A Rotary Equipment Fault Diagnosis System Based on a Large Language Model

Abstract

This invention discloses a fault diagnosis system for rotating equipment based on a large language model, aiming to improve the accuracy and efficiency of fault diagnosis for industrial rotating equipment. The system consists of six parts: data collection, data preprocessing, a neural network model, a time-series label knowledge base, a large language model, and a user unit. The system can not only detect faults in real time but also enable interaction through the user unit, allowing for analysis and prediction of the rotating equipment's operating status based on user needs. The system optimizes input through data preprocessing techniques and enhances the large language model using a time-series label knowledge base. Based on this, and combining equipment operating data, historical records, and neural network diagnostic results, the large language model can accurately identify potential faults and generate actionable maintenance suggestions, thereby significantly improving equipment reliability and the safety of industrial production.

Description

Technical Field

[0001] This invention belongs to the fields of industrial equipment maintenance and artificial intelligence technology, and in particular relates to a fault diagnosis system for rotating equipment based on a large language model. Background Technology

[0002] Rotating equipment, including pumps, fans, and motors, are core components in industrial production, and their operating status directly affects production efficiency and safety. Traditional fault diagnosis methods mainly rely on human experience and preset rules, resulting in low diagnostic accuracy, low efficiency, and difficulty in handling complex, multi-source data. With the development of industrial big data and artificial intelligence technologies, machine learning-based fault diagnosis methods are gradually emerging. However, these methods still face problems such as high data annotation costs and insufficient model generalization ability, especially performing poorly when processing unstructured maintenance records and dynamic sensor data.

[0003] In recent years, Large Language Models (LLMs) have achieved remarkable results in the field of natural language processing due to their powerful semantic understanding and reasoning capabilities. However, the application of LLMs in the field of rotating equipment fault diagnosis remains a gap. Existing technologies lack a system capable of effectively integrating multi-source data, automating analysis, and providing intelligent diagnosis; furthermore, the application of LLMs in industrial fault diagnosis is still unexplored. Therefore, it is necessary to design a rotating equipment fault diagnosis system based on LLMs to address these issues. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a rotating equipment fault diagnosis system based on a large language model. It aims to solve the problems of low accuracy, low efficiency and insufficient data utilization in traditional rotating equipment fault diagnosis methods. By utilizing the semantic understanding and reasoning capabilities of the large language model, combined with time series analysis technology and neural network fault detection technology, it can achieve efficient and accurate diagnosis of rotating equipment faults, thereby improving the reliability and maintenance efficiency of industrial equipment.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A fault diagnosis system for rotating equipment based on a large language model includes: The data collection unit is used to collect historical time-series data of the rotating equipment, as well as maintenance records and expert knowledge. The time-series data is mostly acquired by sensors and includes vibration, temperature, and current, which are used to build a time-series label knowledge base and train models. This unit is also used to collect various types of time-series data from sensors in real time, which is used for fault diagnosis. The data processing unit is used to preprocess maintenance records and time-series data from sensors related to the rotating equipment; The neural network unit, composed of a pre-trained neural network, is used to analyze the time series data of rotating equipment to output fault diagnosis results. The time-series tag knowledge base unit, composed of a time-series tag knowledge base, retrieves and synthesizes prompt words after receiving user questions or neural network unit classification results and time-series data, thereby enhancing the generalization ability and accuracy of the large language model; The large language model unit performs fault diagnosis after receiving prompt words generated by the time-series tag knowledge base; The user unit provides a visual interface for user interaction and outputs large model prediction results and suggestions.

[0006] Preferably, the data collection unit is used to collect historical time-series data of the rotating equipment, as well as maintenance records and expert knowledge of the rotating equipment. The historical time-series data and expert knowledge are used to construct a time-series label knowledge base and train a model. The data collection unit collects time-series data, including vibration, temperature and current, in real time through multi-source sensors. This time-series data is used for fault diagnosis.

[0007] Preferably, the data processing unit uses natural language processing technology to extract key features from maintenance records and expert knowledge for fine-tuning the large language model, performs noise reduction processing according to the sensor data type, and then performs standardization, normalization, discretization, feature transformation and classification on the corresponding data; the data processing unit performs cleaning, deduplication and completion operations on the data used to build the time-series label knowledge base, and extracts the triple data required by the time-series label knowledge base when facing unstructured text data.

[0008] Preferably, the neural network model unit is trained and adjusted based on historical time-series data and operating parameters to output corresponding fault detection results, and the fault prediction results include fault categories.

[0009] Preferably, the time-series tag knowledge base unit is constructed based on relation extraction, entity recognition, and attribute extraction technologies. After receiving user questions, neural network unit classification results, or time-series data, the time-series tag knowledge base unit performs a retrieval operation and synthesizes prompt words to output to the large language model unit. The large language model unit is fine-tuned using a specific domain dataset. After receiving the prompt words output by the time-series tag knowledge base, it begins to analyze and output the corresponding content.

[0010] Furthermore, the user unit has a visual interface that allows for interactive output with the user. It can output fault prediction results and maintenance suggestions generated by the large language model, including fault cause analysis, urgency assessment, and specific repair steps. The units work together through data flow and feedback mechanisms to support real-time diagnosis and long-term status monitoring.

[0011] Preferably, the data processing unit uses the data collection unit to collect time-series data of the rotating device from multiple sources, including real-time data acquired by sensors, to generate a dataset for training a neural network model, specifically including the following steps: Data labeling and classification: Assuming there are n types of faults in rotating equipment, the dataset contains n+1 labels, including n fault categories and 1 normal operation category; based on expert knowledge or historical fault records, each data sample is assigned a corresponding label to represent its corresponding equipment operating status. Data cleaning and noise reduction: If the time series data is vibration data, wavelet transform or high-pass filter is used to remove low-frequency noise and retain signal features related to the fault; if the time series data is temperature or current data, moving average filter is used to smooth the data and remove outliers; other redundant data are deduplicated by removing duplicate data to ensure data consistency. Feature extraction and transformation: Based on the characteristics of time series data, different feature transformation techniques are used to extract key features. If the time series data is vibration data, the Fast Fourier Transform (FFT) is used to convert the time domain signal into the frequency domain, extract frequency components, amplitude and phase information, and calculate statistical features. If the time series data is temperature or current signal, time domain features are extracted. The features of different time series data are concatenated into a unified feature vector for use by large language models and neural network models. Data standardization and normalization: Data is processed through standardization and normalization. Standardization is achieved by standardizing with fixed-length bins, which divides time-series data into fixed time windows. Normalization applies min-max normalization to vibration, temperature, and current data, mapping the numerical range to [0,1]. After standard normalization, modal alignment technology is used to embed multimodal data into a unified feature space. Structured data generation: Generate structured data for use by neural networks or time-series label knowledge bases.

[0012] Preferably, the neural network unit is used to detect anomalies and output fault categories, specifically including the following steps: Model architecture design involves setting up the architecture of the neural network model, including configuring the input layer, hidden layers, and output layer. The input layer receives data generated by the data processing unit, and the output layer outputs fault classification labels. The number and size of the hidden layers are configured according to the task complexity. The hidden layers contain multiple self-attention modules and feedforward neural networks. ReLU or GeLU is selected as the activation function, and a loss function is constructed. Model Training: The neural network model is initially trained using a structured dataset provided by the data processing unit as input. The structured dataset contains structured device operation features. Supervised learning, semi-supervised learning, or unsupervised learning methods are used to train the model, enabling the neural network to recognize device status. Model evaluation: Evaluate whether the model meets the requirements by assessing its performance on the validation set or test set; Parameter optimization: Dynamically adjust the model's hyperparameters or weights to optimize the model through performance evaluation on the validation or test sets; use transfer learning to fine-tune the fully connected layers, transfer knowledge from the pre-trained model, and improve the model's generalization ability; Model output: Generates a trained model that can be used for fault diagnosis, which is used to detect equipment anomalies and provide fault feature support for a large language model.

[0013] Preferably, the construction of the time-series tag knowledge base includes the following steps: The data preprocessing unit performs data cleaning, feature extraction, and standardization operations on the data, and labels the corresponding data. Sequence slicing employs a sliding window method, using a fixed window length L and step size S to slice the original time series data, generating multiple subsequences. Representative segments are extracted by clustering. The subsequences are clustered using k-means clustering or PAM clustering methods. The representative subsequences and their corresponding labels are stored in the time series label knowledge base. Euclidean distance or dynamic time warping (DTW) is used as the distance metric to ensure that the subsequences can accurately represent the fault modes. We formulate cross-modal conversion rules, design reprogramming layers, use pre-trained models to embed temporal subsequences into semantic space, generate text-compatible vector representations, convert numerical sequences into natural language descriptions through reprogramming layers, and design prompt word templates to be imported into the knowledge base to standardize prompt generation. For time series retrieval, if preprocessed time series data is input, the Dynamic Time Warping (DTW) algorithm or the LCSS algorithm is used to calculate the similarity between the input time series data and the subsequences in the knowledge base, and the top K most similar subsequences are selected. The prompt word generation process converts the retrieved time-series subsequences into text descriptions through a reprogramming layer. These descriptions are then combined with the text search results to generate structured prompt words. These prompt words are input into a large language model to enhance the output performance of the large model.

[0014] Preferably, the user unit performs the following steps: The system receives natural language queries or preprocessed time-series data as input through a visual interface and passes them to the large language model. After receiving the user's request, the system preprocesses the required time series data and inputs it into the neural network model. Then, the corresponding diagnostic results and time series data are input into the time series label knowledge base, which retrieves similar sequences and labels. The time-series tag knowledge base first converts the time-series data into natural language form through a reprogramming layer, and then optimizes and integrates it according to the prompt word template to generate corresponding prompt words and send them to the large language model; The large language model receives structured prompts and generates output that meets the user's needs; The large language model outputs answers that match the user's question through a visual interface.

[0015] Preferably, when a user needs to detect whether the device has malfunctioned, the process is as follows: The data collection unit sends the data collected at this time directly to the data processing unit. After the data is processed, it passes through the time-series tag knowledge base, generates prompt words, and sends them to the large language model. The large language model then generates diagnostic results and outputs them. The visualization interface can present preliminary diagnostic results in the form of text or charts; feedback data provided based on actual maintenance experience or on-site verification results are recorded by the large model and used for subsequent fault detection.

[0016] The user unit has a visual interface that allows for interactive output with the user. It can output fault prediction results and maintenance suggestions generated by the large language model, including fault cause analysis, urgency assessment, and specific repair steps.

[0017] The system further includes a feedback mechanism, through which the units work together via data flow and feedback to support real-time diagnostics and long-term status monitoring.

[0018] The beneficial effects of this invention are as follows: This invention utilizes the semantic understanding and reasoning capabilities of large language models, combined with neural networks, to achieve automated and intelligent diagnosis of rotating equipment faults. Compared with existing technologies, it has advantages such as high efficiency, accuracy, predictability, and adaptability, providing technical support for the reliable operation of industrial equipment. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the rotating equipment fault diagnosis system based on a large language model according to the present invention, showing each unit and the data flow; Figure 2 This is a flowchart of the data processing unit of the present invention processing timing data of a rotating device; Figure 3 This is a flowchart illustrating the model training process of the neural network unit of the present invention. Figure 4 This is a flowchart illustrating the graph construction and application of the time-series tag knowledge base unit of the present invention; Figure 5 This is a flowchart of the user unit of the present invention. Detailed Implementation

[0020] Example 1: like Figure 1 As shown, the system architecture consists of six parts: data collection, data preprocessing, neural network model, time-series label knowledge base, large language model, and user unit.

[0021] Module S101 is the data collection unit of this system. It collects historical time-series data of the rotating equipment, as well as maintenance records and expert knowledge of the rotating equipment. This type of data is used to build a time-series label knowledge base, train neural network models, and fine-tune large language models. In addition, time-series data including vibration, temperature, and current are collected in real time through multi-source sensors. This type of data is used for fault diagnosis.

[0022] Module S102 is the data processing unit. This unit uses natural language processing technology to extract key features from maintenance records and expert knowledge for fine-tuning the large language model. It performs noise reduction processing according to the sensor data type, and then standardizes, normalizes, discretizes, transforms and classifies the corresponding data. For data used to build the time-series label knowledge base, it performs cleaning, deduplication and completion operations. For unstructured text data, it needs to extract the triple data required by the time-series label knowledge base.

[0023] Module S103 is a neural network model unit. This unit is trained based on historical fault data and operating parameters. It can analyze the time series data of rotating equipment, predict future states, detect anomalies, and output preliminary fault judgments.

[0024] Module S104 is a time-series labeled knowledge base unit, which is built based on technologies such as relation extraction, entity recognition, and attribute extraction. After receiving user queries, neural network unit classification results, or time-series data, this unit can perform retrieval operations and synthesize prompt words to output to the large language model.

[0025] Module S105 is a large language model unit. This unit can be fine-tuned using a domain-specific dataset as needed. After receiving prompt words from the time-series tag knowledge base, it begins to analyze and output the corresponding content. The model can also be iteratively optimized by receiving user feedback.

[0026] Module S106 is a user unit with a visual interface that allows for interactive output with the user. It can output fault prediction results and maintenance suggestions generated by the large language model, including fault cause analysis, urgency assessment, and specific repair steps. Secondly, this invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned rotating equipment fault diagnosis method based on a large language model.

[0027] Example 2: The following description, in conjunction with the accompanying drawings, illustrates the rotating equipment fault diagnosis system based on a large language model provided by this invention.

[0028] Figure 2 This is a flowchart illustrating how the data processing unit processes time-series data from rotating equipment. The data processing unit utilizes the data collection unit to collect time-series data from multiple sources, including real-time data acquired by sensors (such as vibration, temperature, and pressure), to generate a dataset for training a neural network model. Specifically, it includes the following steps: Step S201: Data labeling and classification. Assuming there are n types of faults in the rotating equipment, the dataset contains n+1 labels (n fault categories and 1 normal operation category). Using expert knowledge or historical fault records, assign a corresponding label to each data sample to represent its corresponding equipment operating status (e.g., normal, bearing wear, inner ring fault, etc.).

[0029] Step S202: Data cleaning and noise reduction. To ensure data quality and consistency, noise, outliers, and redundant data will be removed using the following methods. If the time series data is vibration data, wavelet transform or a high-pass filter will be used to remove low-frequency noise while retaining signal characteristics related to the fault. If the time series data is temperature or current data, a moving average filter will be used to smooth the data and remove outliers (such as readings outside the physical range). Other redundant data will be removed using deduplication algorithms (such as timestamp-based duplicate detection) to ensure data consistency.

[0030] Step S203: Feature Extraction and Transformation. Based on the characteristics of the time-series data, different feature transformation techniques are used to extract key features. If the time-series data is vibration data, a Fast Fourier Transform (FFT) is used to convert the time-domain signal to the frequency domain, extracting frequency components, amplitude, and phase information, and calculating statistical features. If the time-series data is temperature or current signals, time-domain features are extracted. If necessary, features from different time-series data can be concatenated into a unified feature vector for use by large language models and neural network models.

[0031] Step S204: Data Standardization and Normalization. Data standardization and normalization (e.g., fixed-length binning) address feature misalignment issues caused by differences in sampling rates and signal lengths, ensuring consistent model input. Standardization uses fixed-length binning to divide time-series data into fixed time windows (e.g., 1 second), ensuring consistent input length. Normalization applies min-max normalization to vibration, temperature, and current data, mapping the numerical range to [0,1] to improve model training stability. After standard normalization, modal alignment techniques (e.g., linear mapping or embedding layers) are used to embed multimodal data into a unified feature space, ensuring cross-modal consistency.

[0032] Step S205: Structured data generation, the generated structured data can be used by neural networks or time-series label knowledge bases.

[0033] Figure 3 This is a flowchart of the model training process for a neural network unit used to detect anomalies and output fault categories. Specifically, it includes the following steps: Step S301: Model architecture design. Set up the architecture of the neural network model, configuring the input layer, hidden layers, and output layer. The input layer receives data generated by the data processing unit, and the output layer outputs fault classification labels. The number and size of the hidden layers are configured according to the task complexity; this layer often contains multiple self-attention modules and a feedforward neural network. Choose ReLU, GeLU, or other suitable algorithms as the activation function and select an appropriate loss function based on the specific situation.

[0034] Step S302: Model training. The structured dataset mentioned above is used as input to initially train the neural network model. This data, provided by the data processing unit, contains structured device operation features. Supervised learning, semi-supervised learning, or unsupervised learning methods are employed as appropriate (the difference between the three methods lies in whether the training dataset is labeled), enabling the neural network to recognize device status.

[0035] Step S303: Model evaluation, judge whether the model meets the requirements by evaluating the performance of the validation set or test set (such as accuracy, F1 score).

[0036] Step S304: Parameter optimization. Dynamically adjust the model's hyperparameters (such as learning rate, number of layers, optimization algorithm, etc.) or weights based on performance evaluations (such as accuracy, F1 score) on the validation or test sets. If the real-world situation is more complex, use transfer learning to fine-tune the fully connected layers, transferring knowledge from the pre-trained model to improve the model's generalization ability.

[0037] Step S305: Model output, generating a trained model that can be used for fault diagnosis. This model can accurately detect equipment anomalies and provide fault feature support for the large language model.

[0038] Figure 4 This is a flowchart illustrating the construction and application of the time-series tag knowledge base of the present invention. The construction of the time-series tag knowledge base requires a data collection unit to collect time-series data relevant to the field of rotating equipment fault diagnosis. Specifically, it includes the following steps: Step S401: Use the data preprocessing unit to perform data cleaning, feature extraction and standardization, and label the corresponding data.

[0039] Step S402: Sequence slicing. A sliding window method is used, employing a fixed window length L (e.g., 1 second) and a step size S (e.g., 0.1 seconds) to slice the original time-series data, generating multiple subsequences (of length L). This method effectively preserves local information, reduces the risk of information loss when large language models process long sequences, and improves retrieval efficiency.

[0040] Step S403: Clustering to extract representative segments. Clustering methods such as k-means clustering or PAM clustering are used to cluster the subsequences. Representative subsequences are extracted and stored together with their corresponding labels in the time series label knowledge base. Euclidean distance or dynamic time warping (DTW) is used as the distance metric to ensure that the subsequences can accurately represent the fault modes.

[0041] Step S404: Formulate cross-modal conversion rules, design reprogramming layers, use pre-trained models to embed temporal subsequences into the semantic space, generate text-compatible vector representations, convert numerical sequences into natural language descriptions through reprogramming layers, and design prompt word templates to import into the knowledge base to standardize prompt generation.

[0042] Step S405: Time series sequence retrieval. If there is preprocessed time series data input, use the Dynamic Time Warping (DTW) algorithm or LCSS algorithm to calculate the similarity between the query input time series data and the subsequences in the knowledge base, and select the top K most similar subsequences (K is usually 5).

[0043] Step S406: Prompt word generation. The retrieved time-series subsequences are converted into text descriptions through a reprogramming layer and combined with the text retrieval results to generate structured prompts, which are easier for large language models to understand.

[0044] Subsequently, prompt words are input into the large language model to enhance its output. If the output of the large model is still unsatisfactory, expert data can be used to fine-tune the model; similarly, this data needs to cover the target domain as comprehensively as possible.

[0045] Figure 5 This is a flowchart of the user unit of the present invention. Specifically, it includes the following steps: Step S501: Receive natural language queries or preprocessed time-series data as input through a visual interface and pass them to the large language model. User interaction is primarily conducted through the visual interface.

[0046] Step S502: After receiving the user's request, the system preprocesses the required time series data and inputs it into the neural network model. Then, the corresponding diagnostic results and time series data are input into the time series label knowledge base, which retrieves similar sequences and labels.

[0047] Step S503: The time series tag knowledge base first converts the time series data into natural language form through a reprogramming layer, and then optimizes and integrates it according to the prompt word template to generate corresponding prompt words and send them to the large language model.

[0048] Step S504: The large language model receives structured prompts and generates output that meets the user's needs.

[0049] Step S505: The large language model will output answers that match the user's questions through a visual interface. The questions are not limited to fault diagnosis; users can also ask for maintenance suggestions, etc.

[0050] If the user needs to detect whether the device is malfunctioning at this time, the data collection unit will send the data collected at this time directly to the data processing unit. Subsequently, the processed data will pass through the time-series tag knowledge base, generate prompt words and send them to the large language model, which will then generate diagnostic results output.

[0051] The visual interface can present preliminary diagnostic results in the form of text or charts, allowing users to intuitively view the device status and the likelihood of malfunctions, and also receive corresponding suggestions.

[0052] Feedback provided by users based on actual maintenance experience or on-site verification results can be memorized by the large model for subsequent fault detection.

[0053] Example 3: This embodiment provides a fault diagnosis system for rotating equipment based on a large language model, comprising the following parts: The data collection unit collects historical time-series data of the rotating equipment, as well as maintenance records and expert knowledge. This data is used to build a time-series label knowledge base, train neural network models, and fine-tune large language models. In addition, time-series data, including vibration, temperature, and current, are collected in real time through multi-source sensors. This data is used for fault diagnosis.

[0054] The data processing unit uses natural language processing technology to extract key features from maintenance records and expert knowledge for fine-tuning the large language model. It performs noise reduction based on the sensor data type, and then standardizes, normalizes, discretizes, transforms and classifies the corresponding data. For data used to build the time-series label knowledge base, it performs cleaning, deduplication and completion operations. For unstructured text data, it needs to extract the triple data required by the time-series label knowledge base.

[0055] The time-series tag knowledge base unit is built based on technologies such as relation extraction, entity recognition, and attribute extraction. After receiving user queries, neural network unit classification results, or time-series data, this unit can perform retrieval operations and synthesize prompt words to output to the large language model.

[0056] The neural network model unit is trained based on historical fault data and operating parameters. It can analyze the time series data of rotating equipment, predict future states, detect anomalies, and output preliminary fault judgments.

[0057] The large language model unit can be fine-tuned using domain-specific datasets as needed. Upon receiving prompts from the time-series tag knowledge base, it begins analysis and outputs the corresponding content. The model can also iteratively optimize itself by receiving user feedback.

[0058] The user unit has a visual interface that allows for interactive output with the user. It can output fault prediction results and maintenance suggestions generated by a large language model, including fault cause analysis, urgency assessment, and specific repair steps. Secondly, this invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned rotating equipment fault diagnosis method based on a large language model.

Claims

1. A fault diagnosis system for rotating equipment based on a large language model, characterized in that, include: The data collection unit is used to collect historical time-series data of the rotating equipment, as well as maintenance records and expert knowledge. The time-series data is mostly acquired by sensors and includes vibration, temperature, and current, which are used to build a time-series label knowledge base and train models. This unit is also used to collect various types of time-series data from sensors in real time, which is used for fault diagnosis. The data processing unit is used to preprocess maintenance records and time-series data from sensors related to the rotating equipment; The neural network unit, composed of a pre-trained neural network, is used to analyze the time series data of rotating equipment to output fault diagnosis results. The time-series tag knowledge base unit, composed of a time-series tag knowledge base, retrieves and synthesizes prompt words after receiving user questions or neural network unit classification results and time-series data, thereby enhancing the generalization ability and accuracy of the large language model; The large language model unit performs fault diagnosis after receiving prompt words generated by the time-series tag knowledge base; The user unit provides a visual interface for user interaction and outputs large model prediction results and suggestions.

2. The rotating equipment fault diagnosis system based on a large language model according to claim 1, characterized in that, The data collection unit is used to collect historical time-series data of the rotating equipment, as well as maintenance records and expert knowledge of the rotating equipment. The historical time-series data and expert knowledge are used to build a time-series label knowledge base and train models. The data collection unit collects time-series data, including vibration, temperature and current, in real time through multi-source sensors. This time-series data is used for fault diagnosis.

3. The rotating equipment fault diagnosis system based on a large language model according to claim 1, characterized in that, The data processing unit uses natural language processing technology to extract key features from maintenance records and expert knowledge for fine-tuning the large language model. It performs noise reduction processing according to the sensor data type, and then standardizes, normalizes, discretizes, transforms features, and classifies the corresponding data. The data processing unit performs cleaning, deduplication, and completion operations on the data used to build the time-series label knowledge base. When dealing with unstructured text data, it extracts the triple data required by the time-series label knowledge base.

4. The rotating equipment fault diagnosis system based on a large language model according to claim 1, characterized in that, The neural network model unit is trained and adjusted based on historical time-series data and operating parameters to output corresponding fault detection results. The fault prediction results include fault categories.

5. The rotating equipment fault diagnosis system based on a large language model according to claim 1, characterized in that, The time-series tag knowledge base unit is constructed based on relation extraction, entity recognition, and attribute extraction technologies. After receiving user questions, neural network unit classification results, or time-series data, the time-series tag knowledge base unit performs a retrieval operation and synthesizes prompt words to output to the large language model unit. The large language model unit is fine-tuned using a specific domain dataset. After receiving the prompt words output by the time-series tag knowledge base, it begins to analyze and output the corresponding content.

6. The rotating equipment fault diagnosis system based on a large language model according to claim 3, characterized in that, The data processing unit utilizes the data collection unit to collect time-series data of the rotating equipment from multiple sources, including real-time data acquired by sensors, to generate a dataset for training a neural network model. This process specifically includes the following steps: Data labeling and classification: Assuming there are n types of faults in rotating equipment, the dataset contains n+1 labels, including n fault categories and 1 normal operation category; based on expert knowledge or historical fault records, each data sample is assigned a corresponding label to represent its corresponding equipment operating status. Data cleaning and noise reduction: If the time series data is vibration data, wavelet transform or high-pass filter is used to remove low-frequency noise and retain signal features related to the fault; if the time series data is temperature or current data, moving average filter is used to smooth the data and remove outliers; other redundant data are deduplicated by removing duplicate data to ensure data consistency. Feature extraction and transformation: Based on the characteristics of time series data, different feature transformation techniques are used to extract key features. If the time series data is vibration data, the Fast Fourier Transform (FFT) is used to convert the time domain signal into the frequency domain, extract frequency components, amplitude and phase information, and calculate statistical features. If the time series data is temperature or current signal, time domain features are extracted. The features of different time series data are concatenated into a unified feature vector for use by large language models and neural network models. Data standardization and normalization: Data is processed through standardization and normalization. Standardization is achieved by standardizing with fixed-length bins, which divides time-series data into fixed time windows. Normalization applies min-max normalization to vibration, temperature, and current data, mapping the numerical range to [0,1]. After standard normalization, modal alignment technology is used to embed multimodal data into a unified feature space. Structured data generation: Generate structured data for use by neural networks or time-series label knowledge bases.

7. The rotating equipment fault diagnosis system based on a large language model according to claim 4, characterized in that, The neural network unit is used to detect anomalies and output fault categories, specifically including the following steps: Model architecture design involves setting up the architecture of the neural network model, including configuring the input layer, hidden layers, and output layer. The input layer receives data generated by the data processing unit, and the output layer outputs fault classification labels. The number and size of the hidden layers are configured according to the task complexity. The hidden layers contain multiple self-attention modules and feedforward neural networks. ReLU or GeLU is selected as the activation function, and a loss function is constructed. Model Training: The neural network model is initially trained using a structured dataset provided by the data processing unit as input. The structured dataset contains structured device operation features. Supervised learning, semi-supervised learning, or unsupervised learning methods are used to train the model, enabling the neural network to recognize device status. Model evaluation: Evaluate whether the model meets the requirements by assessing its performance on the validation set or test set; Parameter optimization: Dynamically adjust the model's hyperparameters or weights to optimize the model through performance evaluation on the validation or test sets; use transfer learning to fine-tune the fully connected layers, transfer knowledge from the pre-trained model, and improve the model's generalization ability; Model output: Generates a trained model that can be used for fault diagnosis, which is used to detect equipment anomalies and provide fault feature support for a large language model.

8. The rotating equipment fault diagnosis system based on a large language model according to claim 5, characterized in that, The construction of a time-series tag knowledge base includes the following steps: The data preprocessing unit performs data cleaning, feature extraction, and standardization operations on the data, and labels the corresponding data. Sequence slicing employs a sliding window method, using a fixed window length L and step size S to slice the original time series data, generating multiple subsequences. Representative segments are extracted by clustering. The subsequences are clustered using k-means clustering or PAM clustering methods. The representative subsequences and their corresponding labels are stored in the time series label knowledge base. Euclidean distance or dynamic time warping (DTW) is used as the distance metric to ensure that the subsequences can accurately represent the fault modes. We formulate cross-modal conversion rules, design reprogramming layers, use pre-trained models to embed temporal subsequences into semantic space, generate text-compatible vector representations, convert numerical sequences into natural language descriptions through reprogramming layers, and design prompt word templates to be imported into the knowledge base to standardize prompt generation. For time series retrieval, if preprocessed time series data is input, the Dynamic Time Warping (DTW) algorithm or the LCSS algorithm is used to calculate the similarity between the input time series data and the subsequences in the knowledge base, and the top K most similar subsequences are selected. The prompt word generation process converts the retrieved time-series subsequences into text descriptions through a reprogramming layer. These descriptions are then combined with the text search results to generate structured prompt words. These prompt words are input into a large language model to enhance the output performance of the large model.

9. A fault diagnosis system for rotating equipment based on a large language model according to claim 1, characterized in that, The user unit performs the following steps: The system receives natural language queries or preprocessed time-series data as input through a visual interface and passes them to the large language model. After receiving the user's request, the system preprocesses the required time series data and inputs it into the neural network model. Then, the corresponding diagnostic results and time series data are input into the time series label knowledge base, which retrieves similar sequences and labels. The time-series tag knowledge base first converts the time-series data into natural language form through a reprogramming layer, and then optimizes and integrates it according to the prompt word template to generate corresponding prompt words and send them to the large language model; The large language model receives structured prompts and generates output that meets the user's needs; The large language model outputs answers that match the user's question through a visual interface.

10. A fault diagnosis system for rotating equipment based on a large language model according to claim 9, characterized in that, When a user needs to detect whether the device is malfunctioning, the procedure is as follows: The data collection unit sends the data collected at this time directly to the data processing unit. After the data is processed, it passes through the time-series tag knowledge base, generates prompt words, and sends them to the large language model. The large language model then generates diagnostic results and outputs them. The visualization interface can present preliminary diagnostic results in the form of text or charts; feedback data provided based on actual maintenance experience or on-site verification results are recorded by the large model and used for subsequent fault detection.

Images