Power equipment discharge detection and analysis method based on multi-mode large model

Abstract

The invention discloses an electrical equipment discharge detection and analysis method based on a multi-modal large model, and the method specifically comprises the following steps: S1, a signal collection module which aims at collecting multi-source original data, including acoustic signals, environmental parameters and video data, S2, a signal preprocessing module which carries out the feature extraction and optimization processing of the collected data; s3, a discharge source positioning module which realizes accurate positioning of a discharge source; according to the system and the method, errors are likely to occur in a noisy or complex environment, the system is verified through multiple data, the output position information is more reliable, and errors are obviously reduced. The system and the method have the advantages that the system and the method are easy to make mistakes in the noisy or complex environment, and the system has the advantages that the output position information is more reliable, and the errors are obviously reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment operation and maintenance, and in particular to a method for detecting and analyzing discharge of power equipment based on a multimodal large model. Background Art

[0002] The core function of current acoustic positioning detection products is to display the location of the discharge source in real time, but they lack in-depth analysis and intelligent diagnosis capabilities, making it difficult to meet the needs of refined equipment operation and maintenance. On the one hand, the existing system fails to automatically identify discharge modes (such as surface discharge, corona discharge, etc.) and can only provide a rough discharge location, resulting in one-sided detection results. On the other hand, the cause of discharge may involve multiple factors such as insulation aging, equipment aging, and environmental impact, but current products are unable to automatically analyze in combination with operating data. In addition, the existing system lacks trend prediction capabilities and cannot analyze the evolution of discharge based on historical data, resulting in insufficient fault warning capabilities. It can only passively monitor and cannot provide proactive maintenance recommendations.

[0003] After completing the positioning, existing acoustic detection products still require manual data analysis and manual compilation of detection reports, resulting in low efficiency and large subjective errors. At present, the detection results are usually presented in the form of sound field distribution maps or raw numerical data. Inspectors need to manually interpret and write detection reports, which is not only time-consuming, but may also cause misjudgments due to subjective judgment. In addition, because the detection systems of different manufacturers use different data formats, it is difficult to directly connect the detection results to the power operation and maintenance system (such as SCADA, DCS, etc.), affecting information sharing and automated management. If video stream data can be combined with large model analysis, the system can automatically generate standardized detection reports, including discharge location, fault component identification, severity rating, maintenance recommendations and other information, which can greatly reduce manual intervention and improve the intelligence level of detection and engineering application value. Summary of the Invention

[0004] This invention aims to address the problems of insufficient intelligent analysis capabilities, lack of visualization methods, and low efficiency in generating detection reports in existing acoustic positioning detection products. Specifically, the present invention introduces intelligent diagnostic algorithms to achieve automatic recognition of discharge patterns and analysis of fault causes, thereby improving detection accuracy and early warning capabilities. By integrating video stream data, the intuitiveness and visualization level of discharge source positioning are improved, and the convenience of equipment inspections is enhanced. Standardized detection reports are automatically generated through large language models, reducing manual intervention and improving data sharing and operation and maintenance management efficiency. This patent proposes a method for detecting and analyzing discharges of power equipment based on a multimodal large model, which is characterized in that it specifically includes the following steps:

[0005] S1 signal acquisition module

[0006] This module is designed to collect multi-source raw data, including acoustic signals, environmental parameters and video data.

[0007] -S11 Acoustic Signal Acquisition: Deploy a microphone array to record the acoustic signal s(t) with a sampling rate of f s .

[0008] -S12 Environmental and video data acquisition: monitors temperature T, humidity H, and captures video frames V(t) through a camera.

[0009] -Output: dataset {s(t),T,H,V(t)}.

[0010] S2 signal preprocessing module

[0011] This module performs feature extraction and optimization processing on the collected data.

[0012] -S21 signal synchronization: through time difference correction Δt=t i -t j , adjust the signal s(t-Δt).

[0013] -S22 feature extraction:

[0014] -S221 Acoustic Feature Extraction: Calculate frequency f = argmax|S(f)|, where S(f) = ∫s(t)e -j2πft dt; amplitude A = max|s(t)|; phase difference

[0015] -S222 Video feature extraction: Calculate the abnormal area A_v = ∑ (x,y) δ(x,y), where δ(x,y)=1 represents an outlier.

[0016] -S23 data standardization: normalization processing

[0017] -Output: Feature set

[0018] S3 discharge source positioning module

[0019] This module enables precise positioning of the discharge source.

[0020] -S31 acoustic positioning calculation:

[0021] -S311 Time Difference of Attention (TDOA): Based on the distance difference d ij =v·Δt ij (v=343m / s), solve the positioning equation:

[0022]

[0023] -S312 beamforming optimization: direction estimation θ = argmax∑ i si (t-τ i(θ) ).

[0024] -S32 multi-source data correction: combined with video data to verify d = √((xx v ) 2 +(yy v ) 2 )<δ.

[0025] -Output: coordinates (x, y) and sound field map Figure 2 The two-dimensional top view obtained for positioning contains (x, y) coordinates.

[0026] S4 multimodal large model module

[0027] This module uses large vision-language models (such as GPT-4o) for data fusion and analysis.

[0028] -S41 data input and fusion:

[0029] enter {T,H},V(t),(x,y,z),P(x,y), through the attention mechanism Generate representation Z.

[0030] -S42 Image content understanding and component recognition: Generate description D from V(t) image (such as "spark discharge on the surface of the insulator"), identify the element set E = {e1, e2, ...}.

[0031] -S43 acoustic positioning result conversion: convert (x,y,z), P(x,y) to text D location (For example, “the discharge source is located at (x=1.2m, y=0.5m, z=0.8m)”).

[0032] -S44 Discharge pattern recognition and comprehensive analysis: Integrating data to generate prompt P input , identify the discharge mode y i (such as "corona discharge"), analyze the cause P (cause i |X) and trend T trend .

[0033] - Output: {y i ,P(cause i ),T trend ,D image ,E,D location}.

[0034] S5 fault diagnosis and report generation module

[0035] -S51 fault diagnosis: calculate severity S = w1f + w2A + w3P (y i ), combine E and (x, y, z) to determine the faulty component.

[0036] -S52 report generation: integrate S4 output and generate standardized reports.

[0037] -S53 Output and Visualization: Generate SCADA / DCS compatible reports with sound field diagrams and trend curves.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. More Accurate Positioning: Compared to traditional technologies, this invention combines time-of-day positioning of acoustic signals with beamforming technology, supplemented by video data correction, to more accurately determine the location of the discharge source. Traditional methods often rely on a single signal, such as sound or image alone, which can be prone to errors in noisy or complex environments. However, this system uses multiple data sources to verify each other, resulting in more reliable output location information and significantly reduced errors.

[0040] 2. More comprehensive analysis: This invention utilizes a powerful multimodal model to integrate and analyze sound, environmental conditions, and video information. Traditional technologies typically only analyze a single data point, making it difficult to fully understand the situation. However, this system can mine information from multiple perspectives. For example, it can determine the intensity of the discharge through sound and identify the specific faulty component through images. Combined with environmental factors such as temperature and humidity, it can more accurately determine the discharge type, cause, and future trends, significantly enhancing analytical capabilities.

[0041] 3. Smarter and More Efficient Diagnosis: Previous detection systems often required manual intervention, making analysis and report generation time-consuming. This invention automates the entire process from data acquisition to report generation. It leverages a large multimodal model to directly provide image descriptions, location information, and diagnostic results, and automatically generates standardized reports. This not only saves time and effort but also ensures consistent results, making it ideal for industrial scenarios requiring rapid response. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Technology roadmap;

[0043] Figure 2 Discharge source positioning effect. DETAILED DESCRIPTION

[0044] A method for detecting and analyzing discharge of electric power equipment based on a multimodal large model, characterized by comprising the following steps:

[0045] S1 signal acquisition module

[0046] This module is designed to collect multi-source raw data, including acoustic signals, environmental parameters and video data.

[0047] -S11 Acoustic Signal Acquisition: Deploy a microphone array to record the acoustic signal s(t) with a sampling rate of f s .

[0048] -S12 Environmental and video data acquisition: monitors temperature T, humidity H, and captures video frames V(t) through a camera.

[0049] -Output: dataset {s(t),T,H,V(t)}.

[0050] S2 signal preprocessing module

[0051] This module performs feature extraction and optimization processing on the collected data.

[0052] -S21 signal synchronization: through time difference correction Δt=t i -t j , adjust the signal s(t-Δt).

[0053] -S23 feature extraction:

[0054] -S231 Acoustic Feature Extraction: Calculate frequency f = argmax|S(f)|, where S(f) = ∫s(t)e -j2πft dt; amplitude A = max|s(t)|; phase difference

[0055] -S233 video feature extraction: calculate the abnormal area A_v = ∑ (x,y) δ(x,y), where δ(x,y)=1 represents an outlier.

[0056] -S24 data standardization: normalization processing

[0057] -Output: Feature set

[0058] S3 discharge source positioning module

[0059] This module enables precise positioning of the discharge source.

[0060] -S31 acoustic positioning calculation:

[0061] -S311 Time Difference of Attention (TDOA): Based on the distance difference d ij =v·Δt ij (v=343m / s), solve the positioning equation:

[0062]

[0063] -S312 beamforming optimization: direction estimation θ = argmax∑ i s i (t-τ i(θ) ).

[0064] -S32 multi-source data correction: combined with video data to verify d = √((xx v ) 2 +(yy v ) 2 )<δ.

[0065] -Output: coordinates (x, y) and sound field map Figure 2 The two-dimensional top view obtained for positioning contains (x, y) coordinates.

[0066] S4 multimodal large model module

[0067] This module uses large vision-language models (such as GPT-4o) for data fusion and analysis.

[0068] -S41 data input and fusion:

[0069] enter {T,H},V(t),(x,y,z),P(x,y), through the attention mechanism Generate representation Z.

[0070] -S42 Image content understanding and component recognition: Generate description D from V(t) image (such as "spark discharge on the surface of the insulator"), identify the element set E = {e1, e2, ...}.

[0071] -S43 acoustic positioning result conversion: convert (x,y,z), P(x,y) to text D location (For example, “the discharge source is located at (x=1.2m, y=0.5m, z=0.8m)”).

[0072] -S44 Discharge pattern recognition and comprehensive analysis: Integrating data to generate prompt P input , identify the discharge mode y i (such as "corona discharge"), analyze the cause P (cause i |X) and trend T trend .

[0073] - Output: {y i ,P(cause i ),T trend ,D image ,E,D location}.

[0074] S5 fault diagnosis and report generation module

[0075] -S51 fault diagnosis: calculate severity S = w1f + w2A + w3P (y i ), combine E and (x, y, z) to determine the faulty component.

[0076] -S52 report generation: integrate S4 output and generate standardized reports.

[0077] -S53 Output and Visualization: Generate SCADA / DCS compatible reports with sound field diagrams and trend curves.

[0078] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for detecting and analyzing discharge of power equipment based on a multimodal large model, characterized by: The specific steps include: S1 signal acquisition module: This module is designed to collect multi-source raw data, including acoustic signals, environmental parameters and video data; -S11 Acoustic Signal Acquisition: Deploy a microphone array to record the acoustic signal s(t) with a sampling rate of f s ; -S12 Environmental and video data acquisition: monitors temperature T, humidity H, and captures video frames V(t) through a camera; -Output: dataset {s(t),T,H,V(t)}; S2 signal preprocessing module: This module performs feature extraction and optimization processing on the collected data; -S21 signal synchronization: through time difference correction Δt=t i -t j , adjust the signal s(t-Δt); -S22 feature extraction: -S221 Acoustic Feature Extraction: Calculate frequency f = argmax|S(f)|, where S(f) = ∫s(t)e -j2πft dt; amplitude A = max|s(t)|; phase difference -S222 Video feature extraction: Calculate the abnormal area A_v = ∑ (x,y) δ(x,y), where δ(x,y)=1 represents an outlier; -S23 data standardization: normalization processing -Output: Feature set S3 discharge source positioning module: This module realizes the precise positioning of the discharge source; -S31 acoustic positioning calculation: -S311 Time Difference of Attention (TDOA): Based on the distance difference d ij =v·Δt ij (v=343m / s), solve the positioning equation: -S312 beamforming optimization: direction estimation θ = argmax∑ i s i (t-τ i(θ) ); -S32 multi-source data correction: combined with video data to verify d = √((xx v ) 2 +(yy v ) 2 )<δ; -Output: coordinates (x, y) and sound field map Figure 2 is a two-dimensional top view obtained by positioning, including (x, y) coordinates; S4 multimodal large model module: This module uses large vision-language models (such as GPT-4o) for data fusion and analysis; -S41 data input and fusion: enter {T,H},V(t),(x,y,z),P(x,y), through the attention mechanism Generate representation Z; -S42 Image content understanding and component recognition: Generate description D from V(t) image (e.g. "spark discharge on the insulator surface"), identification element set E = {e1, e2, ...}; -S43 acoustic positioning result conversion: convert (x,y,z), P(x,y) to text D location (e.g. "The discharge source is located at (x = 1.2 m, y = 0.5 m, z = 0.8 m)"); -S44 Discharge pattern recognition and comprehensive analysis: Integrating data to generate prompt P input , identify the discharge mode y i (such as "corona discharge"), analyze the cause P (cause i |X) and trend T trend ; - Output: {y i ,P(cause i ),T trend ,D image ,E,D location }; S5 fault diagnosis and report generation module: -S51 fault diagnosis: calculate severity S = w1f + w2A + w3P (y i ), combine E and (x, y, z) to determine the faulty component; -S52 report generation: integrate S4 output and generate standardized reports; -S53 Output and Visualization: Generate SCADA / DCS compatible reports with sound field diagrams and trend curves.

Images