as volume, operating pressure range, 3D geometry of the salt cavern (generated from the BIM surface model after registration with LiDAR measured point clouds), and sealing structure parameters.
[0049] (3) Expansion power generation area: Two multi-stage turbo expanders (each with a rated power of 150MW) are configured with an inlet pressure of 9.5MPa and an outlet pressure of 0.1MPa. During the expansion process, heat needs to be supplemented from the regenerative heat exchanger.
[0050] (4) Heat exchange zone: Multiple shell and tube heat exchangers are configured for compression heat recovery and preheating of air before expansion.
[0051] (5) Control system: including DCS control system, SCADA data acquisition system, and various sensor networks.
[0052] The semantic backbone module adopted the following engineering practice parameters when constructing the above model: the topology graph contains approximately 3,500 entity nodes and approximately 12,000 relational edges; the spatial anchor adopts the unified coordinate system of the power station with millimeter-level precision; the four semantic layers are stored and indexed using the graph database Neo4j; the hop count threshold of the topology attention mask is set to 3, which is determined based on statistical analysis of the actual pipeline length of the power station -- the shortest path hop count between more than 95% of the functionally related devices in the CAES power station does not exceed 3.
[0053] Furthermore, such as Figure 3 As shown, the multimodal coding module includes multiple dedicated encoders for processing text data, image data, time-series data, acoustic data, and point cloud data, respectively. Using device entity identifiers as anchors, the module employs comparative learning to bring different modal features belonging to the same device entity closer together in the feature space, while pushing features belonging to different device entities further apart. Specifically: (1) Text data encoding: Text data such as operation and maintenance logs and technical documents are first preprocessed by word segmentation and standardization. A pre-trained language model is used as the basic encoder, and domain-adaptive pre-training is performed on the CAES power plant domain corpus. The text encoder outputs high-dimensional feature vectors.
[0054] (2) Image data encoding: Visible light images and infrared thermal imagers were acquired separately, and the image resolution was uniformly adjusted to 224×224 pixels. The encoder adopted the Geo-ViT (Geometrically-guided Vision Transformer) architecture. The BIM semantic backbone module generated the ROI mask of the device in the image through perspective projection transformation based on the spatial anchor coordinates of the device, which was superimposed on the attention calculation of the Vision Transformer as an additional attention guidance signal. The Geo-ViT output feature vectors with a dimension of 512.
[0055] (3) Time-series data encoding: The SCADA system collects parameters such as temperature, pressure, flow rate, and vibration amplitude at each measuring point of the power station at a sampling frequency of 1Hz. The time-series encoder adopts the DepthSformer architecture. Based on the MountedOn relationship between sensors and equipment in the BIM semantic backbone, the time-series data of each measuring point is mapped to its corresponding equipment entity. The input sequence length is 3600 time steps (i.e., a 1-hour data window), and the output dimension is a 256-dimensional feature vector.
[0056] (4) Acoustic and vibration data encoding: Accelerometer and microphone sample at high frequency. The raw signal is converted into a time-spectrum map through short-time Fourier transform. The encoder adopts a 1D-CNN architecture with attention mechanism. The spatial anchor coordinates provided by the BIM semantic backbone assist in the localization of sound sources. The 1D-CNN contains multiple convolutional layers and outputs feature vectors of appropriate dimensions.
[0057] (5) Point cloud data encoding: The point density of LiDAR point cloud data is approximately 1000 points / m. 2 The encoder uses the PointNet++ architecture. The point cloud is first registered with the design BIM model using the ICP algorithm to achieve a unified coordinate system, with registration accuracy controlled within 5mm. PointNet++ outputs a 256-dimensional feature vector.
[0058] (6) Cross-modal contrastive alignment: The multimodal coding module uses the device entity identifier as the anchor point. Through contrastive learning, it brings different modal features belonging to the same device entity closer together in the feature space, while pushing features belonging to different device entities further apart. Specifically, the InfoNCE contrastive loss function is used for training. This loss function is defined as: (1) Where q is the feature representation of the query modality. Let be the feature representation of positive samples, k be the feature representation of all positive and negative samples, s(·) be the similarity function, and τ be the temperature hyperparameter. After alignment training, multimodal features of the same entity with a cosine similarity greater than 0.8 in the embedding space are considered successfully aligned, while features of different entities with a cosine similarity less than 0.3 are considered successfully aligned. All modal features are finally merged into a multimodal fusion feature vector of uniform dimension through a learnable fusion layer.
[0059] In existing CAES power plant fault diagnosis technologies, multimodal data is processed in isolation, resulting in low information utilization. While traditional machine learning methods can handle structured data, they cannot effectively integrate unstructured information such as text and images. General-purpose large language models can only process text and cannot utilize multimodal data such as sensor time series and images. This invention uses BIM entities as a unified anchor point and maps five types of heterogeneous data to the same embedding space through InfoNCE contrastive loss. This allows different modal features belonging to the same equipment entity to be closer to each other in the feature space, while features belonging to different entities are further apart, achieving true cross-modal semantic alignment. This mechanism enables the maintenance experience recorded in text, the thermal state of equipment reflected in images, the operating trends implied in time series, the vibration features carried in the spectrum, and the geometric information presented in point clouds to mutually corroborate and complement each other, depicting the true health status of the equipment from multiple dimensions. Compared to the independent processing or simple splicing of modalities in existing technologies, the cross-modal contrastive alignment of this invention achieves deep fusion and complementary enhancement at the information level. After alignment training, all modal features are finally merged into a multimodal fusion feature vector of uniform dimension through a learnable fusion layer.
[0060] Furthermore, such as Figure 4 As shown, the hybrid inference module includes alternating stacked data-driven blocks and physical constraint blocks. In this embodiment, the data-driven blocks adopt a Transformer block architecture, and the physical constraint blocks adopt a PINN (Physical Information Neural Network) block architecture. Each Transformer block contains a multi-head self-attention mechanism, a feedforward neural network, and layer normalization. Each PINN block contains multiple fully connected networks, with each layer followed by an activation function. The alternation order is: input → Transformer block → alternating stacked PINN blocks → output layer.
[0061] The more than 40 CAES governing equations embedded in the hybrid inference module are categorized by physics domain as follows: gas dynamics equations (continuity equation, momentum equation, Bernoulli equation, etc.); thermodynamic equations (ideal / real gas equation of state, energy conservation equation, isentropic efficiency equation, etc.); heat transfer equations (Fourier heat conduction equation, convective heat transfer equation, heat exchanger efficiency-NTU relationship, etc.); and geotechnical equations (salt cave creep equation, three-dimensional thermoelastic equation, etc.). The residuals of these equations are automatically calculated in the PINN block through differentiation, serving as physical constraint losses. .
[0062] This invention achieves a deep fusion of data-driven and physics-driven approaches by alternately stacking Transformer and PINN blocks. The Transformer block learns implicit patterns and complex nonlinear relationships from multimodal fusion features; the PINN block introduces over 40 CAES governing equations as soft constraints into the network training, ensuring the model's output is physically compliant. This alternating stacking structure allows data-driven features and physics-constrained features to interact and merge at each layer, rather than being simply concatenated. Compared to purely data-driven models that easily produce conclusions that violate physical principles, and purely physics-based models that struggle to handle complex nonlinear scenarios, this invention balances the flexibility of data fitting with the rigor of physical laws. like Figure 8 As shown, this embodiment employs a four-stage annealing scheduling strategy: Stage 1 (epoch 0-50): =0.01; Second phase (epochs 51-100): =0.1; Third stage (epochs 101-200): =1.0; Fourth stage (after epoch 201): =10.0. Regularization weight. Fixed at 1×10 -4 The annealing scheduling strategy prioritizes learning data distribution patterns in the early stages of training and strictly follows physical laws in the later stages of training, effectively avoiding the problem of model convergence difficulties caused by excessive physical constraints in the early stages of training.
[0063] In this invention, physical constraints are soft constraints, which refer to the way physical equations are introduced into network training in the form of loss function penalty terms, so that they participate in gradient descent optimization together with data-driven terms.
[0064] The training of the hybrid inference module employs the following strategies: All 40+ CAES governing equations are dimensionless to eliminate dimensional differences between different physical quantities. Gradient clipping is performed on the physical residuals output by each PINN block to prevent gradient explosion during training. A course-based learning strategy is used, introducing physical constraints sequentially according to equation complexity: first, the fundamental equations of mass and energy conservation; then, the gas law and heat transfer equations; and finally, the geotechnical mechanics and thermal stress coupling equations. Training utilizes the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 Cosine annealing learning rate scheduling is used, with weight decay of 1×10. -2The batch size is set to 32. The training data comes from 12 consecutive months of historical operation data from the CAES demonstration power station, containing approximately 800,000 labeled samples covering 20 common fault modes. The training set, validation set, and test set are divided in a 7:1:2 ratio.
[0065] The total loss function of the hybrid backbone network module is: ,in For data fitting loss, For the residual loss of the physical equation, This is the regularization loss; and These are the physical constraint weights and the regularization weights, respectively. The total loss function updates the network parameters using gradients through backpropagation and the chain rule, allowing the model to gradually approximate the constraints of physical laws while fitting the data.
[0066] Furthermore, such as Figure 5 As shown, the knowledge reasoning module includes a two-layer knowledge graph: a hard knowledge graph and a soft knowledge graph. The hard knowledge graph constructs a physical knowledge ontology of the CAES power plant based on the W3C standard's OWL 2 DL outline, containing a large number of concept nodes and relation edges. The soft knowledge graph stores historical fault cases and operation and maintenance experience in a graph database, containing a large number of fault case nodes and operation and maintenance experience nodes, and achieves intelligent retrieval based on a graph-augmented retrieval framework.
[0067] Furthermore, the inference engine is used to execute a structured reasoning process, progressively decomposing complex fault symptoms and performing evidence retrieval and confidence calculation based on a two-layer knowledge graph. The five-step thought chain reasoning process is as follows: (1) Problem decomposition: The composite fault symptoms are broken down into sub-problem chains using a predefined fault decomposition template; (2) Evidence retrieval: Retrieve relevant entities and relationships in parallel from the two-layer knowledge graph; (3) Stepwise reasoning: The confidence of each candidate failure mode is calculated using the Bayesian confidence propagation algorithm; (4) Result verification: Substitute the physical quantities of each intermediate node on the reasoning path into the control equation to calculate the residual norm. If it exceeds the dynamic threshold, backtracking is triggered. (5) Final conclusion: Output a complete fault diagnosis conclusion by combining the reasoning results of each sub-problem.
[0068] It also includes an adversarial training component for enhancing the robustness of the system by generating adversarial examples. The adversarial training component includes an attack unit and a defense unit. The attack unit is used to generate adversarial examples that simulate sensor anomalies, including sensor drift, communication interruption, and environmental noise. The defense model is corrected by retrieving the three most similar historical cases from the soft knowledge graph using GraphRAG.
[0069] In the actual operation of CAES power plants, interference such as sensor drift, communication interruptions, and environmental noise are unavoidable, severely impacting the reliability of fault diagnosis systems. Existing fault diagnosis systems are typically trained and validated under ideal data conditions, lacking the ability to withstand real-world interference. Generalized large language models are more prone to outputting seemingly reasonable but actually erroneous diagnostic conclusions when faced with input noise.
[0070] This invention introduces an adversarial continuous error correction mechanism. The attack unit continuously generates adversarial samples of various sensor anomalies, simulating various interference scenarios that may be encountered at the power plant site. The defense unit detects and corrects the adversarial samples based on historical fault cases in a soft knowledge graph. Alternating training of attack and defense enables the system to maintain stable diagnostic performance even when facing unknown interference. This mechanism fundamentally improves the system's robustness and field adaptability—the system is not passively trained on a static dataset, but actively evolves through continuous adversarial training against attackers. Experiments show that after red-blue adversarial training, the diagnostic accuracy on a test set injected with 20% noise increased from 78.6% to 91.2%, fully demonstrating the effectiveness of this mechanism.
[0071] Furthermore, this invention also includes a verification module that maintains a physical equation checklist containing executable versions of over 40 CAES governing equations after formal transformation. When the inference engine outputs a conclusion, the verification module extracts all physical quantities and their values involved in the conclusion and substitutes them one by one into the relevant physical equations for forward calculation. The threshold values for each verification point are determined based on historical data statistics, with a typical value of 5%-10%. If the residual at any verification point exceeds the threshold, the current conclusion is rejected and a re-inference is triggered.
[0072] The system provided by this invention is deployed on an industrial internet platform of a compressed air energy storage (CAES) power station. The system's computing hardware consists of multiple high-performance GPU servers (configured with NVIDIA A100 or equivalent graphics cards) and a large-capacity distributed storage system. The system communicates in real-time with the power station's SCADA (Supervisory Control and Data Acquisition) system via industrial Ethernet. The data acquisition frequency is dynamically configured according to the signal type (1Hz for analog signals, event-triggered for digital signals). It synchronizes data with a BIM server through an IFC file exchange interface and interfaces with the power station's operation and maintenance management system via a RESTful API.
[0073] According to a second aspect of the present invention, such as Figure 6 As shown, a smart fault diagnosis method for an energy storage power station according to a first aspect embodiment is provided, comprising the following steps: Step S1: Construct a digital semantic model of the energy storage power station according to preset data standards; Step S2: Receive the sensing data of the energy storage power station, extract features from the sensing data, and perform cross-modal feature alignment using the device entities in the semantic backbone as anchor points to generate multimodal fusion features; Step S3: Perform fusion inference on the multimodal fusion features and apply physical law constraints to output deep features that conform to physical laws; Step S4: Perform structured reasoning on the deep features based on the knowledge graph and output the fault diagnosis conclusion.
[0074] Example like Figure 9 As shown, this embodiment uses the method of the present invention to test a simulation model of a 300MW CAES demonstration power station. The method of the present invention is compared with rule-based systems, traditional machine learning methods and general large language model fine-tuning methods. Figure 9 The following bar chart compares the performance of the three methods across three dimensions: diagnostic accuracy, physical compliance, and spatial positioning accuracy. Figure 9 Each dimension was analyzed individually. (1) Diagnostic accuracy dimension. For example... Figure 9 As shown, the fault diagnosis accuracy of the method of the present invention reaches 95.3%, which is higher than the 76.3% of the rule-based system and the 81.2% of the general large language model fine-tuning method, and is about 19.0 percentage points and 14.1 percentage points higher than the two respectively. The technical reason is that the rule system relies on the manually written "IF-THEN" rule base, which has insufficient rule coverage when facing complex faults with multiple factors coupled (such as compressor vibration exceeding the standard and accompanied by abnormal outlet temperature), thus limiting the diagnostic accuracy; although the general large language model fine-tuning method has strong language understanding ability, it lacks physical knowledge and spatial topology cognition in the CAES field, making it difficult to perform causal attribution for complex faults; the present invention provides spatial prior through BIM semantic anchoring, suppresses physical illusions through PINN constraints, and realizes a step-by-step unfolding and traceable reasoning process through knowledge-enhanced thinking chain reasoning, thereby achieving the highest diagnostic accuracy in complex fault scenarios. (2) Physical compliance rate dimension. Figure 9As shown, the physical compliance rate of the method of this invention reaches 99.1%, which is basically the same as that of the rule system (100%), while the fine-tuning method of the general large language model is only 65.8%, and the invention improves by about 33.3 percentage points. This result verifies the effectiveness of embedding more than 40 CAES control equations into the PINN-Transformer hybrid backbone network: the model output is verified by physical residuals, which fundamentally suppresses the problem of the large language model output violating physical laws such as energy conservation and momentum conservation, and makes the diagnostic conclusion strictly satisfy physical laws such as thermodynamics and gas dynamics; at the same time, compared with the rule system, this invention does not sacrifice generalization ability for physical compliance, but achieves a significant improvement in diagnostic accuracy while maintaining a high level of physical compliance, overcoming the inherent contradiction of the rule system that is "compliant but has limited diagnostic ability". (3) Spatial positioning accuracy dimension. Figure 9 As shown, the rule system makes logical judgments based on preset threshold rules, lacks spatial perception capabilities, and cannot output the spatial location of the fault (this indicator is not applicable); the spatial positioning accuracy of the general large language model fine-tuning method is only 42.1%, which is because the general large model lacks an understanding of the three-dimensional spatial structure and equipment topology relationship of the CAES power station, and can only output the fault type but cannot give the fault location; the present invention maps the diagnostic conclusion to the three-dimensional spatial coordinates of the power station through the spatial anchor and topological attention mask of the BIM semantic backbone, and the spatial positioning accuracy reaches 93.7%, which can directly output the specific spatial location of the faulty equipment (such as "the third stage impeller of the compressor No. 2 in the compression area"), providing direct guidance for maintenance personnel to quickly locate the faulty equipment. (4) Real-time response dimension. The average response time of the method of the present invention in the verification is 2.8 seconds, which meets the engineering requirements of real-time monitoring and rapid response of the power station, is better than the rule system investigation process that requires manual intervention, and also avoids the uncontrollable delay of the reasoning process of the general large model.
[0075] In summary, as Figure 9 As shown, this invention leads the way in three dimensions: diagnostic accuracy, physical compliance, and spatial positioning accuracy. It maintains a level of physical compliance comparable to rule-based systems while significantly improving diagnostic accuracy and supplementing the spatial positioning capabilities lacking in rule-based systems. Compared to general large language model fine-tuning methods, this invention comprehensively surpasses existing methods in all three dimensions. These four indicators comprehensively verify the effectiveness of this invention from four dimensions: diagnostic accuracy, physical reliability, spatial positioning capability, and real-time response efficiency. The comprehensive experimental data fully demonstrates that this invention has significant overall performance advantages over existing technical solutions.
[0076] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0077] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0078] Through the above description of the embodiments, those skilled in the art can clearly understand that the above implementation methods can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0079] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A smart fault diagnosis system for energy storage power stations based on semantic anchoring and physical information reasoning, characterized in that, include: The semantic backbone module is used to construct a digital semantic model of the energy storage power station based on preset data standards. A multimodal coding module, connected to the semantic backbone module, is used to receive sensing data from the energy storage power station, extract features from the sensing data, and perform cross-modal feature alignment using the device entities in the semantic backbone module as anchor points to generate multimodal fusion features. The hybrid reasoning module, connected to the multimodal encoding module, is used to perform fusion reasoning on the multimodal fusion features and apply physical law constraints to output deep features that conform to physical laws. The knowledge reasoning module, connected to the hybrid reasoning module, is used to perform structured reasoning on the deep features and output fault diagnosis conclusions.
2. The system according to claim 1, characterized in that, The digital semantic model of the energy storage power station is constructed according to a preset data standard, including: Based on the IFC4.3 standard, the special entity type of the energy storage power station is extended, and the equipment, sensors and pipeline connections of the power station are constructed into a topology map with spatial coordinates; Assign each entity a spatial anchor containing three-dimensional geometric coordinates, partition identifiers, and system identifiers; Four semantic layers are constructed, comprising a process system layer, an equipment layer, a sensor layer, and a space layer, with each layer interconnected through semantic relationships; The topology graph, the spatial anchor, and the four semantic layers together constitute the digital semantic model.
3. The system according to claim 2, characterized in that, The topology graph includes a topology attention mask, which is used to determine the attention weights between different device nodes. Among them, directly connected device nodes receive the highest attention weight, device nodes with the shortest path hop count not exceeding a preset threshold receive attention weight related to the hop count, and device nodes with the shortest path hop count exceeding a preset threshold have their attention weight set to a minimum value.
4. The system according to claim 1, characterized in that, The multimodal coding module includes multiple dedicated encoders for processing text data, image data, time-series data, acoustic data, and point cloud data, respectively. The multimodal coding module uses device entity identifiers as anchors and, through comparative learning, brings different modal features belonging to the same device entity closer together in the feature space, while pushing features belonging to different device entities further apart.
5. The system according to claim 4, characterized in that, The image encoder generates a region of interest mask based on the semantic backbone module to guide the attention mechanism; the temporal encoder maps temporal data to the corresponding device entity based on the installation relationship between the sensor and the device; the acoustic encoder uses the spatial anchor to assist in sound source localization; and the output of the point cloud encoder is registered with the coordinate system of the semantic backbone module.
6. The system according to claim 1, characterized in that, The hybrid inference module includes alternately stacked data-driven blocks and physical constraint blocks; The data-driven block is used to learn the hidden patterns in the data, and the physical constraint block is used to introduce the control equations of the energy storage power station as soft constraints into the network training; the weights of the soft constraints gradually increase from the initial value to the final value according to the training process.
7. The system according to claim 6, characterized in that, The total loss of the hybrid inference module is determined based on the data fitting loss and the physical constraint loss. The physical constraint loss is calculated based on the residuals of the control equations. The weights of the soft constraints are used to dynamically adjust the contribution of the physical constraint loss to the total loss.
8. The system according to claim 6, characterized in that, The knowledge reasoning module includes: A two-layer knowledge graph, comprising a hard knowledge graph and a soft knowledge graph, wherein the hard knowledge graph is used to store formalized physical equation knowledge, and the soft knowledge graph is used to store historical fault cases and operation and maintenance experience; The inference engine is used to execute the structured inference process, which decomposes the complex fault symptoms step by step and performs evidence retrieval and confidence calculation based on the two-layer knowledge graph. An adversarial training component is used to enhance the robustness of the system by generating adversarial examples. The adversarial training component includes an attack unit and a defense unit. The attack unit is used to generate adversarial examples that simulate sensor anomalies, and the defense unit is used to correct the adversarial examples based on historical failure cases in the soft knowledge graph.
9. The system according to claim 8, characterized in that, It also includes a verification module connected to the knowledge reasoning module. The verification module is used to verify the reasoning result according to the soft constraints. If the verification fails, it triggers a re-reasoning.
10. A method for intelligent fault diagnosis of an energy storage power station based on the system described in any one of claims 1-9, characterized in that, Includes the following steps: Step S1: Construct a digital semantic model of the energy storage power station according to preset data standards; Step S2: Receive the sensing data of the energy storage power station, extract features from the sensing data, and perform cross-modal feature alignment using the device entities in the semantic backbone as anchor points to generate multimodal fusion features; Step S3: Perform fusion inference on the multimodal fusion features and apply physical law constraints to output deep features that conform to physical laws; Step S4: Perform structured reasoning on the deep features based on the knowledge graph and output the fault diagnosis conclusion.
