Power equipment monitoring and identification method and system based on AI large model analysis

Abstract

The application discloses a kind of power equipment monitoring and identification method and system based on AI big model analysis, through edge side model to the original data collected is preprocessed, in combination with noise filtering and normalization technology improves data quality, subsequently according to abnormal fluctuation threshold value judges whether to upload cloud analysis.In cloud, fusion analysis model combines knowledge graph to carry out multi-source data analysis, accurately identify defect type, and construct defect evolution path by fine-tuning technology integration historical data, finally generate risk assessment and prediction results.For high-risk situation, output disposal suggestion, ensure timely intervention.The application realizes the full-process intelligentization from data acquisition to defect prediction, improves the safety and reliability of power equipment operation, significantly reduces the risk of failure.

Description

Technical Field

[0001] This invention relates to the field of power equipment management technology, and in particular discloses a method and system for monitoring and identifying power equipment based on AI large model analysis. Background Technology

[0002] In the field of power equipment management, ensuring the stability and safety of equipment operation is of paramount importance to socio-economic operations. As a core component of infrastructure, the health of power system equipment directly affects the reliability and continuity of power supply. However, with the expansion of equipment scale and the increasing complexity of operating environments, how to promptly identify and address potential equipment hazards has become a pressing challenge for the industry. Research in this field is not only related to technological advancement but also closely linked to people's well-being.

[0003] Currently, although many solutions have been applied in power equipment monitoring, a common problem is the lack of data integration and responsiveness. Existing methods often struggle to effectively integrate information from different sources, resulting in incomplete assessments of equipment status and a tendency to miss potential risks. Furthermore, existing technologies exhibit significant shortcomings in real-time performance and accuracy when facing complex environments, particularly in their weak adaptability to varying operating conditions. These limitations often leave equipment management in a reactive, reactive state.

[0004] A deeper technical challenge lies in how to achieve collaborative processing of multi-source information. This core factor directly determines whether the system can extract valuable information from complex data. Because the types and characteristics of data generated by different devices and environments vary greatly—for example, images, sounds, and operating parameters each have their own characteristics—if it is not technically possible to achieve unified analysis and in-depth mining of this information, it will lead to biased judgments about the equipment's status. For instance, in actual operation, equipment in a substation may show surface wear through images and reflect internal vibration anomalies through sound signals. However, if the system cannot correlate these different signals, it may overlook a critical potential fault that both signals point to.

[0005] Therefore, how to effectively integrate and collaboratively analyze multi-source information at the technical level, so as to accurately identify hidden problems in equipment status, has become a key issue that urgently needs to be addressed in the current management of power equipment. Summary of the Invention

[0006] This invention provides a method and system for monitoring and identifying power equipment based on AI large model analysis, aiming to solve at least one of the defects in the above-mentioned prior art.

[0007] One aspect of this invention relates to a method for monitoring and identifying power equipment based on AI large-scale model analysis, comprising the following steps:

[0008] S100: Collect time-series operational data of power equipment image and acoustic print through a data acquisition device to obtain a raw multi-source dataset, which contains signals from various sensors.

[0009] S200: The edge-side model is used to perform real-time preprocessing on the original multi-source dataset to obtain preprocessed data. The preprocessed data is then subjected to noise filtering and normalization.

[0010] S300: Determine whether the abnormal fluctuations in the preprocessed data exceed the threshold. If they do, transmit the preprocessed data to the cloud model; otherwise, archive the preprocessed data locally.

[0011] S400: For the pre-processed data transmitted, multi-source data analysis is performed by combining a fusion analysis model with a knowledge graph to determine the type of equipment defect. The knowledge graph includes equipment relationships and historical patterns.

[0012] S500. Based on the determined equipment defect type, use fine-tuning technology to integrate historical data to obtain the defect evolution path used to describe the fault development process.

[0013] S600: By processing the defect evolution path through prompting engineering, a risk assessment is generated to obtain predictive results that indicate the failure trend;

[0014] S700: Extract indicators from the prediction results to determine the risk level. If the risk is high, output disposal suggestions; otherwise, continue to monitor the preprocessed data.

[0015] Further, step S100 includes:

[0016] S110. Acquire multi-channel sensor analog signals from power equipment, and generate a time-aligned multimodal raw data stream based on the timestamps of the multi-channel sensor analog signals;

[0017] S120. Extract the image spatial features and speaker frequency domain features of the time-aligned multimodal raw data stream, and concatenate them to obtain a multi-source heterogeneous joint feature vector;

[0018] S130. Dimensionality reduction is performed on the multi-source heterogeneous joint feature vector to obtain a structured original multi-source dataset containing signals from multiple sensors.

[0019] Further, step S200 includes:

[0020] S210. Obtain the original multi-source dataset containing image and voiceprint information. The original multi-source dataset is analyzed by the edge-side model to obtain the spectral distribution features.

[0021] S220. Based on the spectral distribution characteristics, high-frequency environmental interference components are removed to obtain the pure signal spectrum;

[0022] S230. Perform outlier smoothing replacement on the pure signal spectrum to obtain a smooth signal sequence;

[0023] S240. Obtain the global extrema of the smoothed signal sequence and construct a linear mapping function. Map the smoothed signal sequence to the standard numerical range through the linear mapping function to generate preprocessed data that has completed noise filtering and normalization.

[0024] Further, step S300 includes:

[0025] S310. Obtain the preprocessed data after noise filtering and normalization, and calculate the temporal volatility of the preprocessed data in a continuous time slice.

[0026] S320. Calculate the deviation between the time series volatility and the benchmark feature template, and determine whether the deviation is greater than the preset volatility threshold.

[0027] S330. If the deviation value is greater than the fluctuation threshold, the encrypted transmission channel will be activated to transmit the preprocessed data to the cloud model.

[0028] S340. If the deviation value is less than or equal to the fluctuation threshold, then perform lossless compression encoding on the preprocessed data and generate an archive index to complete the local archiving of the preprocessed data.

[0029] Further, step S400 includes:

[0030] S410. Use a multi-dimensional feature mapping algorithm to convert the preprocessed data transmitted to the cloud into multi-source feature vectors, and retrieve the corresponding device entity nodes from the knowledge graph;

[0031] S420. Extract adjacent nodes and association strength values ​​based on the device entity nodes, and calculate the Euclidean distance between the multi-source feature vector and historical pattern data to generate a pattern matching similarity matrix.

[0032] S430. Input the adjacent nodes, the association strength values ​​and the pattern matching similarity matrix into the fusion analysis model to aggregate and generate the device status representation vector;

[0033] S440. Perform semantic decoding on the equipment state representation vector, locate the matching fault concept in the fault ontology library, and determine the equipment defect type.

[0034] Further, step S500 includes:

[0035] S510. Assemble a specific defect sample set according to the determined equipment defect type, and convert the specific defect sample set into a degradation feature sequence.

[0036] S520. Update the general model parameters based on the degradation feature sequence to generate a specific defect evolution analysis model, and input the current equipment monitoring data into the specific defect evolution analysis model to calculate the state transition probability;

[0037] S530. If the state transition probability exceeds the threshold, the evolution node is determined. The evolution nodes are then connected to form a directed acyclic graph. The directed acyclic graph is then mapped to physical meaning to obtain the defect evolution path used to describe the fault development process.

[0038] Further, step S600 includes:

[0039] S610. Obtain the defect evolution path and serialize the evolution nodes and topology in the defect evolution path into a structured text sequence.

[0040] S620. Embed the structured text sequence into a pre-built prompt word template to construct a prompt engineering input vector, and input the prompt engineering input vector into a large language model to output a risk level distribution matrix;

[0041] S630. Calculate the cumulative failure probability value for the risk level distribution matrix to generate a time-series failure probability dataset, and decode the time-series failure probability dataset to obtain prediction results for indicating failure trends.

[0042] Further, step S700 includes:

[0043] S710. Obtain the prediction results used to indicate the fault trend, and perform feature deconstruction on the prediction results to extract the quantified feature numerical sequence.

[0044] S720. Determine the risk level judgment matrix based on the Euclidean distance of the quantified feature numerical sequence in the safety feature space;

[0045] S730. If the risk level determination matrix indicates high risk, the key abnormal feature vectors extracted from the risk level determination matrix will be mapped to the operation and maintenance knowledge graph to generate structured handling text as handling suggestions.

[0046] S740. If the risk level determination matrix indicates that it is not high risk, then a continuous monitoring instruction is generated to continue monitoring the preprocessed data input later.

[0047] Another aspect of the present invention relates to a power equipment monitoring and identification system based on AI large model analysis, for performing the above-described power equipment monitoring and identification method based on AI large model analysis, comprising:

[0048] The original multi-source dataset acquisition module is used to acquire image and acoustic time-series operation data of power equipment through a data acquisition device to obtain the original multi-source dataset, which contains signals from multiple sensors.

[0049] The preprocessing data acquisition module is used to perform real-time preprocessing on the original multi-source dataset using the edge-side model to obtain preprocessed data. The preprocessed data undergoes noise filtering and normalization.

[0050] The data anomaly fluctuation judgment module is used to determine whether the data anomaly fluctuation in the preprocessed data exceeds the threshold. If it does, the preprocessed data is transmitted to the cloud model; otherwise, the preprocessed data is archived locally.

[0051] The equipment defect type determination module is used to determine the equipment defect type by performing multi-source data analysis on the transmitted preprocessed data through a fusion analysis model combined with a knowledge graph. The knowledge graph includes equipment relationships and historical patterns.

[0052] The defect evolution path acquisition module is used to integrate historical data using fine-tuning technology based on the determined equipment defect type to obtain a defect evolution path that describes the fault development process.

[0053] The prediction result acquisition module is used to process the defect evolution path through prompting engineering, generate risk assessment, and obtain prediction results to indicate the failure trend;

[0054] The risk level assessment module is used to extract indicators from the prediction results to determine the risk level. If the risk is high, it outputs disposal suggestions; otherwise, it continues to monitor the preprocessed data.

[0055] The beneficial effects achieved by this invention are as follows:

[0056] This invention provides a method and system for monitoring and identifying power equipment based on AI large-scale model analysis, offering a complete solution to the business scenario of real-time monitoring of multi-source data, anomaly detection, and defect evolution prediction in power equipment operation. This problem involves how to extract effective information from multi-source data such as images and soundprints to determine equipment anomalies and predict fault trends. This invention preprocesses the collected raw data using an edge-side model, combining noise filtering and normalization techniques to improve data quality. Then, it determines whether to upload the data to the cloud for analysis based on anomaly fluctuation thresholds. In the cloud, a fusion analysis model, combined with a knowledge graph, performs multi-source data analysis to accurately identify defect types. Through fine-tuning techniques, it integrates historical data to construct defect evolution paths, ultimately generating risk assessment and prediction results. For high-risk situations, this invention outputs handling suggestions to ensure timely intervention. The overall technical effect of this invention's method and system for monitoring and identifying power equipment based on AI large-scale model analysis is that it achieves intelligent management of the entire process from data acquisition to defect prediction, improving the safety and reliability of power equipment operation and significantly reducing fault risks. Attached Figure Description

[0057] Figure 1This is a flowchart illustrating an embodiment of the power equipment monitoring and identification method based on AI large model analysis of the present invention.

[0058] Figure 2 This is a functional block diagram of an embodiment of the power equipment monitoring and identification system based on AI large model analysis of the present invention.

[0059] Explanation of icon numbers:

[0060] 10. Raw multi-source dataset acquisition module; 20. Preprocessed data acquisition module; 30. Data anomaly fluctuation judgment module; 40. Equipment defect type determination module; 50. Defect evolution path acquisition module; 60. Prediction result acquisition module; 70. Risk level judgment module. Detailed Implementation

[0061] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0062] like Figure 1 As shown, the first embodiment of the present invention proposes a method for monitoring and identifying power equipment based on AI large model analysis, including the following steps:

[0063] Step S100: Collect time-series operation data of power equipment image and acoustic print through a data acquisition device to obtain the original multi-source dataset, which contains signals from various sensors.

[0064] 1. Deployment and Selection of Data Acquisition Devices: For key monitoring parts of power equipment (such as transformers and circuit breakers), deploy multiple types of data acquisition devices, including high-definition industrial cameras (for acquiring image data such as equipment appearance, oil level, and insulators), acoustic fingerprint sensors (for acquiring vibration acoustic fingerprint and partial discharge acoustic fingerprint data during equipment operation), electrical parameter sensors (for acquiring electrical operating data such as current, voltage, and temperature), and environmental sensors (for acquiring temperature, humidity, and dust concentration data of the equipment operating environment).

[0065] 2. Multi-source data acquisition process: Each acquisition device acquires data synchronously at a preset frequency, with an image data acquisition frame rate of no less than 15 frames / second, a voiceprint data sampling rate of no less than 44.1kHz, and an electrical parameter and environmental data acquisition interval of no more than 1 second; during the acquisition process, all data are bound with unified timestamps, device numbers, acquisition locations, and other identification information to ensure the temporal consistency and traceability of the data.

[0066] 3. Construction of the original multi-source dataset: The collected image data, voiceprint time series data, electrical operation data and environmental data are integrated to form an original multi-source dataset covering "visual-voiceprint-electrical-environment". This dataset contains multiple sensor signals and comprehensively reflects the operating status of power equipment.

[0067] Step S200: Perform real-time preprocessing on the original multi-source dataset using the edge-side model to obtain preprocessed data. The preprocessed data undergoes noise filtering and normalization.

[0068] 1. Edge-side model deployment: The lightweight preprocessing model is deployed on the edge computing nodes (such as edge gateways and edge servers) at the power equipment site. The edge-side model has the characteristics of low latency and low computing power consumption, which is suitable for the real-time processing needs of the site.

[0069] 2. Noise Filtering: Differentiated filtering algorithms are adopted according to the noise characteristics of different types of data: median filtering is used to remove salt-and-pepper noise for image data, combined with Gaussian filtering to smooth the image; wavelet threshold denoising algorithm is used for voiceprint time series data to eliminate environmental noise (such as wind and rain sound, people walking sound) and electromagnetic interference; Kalman filtering algorithm is used for electrical and environmental data to filter random noise caused by sensor inherent errors.

[0070] 3. Normalization: The filtered data are normalized to map data of different dimensions and numerical ranges to the interval [0, 1]. Image data is normalized by pixel value (pixel value / 255), and voiceprint and electrical data are normalized by min-max ((original value - minimum value) / (maximum value - minimum value)) to ensure the comparability of various data and lay the foundation for subsequent fusion analysis.

[0071] 4. Preprocessed data output: After noise filtering and normalization, the output is preprocessed data with a regular structure and clear features, while retaining the temporal correlation and identification information of the data.

[0072] Step S300: Determine whether the abnormal fluctuations in the preprocessed data exceed the threshold. If they do, transmit the preprocessed data to the cloud model; otherwise, archive the preprocessed data locally.

[0073] 1. Abnormal fluctuation threshold calibration: Based on the historical normal operation data of power equipment, statistical analysis methods (such as the 3σ principle) are used to calibrate the normal fluctuation range of various data and determine the abnormal fluctuation threshold; for example, the normal operation temperature fluctuation threshold of transformer is set to ±5℃, and the amplitude fluctuation threshold of acoustic fingerprint signal is set to ±20% of the historical average.

[0074] 2. Real-time fluctuation judgment: The edge computing node calculates the time-series fluctuation value of the preprocessed data in real time (such as the numerical difference between adjacent acquisition cycles, the standard deviation within the sliding window), and compares it with the preset abnormal fluctuation threshold; if the fluctuation value of a certain type of data exceeds the threshold, it is judged that there is abnormal fluctuation; if the fluctuation value of all data is within the threshold range, it is judged that the operation is normal.

[0075] 3. Data tiered processing: If abnormal fluctuations are detected, the edge computing node transmits the preprocessed data to the cloud AI big model in real time through encrypted communication links (such as 5G, fiber optic); if the operation is normal, in order to reduce transmission and storage pressure, the preprocessed data is archived and stored on the local edge node (the archiving period can be set to 30 days, and normal data exceeding the period can be compressed and backed up).

[0076] Step S400: For the pre-processed data transmitted, multi-source data analysis is performed using a fusion analysis model combined with a knowledge graph to determine the equipment defect type. The knowledge graph includes equipment relationships and historical patterns.

[0077] 1. Construction of Fusion Analysis Model: The cloud-based fusion analysis model is built on a large AI model (such as the Transformer architecture model). The model input layer is adapted to multiple data formats (image feature vectors, audioprint time-series features, electrical parameter features). The attention mechanism is used to mine the correlation features between different types of data (such as the coupling relationship between partial discharge audioprints and insulation aging images).

[0078] 2. Knowledge Graph Construction and Application: The knowledge graph contains knowledge in the field of power equipment, specifically covering: equipment relationships (such as the connection relationship between transformers and switchgear, and the association relationship between similar equipment), historical patterns (the matching pattern of historical defect data and corresponding multi-source features, and the typical time sequence pattern of defect development), and defect judgment rules (such as the feature threshold range corresponding to different defects). During the analysis process, the analysis model calls on the domain knowledge in the knowledge graph to constrain and verify the features of multi-source data, thereby improving the accuracy of the analysis.

[0079] 3. Multi-source data analysis and defect determination: The fusion analysis model extracts and fuses features from preprocessed data, and combines historical patterns and judgment rules in the knowledge graph to accurately identify equipment defect types. For example, if image data shows cracks on the surface of the insulator, acoustic data detects partial discharge signals, and electrical data shows an increase in leakage current, the knowledge graph can be used to determine that the insulator is an insulation aging defect. Finally, a clear equipment defect type (such as insulation aging, mechanical jamming, partial discharge, low oil level, etc.) and the corresponding confidence level (confidence level ≥ 90% is a valid judgment result) are output.

[0080] Step S500: Based on the determined equipment defect type, integrate historical data using fine-tuning technology to obtain a defect evolution path used to describe the fault development process.

[0081] 1. Historical data screening and integration: Based on the identified equipment defect types, historical data of similar defects are screened from the power equipment operation and maintenance database, including multi-source time-series data corresponding to historical defects, characteristic changes at each stage of defect development, defect handling records, and final consequences of failures; the screened historical data is cleaned and standardized to ensure data quality.

[0082] 2. Model fine-tuning and feature alignment: Fine-tuning techniques are used to optimize the cloud-based fusion analysis model. Selected historical data is used as incremental training samples to adjust the feature extraction weights of the model, enabling the model to more accurately capture the temporal evolution characteristics of similar defects. At the same time, through feature alignment algorithms, the initial features of the current defect are matched with the features of each stage of the development of historical defects.

[0083] 3. Defect Evolution Path Construction: Based on the analysis results of the fine-tuned model and the temporal patterns of historical data, a defect evolution path is constructed. The path is time-axis-based, clearly describing the development process of the defect from the initial state to the fault state, including key characteristic indicators of each stage (such as changes in acoustic signature amplitude, growth of image defect area, and fluctuation trends of electrical parameters), the duration of each stage, and possible secondary defects. For example, the evolution path of insulation aging defects can be described as: initial partial discharge (acoustic signature amplitude 10-20dB) → insulation crack generation (image feature) → increased leakage current (electrical feature) → insulation breakdown (fault state).

[0084] Step S600: Process the defect evolution path through prompting engineering, generate a risk assessment, and obtain prediction results to indicate the failure trend.

[0085] 1. Prompt Engineering Design: Based on knowledge of power equipment operation and maintenance, design targeted prompt word templates. The prompt words must clearly require the AI ​​model to combine the characteristics, duration, and secondary defect risks of each stage of the defect evolution path, and output risk assessment content including "probability of failure occurrence, estimated time of failure occurrence, and scope of risk impact". Prompt word template example: "Based on the following defect evolution path: {details of evolution path}, analyze the probability of the defect developing into a failure, the estimated time of occurrence, assess the impact of the failure on surrounding equipment and power grid supply, and generate a structured risk assessment report."

[0086] 2. Risk Assessment Generation: The constructed defect evolution path is input into the cloud-based AI model. The model is guided to conduct in-depth analysis through pre-designed prompts. The model combines historical fault impact data in the knowledge graph to complete the quantitative risk assessment. The final output includes the probability of fault occurrence (e.g., high / medium / low levels), the estimated fault occurrence time window, and the scope of risk impact (e.g., single device / local power grid / overall power grid). This result can intuitively indicate the fault development trend.

[0087] Step S700: Extract indicators from the prediction results to determine the risk level. If the risk is high, output disposal suggestions; otherwise, continue monitoring the preprocessed data.

[0088] 1. Risk assessment indicator extraction: Extract core assessment indicators from the prediction results, including the probability of failure, the estimated time of failure, and the scope of risk impact, as the core basis for determining the risk level.

[0089] 2. Risk Level Determination: Preset risk level determination rules, such as: "fault occurrence probability ≥ 70%, and estimated fault occurrence time ≤ 72 hours" or "risk impact range covers the local power grid or above", either condition is determined as high risk; if the fault occurrence probability < 50%, and the estimated fault occurrence time > 168 hours, it is determined as low risk; the rest are determined as medium risk.

[0090] 3. Tiered handling and continuous monitoring: If the risk is determined to be high, the cloud system immediately generates targeted handling suggestions. The handling suggestions combine historical defect handling cases in the knowledge graph, including defect intervention measures (such as shutdown for maintenance, partial power outage handling), handling priority, required operation and maintenance resources, etc., and are pushed to the operation and maintenance personnel's terminals through the operation and maintenance platform. If the risk is determined to be medium or low, the handling suggestion output is not triggered, and the system returns to the edge side monitoring link corresponding to step S300 to continue to monitor the preprocessed data in real time and dynamically track the development status of the defect.

[0091] Furthermore, the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment includes step S100 as follows:

[0092] Step S110: Acquire multi-channel sensor analog signals of power equipment, and generate time-aligned multimodal raw data stream based on the timestamps of the multi-channel sensor analog signals.

[0093] The time-aligned multimodal raw data stream is derived using the following formula:

[0094] (1)

[0095] In formula (1), This represents a time-aligned multimodal raw data stream. Indicates the first Sensor analog signal after channel time alignment Indicates the total number of channels.

[0096] The process of acquiring multi-channel sensor analog signals from power equipment begins with various sensors deployed in substations, such as temperature sensors, vibration sensors, and sound sensors. These sensors generate a continuous stream of analog signals. Specifically, for a high-voltage transformer, the temperature sensor collects temperature data once per second, the vibration sensor captures the waveform of mechanical vibrations, and the sound sensor records the noise signals generated during equipment operation. These signals all carry precise timestamps, such as those in UTC (Universal Time Coordinated) at the millisecond level. These timestamps are used to align the signals from different channels. For example, the timestamps of the temperature signal are matched with those of the vibration signal. If the temperature signal is recorded as 50°C at t=1000ms, while the vibration signal has a peak at t=1002ms, the vibration signal is adjusted to t=1000ms using interpolation, thus generating a time-aligned multimodal raw data stream. This data stream integrates multimodal information such as images, sound, and physical parameters, ensuring the synchronization of subsequent analysis. In practical operations, this helps to monitor equipment status in real time and avoid misjudgments of faults due to time discrepancies. For example, in wind power applications, multi-channel sensor analog signals may include signals collected by wind speed sensors, blade image cameras, and microphone arrays.

[0097] When generating time-aligned multimodal raw data streams, the timestamps of these signals are first collected. For example, the timestamp sequence of wind speed signals is t1, t2, t3, while the timestamps of image signals are t1.5, t2.5. Then, a linear interpolation algorithm is used to align the image signals to points t1, t2, etc., forming a unified streaming data structure. This alignment not only preserves the integrity of the original signals but also provides a foundation for multi-source data fusion. In business applications, it can improve the accuracy of fault prediction because time-synchronized multimodal data can more accurately capture precursors to equipment anomalies, such as the correlation between wind turbine vibration and noise.

[0098] Step S120: Extract the image spatial features and acoustic frequency domain features of the time-aligned multimodal raw data stream, and concatenate them to obtain a multi-source heterogeneous joint feature vector.

[0099] The multi-source heterogeneous joint feature vector is obtained through the following formula:

[0100] (2)

[0101] In formula (2), Represents a multi-source heterogeneous joint feature vector. Representing image spatial features, This represents the frequency domain characteristics of the voiceprint.

[0102] Image spatial features are derived using the following formula:

[0103] (3)

[0104] In formula (3), Represents the image spatial feature extraction function. The image portion representing the time-aligned multimodal raw data stream.

[0105] The frequency domain characteristics of the acoustic signature are derived using the following formula:

[0106] (4)

[0107] In formula (4), Represents the frequency domain transform function. The voiceprint portion represents the time-aligned multimodal raw data stream.

[0108] The process of extracting image spatial features and acoustic frequency domain features from time-aligned multimodal raw data streams involves feature engineering techniques. Specifically, for image spatial features, a convolutional neural network model is used to extract spatial information such as edges and textures from images captured by equipment monitoring cameras, for example, identifying the pixel distribution characteristics of hotspot areas in transformer images. For acoustic frequency domain features, Fourier transform is used to convert the sound signal into a spectrum, extracting the dominant frequency components and harmonic features, such as the 50Hz fundamental frequency peak in equipment operating noise. These features are then concatenated into a multi-source heterogeneous joint feature vector, for example, concatenating the 128-dimensional spatial vector of the image with the 64-dimensional frequency domain vector of the sound to form a 192-dimensional vector. This concatenation method integrates the advantages of heterogeneous data sources and, in power equipment monitoring, can reveal implicit relationships between multiple modalities, such as the correspondence between hot spots in images and sound anomalies, thereby improving the overall feature expressiveness. For example, in a transmission line inspection system, the time-aligned data stream may contain line images captured by drones and arc sounds recorded by ground microphones.

[0109] When extracting spatial features from an image, grayscale processing and edge detection are first performed to obtain spatial vectors such as the curvature of a conductor. Audioprint frequency domain feature extraction involves short-time Fourier transform to calculate the Mel-frequency cepstral coefficients of the sound signal as features. These vectors are then concatenated to form a joint feature vector. This allows for more robust anomaly detection models in practical applications because the fusion of multi-source features reduces noise interference from single modalities, achieving higher diagnostic accuracy.

[0110] Step S130: Perform dimensionality reduction on the multi-source heterogeneous joint feature vector to obtain a structured original multi-source dataset containing signals from multiple sensors.

[0111] The structured raw multi-source dataset containing signals from multiple sensors is derived using the following formula:

[0112] (5)

[0113] In formula (5), Represents the original multi-source dataset. Indicates the first dimensionality-reduced feature vectors of each sample Indicates the first Multiple sensor signal matrices corresponding to each sample This represents the total number of samples.

[0114] Dimensionality reduction of multi-source heterogeneous joint feature vectors is achieved through principal component analysis (PCA), such as projecting high-dimensional vectors into a low-dimensional space while retaining the main variance components. Specifically, for a 192-dimensional joint vector, the covariance matrix is ​​calculated and the eigenvalues ​​are solved. The vectors corresponding to the top 50 largest eigenvalues ​​are selected as the new space, thus obtaining a structured original multi-source dataset. This original multi-source dataset contains refined representations of various sensor signals, such as integrated features of temperature, vibration, and sound. In power equipment maintenance, it can significantly reduce computational complexity while maintaining information integrity, which is helpful for subsequent machine learning tasks such as fault classification. For example, in a smart grid monitoring platform, the original multi-source dataset obtained after dimensionality reduction is used to build equipment health models. Assuming the initial joint feature vector is high-dimensional, 90% of the variance is retained during dimensionality reduction through singular value decomposition, generating a structured dataset containing temperature fluctuation patterns, image defect markers, and abnormal acoustic frequency bands. This not only optimizes storage efficiency but also improves the speed of real-time analysis in business applications, such as quickly identifying transformer overload risks to prevent large-scale power outages.

[0115] Preferably, in the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment, step S200 includes:

[0116] Step S210: Obtain the original multi-source dataset containing image and voiceprint information. The original multi-source dataset is analyzed by the edge-side model to obtain the spectral distribution features.

[0117] The original multi-source dataset is obtained using the following formula:

[0118] (6)

[0119] In formula (6), Represents the original multi-source dataset. Indicates the first Image information, Indicates the first Voiceprint information This represents the total number of samples in the dataset. This indicates an intermediate condition.

[0120] The spectral distribution characteristics are derived using the following formula:

[0121] (7)

[0122] In formula (7), Indicates spectral distribution characteristics, This represents the edge-side model.

[0123] The process of acquiring the raw multi-source dataset containing image and acoustic signature information begins with monitoring equipment deployed in substations. This equipment, such as high-definition cameras and microphone arrays, captures real-time images of high-voltage circuit breakers in operation, along with the accompanying acoustic signature signals of switching actions. Specifically, a series of static image sequences are first acquired via cameras, each recording the surface state of the circuit breaker casing. Simultaneously, the microphone array records the acoustic signature waveform. This data is integrated into the raw multi-source dataset, where the image portion may be stored in JPEG format, and the acoustic signature portion in WAV audio files. Next, the raw multi-source dataset is parsed using an edge-side model to obtain spectral distribution features. Here, the edge-side model refers to a lightweight neural network framework deployed on a local computing node, such as a convolutional model based on TensorFlow Lite. It processes data directly on the device rather than uploading it to the cloud. The parsing process involves extracting color histogram features from the image input model and simultaneously performing a Fast Fourier Transform on the acoustic signature signal to generate a spectrum, thereby calculating the spectral distribution, such as the power spectral density curve. The dominant frequency band is shown as the 20-500Hz range during normal equipment operation, while the high-frequency portion may be mixed with environmental noise.

[0124] Step S220: Based on the spectral distribution characteristics, remove the high-frequency environmental interference components to obtain the pure signal spectrum.

[0125] The following formula is used to directly remove high-frequency environmental interference components above a threshold, yielding a clean spectrum:

[0126] (8)

[0127] In formula (8), Represents the spectrum of a pure signal. Represents the spectrum of the original signal. This represents the high-frequency threshold determined based on spectral distribution characteristics. Represents a frequency variable.

[0128] In subsequent processing, to obtain a clean signal stream by stripping away high-frequency environmental interference components based on spectral distribution characteristics, a filtering algorithm is applied to the spectrum. Specifically, high-frequency thresholds in the spectral distribution, such as those exceeding 1kHz, are considered interference. These components are then removed using a Butterworth low-pass filter. The process involves calculating the filter's transfer function and applying it to the signal, for example, multiplying the acoustic spectrum point by point by the filter response, preserving the core low-frequency information, thereby outputting a clean signal stream. This ensures signal clarity in power line inspection operations.

[0129] Step S230: Perform outlier smoothing replacement on the pure signal spectrum to obtain a smooth signal sequence.

[0130] Outliers are detected based on the Z-score using the following formula and replaced with the neighborhood mean to smooth the signal:

[0131] (9)

[0132] In formula (9), Represents the smoothed signal sequence. Spectral points, The spectrum of the pure signal is represented by the first... The nth point represents the nth... The neighborhood window of a point Represents the median of the neighborhood. Indicates the neighborhood standard deviation. The standardization threshold is represented by Z-score. The control logic of formula (9) is to use Z-score to determine whether a spectral point is an outlier. If it is an outlier, it is replaced with the neighborhood mean; otherwise, the original value is retained, thereby achieving signal smoothing.

[0133] The process of smoothing outliers in a clean signal stream to obtain a smooth signal sequence involves anomaly detection techniques. For example, median filtering can be used to scan the signal stream, identify outliers such as sudden spikes that deviate from the mean, and replace them with the median of neighboring points. Specifically, for a signal sequence containing 1000 sampling points, if the value of the 500th point is 10 times the mean (an anomaly), the median of the five points before and after it (e.g., 2.5) is calculated and replaced. This process is repeated until all outliers are smoothed, forming a continuous smooth signal sequence.

[0134] Step S240: Obtain the global extrema of the smoothed signal sequence and construct a linear mapping function. Map the smoothed signal sequence to the standard numerical range through the linear mapping function to generate preprocessed data that has completed noise filtering and normalization.

[0135] The linear mapping function is defined by the following formula and is used to perform noise filtering and normalization:

[0136] (10)

[0137] In formula (10), This represents the preprocessed data after mapping to a standard numerical range. One value, Represents the smoothed signal sequence. One value, This represents the global minimum of a smoothed signal sequence. This represents the global maximum value of a smoothed signal sequence.

[0138] The global minimum of a smoothed signal sequence is obtained using the following formula:

[0139] (11)

[0140] In formula (11), Indicates the length of the smoothed signal sequence. This indicates minimization.

[0141] The global maximum value of a smoothed signal sequence is obtained by the following formula:

[0142] (12)

[0143] In formula (12), This indicates maximization.

[0144] The process of constructing a linear mapping function to obtain the global extrema of a smoothed signal sequence first scans the entire sequence to find the maximum value (e.g., 5.0) and the minimum value (e.g., 0.1). Then, the mapping function is constructed as y = (x - min) / (max - min), where x is the original value and y is the mapped value. Specifically, this function is applied to each point. For example, the original value 2.0 is mapped to (2.0 - 0.1) / (5.0 - 0.1) ≈ 0.388. The smoothed signal sequence is mapped to a standard numerical range (e.g., [0, 1]) through the linear mapping function, generating preprocessed data that has undergone noise filtering and normalization. This provides standardized input for subsequent analysis in substation equipment monitoring.

[0145] Furthermore, in the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment, step S300 includes:

[0146] Step S310: Obtain the preprocessed data after noise filtering and normalization, and calculate the temporal volatility of the preprocessed data in a continuous time slice.

[0147] Time series volatility is derived using the following formula:

[0148] (13)

[0149] In formula (13), Represents time series volatility. Indicates the number of consecutive time slices. Indicates the first Preprocessed data for each time slice, This represents the average value of the preprocessed data.

[0150] After obtaining the preprocessed data through noise filtering and normalization, it is necessary to calculate its temporal volatility over consecutive time slices. Temporal volatility refers to the amplitude and frequency of data changes over a period of time, reflecting the stability of the data. For example, assuming the preprocessed data comes from signals from an indoor temperature sensor, and each time slice is set to 5 minutes, the amplitude of temperature value changes over multiple time slices is analyzed. If the temperature value fluctuates frequently within a short period of time, it indicates an anomaly in the environmental control system.

[0151] Step S320: Calculate the deviation between the time series volatility and the benchmark feature template, and determine whether the deviation is greater than the preset volatility threshold.

[0152] The deviation value is obtained using the following formula:

[0153] (14)

[0154] In formula (14), Indicates the deviation value. Represents the reference feature template.

[0155] When calculating the deviation between time-series volatility and a benchmark feature template, the benchmark feature template is a pre-defined standard data model used to measure the degree of deviation of the current data. For example, the benchmark feature template might be an ideal temperature fluctuation range generated based on historical data. The deviation is calculated by comparing the current time-series volatility with the standard value defined in the template. If the standard fluctuation range is ±0.5 degrees Celsius, and the temperature fluctuation within the current time slice reaches 1.2 degrees Celsius, then the deviation significantly exceeds expectations.

[0156] Determining whether the deviation value exceeds the preset fluctuation threshold is crucial in deciding the data processing path. The fluctuation threshold is a preset critical value used to distinguish whether data requires further processing. Assuming the fluctuation threshold is set to 0.8 degrees, when the deviation value reaches 1.2 degrees, it is significantly greater than the threshold. At this point, the system will trigger a subsequent special processing mechanism to ensure that data anomalies are responded to promptly.

[0157] Step S330: If the deviation value is greater than the fluctuation threshold, the encrypted transmission channel is activated to transmit the preprocessed data to the cloud model.

[0158] The activation conditions for the encrypted transmission channel are defined using the following formula:

[0159] (15)

[0160] In formula (15), This indicates that the result was received in the cloud. This indicates preprocessed data. Indicates encryption. Indicates decryption. This indicates that the cloud-based model receives and processes the data. Indicates the deviation value. The value represents the fluctuation threshold. The control logic of formula (15) is that when the deviation value exceeds the fluctuation threshold, the encrypted transmission channel of "encryption → decryption → cloud processing" is used; otherwise, the channel is not activated and the received result is 0.

[0161] If the deviation exceeds the fluctuation threshold, the system will activate the encrypted transmission channel to transmit the preprocessed data to the cloud model. The encrypted transmission channel is a mechanism that protects data transmission security through a specific encryption protocol. For example, when temperature data fluctuates abnormally, the system will upload the data to the cloud server via the encrypted channel. The cloud model will then further analyze whether there is equipment failure or external interference, ensuring that the data is not stolen or tampered with during transmission.

[0162] Step S340: If the deviation value is less than or equal to the fluctuation threshold, perform lossless compression encoding on the preprocessed data and generate an archive index to complete the local archiving of the preprocessed data.

[0163] The following formula is used for local archiving of preprocessed data:

[0164] (16)

[0165] In formula (16), This indicates the completed local archive. Indicates the archiving function. This indicates the generated archive index.

[0166] If the deviation is less than or equal to the fluctuation threshold, lossless compression encoding is performed on the preprocessed data, and an archive index is generated to complete local archiving. Lossless compression encoding is a technique that reduces storage space without losing data information, while the archive index is an identification system designed for rapid data retrieval later. When temperature fluctuations are within the normal range, the system compresses the data and stores it on the local device, generating an index combining a timestamp and a device ID for easy historical record lookup later. This approach saves storage space while ensuring data integrity, facilitating subsequent analysis.

[0167] Preferably, in the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment, step S400 includes:

[0168] Step S410: Use a multi-dimensional feature mapping algorithm to convert the preprocessed data transmitted to the cloud into a multi-source feature vector, and retrieve the corresponding device entity node from the knowledge graph.

[0169] The corresponding device entity node is obtained through the following formula:

[0170] (17)

[0171] In formula (17), This represents the corresponding device entity node. Representation of knowledge graph Entity nodes in This represents the transpose of the query feature vector. Represents a knowledge graph.

[0172] To address the process of converting preprocessed data transmitted to the cloud into multi-source feature vectors, a multi-dimensional feature mapping algorithm is used to achieve multi-dimensional representation of the data. Specifically, the multi-dimensional feature mapping algorithm is a technique that decomposes raw data into multiple feature dimensions to extract hidden correlation information within the data. Assuming the data transmitted to the cloud is signal data from an indoor temperature and humidity sensor, the system will decompose this data into multiple feature dimensions such as temperature change trends and humidity fluctuation frequency, forming a multi-source feature vector to facilitate subsequent analysis of device status. For example, when retrieving corresponding device entity nodes from a knowledge graph, which is a structured network of device information containing relationships and attribute information between devices, assuming the knowledge graph stores the association between temperature and humidity sensors and electrical equipment, the system will retrieve the corresponding device entity node based on the sensor's identifier, such as a specific model of electrical equipment, and further mine relevant information based on this.

[0173] Step S420: Extract adjacent nodes and association strength values ​​based on the device entity nodes, and calculate the Euclidean distance between the multi-source feature vector and the historical pattern data to generate a pattern matching similarity matrix.

[0174] The association strength between a device entity node and its adjacent nodes is calculated using the following formula:

[0175] (18)

[0176] In formula (18), Represents device entity nodes Its adjacent nodes The correlation strength value, Represents a node and The original correlation strength between them Represents device entity nodes The set of adjacent nodes.

[0177] The Euclidean distance between multi-source feature vectors and historical pattern data is obtained by the following formula:

[0178] (19)

[0179] In formula (19), Represents multi-source feature vectors Compared with historical pattern data Euclidean distance, Representing feature dimension, and Indicates the first 1D feature components.

[0180] The pattern matching similarity matrix is ​​obtained using the following formula:

[0181] (20)

[0182] In formula (20), This represents the elements in the pattern matching similarity matrix. Represents the Euclidean distance norm. This represents an exponential function.

[0183] The steps of extracting adjacent nodes and their association strength values ​​are primarily for analyzing the connection strength between device entities and other related nodes. Adjacent nodes are power modules or control panels associated with electrical equipment, while the association strength values ​​reflect the dependence of these nodes in historical operation. For example, the influence weight of a power module on the operation of electrical equipment is 0.7, and the influence weight of a control panel is 0.3. The system constructs a relationship network between devices based on this information. Similarly, the process of calculating the Euclidean distance between multi-source feature vectors and historical pattern data to generate a pattern matching similarity matrix is ​​mainly to measure the closeness of the current data to historical normal operating patterns. Euclidean distance is a geometric method for measuring the difference between two vectors. The system compares the current temperature and humidity feature vector with historical normal operating data to generate a similarity matrix used to determine whether the current equipment operation deviates from the normal state.

[0184] Step S430: Input the adjacent nodes, the association strength values ​​and the pattern matching similarity matrix into the fusion analysis model to aggregate and generate the device status representation vector.

[0185] The device state representation vector is derived using the following formula:

[0186] (twenty one)

[0187] In formula (21), This represents the device state representation vector.

[0188] The process of inputting adjacent nodes, association strength values, and pattern matching similarity matrices into a fusion analysis model to generate a device status representation vector is a technical means of comprehensively analyzing multi-source information. The fusion analysis model is an analytical framework capable of integrating heterogeneous data. It combines device relationship networks and similarity matrices to generate a vector that comprehensively represents the device's operating status, reflecting whether the device is in an abnormal state.

[0189] Step S440: Semantically decode the equipment state representation vector and locate the matching fault concept in the fault ontology library to determine the equipment defect type.

[0190] The following formula is used to determine the type of equipment defect:

[0191] (twenty two)

[0192] In formula (22), Indicates a specific type of equipment defect. This represents the concept of a matching fault. This represents a mapping function from the concept of a fault to a type of defect.

[0193] The concept of a matched fault is derived from the following formula:

[0194] (twenty three)

[0195] In formula (23), Represents the fault ontology library, This represents the input semantic representation vector. Representing the concept of fault Embedded vector, This represents the transpose of the input semantic representation vector.

[0196] The steps of semantically decoding the equipment state representation vector and locating and matching fault concepts in the fault ontology are to transform abstract vector information into specific fault types. The fault ontology is a knowledge base that stores equipment fault features and classifications. The system matches the semantic information decoded from the representation vector to the corresponding fault concepts in the fault ontology, such as a fault in the cooling module of a power equipment, thereby determining the specific equipment defect type and providing guidance for subsequent maintenance.

[0197] Furthermore, in the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment, step S500 includes:

[0198] Step S510: Assemble a specific defect sample set according to the determined equipment defect type, and convert the specific defect sample set into a degradation feature sequence.

[0199] A specific defect sample set is obtained using the following formula:

[0200] (twenty four)

[0201] In formula (24), Represents a specific set of defect samples. Indicates a specific type of equipment defect. This represents a single defect sample. This indicates the defect type of the defect sample.

[0202] The degenerative feature sequence is derived using the following formula:

[0203] (25)

[0204] In formula (25), Represents the degenerate feature sequence. This represents the transformation function from a sample set to a sequence. Indicates the size of the sample set.

[0205] The process of assembling a specific defect sample set based on the identified equipment defect type is achieved by filtering relevant data from a historical database. Specifically, once the defect type of the cooling module of the power equipment has been determined, the system extracts all sample data related to that defect from the stored equipment logs, including records of abnormal temperature fluctuations, power equipment operation logs, and sensor feedback signals. This data is organized into a specific defect sample set for subsequent analysis.

[0206] The step of converting a specific defect sample set into a degradation feature sequence is mainly to capture the gradual process of defects from initial to severe. The system analyzes the time series data in the sample set, such as the trajectory of the cooling efficiency of power equipment gradually decreasing from the normal value to the fault threshold. Through time series decomposition algorithms, this data is converted into a degradation feature sequence, where each point in the sequence represents the efficiency degradation rate, thus forming an ordered feature chain, which facilitates the model's learning of the defect evolution law.

[0207] Step S520: Update the general model parameters based on the degradation feature sequence to generate a specific defect evolution analysis model, and input the current equipment monitoring data into the specific defect evolution analysis model to calculate the state transition probability.

[0208] The following formula is used to update the general model parameters through the specific defect evolution analysis model parameters to generate the specific defect evolution analysis model:

[0209] (26)

[0210] In formula (26), This represents the parameters of a specific defect evolution analysis model. Indicates general model parameters, Indicates the learning rate. Represents the gradient. Let represent the loss function. The control logic of formula (26) is to control the update magnitude with the learning rate along the direction of decreasing loss function (gradient inversion), and to obtain model parameters adapted to specific defects through iteration based on general model parameters. The core function of formula (26) is to optimize the general model parameters in the direction of adapting to the evolution analysis of specific defects, generate a special analysis model for specific defects, and improve the analysis accuracy of this type of defect.

[0211] The state transition probability is calculated using the following formula:

[0212] (27)

[0213] In formula (27), Represents the state transition probability. Indicates the next state. Indicates the current state. This indicates the current equipment monitoring data. Indicates the calculation of the next state The exponential result of the inner product of the corresponding weight vector and the monitoring data. This represents the summation of the exponential results for all states. The control logic of formula (27) uses the inner product exponent of the current device monitoring data and the corresponding state weight as weights, and calculates the probability of the current state transitioning to a specific next state through softmax normalization.

[0214] The process of updating general model parameters based on degradation feature sequences to generate a specific defect evolution analysis model involves fine-tuning the base model using transfer learning techniques. Specifically, the general model is a pre-trained neural network framework used to handle various equipment state predictions. In the case of power equipment defects, the system inputs degradation feature sequences and adjusts the model parameters, such as the weight matrix, to adapt them to the specific degradation pattern of the refrigeration module, thereby generating a dedicated evolution analysis model that can simulate the dynamic transition of defects from mild to severe.

[0215] The step of inputting current power equipment monitoring data into a specific defect evolution analysis model to calculate state transition probabilities is to quantify the likelihood of defect development. In practical applications, assuming that the current monitoring data includes the real-time temperature and power consumption of the power equipment, the specific defect evolution analysis model processes these inputs and evaluates the probability of transitioning from the current state to a more severe state through its internal probability calculation layer. For example, it calculates the transition value from slight overheating to complete failure, helping to identify potential risks.

[0216] Step S530: If the state transition probability exceeds the threshold, the evolution node is determined. Then, the evolution nodes are connected to form a directed acyclic graph, and the directed acyclic graph is mapped with physical meaning to obtain the defect evolution path used to describe the fault development process.

[0217] The following formula is used to map the physical meaning of the directed acyclic graph to obtain the defect evolution path describing the fault development process:

[0218] (28)

[0219] In formula (28), Indicates the defect evolution path, Represents a mapping function for physical meaning. This represents a directed acyclic graph.

[0220] The directed acyclic graph formed by connecting the evolving nodes is derived using the following formula:

[0221] (29)

[0222] In formula (29), Represents the set of evolutionary nodes. Represents an edge set. and Let represent the evolution nodes. Connect the evolution nodes that satisfy the transition probability greater than 0 to form a directed acyclic graph. The control logic of formula (29) is to only connect the evolution nodes that have a transition probability greater than 0, and use these nodes and edges to form a directed acyclic graph to reflect the effective transition relationship between states.

[0223] The process of determining evolution nodes when the state transition probability exceeds a threshold emphasizes the role of the threshold as a decision boundary. Specifically, if the calculated transition probability exceeds a preset threshold of 0.5, the system marks the current state as an evolution node. For example, when the probability of power equipment transitioning from normal operation to overload state is higher than the threshold, this point is identified as a critical node for constructing the defect path. The step of connecting evolution nodes to form a directed acyclic graph aims to construct an acyclic graphical structure to represent defect progression. In power equipment defect analysis, the system connects multiple nodes, such as the initial overheating node, the power equipment addition node, and the final failure node, to form a directed graph from the starting point to the end point, ensuring that the path is acyclic to reflect the unidirectional evolution process.

[0224] The process of mapping the physical meaning of a directed acyclic graph (DAG) to obtain a defect evolution path that describes the fault development process is achieved by associating it with actual equipment components. Specifically, the system maps nodes in the graph to physical meanings. In the refrigeration system of power equipment, for example, the initial node corresponds to refrigerant leakage, the intermediate node corresponds to motor wear, and the final path describes the entire development process from efficiency degradation caused by leakage to complete failure, thus providing a clear view of fault evolution.

[0225] Preferably, in the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment, step S600 includes:

[0226] Step S610: Obtain the defect evolution path and serialize the evolution nodes and topology in the defect evolution path into a structured text sequence.

[0227] The following formula describes the aggregation framework of evolution nodes and topological joint serialization in the defect evolution path:

[0228] (30)

[0229] In formula (30), This represents the generated structured text sequence. Indicates the number of evolutionary nodes. Indicates the first An evolutionary node, Indicates the first A topological structure, This function encodes nodes and topology into text units.

[0230] The process of acquiring the defect evolution path and serializing its evolution nodes and topology into a structured text sequence is primarily aimed at converting complex graphical data into a processable text format, facilitating subsequent model input. Specifically, this serialization step involves traversing each node and edge in a directed acyclic graph (DAG). For example, for a defect path in a power equipment refrigeration module, the system first identifies nodes such as the initial temperature anomaly node, the compressor overload node, and the final failure node. Then, these nodes and their connections are encoded into a text sequence, such as "Node 1: Temperature Anomaly -> Node 2: Compressor Overload -> Node 3: System Failure," including topological details such as edge weights representing transition probabilities. This structured text sequence fully preserves the logical structure of the path, providing foundational data for the next step of constructing prompts. Furthermore, serialization ensures data consistency and prevents information loss, thereby improving the accuracy of fault prediction in business operations.

[0231] Step S620: Embed the structured text sequence into the pre-built prompt word template to construct the prompt engineering input vector, and input the prompt engineering input vector into the large language model to output the risk level distribution matrix.

[0232] The risk level distribution matrix is ​​derived using the following formula:

[0233] (31)

[0234] In formula (31), Represents the risk level distribution matrix. This indicates the input vector for the project. A function that represents a large language model, processing input vectors to output matrices.

[0235] The input vector for the project is derived using the following formula:

[0236] (32)

[0237] In formula (32), This refers to a pre-built prompt word template. This represents the operation function that embeds a structured text sequence into a pre-built prompt word template.

[0238] The step of embedding structured text sequences into pre-built cue word templates to construct cue-engineered input vectors emphasizes the role of cue engineering in optimizing model input. Cue engineering is a technique for designing effective input cues to guide large language models to generate specific outputs. Here, the pre-built template is a standardized string framework, such as "Based on the following defect path: [sequence], analyze risk distribution." The system embeds the structured text sequence into it to form a complete input vector. For example, in the scenario of power equipment defects, the input vector becomes "Based on the following defect path: Node 1: Temperature anomaly -> Node 2: Compressor overload -> Node 3: System failure, analyze risk distribution." This engineered input vector encapsulates path information in natural language form, helping large language models better understand the context.

[0239] The process of inputting the input vector to a large language model to output a risk level distribution matrix involves leveraging the natural language processing capabilities of the large language model to quantify risk. A large language model is a deep learning-based AI framework, such as the Transformer architecture, capable of processing complex text and generating structured output. After the system inputs the aforementioned vector, the large language model analyzes the node transitions in the path and outputs a risk level distribution matrix. For example, rows represent risk levels (low, medium, high), columns represent evolution stages, and matrix elements represent the probability distribution of each stage, such as low risk 0.7, medium risk 0.2, and high risk 0.1. This generated risk level distribution matrix provides a multi-dimensional view of risk, facilitating subsequent calculations.

[0240] Step S630: Calculate the cumulative failure probability value for the risk level distribution matrix to generate a time-series failure probability dataset, and decode the time-series failure probability dataset to obtain prediction results for indicating failure trends.

[0241] The cumulative failure probability value is obtained by using the sum of matrix elements as an exponential form of the total risk intensity to calculate the survival probability:

[0242] (33)

[0243] In formula (33), This represents the cumulative failure probability value. The risk level distribution matrix represents the first... Line number Column elements, This indicates the row number of the risk level distribution matrix. The number of columns in the risk level distribution matrix is ​​represented by . The control logic of formula (33) is to add up all the elements of the risk level distribution matrix to get the total risk, and then subtract the negative exponent of the total risk from 1 to get the cumulative failure probability value.

[0244] The following formula is used to decode the dataset through linear transformation and the softmax function to obtain fault trend predictions:

[0245] (34)

[0246] In formula (34), This represents the prediction results used to indicate failure trends. This represents a dataset representing the probability of time-series failures. Represents the weight matrix. This represents the bias vector.

[0247] The step of calculating cumulative failure probability values ​​for the risk level distribution matrix to generate a time-series failure probability dataset focuses on cumulative calculation to capture failure trends over time. Specifically, the system iterates through the risk level distribution matrix, accumulating the failure probability for each stage. For example, it extracts the values ​​of the high-risk column from the risk level distribution matrix, such as stage 1: 0.1, stage 2: 0.3, and stage 3: 0.6, and then calculates the cumulative values, such as stage 1: 0.1, stage 2: 0.4, and stage 3: 1.0, forming a time-series dataset. This time-series dataset reflects the dynamic evolution of failure probability over time, helping to identify potential outbreak points.

[0248] The process of decoding a time-series failure probability dataset to obtain predictions indicating failure trends involves interpreting the probability sequences within the dataset to derive trend insights. For example, if the cumulative probability rapidly increases from 0.1 to 0.8 after the system decodes the dataset, the prediction would be "accelerating failure trend, expected to fail within 48 hours." This decoding, combined with threshold judgment and trend analysis, not only provides early warnings but also enables preventative maintenance in business operations, reducing equipment downtime.

[0249] Furthermore, in the power equipment monitoring and identification method based on AI large model analysis provided in this embodiment, step S700 includes:

[0250] Step S710: Obtain the prediction results used to indicate the fault trend, and perform feature deconstruction on the prediction results to extract the quantitative feature value sequence.

[0251] The following formula is used to extract the quantized feature numerical sequence from the prediction results through principal component projection:

[0252] (35)

[0253] In formula (35), Indicates the extracted first Each quantized feature value, Indicates the first Transpose of principal component vectors The prediction result vector is represented by the formula (35). The control logic of formula (35) is to project (inner product) the prediction result vector onto the principal component vector to obtain the corresponding quantized feature values.

[0254] After obtaining the predicted results that indicate fault trends, the process of deconstructing them to extract quantifiable feature numerical sequences mainly involves decomposing the core elements of the predicted results to obtain quantifiable data sequences, thus providing a foundation for subsequent risk assessment. Specifically, this feature deconstruction involves breaking down the predicted results, such as fault probability curves or trend indicators, into multiple dimensions. For example, in a smart grid system, for the predicted results of transformer overheating trends, the system first identifies features such as peak values, slopes, and fluctuation amplitudes in the curve, and then extracts numerical sequences using standardization methods, such as a peak value of 0.85, a slope of 0.12, and a fluctuation of 0.05, forming a sequence [0.85, 0.12, 0.05]. This sequence captures the dynamic changes in the trend, ensuring that the data can be used for spatial computation. Furthermore, feature deconstruction can preserve the integrity of the original prediction, avoiding information loss and thus improving the accuracy of risk assessment in business operations.

[0255] Step S720: Determine the risk level judgment matrix based on the Euclidean distance of the quantified feature numerical sequence in the safety feature space.

[0256] The Euclidean distance between the quantized feature numerical sequence and the secure feature space is obtained by the following formula:

[0257] (36)

[0258] In formula (36), Indicates the first The Euclidean distance from a quantized feature numerical sequence to a safe reference point in the safe feature space. Indicates the first The k-th dimension quantized feature value of the sequence. Representing the safety feature space reference point coordinate, The dimension of the safety feature space is represented by the control logic of formula (36), which is to calculate the sum of squares of the deviations between each dimension of the feature sequence and the reference point, and then take the square root to obtain the straight-line distance between the two in the safety feature space.

[0259] The risk level determination matrix is ​​determined using the following formula:

[0260] (37)

[0261] In formula (37), The risk level determination matrix represents the first... Line number Column element, number Sequence risk level The judgment criteria, Indicates the first The Euclidean distance threshold for risk level is 1, indicating compliance with the risk level, and 0 otherwise. The control logic of formula (37) is to compare the Euclidean distance from the feature sequence to the safety reference point with the threshold of the corresponding risk level. If the distance is large enough, it is marked as 1 (belonging to the risk level), and if it is not large enough, it is marked as 0.

[0262] The step of determining the risk level judgment matrix based on the Euclidean distance of the quantified feature numerical sequence in the safety feature space emphasizes the use of geometric distance to quantify the risk level. The safety feature space is a multi-dimensional vector space used to represent the distribution of various safety indicators, while Euclidean distance is a metric for calculating the straight-line distance between two points. Here, the system maps the extracted quantified feature numerical sequence into this space and compares the distance with a predefined safety threshold vector. Specifically, in the transformer scenario of a smart grid, assuming the quantified feature numerical sequence is [0.85, 0.12, 0.05] and the safety threshold vector is [0.50, 0.05, 0.02], the system calculates the Euclidean distance as sqrt((0.85-0.50)). 2 +(0.12-0.05) 2 +(0.05-0.02) 2 The distance threshold is approximately 0.36. Then, a risk level determination matrix is ​​generated based on the distance threshold. For example, the matrix rows represent the risk level (low, medium, high), the columns represent the feature dimensions, and the elements are normalized probabilities. For example, the high-risk row is [0.6, 0.4, 0.3]. This risk level determination matrix provides a comprehensive view of the risk, which is convenient for subsequent decision-making.

[0263] Step S730: If the risk level determination matrix indicates high risk, the key abnormal feature vectors extracted from the risk level determination matrix will be mapped to the operation and maintenance knowledge graph to generate structured disposal text as disposal suggestions.

[0264] The following formula is used to define the criteria for determining a high-risk indicator:

[0265] (38)

[0266] In formula (38), Indicates a high-risk indicator. This represents the risk level determination matrix. Describing the Frobenius norm, Indicates a high-risk threshold, if for This indicates high risk. The control logic of formula (38) is to measure the overall intensity of risk using the Frobenius norm of the risk level judgment matrix. If the threshold is exceeded, it is judged as high risk; otherwise, it is not.

[0267] The following formula is used to extract key anomaly feature vectors from the risk level determination matrix:

[0268] (39)

[0269] In formula (39), Represents the key anomaly feature vector. This represents the risk level determination matrix. This represents element-wise multiplication. The key anomaly binary mask matrix is ​​represented. The control logic of formula (39) is to "filter" the risk level judgment matrix using the key anomaly binary mask matrix - the original matrix elements are retained at the positions of 1 in the mask, and the positions of 0 are set to 0, thereby extracting the key anomaly features.

[0270] If the risk level assessment matrix indicates high risk, the process involves mapping key anomaly feature vectors extracted from the matrix to the operations and maintenance (O&M) knowledge graph to generate structured handling text as a handling suggestion. This process utilizes the knowledge graph's matching mechanism. The O&M knowledge graph is a structured knowledge base containing nodes such as equipment type, fault mode, and handling strategy, with edges representing relationships. Here, the system extracts anomaly vectors from the matrix, such as [0.6, 0.4], and maps them to relevant nodes in the graph. For example, in a high-risk situation involving a power grid transformer, the vector is mapped to the "overheating fault" node, connected to the "cooling system check" and "load adjustment" strategies, generating text such as "It is recommended to immediately check the cooling fans and reduce the load to 80%." This handling suggestion, through the semantic association of the graph, ensures the relevance and operability of the suggestion, enabling rapid response in business operations and reducing power outage time.

[0271] Step S740: If the risk level determination matrix indicates that it is not high risk, then generate a continuous monitoring instruction to continue monitoring the preprocessed data input later.

[0272] The following formula is used to generate continuous monitoring instructions for subsequent data:

[0273] (40)

[0274] In formula (40), Indicates a continuous monitoring instruction. This indicates the function for generating continuous monitoring instructions. This indicates the subsequent input preprocessed data. The control logic of formula (40) is to pass the preprocessed data to be processed to the function that generates continuous monitoring instructions, and the function that generates continuous monitoring instructions directly generates the corresponding continuous monitoring instructions.

[0275] If the risk level assessment matrix indicates a non-high risk, a continuous monitoring instruction is generated to continue monitoring the preprocessed data input later. The focus is on maintaining the monitoring cycle. The continuous monitoring instruction is an automated sequence of instructions used to trigger data acquisition and analysis. Here, the system generates the instruction "Collect temperature data every hour, preprocess to filter noise, and continue trend prediction" based on the matrix, such as medium risk [0.3, 0.2]. In the power grid scenario, this ensures continuous tracking of transformer temperature, avoids the escalation of potential risks, and thus supports preventative maintenance strategies in business operations.

[0276] Please see Figure 2 This invention provides a power equipment monitoring and identification system based on AI large-scale model analysis, used to execute the aforementioned power equipment monitoring and identification method based on AI large-scale model analysis. It includes a raw multi-source dataset acquisition module 10, a preprocessing data acquisition module 20, a data anomaly fluctuation judgment module 30, an equipment defect type determination module 40, a defect evolution path acquisition module 50, a prediction result acquisition module 60, and a risk level judgment module 70. The raw multi-source dataset acquisition module 10 is used to acquire time-series operational data of power equipment images and acoustic fingerprints through a data acquisition device to obtain a raw multi-source dataset containing signals from various sensors. The preprocessing data acquisition module 20 is used to perform real-time preprocessing on the raw multi-source dataset using an edge-side model to obtain preprocessed data, which undergoes noise filtering and normalization. The data anomaly fluctuation judgment module 30... The system is used to determine whether abnormal fluctuations in preprocessed data exceed a threshold. If so, the preprocessed data is transmitted to the cloud model; otherwise, the preprocessed data is archived locally. The equipment defect type determination module 40 uses a fusion analysis model combined with a knowledge graph to perform multi-source data analysis on the transmitted preprocessed data to determine the equipment defect type. The knowledge graph includes equipment relationships and historical patterns. The defect evolution path acquisition module 50 integrates historical data using fine-tuning techniques based on the determined equipment defect type to obtain a defect evolution path describing the fault development process. The prediction result acquisition module 60 processes the defect evolution path through a prompting process, generates a risk assessment, and obtains prediction results indicating fault trends. The risk level judgment module 70 extracts indicators from the prediction results to judge the risk level. If the risk is high, a handling suggestion is output; otherwise, the preprocessed data continues to be monitored.

[0277] The power equipment monitoring and identification method based on AI large model analysis provided in this embodiment has the following beneficial effects compared with the prior art:

[0278] 1. Comprehensive monitoring of multi-source data to enhance status perception capabilities: By collecting multi-dimensional data such as images, soundprints, electrical data, and environmental data, the system comprehensively covers the factors influencing the operating status of power equipment, overcomes the limitations of single-data monitoring, and achieves all-round perception of equipment status.

[0279] 2. Real-time preprocessing at the edge reduces transmission pressure and latency: Data denoising and normalization are completed locally at the edge, and only abnormal data is transmitted to the cloud, which greatly reduces network bandwidth usage and enables rapid response to abnormal data, meeting the real-time monitoring needs of power equipment.

[0280] 3. AI large model + knowledge graph fusion to improve defect identification accuracy: Relying on the multi-source feature mining capability of AI large model, combined with the domain knowledge and historical experience of knowledge graph, the accuracy and generalization ability of defect identification are effectively improved, and false positives and false negatives are reduced.

[0281] 4. Defect evolution path prediction for proactive operation and maintenance: By integrating historical data and model fine-tuning, defect evolution paths are constructed to predict the fault development process, providing operation and maintenance personnel with a basis for early intervention, transforming passive operation and maintenance into proactive prevention and control, and reducing fault handling costs.

[0282] 5. Adaptable to various types of power equipment and highly versatile: By adjusting model parameters and knowledge graph content, it can adapt to the monitoring needs of various power equipment such as transformers, circuit breakers, and switchgear, and has broad engineering application value.

[0283] 6. Risk prediction and tiered handling to improve the targeted nature of operation and maintenance: The addition of prompts and risk assessment steps enables accurate prediction of failure trends. Combined with risk level tiers, it outputs handling suggestions to avoid blind operation and maintenance, improve the efficiency of operation and maintenance resource utilization, and further reduce the probability of failure and handling costs.

[0284] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for monitoring and identifying power equipment based on AI large-scale model analysis, characterized in that, Includes the following steps: S100. Collect time-series operational data of power equipment image and acoustic print through a data acquisition device to obtain a raw multi-source dataset, wherein the raw multi-source dataset contains multiple sensor signals. S200. The original multi-source dataset is preprocessed in real time using an edge-side model to obtain preprocessed data, which is then subjected to noise filtering and normalization. S300. Determine whether the abnormal fluctuation of the preprocessed data exceeds the threshold. If it does, transmit the preprocessed data to the cloud model; otherwise, archive the preprocessed data locally. S400. For the transmitted preprocessed data, multi-source data analysis is performed by combining a fusion analysis model with a knowledge graph to determine the equipment defect type. The knowledge graph includes equipment relationships and historical patterns. S500. Based on the determined equipment defect type, historical data is integrated using fine-tuning technology to obtain a defect evolution path used to describe the fault development process. S600: The defect evolution path is processed through the prompting process to generate a risk assessment and obtain prediction results to indicate the failure trend. S700. Extract indicators from the prediction results to determine the risk level. If the risk is high, output a disposal suggestion; otherwise, continue to monitor the preprocessed data. Step S500 includes: S510. Assemble a specific defect sample set according to the determined equipment defect type, and convert the specific defect sample set into a degradation feature sequence. S520. Update the general model parameters according to the degradation feature sequence to generate a specific defect evolution analysis model, and input the current equipment monitoring data into the specific defect evolution analysis model to calculate the state transition probability; S530. If the state transition probability exceeds the threshold, an evolution node is determined. The evolution nodes are then connected to form a directed acyclic graph, and the directed acyclic graph is mapped to physical meaning to obtain a defect evolution path used to describe the fault development process.

2. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 1, characterized in that, Step S100 includes: S110. Acquire multi-channel sensor analog signals from power equipment, and generate a time-aligned multimodal raw data stream based on the timestamps of the multi-channel sensor analog signals; S120. Extract the image spatial features and speaker frequency domain features of the time-aligned multimodal raw data stream, and concatenate them to obtain a multi-source heterogeneous joint feature vector; S130. Perform dimensionality reduction processing on the multi-source heterogeneous joint feature vector to obtain a structured original multi-source dataset containing signals from multiple sensors.

3. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 1, characterized in that, Step S200 includes: S210. Obtain the original multi-source dataset containing image and voiceprint information, and obtain the spectral distribution features of the original multi-source dataset through edge-side model analysis; S220. Based on the spectral distribution characteristics, remove the high-frequency environmental interference components to obtain the clean signal spectrum; S230. Perform outlier smoothing replacement on the spectrum of the pure signal to obtain a smooth signal sequence; S240. Obtain the global extrema of the smoothed signal sequence to construct a linear mapping function, and map the smoothed signal sequence to a standard numerical range through the linear mapping function to generate preprocessed data that has completed noise filtering and normalization.

4. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 1, characterized in that, Step S300 includes: S310. Obtain the preprocessed data after noise filtering and normalization, and calculate the temporal volatility of the preprocessed data in a continuous time slice. S320. Calculate the deviation value between the time series volatility and the benchmark feature template, and determine whether the deviation value is greater than the preset volatility threshold. S330. If the deviation value is greater than the fluctuation threshold, the encrypted transmission channel is activated to transmit the preprocessed data to the cloud model. S340. If the deviation value is less than or equal to the fluctuation threshold, then perform lossless compression encoding on the preprocessed data and generate an archive index to complete the local archiving of the preprocessed data.

5. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 1, characterized in that, Step S400 includes: S410. The preprocessed data transmitted to the cloud is converted into a multi-source feature vector using a multi-dimensional feature mapping algorithm, and the corresponding device entity node is retrieved from the knowledge graph. S420. Extract adjacent nodes and association strength values ​​based on the device entity nodes, and calculate the Euclidean distance between the multi-source feature vector and historical pattern data to generate a pattern matching similarity matrix. S430. Input the adjacent nodes, the association strength values ​​and the pattern matching similarity matrix into the fusion analysis model to aggregate and generate the device status representation vector; S440. Semantically decode the device state representation vector and locate the matching fault concept in the fault ontology library to determine the device defect type.

6. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 1, characterized in that, Step S600 includes: S610. Obtain the defect evolution path, and serialize the evolution nodes and topology in the defect evolution path into a structured text sequence. S620. Embed the structured text sequence into a pre-built prompt word template to construct a prompt engineering input vector, and input the prompt engineering input vector into a large language model to output a risk level distribution matrix; S630. Calculate the cumulative failure probability value for the risk level distribution matrix to generate a time-series failure probability dataset, and decode the time-series failure probability dataset to obtain prediction results for indicating failure trends.

7. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 1, characterized in that, Step S700 includes: S710. Obtain the prediction results used to indicate the fault trend, and perform feature deconstruction on the prediction results to extract a quantized feature value sequence. S720. Determine the risk level determination matrix based on the Euclidean distance of the quantized feature numerical sequence in the safety feature space; S730. If the risk level determination matrix indicates high risk, the key abnormal feature vector extracted from the risk level determination matrix will be mapped to the operation and maintenance knowledge graph to generate structured handling text as handling suggestions. S740. If the risk level determination matrix indicates that the risk is not high, a continuous monitoring instruction is generated to continue monitoring the preprocessed data input later.

8. The method for monitoring and identifying power equipment based on AI large model analysis according to claim 7, characterized in that, In step S730, the following formula is used to define the criteria for determining a high-risk indication: ； in, Indicates a high-risk indicator. This represents the risk level determination matrix. Denotes the Frobenius norm. Indicates a high-risk threshold, if A value of 1 indicates high risk; The following formula is used to extract key anomaly feature vectors from the risk level determination matrix: ； in, Represents the key anomaly feature vector. This represents the risk level determination matrix. This represents element-wise multiplication. This represents the binary mask matrix for key anomalies.

9. A power equipment monitoring and identification system based on AI large model analysis, used to execute the power equipment monitoring and identification method based on AI large model analysis as described in any one of claims 1 to 8, characterized in that, include: The original multi-source dataset acquisition module (10) is used to acquire the image acoustic text time-series operation data of power equipment through the data acquisition device to obtain the original multi-source dataset, which contains multiple sensor signals. The preprocessing data acquisition module (20) is used to perform real-time preprocessing on the original multi-source dataset using the edge-side model to obtain preprocessed data, which is subjected to noise filtering and normalization. The data anomaly fluctuation judgment module (30) is used to determine whether the data anomaly fluctuation in the preprocessed data exceeds the threshold. If it exceeds the threshold, the preprocessed data is transmitted to the cloud model; otherwise, the preprocessed data is archived locally. The equipment defect type determination module (40) is used to determine the equipment defect type by performing multi-source data analysis on the transmitted preprocessed data through a fusion analysis model combined with a knowledge graph. The knowledge graph includes equipment relationships and historical patterns. The defect evolution path acquisition module (50) is used to integrate historical data using fine-tuning technology based on the determined equipment defect type to obtain a defect evolution path that describes the fault development process. The prediction result acquisition module (60) is used to process the defect evolution path through prompting engineering, generate a risk assessment, and obtain prediction results to indicate the failure trend; The risk level judgment module (70) is used to extract indicators from the prediction results and judge the risk level. If the risk is high, it outputs a disposal suggestion; otherwise, it continues to monitor the preprocessed data.

Images