Explanatable fault diagnosis method, system and device for fan blade and medium

Abstract

The invention provides an interpretable fault diagnosis method, system and device for a fan blade and a medium, and relates to the technical field of wind power plants. According to the method, multi-modal sensing data and priori knowledge of material science are integrated, and a pre-trained multi-modal big language model is utilized to perform data fusion and causal reasoning, so that accurate diagnosis of fault types of the fan blades, accurate positioning of damage positions and effective tracking of damage development paths are realized. The method not only improves the accuracy and robustness of fault diagnosis, but also provides an explanatory result which is easy to understand, thereby providing powerful support for engineering technicians, and enabling the fault diagnosis process to be more transparent and reliable. The method effectively overcomes the problems that a traditional single-mode diagnosis method is weak in generalization ability and lacks semantic interpretation.

Description

Technical Field

[0001] This invention relates to the field of wind farm technology, and in particular to a method, system, device and medium for diagnosing interpretable faults in wind turbine blades. Background Technology

[0002] Wind turbine blades have complex structures, operate under diverse conditions, and have long service lives, making them prone to fatigue cracks, interlaminar debonding, and material aging. Traditional diagnostic models based on single modes, such as vibration or acoustic emission signals, suffer from poor generalization and robustness, making it difficult to cover multiple operating conditions and damage types. Furthermore, they lack sufficient semantic interpretation capabilities, failing to meet the needs of practical engineering applications. In recent years, although multimodal large language models have demonstrated powerful cross-modal semantic understanding and reasoning capabilities in areas such as image-text question answering and semantic retrieval, applying them to wind turbine blade fault diagnosis still faces challenges, including how to effectively integrate materials science knowledge, sensor data, and construct interpretable diagnostic systems. Summary of the Invention

[0003] The purpose of this invention is to provide a method, system, device, and medium for diagnosing interpretable faults in wind turbine blades, in order to solve one or more technical problems existing in the prior art, or at least provide a beneficial option or create conditions.

[0004] The solution to the technical problem of this invention is as follows: On the one hand, this invention provides a method for interpretable fault diagnosis of wind turbine blades, comprising the following steps: Acquire multimodal sensing data and prior knowledge of materials science for wind turbine blades; The multimodal sensing data includes sensor topology, vibration signals, strain data, acoustic emission waveforms, thermal images, operating parameters, and manual inspection text records; the prior knowledge of materials science includes prior parameters of material damage for the wind turbine blades, a material damage knowledge graph, a materials science semantic prompting mechanism, and a causal path diagram of material mechanisms. The multimodal sensing data is preprocessed to construct a multimodal dataset; Based on the aforementioned multimodal dataset and prior knowledge of materials science, a pre-trained multimodal large language model is used for multimodal fusion and causal reasoning, outputting interpretable diagnostic results including fault type, damage location, and damage development path, specifically including: By combining the aforementioned prior parameters of material damage, the multimodal dataset is transformed into a multimodal representation; Based on the material science semantic prompting mechanism and the material mechanism causal path graph, cross-modal matching and fusion are performed on the multimodal representation to obtain diagnostic features for cross-modal alignment; Based on the material mechanism causal path diagram, the material damage knowledge graph, and the diagnostic features, fault type identification and diagnostic reasoning are performed to generate the interpretable diagnostic results.

[0005] Furthermore, the prior parameters of material damage are obtained in the following manner: Based on the design drawings and process parameters of the wind turbine blades, a model is created, and the stress distribution of the wind turbine blades under fatigue load is simulated through finite element simulation. The damage threshold of key parts is extracted, and the material damage prior parameters are integrated to obtain the results.

[0006] Furthermore, the materials science semantic prompting mechanism is implemented in the following ways: Collect the entire life cycle data of the wind turbine blades, including composite material manuals, wind farm fault case libraries, and sensor signal samples, and construct the material damage knowledge graph; wherein, the material damage knowledge graph includes a variety of typical damage types of the wind turbine blades; A structured prompt template is generated based on the material damage knowledge graph; wherein, the structured prompt template includes a triplet logical chain of "damage type - material property change - sensor signal feature"; The structured cue template is adjusted by low-rank adaptation technique and injected into the underlying semantic space of the multimodal large language model, thereby injecting material terms and damage mechanism descriptions into the multimodal representation.

[0007] Furthermore, the causal path diagram of the material mechanism is constructed using a Bayesian network approach, including the following steps: Determine the node set of the Bayesian network; in the node set, the input nodes are wind speed, rotational speed and ambient temperature, the intermediate nodes are fiber volume fraction, interfacial strength and vibration amplitude, and the output nodes are delamination area, crack length and remaining lifetime; Based on the theory of composite material mechanics and previous experimental data, the conditional probability table between nodes in the node set is calculated by the maximum likelihood estimation method. Based on the conditional probability table, the network structure of the Bayesian network is optimized using the Markov chain Monte Carlo method, and the confidence of each causal path in the Bayesian network is calculated. The causal paths in the Bayesian network with confidence levels greater than a preset confidence threshold are retained to generate a causal path graph of material mechanism.

[0008] Furthermore, the first module of the multimodal large language model includes a graph structure encoder, a temporal signal encoder, an image encoder, a text encoder, and a multimodal fusion processor; The graph structure encoder is used to extract sensor topology features based on the sensor topology. The timing signal encoder is used to extract timing signal features based on the vibration signal, the strain data, and the acoustic emission waveform; The image encoder is used to extract image features based on the thermal image; The text encoder is used to extract text features based on the operating parameters and the manual inspection text records; The multimodal fusion processor is used to concatenate the sensor topological features, the time-series signal features, the image features, and the text features in the feature dimension to form a combined feature. Through a stacked self-attention mechanism, the material damage prior parameters are introduced to perform cross-modal joint modeling of the combined feature and output the multimodal representation.

[0009] Furthermore, the second module of the multimodal large language model includes a materials science semantic cue layer, a cross-attention and causal path layer, and a matching optimization layer; The materials semantic hint layer is used to inject materials terminology and damage mechanism descriptions into the multimodal representation through a materials semantic hint mechanism; The cross-attention and causal path layer is used to match and fuse the multimodal representation through the cross-attention mechanism and the material mechanism causal path graph to obtain the cross-modal aligned diagnostic features; The matching optimization layer is used to perform comparative learning and modal adversarial training optimization on the diagnostic features.

[0010] Furthermore, the third module of the multimodal large language model includes a shared perception layer, a fault category identification layer, a diagnostic reasoning layer, and a natural language interpretation layer; The shared perception layer is used to extract the hierarchical representation of the diagnostic features through a deep residual network, and combined with the attention distillation mechanism, to compress redundant information and output the shared perception feature vector. The fault category identification layer is used to identify fault types based on the shared perceptual feature vector, using a classification network that integrates bidirectional gated recurrent units and an attention mechanism, and outputs the fault category. The diagnostic reasoning layer is used to combine the material mechanism causal path diagram, learn the correlation between the sensor topology and fault propagation through graph neural network, perform damage location localization and damage development path reasoning, and output the damage location and damage development path. The natural language interpretation layer is used to generate the interpretable diagnostic results based on the fault type, the damage location, and the damage development path, combined with the material damage knowledge graph.

[0011] On the other hand, this application provides an interpretable fault diagnosis system for wind turbine blades, including a data acquisition module, a data preprocessing module, and an interpretable diagnosis module; The data acquisition module is used to acquire multimodal sensing data and prior knowledge of materials science for wind turbine blades; The multimodal sensing data includes sensor topology, vibration signals, strain data, acoustic emission waveforms, thermal images, operating parameters, and manual inspection text records; the prior knowledge of materials science includes prior parameters of material damage for the wind turbine blades, a material damage knowledge graph, a materials science semantic prompting mechanism, and a causal path diagram of material mechanisms. The data preprocessing module is used to preprocess the multimodal sensing data to construct a multimodal dataset; The interpretable diagnostic module is used to perform multimodal fusion and causal reasoning based on the multimodal dataset and the prior knowledge of materials science, using a pre-trained multimodal large language model, and output interpretable diagnostic results including fault type, damage location, and damage development path, specifically including: By combining the aforementioned prior parameters of material damage, the multimodal dataset is transformed into a multimodal representation; Based on the material science semantic prompting mechanism and the material mechanism causal path graph, cross-modal matching and fusion are performed on the multimodal representation to obtain diagnostic features for cross-modal alignment; Based on the material mechanism causal path diagram, the material damage knowledge graph, and the diagnostic features, fault type identification and diagnostic reasoning are performed to generate the interpretable diagnostic results.

[0012] On the other hand, this application provides a wind turbine blade interpretable fault diagnosis device, including a vibration sensor, an acoustic emission sensor, a strain gauge array, a thermal imager, an edge computing unit, a communication unit, and a cloud computing unit; The vibration sensor is used to collect vibration signals from the wind turbine blades; The acoustic emission sensor is used to collect the acoustic emission waveform of the wind turbine blades; The strain gauge array is used to collect strain data of the wind turbine blades; The thermal imager is used to acquire thermal images of the wind turbine blades; The edge computing unit is used to preprocess the sensor topology, vibration signal, strain data, acoustic emission waveform, thermal image, operating parameters and manual inspection text records to construct a multimodal dataset. The communication unit is used to realize the communication connection between the edge computing unit and the cloud computing unit; The cloud computing unit includes a processor and a memory; the memory is used to store programs; when the program is executed by the processor, the processor implements the aforementioned wind turbine blade interpretable fault diagnosis method based on the multimodal dataset.

[0013] On the other hand, this application provides a computer storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the aforementioned wind turbine blade interpretable fault diagnosis method.

[0014] The beneficial effects of this invention are as follows: This application provides an interpretable fault diagnosis method for wind turbine blades. By integrating multimodal sensor data and prior knowledge of materials science, and utilizing a pre-trained multimodal large language model for data fusion and causal reasoning, it achieves accurate diagnosis of wind turbine blade fault types, accurate location of damage, and effective tracking of damage development paths. This method not only improves the accuracy and robustness of fault diagnosis but also provides easily understandable interpretable results, thus providing strong support for engineering technicians and making the fault diagnosis process more transparent and reliable. This method effectively overcomes the problems of weak generalization ability and lack of semantic interpretation in traditional single-modal diagnostic methods. This application also provides corresponding systems, devices, and media. The beneficial effects of the systems, devices, and media are similar to those of the method and will not be elaborated here.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0017] Figure 1 This is a flowchart of the wind turbine blade interpretable fault diagnosis method provided in this application; Figure 2 This is a structural diagram of the multimodal large language model provided in this application; Figure 3 This is a structural diagram of the wind turbine blade interpretable fault diagnosis system provided in this application; Figure 4 This is a structural diagram of the wind turbine blade fault diagnosis device provided in this application; Detailed Implementation To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] This application relates to the technical field of wind turbine blade fault diagnosis in wind power generation equipment. As a key load-bearing component of wind turbines, wind turbine blades have complex structures, operate under variable conditions, and have long service lives, making them susceptible to various problems such as fatigue cracks, interlaminar debonding, and material aging. To ensure reliable wind turbine operation, effective health monitoring and fault diagnosis of wind turbine blades are crucial.

[0022] Traditional wind turbine blade fault diagnosis primarily relies on single-mode data analysis methods, such as vibration analysis and acoustic emission (AE) detection. These methods capture specific types of signals using sensors and identify potential damage based on changes in these signals. However, with the development of the wind power industry, wind turbine designs are becoming increasingly complex, and operating environments are becoming more diverse, gradually revealing the limitations of traditional single-mode diagnostic methods.

[0023] Traditional methods mostly utilize only one type of sensing data (such as vibration or acoustic emission), which proves inadequate when dealing with complex damage modes. Because wind turbine blades exhibit diverse damage forms, different types of damage may show anomalies in different physical quantities, making it difficult for a single mode to comprehensively cover all possible damage types.

[0024] Existing diagnostic models often only provide a simple "normal" or "abnormal" judgment, without offering detailed explanations such as the specific type of damage, its exact location, and how it developed. This lack of interpretability hinders engineers from making further maintenance decisions.

[0025] Moreover, traditional diagnostic systems based on fixed rules or single models perform poorly when faced with variable working conditions or novel types of damage. This is because they do not fully utilize the complementarity between multiple information sources, limiting the accuracy and reliability of diagnostic results. Furthermore, wind turbine blades are typically made of composite materials, whose internal damage mechanisms are highly complex. Existing diagnostic technologies rarely incorporate materials science knowledge, such as damage evolution laws and fracture mechanics theories, thus limiting the depth and accuracy of the diagnostic system.

[0026] To address the aforementioned issues, this application proposes a method, system, device, and medium for interpretable fault diagnosis of wind turbine blades. This method not only comprehensively processes information from multiple sources such as vibration, strain, acoustic emission, and thermal imaging, but also incorporates rich prior knowledge in materials science, including material damage parameters, knowledge graphs, semantic hint mechanisms, and causal path graphs. This approach not only improves the accuracy and robustness of fault diagnosis but also generates highly interpretable diagnostic reports, helping engineers better understand the nature of the fault and formulate corresponding maintenance strategies. Furthermore, this invention emphasizes the importance of cross-modal feature fusion, aiming to enhance the model's adaptability and transferability across different application scenarios.

[0027] First, the wind turbine blade interpretable fault diagnosis method provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0028] Reference Figure 1 The implementation process of the wind turbine blade interpretable fault diagnosis method provided in this application embodiment includes, but is not limited to, the following steps.

[0029] Step S110: Obtain multimodal sensing data and prior knowledge of materials science for the wind turbine blades.

[0030] The multimodal sensing data includes sensor topology, vibration signals, strain data, acoustic emission waveforms, thermograms, operating parameters, and manual inspection records. Prior materials knowledge includes prior parameters for material damage in wind turbine blades, a material damage knowledge graph, materials semantic hint mechanisms, and causal path diagrams of material mechanisms.

[0031] In step S110, comprehensive information on the health status of the wind turbine blades is collected, including data from various sensors (such as sensor topology, vibration signals, strain data, acoustic emission waveforms, and thermograms), as well as operating parameters and manual inspection records. Simultaneously, prior material science knowledge related to the wind turbine blades is incorporated, such as prior parameters of material damage, material damage knowledge graphs, material semantic hint mechanisms, and causal path diagrams of material mechanisms. This information collectively forms the basis for assessing the health status of the wind turbine blades. By integrating different types of sensor data and rich materials science knowledge, the internal and external state changes of the blades can be more accurately reflected, providing comprehensive and in-depth data support for subsequent fault diagnosis.

[0032] Step S120: Preprocess the multimodal sensing data to construct a multimodal dataset.

[0033] In step S120, the collected raw multimodal sensing data undergoes a series of preprocessing operations to ensure its suitability for further analysis and modeling. This includes noise removal, data format standardization, and aligning timestamps from different sources. The goal of preprocessing is to improve data quality, making it more consistent and reliable, thus laying the foundation for building a high-quality multimodal dataset. Furthermore, this process includes converting various types of data into a unified representation to facilitate subsequent multimodal fusion. A well-designed preprocessing workflow can effectively reduce data inconsistencies, enhance the effectiveness of model learning, and ultimately improve the accuracy of fault diagnosis.

[0034] Step S130: Based on the multimodal dataset and prior knowledge of materials science, multimodal fusion and causal reasoning are performed using a pre-trained multimodal large language model to output interpretable diagnostic results including fault type, damage location and damage development path.

[0035] In step S130, a pre-trained multimodal large language model is used in conjunction with a previously prepared multimodal dataset and prior material science knowledge to perform deep data fusion and causal reasoning. First, multimodal data is transformed into a unified representation by incorporating prior material damage parameters. Then, these representations are fused across modalities using a material science semantic prompting mechanism and a material mechanism causal path graph to obtain aligned diagnostic features. Next, fault type identification and diagnostic reasoning are performed based on the material damage knowledge graph and diagnostic features, generating interpretable diagnostic results including fault type, damage location, and damage development path. This method not only improves the accuracy of fault diagnosis but also provides detailed explanations, enabling engineers to better understand and address problems in wind turbine blades. In this way, precise monitoring and intelligent diagnosis of wind turbine blade health are achieved.

[0036] In some embodiments of this application, in step S130, based on multimodal datasets and prior knowledge of materials science, a pre-trained multimodal large language model is used to perform multimodal fusion and causal reasoning, outputting interpretable diagnostic results including fault type, damage location, and damage development path, specifically including: Step S210: Combine the prior parameters of material damage to transform the multimodal dataset into a multimodal representation.

[0037] In step S210, the preprocessed multimodal sensing data (such as vibration, strain, acoustic emission, thermograms, etc.) is combined with the physical and mechanical properties of the wind turbine blade material. Prior material damage parameters (such as fatigue limit, fracture toughness, modulus degradation law, interlaminar shear strength threshold, etc.) are used to semantically enhance and physically map the original data, thereby transforming the original signal into a multimodal representation with material mechanism implications. This representation is not merely a mathematical abstraction of the data, but rather imbues each type of sensor data with a potential correlation to the material damage mechanism.

[0038] For example, by combining strain data with a material fatigue damage evolution model, a mapping relationship can be established between strain amplitude and cumulative damage; the energy of acoustic emission waveforms can be correlated with the microcrack propagation rate of the material. This transformation process based on prior knowledge enables the model to rise from perceived data to a state of understanding, providing a physically interpretable feature foundation for subsequent cross-modal fusion and causal inference, and avoiding the limitations of traditional black-box models that rely solely on statistical correlation.

[0039] Step S220: Based on the material science semantic prompting mechanism and the material mechanism causal path graph, cross-modal matching and fusion are performed on the multimodal representation to obtain cross-modal aligned diagnostic features.

[0040] In step S220, the system utilizes a materials science semantic cue mechanism to guide the multimodal large language model to focus on semantic information related to specific material damage mechanisms in different modal data. For example, when the system suspects matrix cracking, the semantic cue activates joint attention to low-frequency vibration modes, strain concentration in specific regions, and low-energy events in acoustic emission signals. Simultaneously, by using a material mechanism causal path diagram (i.e., describing the physical evolution path from initial micro-damage to macroscopic failure, such as "fatigue loading → matrix micro-cracks → crack propagation → delamination → structural instability"), the system can establish causal logical relationships between different modal signals.

[0041] For example, a localized temperature rise in a thermogram may correspond to a region of strain concentration, and an increase in the frequency of acoustic emission events in that region can be interpreted as a precursor to crack propagation. Through this guided fusion based on causal paths, the system not only achieves spatial and temporal alignment of different modes but also semantic alignment at the "damage mechanism" level, thereby generating highly consistent and interpretable cross-modal aligned diagnostic features. This fusion approach effectively improves the model's robustness under complex operating conditions and the logical consistency of diagnosis.

[0042] Step S230: Based on the material mechanism causal path diagram, material damage knowledge graph and diagnostic features, perform fault type identification and diagnostic reasoning to generate interpretable diagnostic results.

[0043] In step S230, the system inputs the cross-modal aligned diagnostic features obtained in the previous step into a pre-trained multimodal large language model, and combines it with a material damage knowledge graph (containing structured knowledge such as material type, damage mode, failure mode, and environmental impact factors) and a material mechanism causal path graph to perform in-depth causal reasoning and pattern recognition. The model not only determines the most likely fault type (such as "tail-edge adhesive layer debonding" or "fiber fracture caused by leading-edge corrosion"), but also traces the starting point of the damage, infers its development path (such as "local delamination caused by lightning strike damage, gradually expanding to the main beam area"), and predicts its future evolution trend.

[0044] More importantly, the multimodal large language model can generate diagnostic reports in natural language, clearly explaining the reasoning process. For example, based on the leading-edge temperature rise shown in the thermal image, the aggregation of high-frequency events in the acoustic emission signal, and the abnormal fluctuations in strain data at the 0.7R position, combined with the mechanism that GFRP materials are prone to interface debonding in humid and hot environments, it is determined that the leading-edge adhesive layer is debonding due to moisture, and is currently in the early stage of expansion. It is recommended to perform glue injection repair during the next shutdown. This generative and interpretable output greatly improves the credibility and engineering practicality of the diagnostic results, realizing an intelligent diagnostic paradigm that integrates "data-driven" and "knowledge-driven + data-driven" approaches.

[0045] In some embodiments of this application, the prior parameters of material damage are obtained in the following ways: modeling is performed based on the design drawings and process parameters of the wind turbine blades, the stress distribution of the wind turbine blades under fatigue load is simulated by finite element simulation, the damage threshold of key parts is extracted, and the prior parameters of material damage are integrated.

[0046] By integrating design information from wind turbine blades with physical simulation methods, a priori knowledge base with engineering reliability and physical significance is constructed. Specifically, a high-fidelity three-dimensional finite element model is first established based on the original design drawings of the wind turbine blades (including geometric configuration, layup sequence, material distribution, etc.) and manufacturing process parameters (such as curing temperature, pressure, fiber volume fraction, etc.). On this basis, the complex fatigue load spectrum that the wind turbine blades may experience throughout their entire life cycle (including normal operating loads, extreme wind conditions, gusts, yaw errors, etc.) is simulated, and structural mechanics simulation analysis is performed to accurately calculate the stress, strain, and vibration response distribution of the blades under different operating conditions. Through long-term fatigue simulation or damage evolution simulation based on Miner's linear cumulative damage theory, key sensitive areas such as stress concentration regions and high-cycle fatigue zones are identified, and critical damage indices for these locations are extracted, such as the equivalent von Mises stress threshold, strain energy density threshold, acoustic emission energy accumulation threshold, and abnormal heat dissipation level. Finally, these quantitative damage criteria obtained from simulation are systematically integrated to form prior material damage parameters closely related to the actual blade structure and material properties.

[0047] The significance of this process lies in its transformation of abstract material properties (such as fatigue life and fracture toughness) into specific numerical parameters that can be directly used for data analysis and model judgment. This provides a solid foundation for mapping the raw sensor data into a physically meaningful multimodal representation, ensuring that the entire diagnostic system not only relies on data-driven approaches but also has a solid engineering physics basis, thereby effectively improving the accuracy, interpretability, and generalization ability of the diagnostic results.

[0048] In some embodiments of this application, the materials science semantic prompting mechanism is implemented in the following ways: Step S310: Collect full life cycle data of wind turbine blades, including composite material manuals, wind farm fault case libraries and sensor signal samples, and construct a material damage knowledge graph.

[0049] The material damage knowledge graph includes various typical damage types of wind turbine blades.

[0050] In step S310, by systematically collecting various data throughout the entire lifecycle of wind turbine blades—from design and manufacturing to operation and decommissioning—a comprehensive knowledge system reflecting the material damage modes of wind turbine blades is established—namely, a material damage knowledge graph. This includes, but is not limited to, basic material properties (such as elastic modulus and ultimate tensile strength) provided in composite material handbooks, actual failure cases occurring in wind farms (such as specific phenomena like delamination and crack propagation), and the original signal characteristics recorded by sensors under different operating conditions. By integrating this information, a detailed knowledge graph covering multiple typical damage types can be formed, encompassing not only damage mechanism descriptions from materials science theory but also their manifestations in the actual operating environment. The importance of this process lies in providing rich background knowledge support for subsequent diagnosis, enabling the model to make judgments from a broader perspective during analysis and improving its ability to understand unknown or complex damage modes.

[0051] Step S320: Generate a structured prompt template based on the material damage knowledge graph.

[0052] The structured prompt template includes a triplet logic chain of "damage type - material property change - sensor signal characteristics".

[0053] In step S320, based on the constructed material damage knowledge graph, a series of structured prompt templates are further extracted. These templates exist in the form of a triplet logical chain of "damage type - material property change - sensor signal characteristics". For example, the damage type "fiber fracture" may correspond to a significant decrease in the tensile strength of the material and an enhancement of the high-frequency acoustic emission signal. Such structured prompt templates are not only a simplified expression of the material damage mechanism, but also a key step in transforming abstract physical concepts into a form that machines can understand and process. They act as bridges, connecting specific material damage phenomena with their representation in sensor data, enabling the multimodal large language model to automatically associate relevant materials science background knowledge while understanding the input data, thereby enhancing the accuracy and specificity of the model in identifying specific types of damage.

[0054] Step S330: The structured cue template is adjusted using low-rank adaptation technology and injected into the underlying semantic space of the multimodal large language model to inject material terms and damage mechanism descriptions into the multimodal representation.

[0055] In step S330, the meticulously designed structured prompt template is effectively integrated into the multimodal large language model. Low-rank adaptation (LoRA) technology is used to fine-tune the model parameters, ensuring that the structured prompts are accurately embedded into the model's underlying semantic space. This approach not only preserves the original model's strong generalization ability but also specifically enhances its understanding of terminology and mechanisms related to wind turbine blade material damage. When multimodal data is encoded and fed into the model, these prompts, rich in materials science knowledge, guide the model to focus more on feature representations related to specific damage types, thereby improving the interpretability of the final output. The advantage of this method is that it allows the model to effectively integrate domain-specific knowledge while maintaining high efficiency, resulting in diagnostic results that are not only highly accurate but also highly readable and interpretable, facilitating understanding and application by engineering technicians.

[0056] In some embodiments of this application, the material mechanism causal path graph is constructed using a Bayesian network approach, including the following steps: Step S410: Determine the node set of the Bayesian network.

[0057] The nodes are concentrated, with input nodes being wind speed, rotational speed, and ambient temperature; intermediate nodes being fiber volume fraction, interfacial strength, and vibration amplitude; and output nodes being delamination area, crack length, and remaining life.

[0058] In step S410, a structured causal model reflecting the damage evolution process of wind turbine blade materials is established by clarifying the variables (i.e., nodes) and their hierarchical relationships contained in the Bayesian network. Input nodes (such as wind speed, rotational speed, and ambient temperature) represent external operational and environmental factors affecting the blade's operating state; these are the initial driving forces inducing material damage. Intermediate nodes (such as fiber volume fraction, interfacial strength, and vibration amplitude) represent the material's microstructural characteristics and dynamic response; they are key transitional variables connecting external loads and the final damage state, reflecting the physical degradation mechanism of the material during service. Output nodes (such as delamination area, crack length, and remaining life) represent the final manifestation of damage development, directly reflecting the blade's health status and failure risk. By scientifically classifying these three types of nodes, a logical chain from "external excitation → material response → damage evolution" is constructed, laying a clear topological foundation for subsequent quantification of the causal relationships between various factors.

[0059] Step S420: Based on the mechanics theory of composite materials and previous experimental data, calculate the conditional probability table between nodes in the node set using the maximum likelihood estimation method.

[0060] In step S420, the system combines theoretical analysis with previous experimental data to quantitatively model the dependencies between nodes in the Bayesian network. Specifically, it uses composite material mechanics theories (such as laminate theory, fracture mechanics, and fatigue damage models) to guide the formal assumptions of causal relationships between variables, and combines previous experimental data (such as interface debonding initiation time, crack propagation rate, modulus degradation curves, etc. recorded under different stress levels) to fit and calculate the conditional probability distribution (i.e., conditional probability table, CPT) of each node under the conditions of its parent node.

[0061] For example, it can estimate the probability that a specific combination of wind speed and rotational speed will lead to a decrease in interface strength of more than 10%, or the likelihood of detectable delamination defects occurring within a certain vibration amplitude range. This process transforms abstract physical mechanisms into computable probabilistic expressions, enabling Bayesian networks to not only reflect qualitative causal relationships of "what causes what," but also provide quantitative assessments of "how likely it is to happen," thus enhancing the model's reasoning ability in uncertain environments.

[0062] Step S430: Based on the conditional probability table, optimize the network structure of the Bayesian network using the Markov chain Monte Carlo method, and calculate the confidence of each causal path in the Bayesian network.

[0063] In step S430, after initially setting the network topology and filling the conditional probability table, the initial structure may contain redundant or insignificant connections, requiring further structure learning and optimization through a data-driven approach. The Markov Chain Monte Carlo (MCMC) method is employed to sample in the parameter space and iteratively search for the optimal network structure configuration, maximizing the overall model's fit to the observed data (i.e., posterior probability). This process not only removes weakly correlated or meaningless edges but also discovers potential hidden causal paths. More importantly, the MCMC method can calculate the frequency or posterior probability (i.e., confidence level) of each possible causal path (e.g., "high wind speed → high vibration amplitude → interface strength degradation → hierarchical expansion") with data support, thereby ranking the importance and reliability of different damage evolution paths and providing a scientific basis for subsequent screening.

[0064] Step S440: Retain causal paths in the Bayesian network with confidence levels greater than a preset confidence threshold, and generate a causal path diagram of material mechanism.

[0065] In step S440, the preceding modeling results are refined and output to generate a concise, reliable, and engineering-interpretive causal path diagram of material mechanisms. By setting a reasonable confidence threshold (e.g., 90% or 95%), only those causal paths that occur frequently in MCMC sampling and are well supported by data are retained, while low-confidence or accidental connections are eliminated, thereby ensuring that the final causal diagram has high statistical robustness and physical interpretability.

[0066] The causal path diagram of material mechanisms not only intuitively displays the complete evolution chain from external working conditions to material degradation and then to macroscopic damage, but also quantitatively labels the confidence level of each path, becoming an important knowledge guidance tool for multimodal large language models in fault diagnosis reasoning. In practical applications, this causal path diagram can be used to explain the causal chain of diagnostic results (such as "this stratification expansion is mainly caused by fatigue accumulation due to continuous high wind speeds"), and can also be used for predictive maintenance decision support, effectively improving the transparency, credibility, and engineering practicality of the entire fault diagnosis system.

[0067] In some embodiments of this application, reference is made to Figure 2 The multimodal large language model comprises three modules. The first module transforms the multimodal dataset into a multimodal representation by incorporating prior parameters of material damage. The second module performs cross-modal matching and fusion of the multimodal representation based on materials science semantic prompting mechanisms and material mechanism causal path graphs to obtain cross-modal aligned diagnostic features. The third module identifies fault types and performs diagnostic reasoning based on the material mechanism causal path graph, material damage knowledge graph, and diagnostic features, generating interpretable diagnostic results.

[0068] In some embodiments of this application, reference is made to Figure 2 The first module of the multimodal large language model comprises a graph structure encoder, a temporal signal encoder, an image encoder, a text encoder, and a multimodal fusion processor. The primary task of this first module is to transform data collected from multiple sensors (i.e., the multimodal dataset) into a unified and easily processed multimodal representation by incorporating prior material damage parameters. This process begins by extracting features from heterogeneous data such as sensor topology, vibration signals, strain data, acoustic emission waveforms, thermograms, operating parameters, and manual inspection text records using different encoders, generating their respective low-dimensional representations. These features are then concatenated in the multimodal fusion processor and further integrated through a stacked self-attention mechanism, while introducing prior material damage parameters to enhance semantic consistency. This is done to ensure that the final multimodal representation not only contains all the information from the original data but also reflects prior knowledge related to specific material damage, providing a solid foundation for subsequent cross-modal matching and diagnostic inference.

[0069] In some embodiments of this application, a graph structure encoder is used to extract sensor topological features based on the sensor topology.

[0070] The role of a graph encoder is to abstract the sensor network deployed on a wind turbine blade into a graph structure, where nodes represent the locations of individual sensors, and edges represent the physical connections or spatial proximity relationships between sensors. This graph modeling effectively captures the topological dependencies and spatial distribution characteristics between sensors. For example, when localized damage occurs on the blade, the surrounding sensors exhibit specific response patterns. A graph encoder can utilize techniques such as graph neural networks (GNNs) to extract these spatial correlation features, known as "sensor topological features." These features not only reflect the geometric information of the sensor layout but also implicitly contain the propagation paths and mutual influence mechanisms of the structural response, providing spatial context support for subsequent fusion of other modal data and enhancing the model's ability to model the sensitivity of damage locations.

[0071] In some embodiments of this application, a timing signal encoder is used to extract timing signal features based on vibration signals, strain data, and acoustic emission waveforms.

[0072] Timing signal encoders are specifically designed to process dynamic signals acquired from devices such as vibration sensors, strain gauges, and acoustic emission (AE) probes, including vibration signals, strain data, and acoustic emission waveforms. These signals are essentially high-frequency, continuously varying time series, containing rich information about the structural dynamic response. Timing signal encoders typically employ timing modeling architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or Transformers to extract representative timing signal features, such as spectral components, impact energy, strain accumulation rate, and signal entropy. These features reflect changes in the mechanical state of the structure during operation, and are particularly important for identifying dynamic damage processes such as fatigue crack initiation and debonding propagation, making them one of the most direct data sources for fault diagnosis.

[0073] In some embodiments of this application, an image encoder is used to extract image features based on a thermal image.

[0074] Image encoders are responsible for processing thermal image data acquired by infrared thermal imagers. Thermal images reflect the temperature distribution on the surface of wind turbine blades, and temperature anomalies are often precursors to internal friction, localized overload, or material failure. Image encoders typically use deep convolutional networks (such as ResNet and EfficientNet) to extract features layer by layer from thermal images, capturing multi-level "image features" from low-level edge textures to high-level semantic regions (such as hot spots and cold spots). These features can not only be used to identify potential slippage or delamination regions caused by frictional heating, but also, when combined with structural models, to invert the distribution of internal heat sources, thereby helping to determine the location and severity of internal defects in composite materials and providing important non-contact sensing information for multimodal fusion.

[0075] In some embodiments of this application, a text encoder is used to extract text features based on operating parameters and manual inspection text records.

[0076] The text encoder is used to parse and understand operating parameters (such as structured data like wind speed, RPM, power, and yaw angle, typically recorded in text form) and unstructured text records generated during manual inspections (such as "obvious scratches on the leading edge of the blade" or "oil seepage at the root of the B-phase blade"). This encoder is usually based on a pre-trained language model (such as BERT, RoBERTa, or the text branch of a multimodal large language model) to transform this textual information into semantically rich vector representations, i.e., "text features." This process not only extracts keywords and entity information but also understands contextual semantics, such as distinguishing the severity difference between "slight noise" and "severe abnormal noise." The introduction of text features allows the system to integrate the experience and judgment of human experts with operational background information, compensating for the semantic deficiencies of pure sensor data and improving the comprehensive judgment capability of the diagnostic system.

[0077] In some embodiments of this application, the multimodal fusion processor is used to concatenate sensor topological features, time-series signal features, image features and text features in the feature dimension to form a combined feature, and through a stacked self-attention mechanism, introduce material damage prior parameters to perform cross-modal joint modeling of the combined feature and output a multimodal representation.

[0078] The multimodal fusion processor is the core integration unit of the first module. Its function is to perform unified modeling and deep fusion of sensor topological features, temporal signal features, image features, and text features extracted by the graph encoder, temporal signal encoder, image encoder, and text encoder, respectively. First, it concatenates these four types of heterogeneous features along the feature dimension to form a high-dimensional "combined feature" vector, initially achieving the integration of multi-source information. However, simple concatenation is insufficient to eliminate semantic gaps and redundant information between modalities. To address this, the processor further introduces a stacked self-attention mechanism. By calculating the attention weights between each feature element, it dynamically adjusts the importance of features from different modalities and locations, achieving cross-modal information interaction and alignment. Crucially, the processor also injects "material damage prior parameters" as guiding signals into the self-attention mechanism. For example, by adjusting the attention distribution, it makes the model pay more attention to feature responses related to known fatigue thresholds or debonding sensitive frequency bands. This fusion approach not only enables joint representation learning of multimodal data, but also endows the feature representations with clear material physics meanings, ultimately outputting a unified "multimodal representation" that integrates spatial structure, dynamic response, thermal field distribution, operational context, and material mechanism, laying a solid foundation for cross-modal matching and causal reasoning in subsequent modules.

[0079] In some embodiments of this application, reference is made to Figure 2 The second module of the multimodal large language model includes a material science semantic cue layer, a cross-attention and causal path layer, and a matching optimization layer.

[0080] The second module of the multimodal large language model builds upon the multimodal representations generated by the first module. It aims to leverage materials science semantic cues and causal path graphs of materials mechanisms to perform deep cross-modal matching and fusion of these representations. Specifically, this process first injects professional materials terminology and damage mechanism descriptions into the multimodal representations through a materials science semantic cues layer, enhancing the model's understanding of complex materials phenomena. Then, using cross-attention and causal path layers, based on predefined causal path graphs of materials mechanisms, it performs deep cross-modal interactions to achieve information alignment and complementarity between different modalities, resulting in more accurate and consistent diagnostic features. Furthermore, the matching optimization layer further improves the quality of the diagnostic features through contrastive learning and modal adversarial training, ensuring good discriminative and generalization capabilities.

[0081] In some embodiments of this application, the materials semantic cue layer is used to inject materials terminology and damage mechanism descriptions into the multimodal representation through a materials semantic cue mechanism.

[0082] The core function of the materials science semantic cue layer is to inject specialized materials science knowledge into the deep semantic space of the multimodal large language model in the form of "semantic cues," thereby guiding the model to possess domain-prior cognitive abilities when processing and understanding multimodal data. This layer, based on a pre-constructed material damage knowledge graph and structured cue templates (such as the "damage type-material property change-sensor signal feature" triple), uses efficient parameter fine-tuning techniques such as LoRA (Low-Rank Adaptation) to embed these structured material terms and damage mechanism descriptions (e.g., "matrix cracking leads to increased acoustic emission energy," "interface debonding accompanied by local temperature rise") into the model's underlying representation. When the input passes through the multimodal representation generated by the first module, the semantic cue layer activates the material mechanism description most relevant to the current data features, enabling the model to correlate with physical behavior under specific damage modes. This mechanism effectively enhances the model's physical interpretation ability of sensor signals, enabling it to move beyond relying solely on statistical correlation and instead perform directional feature focus and reasoning guided by material failure mechanisms, thus improving the accuracy and interpretability of diagnosis.

[0083] In some embodiments of this application, cross-attention and causal path layers are used to match and fuse multimodal representations through cross-attention mechanisms and material mechanism causal path graphs to obtain cross-modal aligned diagnostic features.

[0084] This layer is a key module for achieving cross-modal depth alignment and causal fusion. Its function is to perform fine-grained matching and fusion of representations from different modes by utilizing cross-attention mechanisms and material mechanism causal path graphs. Specifically, the cross-attention mechanism allows a feature of one mode (such as vibration signal) to be used as a query to retrieve relevant responses, i.e., keys and values, in other modes (such as thermograms and strain data), thereby establishing dynamic correlations between modes.

[0085] For example, when vibration energy in a certain frequency band is abnormal, the model can automatically monitor whether hot spots appear in the thermal image of the corresponding region at that moment, or whether strain sensors record abrupt changes. More importantly, this process is constrained and guided by the causal path graph of the material mechanism—the causal path graph defines the reasonable evolutionary logic between different physical variables (such as "high stress → interface fatigue → debonding → frictional heating → temperature rise"). This layer uses the causal path as prior knowledge and regularizes or gates the weight distribution of cross-attention to ensure that the fusion process follows the physical laws of material damage and avoids generating false associations that violate the mechanism. The final output of cross-modal aligned diagnostic features not only achieves deep fusion of multimodal information mathematically, but also maintains causal consistency in a physical sense, providing a high-fidelity and traceable feature foundation for subsequent accurate diagnosis.

[0086] In some embodiments of this application, the matching optimization layer is used to perform contrastive learning and modality adversarial training optimization on diagnostic features.

[0087] The matching optimization layer further enhances the quality and robustness of the diagnostic features obtained through the initial fusion of cross-attention and causal path layers. This is primarily achieved through two strategies: contrastive learning and modality adversarial training. In contrastive learning, the system constructs positive sample pairs (such as responses to the same injury event in different modalities) and negative sample pairs (such as features of normal and abnormal states). By optimizing the loss function (such as InfoNCE), the distance between positive sample features is narrowed, and the distance between negative sample features is widened, thereby enhancing the discriminative power of the diagnostic features and enabling them to form clear intra-class clusters and inter-class separations in high-dimensional space. Modality adversarial training aims to eliminate representational biases between different modalities caused by sensor differences, noise levels, or data distribution shifts.

[0088] By introducing a learnable modality discriminator, the diagnostic features are forced to be indistinguishable in terms of modality identity, thereby extracting common features that truly reflect the essence of the damage, rather than modality-specific noise or artifacts. The synergistic effect of these two optimization strategies results in diagnostic features that not only possess stronger classification performance but also greater cross-device and cross-operating-condition transferability and anti-interference capabilities, providing crucial support for building a stable and reliable wind turbine blade fault diagnosis system.

[0089] In some embodiments of this application, reference is made to Figure 2 The third module of the multimodal large language model includes a shared perception layer, a fault category identification layer, a diagnostic reasoning layer, and a natural language interpretation layer.

[0090] The third module is responsible for performing specific fault type identification and diagnostic reasoning tasks based on the material mechanism causal path diagram, material damage knowledge graph, and refined diagnostic features provided by the preceding two modules. This module first uses a shared perception layer to extract hierarchical representations from the diagnostic features and compresses redundant information to form a compact and comprehensive shared perception feature vector. Next, the fault category identification layer employs advanced machine learning algorithms (such as a classification network combining bidirectional gated recurrent units and attention mechanisms) to determine the specific fault type based on the aforementioned feature vector. Simultaneously, the diagnostic reasoning layer combines graph neural network technology to explore the correlation between sensor topology and fault propagation, accurately locating the damage site and inferring the damage development path. Finally, the natural language interpretation layer generates a detailed and easy-to-understand diagnostic report based on all the obtained information, enabling engineers to clearly understand the root cause of the fault and its potential impact. In this way, the entire system can not only accurately identify potential problems with wind turbine blades but also provide scientifically sound maintenance recommendations.

[0091] In some embodiments of this application, the shared perception layer is used to extract hierarchical representations of diagnostic features through a deep residual network, and combined with an attention distillation mechanism to compress redundant information and output a shared perception feature vector.

[0092] The shared perception layer serves to deeply abstract and refine the diagnostic features obtained after cross-modal fusion, aiming to extract a unified representation that can represent the common essence of multimodal data—namely, the shared perception feature vector. This layer uses a deep residual network as its backbone architecture, extracting hierarchical representations from the diagnostic features layer by layer by stacking multiple convolutional or fully connected residual blocks: the lower layers capture local details and original response patterns, the middle layers integrate cross-modal collaborative responses, and the higher layers form highly abstract semantic representations of damage states.

[0093] To further enhance the compactness and interpretability of the features, this layer introduces an attention distillation mechanism. This mechanism uses attention weights to weighted filter the feature map, actively suppressing redundant information unrelated to the fault (such as environmental noise and fluctuations in normal operating conditions), and distilling and concentrating key damage-related information into the core dimensions. This process not only reduces the feature dimensionality and improves computational efficiency, but more importantly, it enhances the model's ability to focus on key damage signals. This results in a shared perceptual feature vector with stronger discriminative power and generalization ability, providing high-quality input for subsequent classification and inference tasks.

[0094] In some embodiments of this application, the fault category identification layer is used to identify fault types based on shared perceptual feature vectors, using a classification network that integrates bidirectional gated recurrent units and attention mechanisms, and outputs the fault category.

[0095] The core function of the fault category identification layer is to accurately determine the specific fault type occurring on the wind turbine blades based on shared sensing feature vectors, such as leading-edge corrosion, trailing-edge adhesive layer debonding, main beam cracks, and lightning damage. This layer employs a classification network structure that integrates a bidirectional gated recurrent unit (Bi-GRU) and an attention mechanism. The Bi-GRU effectively models the time dependence of fault evolution, capturing the dynamic trends in damage development from both forward and backward directions, making it particularly suitable for handling progressive damage reflected by continuous monitoring data. Building upon this, an attention mechanism is introduced to automatically identify the most discriminative feature segments or time windows in fault determination (such as strain abrupt changes after a severe gust of wind or sustained vibration enhancement in a specific frequency band), assigning them higher weights. This combination not only improves classification accuracy but also reveals key evidence leading to fault judgment, making the classification results more traceable. Finally, this layer outputs a clear fault category label, serving as one of the core judgment results of the diagnostic system.

[0096] In some embodiments of this application, the diagnostic inference layer is used to combine the material mechanism causal path diagram, learn the correlation between sensor topology and fault propagation through graph neural network, perform damage location localization and damage development path inference, and output the damage location and damage development path.

[0097] The diagnostic inference layer is responsible for inferring the spatial location and development path of damage based on known fault categories, achieving deep reasoning. This layer combines material mechanism causal path graphs and graph neural network (GNN) techniques to model the sensor topology as a graph structure. It then uses GNN for message passing and aggregation on the graph to learn the functional relationships and response propagation patterns between sensor nodes. By mapping shared sensing features onto this graph structure, the model can identify which sensor regions exhibit the most significant anomalous responses, thereby accurately locating the physical location of the damage, such as the "tail edge at 0.7R" and the "root adhesion zone."

[0098] Furthermore, by combining the damage evolution logic defined in the causal path diagram (such as surface erosion → resin aging → fiber exposure → fracture), the model can trace back to the starting point of damage and predict its possible future development direction and speed, generating a damage development path that conforms to the material physics mechanism. This reasoning process not only improves the spatial accuracy of diagnosis but also provides a scientific basis for predictive maintenance.

[0099] In some embodiments of this application, the natural language interpretation layer is used to generate interpretable diagnostic results based on the fault type, damage location, and damage development path, combined with a material damage knowledge graph.

[0100] The natural language interpretation layer transforms the structured diagnostic results (including fault type, damage location, and damage development path) generated by the preceding modules into human-readable, logically clear, and professionally in-depth natural language reports, achieving complete interpretability of the diagnostic process. Based on a pre-trained multimodal large language model and combined with domain knowledge from the material damage knowledge graph (such as material properties, failure modes, and maintenance recommendations), this layer automatically organizes language to generate reports such as: "In the 0.7R region of the B-phase blade, signs of adhesive layer debonding were detected, manifested by a temperature rise of approximately 3.2°C and a 45% increase in acoustic emission event density in the thermal image. Based on the interface strength degradation mechanism of GFRP material under humid and hot conditions, this is determined to be early delamination caused by moisture penetration. The current expansion rate is low, and it is recommended to perform sealing repair and strengthen drainage checks during the next downtime window." These reports not only state the facts but also clarify causal logic, cite material mechanisms, and propose maintenance recommendations, effectively improving the credibility and engineering applicability of the diagnostic results.

[0101] In some embodiments of this application, the training process of a multimodal large language model includes the following steps.

[0102] (1) Different encoders are initialized differently for different modal data.

[0103] For text encoders (such as MC-BERT): a pre-trained model in the Chinese medical / industrial semantic domain is adopted, and fine-tuned using wind turbine blade fault diagnosis and maintenance records and expert corpus to enhance the semantic understanding of material terms and crack patterns.

[0104] For image encoders (such as ViT), based on ImageNet or MAE pre-trained models, low-level parameters are frozen, and only high-level structures are fine-tuned to retain general texture extraction capabilities and adapt to local features such as microcracks and heat distribution.

[0105] For time-series signal encoders (such as 1D-ResNet+TCN), due to the strong correlation between vibration signals and tasks, training is conducted from scratch using supervised training on a wind turbine real-world test dataset.

[0106] For graph-structured encoders (such as GAT), random initialization is performed, and no separate pre-training is performed. In the subsequent joint training phase, they are optimized together with other modules. The input is the sensor space topology graph to capture the coupling relationship of structural regions.

[0107] (2) Project each modal output onto the same embedding dimension, satisfying the following calculation formula (1): (1); In formula (1), For modality The encoder output, It is a linear projection function that ensures that text, images, time-series signals, and graph structure features are aligned in the same semantic space.

[0108] (3) A cross-modal attention mechanism is adopted to achieve heterogeneous information fusion. The text feature is used as the query to guide other modal features (key / value) to participate in the fusion, satisfying the following calculation formula (2): (2); In formula (2), Representing text features, Representing image features, Represents the characteristics of time-series signals. Indicates sensor topological characteristics; For text feature query matrix, For the bond matrix of other modalities, For the value matrix of other modes, For a unified feature dimension.

[0109] (4) Design independent output headers and loss functions for different tasks.

[0110] For the fault category identification task, cross-entropy loss is used. It satisfies the following calculation formula (3): (3); In formula (3), Given the number of fault categories (e.g., cracks, delamination, fatigue, etc.), the system iterates through all categories to obtain the true label for each category. With model predicted probability The logarithmic multiplication is accumulated and summed to finally obtain the loss value for the fault category identification task.

[0111] For damage localization tasks, mean square error loss is used. It satisfies the following calculation formula (4): (4); In formula (4), For the number of damage samples, For the first The true geometric coordinates of a damaged sample For the first Predicted coordinates of each damaged sample.

[0112] For the task of ensuring consistency in damage development paths, KL divergence loss is used. To align it with the expert path distribution, satisfying the following calculation formula (5): (5); In formula (5), For the damage development path predicted by the model, The path distribution defined for expert knowledge.

[0113] To align the embedding representations of different modalities in the semantic space, a modality consistency task was designed, employing contrastive loss. (NT-Xent or InfoNCE), satisfying the following calculation formula (6): Formula (6); In formula (6), This represents the anchor modal embedding, which is the modal feature vector that serves as a reference; for example, if the text modality is dominant, then... It can be the embedding vector output by a text encoder; This refers to positive sample modality embeddings, which are different modality features belonging to the same input sample as the anchor modality. For example, the image modality embedding vector corresponding to the text modality is used as... Both describe the fault state of the same wind turbine blade and need to be semantically related to it. "Closer"; This represents negative sample mode embeddings, which are mode features belonging to different input samples than the anchor mode. For example, the time-series signal mode embedding vectors or irrelevant mode features of other wind turbine blades can be used as... It needs to be in the semantic space with "Push away"; This represents the function for calculating cosine similarity. The temperature parameter is used to align different modal embedding representations.

[0114] For the task of generating interpretable results, the token cross-entropy loss of an autoregressive language model is adopted. It satisfies the following calculation formula (7): (7); In formula (7), The length of the text sequence. For the first One token, This indicates that the autoregressive language model is based on the preceding sequence. Predict the next token as The conditional probability.

[0115] (5) The total loss function is constructed by weighted summation of the multi-task losses.

[0116] Total loss function The following calculation formula (8) must be satisfied: (8); In formula (8), Loss of the fault category identification task The weighting coefficients, Loss for damage localization task The weighting coefficients, Loss of the task of generating interpretability results The weighting coefficients, Loss for modal consistency tasks The weighting coefficients, Loss of tasks related to the consistency of damage development path The weighting coefficients.

[0117] (6) Perform end-to-end fine-tuning.

[0118] After the multi-task model initially converges, some encoder parameters (such as the first few layers of ViT and the first few layers of MC-BERT) are unfrozen, and the entire model is fine-tuned with a small learning rate to improve modal co-operation performance and task generalization ability.

[0119] Secondly, refer to Figure 3 This application provides an interpretable fault diagnosis system for wind turbine blades, including a data acquisition module, a data preprocessing module, and an interpretable diagnosis module.

[0120] The data acquisition module is used to acquire multimodal sensing data and prior knowledge of materials science for wind turbine blades.

[0121] The multimodal sensing data includes sensor topology, vibration signals, strain data, acoustic emission waveforms, thermograms, operating parameters, and manual inspection records. Prior materials knowledge includes prior parameters for material damage in wind turbine blades, a material damage knowledge graph, materials semantic hint mechanisms, and causal path diagrams of material mechanisms.

[0122] The data preprocessing module is used to preprocess multimodal sensor data and construct a multimodal dataset.

[0123] The interpretable diagnostic module is used to perform multimodal fusion and causal reasoning based on multimodal datasets and prior knowledge of materials science, utilizing a pre-trained multimodal large language model. It outputs interpretable diagnostic results including fault type, damage location, and damage development path, specifically including: By incorporating prior parameters of material damage, the multimodal dataset is transformed into a multimodal representation.

[0124] Based on the semantic prompting mechanism of materials science and the causal path graph of materials mechanism, cross-modal matching and fusion are performed on multimodal representations to obtain diagnostic features for cross-modal alignment.

[0125] Based on the material mechanism causal path diagram, material damage knowledge graph, and diagnostic features, fault type identification and diagnostic reasoning are performed to generate interpretable diagnostic results.

[0126] Furthermore, refer to Figure 4 This application provides a wind turbine blade interpretable fault diagnosis device, including a vibration sensor, an acoustic emission sensor, a strain gauge array, a thermal imager, an edge computing unit, a communication unit, and a cloud computing unit.

[0127] Vibration sensors are used to collect vibration signals from wind turbine blades. These sensors acquire vibration signals in real time under excitation or operating conditions, including dynamic response information such as acceleration, velocity, and displacement. These signals directly reflect the overall and local structural dynamics of the blade, containing rich information on changes in modal parameters (such as natural frequency, damping ratio, and mode shape). When the blade experiences faults such as cracks, mass imbalance, or loose connections, its vibration characteristics will change significantly, for example, with energy enhancement at specific frequencies, modal coupling, or resonance point shift. The time-domain and frequency-domain characteristics obtained through high-precision vibration sensors provide core data support for subsequent identification of faults such as structural stiffness degradation and fatigue damage, making it one of the most fundamental and critical sensing units in a multimodal diagnostic system.

[0128] Acoustic emission sensors are used to acquire acoustic emission waveforms from wind turbine blades. These sensors capture transient elastic wave signals, or acoustic emission waveforms, generated during the stress process of wind turbine blades due to microscopic damage within the material (such as fiber breakage, matrix cracking, and interface debonding). These signals possess high temporal resolution and local sensitivity, allowing detection instantaneously upon damage occurrence. By analyzing parameters such as energy, rise time, duration, and frequency distribution of acoustic emission events, damage types and their activity levels can be effectively identified, and source localization can be achieved in conjunction with sensor arrays. This sensor is particularly suitable for monitoring highly concealed and rapidly developing internal defects in composite material blades, serving as an important supplement to macroscopic response measurements such as vibration and strain, and significantly enhancing the ability to detect early micro-damage.

[0129] Strain gauge arrays are used to acquire strain data from wind turbine blades. Composed of multiple distributed strain sensors deployed in critical stress areas of the blades (such as the root, main beam, and trailing edge), the array continuously measures local strain data under operating or excitation loads. Strain is one of the most direct physical quantities reflecting the stress state of a structure; its abnormal concentration often indicates stress concentration areas, fatigue hotspots, or a decrease in structural load-bearing capacity. Through array-based arrangement, not only can single-point strain values ​​be obtained, but the strain field distribution can also be reconstructed through spatial interpolation, identifying regions of abrupt strain gradient changes and thus inferring potential damage locations. Combined with the stress-strain relationship of the material, strain data can also be used to assess the residual strength and cumulative fatigue damage of the structure, providing crucial mechanical boundary condition information for multimodal fusion.

[0130] Thermal imagers are used to acquire thermal images of wind turbine blades. They are used for non-contact acquisition of infrared thermal images of the wind turbine blade surface, reflecting its temperature field distribution. During active excitation (such as vibration) or long-term operation, internal friction, plastic deformation, viscoelastic dissipation, or localized electrical faults generate heat, leading to abnormal temperature rises in the damaged area. Thermal imagers can quickly and over large areas capture these thermal anomalies, making them particularly suitable for detecting fault types accompanied by energy dissipation, such as debonding, stratification, and lightning-induced ablation. By analyzing hotspot locations, heating rates, and thermal diffusion patterns in the thermal images, the damaged area can be visualized and located. This sensor provides physical dimension information complementary to the mechanical response, enhancing the spatial perception capability of the diagnostic system and the detection rate of hidden defects.

[0131] The edge computing unit is used to preprocess sensor topology, vibration signals, strain data, acoustic emission waveforms, thermal images, operating parameters, and manual inspection text records to construct a multimodal dataset.

[0132] The edge computing unit is deployed at the wind turbine site and is responsible for real-time preprocessing and initial integration of raw data from exciters, various sensors, and external systems (such as SCADA). Its main functions include: filtering, denoising, detrending, and normalizing time-series signals such as vibration, strain, and acoustic emission; enhancing and registering thermal images; structurally encoding operating parameters (wind speed, rotational speed, etc.) with manual inspection records; and constructing sensor topology information based on the physical layout of the sensors. Finally, the edge computing unit uniformly formats and times-aligns this heterogeneous data to build a high-quality multimodal dataset. This unit performs data preprocessing locally, reducing the data burden of long-distance transmission and enabling low-latency preliminary analysis and anomaly warning, serving as a hub connecting the physical sensing layer and the cloud-based intelligent diagnostic layer.

[0133] The communication unit is used to establish communication connections between the edge computing unit and the cloud computing unit. It is responsible for establishing a stable and secure data transmission channel between the two units. It supports wired (e.g., fiber optic) or wireless (e.g., 5G, LoRa, Wi-Fi) communication protocols, ensuring efficient and reliable interaction of multimodal datasets, operational information, and diagnostic commands between the field and the cloud. This unit must be resistant to electromagnetic interference, adaptable to complex weather environments, and support data encryption and authentication to guarantee the security and integrity of system communication. As an information bridge in the system architecture, the communication unit enables the timely uploading of massive amounts of locally collected data to the cloud for in-depth analysis, while simultaneously allowing cloud-based diagnostic results and control strategies to be fed back to the field, achieving a closed-loop intelligent diagnostic system that integrates the edge and cloud.

[0134] The cloud computing unit includes a processor and memory. The memory is used to store programs. When the program is executed by the processor, it enables the processor to implement the aforementioned interpretable fault diagnosis method for wind turbine blades based on the multimodal dataset.

[0135] The cloud computing unit is the intelligent core of the entire diagnostic system, deployed in a data center or cloud platform, and consists of a high-performance processor and a large-capacity memory. The memory stores the complete program code implementing the method described in this application, including key knowledge assets such as a multimodal large language model, a material damage knowledge graph, a material mechanism causal path graph, and pre-trained model parameters. When the program is executed by the processor, the cloud unit can receive multimodal datasets from the edge and invoke complex deep learning and causal inference algorithms to perform the entire computational process from multimodal representation construction, cross-modal fusion, fault identification to the generation of interpretable diagnostic reports. Thanks to the powerful computing resources of the cloud, the system can run highly complex models (such as Transformer, GNN, and Bayesian networks) for large-scale data training and knowledge updates, continuously optimizing diagnostic performance. Simultaneously, the cloud also supports centralized management and horizontal comparative analysis of data from multiple wind turbines, achieving fleet-level health status assessment and knowledge sharing, serving as the decision-making center for interpretable, high-precision, and self-evolving intelligent diagnosis.

[0136] In some embodiments of this application, the wind turbine blade interpretable fault diagnosis device further includes an exciter for applying an excitation signal to the wind turbine blade. An electromagnetic or piezoelectric exciter is installed at the blade root or mid-section to excite the blade structure with sweep frequency, step, or white noise signals, covering a frequency range of 0–10 kHz, to stimulate multi-mode responses and nonlinear behavior of the local structure.

[0137] A vibrator is a key device for actively exciting the structural response of wind turbine blades. Its function is to apply a controllable excitation signal to the blades. This excitation signal varies continuously within a preset frequency range, systematically scanning the dynamic characteristics of the blades and exciting their vibration modes at different frequencies. Through active excitation, the nonlinear response or local stiffness changes caused by minor structural damage (such as early cracks or debonding) can be enhanced, making them more apparent in sensor signals, thereby improving the sensitivity and signal-to-noise ratio of fault detection. Compared to traditional monitoring methods that rely solely on passive excitations such as wind loads, using a vibrator enables controllable and repeatable testing conditions, making it particularly suitable for shutdown inspections or periodic health assessments. It provides a physical basis for obtaining high-quality, high-signal-to-noise-ratio multimodal response data, a prerequisite for accurate diagnosis.

[0138] Furthermore, embodiments of this application provide a computer storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the aforementioned wind turbine blade interpretable fault diagnosis method.

[0139] In summary, the wind turbine blades provided in this application embodiment can explain the fault diagnosis method, system, device and medium, and have the following technical effects.

[0140] This application's embodiments achieve precise monitoring and intelligent diagnosis of the health status of wind turbine blades by integrating multi-source sensor data and prior knowledge of materials science. These embodiments deeply integrate multi-modal data such as vibration, strain, acoustic emission, and thermography, and combine them with professional knowledge such as material damage parameters, knowledge graphs, and causal path diagrams. Utilizing a multi-modal large language model for feature extraction and causal reasoning not only improves the accuracy and robustness of the diagnosis but also generates interpretable diagnostic results with physical causal logic. Furthermore, the edge-cloud collaborative architecture ensures a balance between real-time processing and in-depth analysis, while the closed-loop knowledge-driven system endows the system with the ability to continuously learn and evolve, enabling it to cope with complex operating conditions and novel damage modes.

[0141] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0142] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0143] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0144] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0145] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or, if necessary, processing in a suitable manner, and then stored in computer memory.

[0146] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0147] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0148] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0149] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for diagnosing interpretable faults in wind turbine blades, characterized in that, Includes the following steps: Acquire multimodal sensing data and prior knowledge of materials science for wind turbine blades; The multimodal sensing data includes sensor topology, vibration signals, strain data, acoustic emission waveforms, thermal images, operating parameters, and manual inspection text records; the prior knowledge of materials science includes prior parameters of material damage for the wind turbine blades, a material damage knowledge graph, a materials science semantic prompting mechanism, and a causal path diagram of material mechanisms. The multimodal sensing data is preprocessed to construct a multimodal dataset; Based on the aforementioned multimodal dataset and prior knowledge of materials science, a pre-trained multimodal large language model is used for multimodal fusion and causal reasoning, outputting interpretable diagnostic results including fault type, damage location, and damage development path, specifically including: By combining the aforementioned prior parameters of material damage, the multimodal dataset is transformed into a multimodal representation; Based on the material science semantic prompting mechanism and the material mechanism causal path graph, cross-modal matching and fusion are performed on the multimodal representation to obtain diagnostic features for cross-modal alignment; Based on the material mechanism causal path diagram, the material damage knowledge graph, and the diagnostic features, fault type identification and diagnostic reasoning are performed to generate the interpretable diagnostic results.

2. The wind turbine blade interpretable fault diagnosis method according to claim 1, characterized in that, The prior parameters for material damage are obtained through the following methods: Based on the design drawings and process parameters of the wind turbine blades, a model is created, and the stress distribution of the wind turbine blades under fatigue load is simulated through finite element simulation. The damage threshold of key parts is extracted, and the material damage prior parameters are integrated to obtain the results.

3. The wind turbine blade interpretable fault diagnosis method according to claim 1, characterized in that, The materials science semantic prompting mechanism is implemented in the following ways: Collect the entire life cycle data of the wind turbine blades, including composite material manuals, wind farm fault case libraries, and sensor signal samples, and construct the material damage knowledge graph; wherein, the material damage knowledge graph includes a variety of typical damage types of the wind turbine blades; A structured prompt template is generated based on the material damage knowledge graph; wherein, the structured prompt template includes a triplet logical chain of "damage type - material property change - sensor signal feature"; The structured cue template is adjusted by low-rank adaptation technique and injected into the underlying semantic space of the multimodal large language model, thereby injecting material terms and damage mechanism descriptions into the multimodal representation.

4. The wind turbine blade interpretable fault diagnosis method according to claim 1, characterized in that, The material mechanism causal path diagram is constructed using a Bayesian network approach, including the following steps: Determine the node set of the Bayesian network; in the node set, the input nodes are wind speed, rotational speed and ambient temperature, the intermediate nodes are fiber volume fraction, interfacial strength and vibration amplitude, and the output nodes are delamination area, crack length and remaining lifetime; Based on the theory of composite material mechanics and previous experimental data, the conditional probability table between nodes in the node set is calculated by the maximum likelihood estimation method. Based on the conditional probability table, the network structure of the Bayesian network is optimized using the Markov chain Monte Carlo method, and the confidence of each causal path in the Bayesian network is calculated. The causal paths in the Bayesian network with confidence levels greater than a preset confidence threshold are retained to generate a causal path graph of material mechanism.

5. The wind turbine blade interpretable fault diagnosis method according to claim 1, characterized in that, The first module of the multimodal large language model includes a graph structure encoder, a temporal signal encoder, an image encoder, a text encoder, and a multimodal fusion processor; The graph structure encoder is used to extract sensor topology features based on the sensor topology. The timing signal encoder is used to extract timing signal features based on the vibration signal, the strain data, and the acoustic emission waveform; The image encoder is used to extract image features based on the thermal image; The text encoder is used to extract text features based on the operating parameters and the manual inspection text records; The multimodal fusion processor is used to concatenate the sensor topological features, the time-series signal features, the image features, and the text features in the feature dimension to form a combined feature. Through a stacked self-attention mechanism, the material damage prior parameters are introduced to perform cross-modal joint modeling of the combined feature and output the multimodal representation.

6. The wind turbine blade interpretable fault diagnosis method according to claim 1, characterized in that, The second module of the multimodal large language model includes a materials science semantic prompting layer, a cross-attention and causal path layer, and a matching optimization layer; The materials semantic hint layer is used to inject materials terminology and damage mechanism descriptions into the multimodal representation through a materials semantic hint mechanism; The cross-attention and causal path layer is used to match and fuse the multimodal representation through the cross-attention mechanism and the material mechanism causal path graph to obtain the cross-modal aligned diagnostic features; The matching optimization layer is used to perform comparative learning and modal adversarial training optimization on the diagnostic features.

7. The wind turbine blade interpretable fault diagnosis method according to claim 1, characterized in that, The third module of the multimodal large language model includes a shared perception layer, a fault category identification layer, a diagnostic reasoning layer, and a natural language interpretation layer. The shared perception layer is used to extract the hierarchical representation of the diagnostic features through a deep residual network, and combined with the attention distillation mechanism, to compress redundant information and output the shared perception feature vector. The fault category identification layer is used to identify fault types based on the shared perceptual feature vector, using a classification network that integrates bidirectional gated recurrent units and an attention mechanism, and outputs the fault category. The diagnostic reasoning layer is used to combine the material mechanism causal path diagram, learn the correlation between the sensor topology and fault propagation through graph neural network, perform damage location localization and damage development path reasoning, and output the damage location and damage development path. The natural language interpretation layer is used to generate the interpretable diagnostic results based on the fault type, the damage location, and the damage development path, combined with the material damage knowledge graph.

8. A wind turbine blade fault diagnosis system, characterized in that, It includes a data acquisition module, a data preprocessing module, and an interpretable diagnostic module; The data acquisition module is used to acquire multimodal sensing data and prior knowledge of materials science for wind turbine blades; The multimodal sensing data includes sensor topology, vibration signals, strain data, acoustic emission waveforms, thermal images, operating parameters, and manual inspection text records; the prior knowledge of materials science includes prior parameters of material damage for the wind turbine blades, a material damage knowledge graph, a materials science semantic prompting mechanism, and a causal path diagram of material mechanisms. The data preprocessing module is used to preprocess the multimodal sensing data to construct a multimodal dataset; The interpretable diagnostic module is used to perform multimodal fusion and causal reasoning based on the multimodal dataset and the prior knowledge of materials science, using a pre-trained multimodal large language model, and output interpretable diagnostic results including fault type, damage location, and damage development path, specifically including: By combining the aforementioned prior parameters of material damage, the multimodal dataset is transformed into a multimodal representation; Based on the material science semantic prompting mechanism and the material mechanism causal path graph, cross-modal matching and fusion are performed on the multimodal representation to obtain diagnostic features for cross-modal alignment; Based on the material mechanism causal path diagram, the material damage knowledge graph, and the diagnostic features, fault type identification and diagnostic reasoning are performed to generate the interpretable diagnostic results.

9. A wind turbine blade fault diagnosis device, characterized in that, It includes vibration sensors, acoustic emission sensors, strain gauge arrays, thermal imagers, edge computing units, communication units, and cloud computing units; The vibration sensor is used to collect vibration signals from the wind turbine blades; The acoustic emission sensor is used to collect the acoustic emission waveform of the wind turbine blades; The strain gauge array is used to collect strain data of the wind turbine blades; The thermal imager is used to acquire thermal images of the wind turbine blades; The edge computing unit is used to preprocess the sensor topology, vibration signal, strain data, acoustic emission waveform, thermal image, operating parameters and manual inspection text records to construct a multimodal dataset. The communication unit is used to realize the communication connection between the edge computing unit and the cloud computing unit; The cloud computing unit includes a processor and a memory; the memory is used to store a program; when the program is executed by the processor, the processor implements the wind turbine blade interpretable fault diagnosis method as described in any one of claims 1 to 7 based on the multimodal dataset.

10. A computer storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to implement the wind turbine blade interpretable fault diagnosis method as described in any one of claims 1 to 7.

Images