A transformer insulating oil photoacoustic spectrum on-line monitoring system

Abstract

The present application relates to a kind of transformer insulating oil photoacoustic spectrum online monitoring system, belong to power equipment condition monitoring and fault diagnosis technical field.The system includes oil-gas separation unit, photoacoustic spectrum detection unit, signal acquisition and preprocessing unit, feature extraction unit, intelligent diagnosis and early warning unit, temperature compensation unit, pressure compensation unit, data storage and management unit and remote communication and display unit.Feature extraction unit extracts gas concentration, change rate, component ratio and spectral waveform etc multidimensional features;Intelligent diagnosis and early warning unit is built in graph neural network fault diagnosis model and time series trend prediction model, realizes fault identification and risk early warning;System also constructs the linkage closed loop of diagnosis result drive oil-gas separation parameter and detection parameter dynamic adjustment, and combines temperature pressure adaptive compensation.Realize the high-precision online monitoring and intelligent diagnosis of dissolved gas in transformer oil, significantly improve the accuracy of fault identification and the timeliness of early warning.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, specifically to an online monitoring system for the photoacoustic spectrum of transformer insulating oil. Background Technology

[0002] Dissolved gas analysis (DGA) in transformer oil is a core technical means for diagnosing latent faults in oil-immersed power transformers. During long-term operation, transformers are subjected to heat and electrical stress, causing the insulating oil and solid insulation materials to decompose and produce characteristic fault gases such as hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), ethane (C2H6), carbon monoxide (CO), and carbon dioxide (CO2). By analyzing the types, contents, and variation patterns of these gas components, early faults such as partial discharge, overheating, and spark discharge can be effectively identified, preventing further escalation of the fault. Traditional dissolved gas detection in oil mainly relies on offline gas chromatography. While this method offers high accuracy, it suffers from inherent limitations such as long detection cycles, numerous manual steps, poor repeatability, and high labor intensity. Furthermore, offline analysis involves a lengthy process from sampling and transportation to laboratory testing, making it difficult to meet the practical needs of real-time, continuous monitoring of transformer operating conditions. In recent years, with the rapid development of optical sensing technology, online monitoring methods based on spectral principles have received increasing attention. Photoacoustic spectroscopy (PAS), based on the photoacoustic effect, quantitatively analyzes gas concentration by detecting the acoustic signal generated after a gas absorbs and modulates a laser. It offers significant advantages such as requiring no carrier gas, not consuming gas samples, high detection sensitivity, and the ability to simultaneously detect multiple gas components. This technology shows promising application prospects in the field of online monitoring of dissolved gases in transformer oil.

[0003] Although photoacoustic spectroscopy technology has been initially applied in online monitoring of dissolved gases in transformer oil, existing systems still face many technical challenges that urgently need to be addressed. First, existing photoacoustic spectroscopy detection systems mostly employ fixed-parameter separation strategies in the oil-gas separation stage, failing to dynamically adjust separation efficiency based on real-time changes in gas concentration. This results in a delayed system response when the concentration of fault-characteristic gases abnormally increases, making it difficult to meet the timeliness requirements for early warning. Second, spectral cross-interference between multiple gas components during photoacoustic spectroscopy detection is one of the main factors affecting detection accuracy. Existing systems have relatively simple data processing methods, lacking deep decoupling and feature mining capabilities for spectral data, making it difficult to accurately extract independent concentration information of each gas component from complex spectral signals. Third, existing photoacoustic spectroscopy online monitoring systems are mostly open-loop architectures, with the photoacoustic spectroscopy detection unit, oil-gas separation unit, and data analysis unit operating independently, lacking dynamic feedback and adaptive adjustment mechanisms based on detection results. When diagnostic results are in an uncertain range, the system cannot proactively adjust detection parameters for secondary verification, making it difficult to guarantee the confidence level of diagnostic decisions. Furthermore, existing systems largely rely on single threshold judgments or traditional shallow machine learning models in data processing and fault diagnosis, failing to fully utilize multi-dimensional feature parameters for comprehensive diagnosis and lacking the ability to predict gas concentration change trends, making it difficult to achieve the leap from "current state diagnosis" to "future risk warning." Meanwhile, photoacoustic spectroscopy detection systems operate long-term in outdoor substation environments, where fluctuations in environmental parameters such as temperature and pressure significantly affect the stability of photoacoustic signals and the consistency of detection results. Existing systems have limited compensation methods for such environmental interference, making it difficult to maintain high-precision detection under all-weather conditions. In summary, how to construct an online photoacoustic spectroscopy monitoring system for transformer insulating oil that integrates efficient oil-gas separation, high-precision spectral detection, multi-dimensional feature extraction, intelligent fault diagnosis and trend prediction, environmental adaptive compensation, and multi-parameter linkage closed-loop control has become an urgent technical problem to be solved in this field. Summary of the Invention

[0004] To address the problems of the prior art, this invention provides an online monitoring system for the photoacoustic spectrum of transformer insulating oil.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, an online monitoring system for the photoacoustic spectrum of transformer insulating oil, comprising: An oil-gas separation unit is used to separate dissolved gases from transformer insulating oil and transport the dissolved gases to a photoacoustic spectroscopy gas chamber. A photoacoustic spectroscopy detection unit is installed in the photoacoustic spectroscopy chamber to perform photoacoustic spectroscopy detection on the dissolved gas and obtain photoacoustic spectral signals that reflect the concentration information of various fault characteristic gases in transformer oil. The signal acquisition and preprocessing unit is used to acquire the photoacoustic spectral signal, and to amplify, filter and perform analog-to-digital conversion preprocessing on the photoacoustic spectral signal to generate preprocessed photoacoustic spectral data. The feature extraction unit is used to extract multi-dimensional feature parameters related to fault diagnosis from the preprocessed photoacoustic spectral data. The multi-dimensional feature parameters include the concentration values ​​of each fault characteristic gas, the concentration change rate calculated with a 24-hour time window, the gas component ratio characteristics, and the spectral waveform characteristics. The intelligent diagnostic and early warning unit includes: The fault diagnosis module has a built-in pre-trained fault diagnosis model, which is used to identify the fault type and assess the fault severity of the transformer based on the multi-dimensional feature parameters, and output the diagnosis results, which include the fault type and its corresponding probability value. The trend prediction module is used to predict the concentration change trend of each fault characteristic gas in the next 72 hours or 168 hours based on the time series data of the multi-dimensional characteristic parameters, and output the prediction results. The early warning decision module is used to generate a graded early warning signal based on the diagnostic results and the prediction results, combined with preset multi-level early warning thresholds; The data storage and management unit is used to store historical photoacoustic spectral data, historical diagnostic results, and early warning records; And a remote communication and display unit, used to transmit the diagnostic results, the prediction results and the graded early warning signals to a remote monitoring terminal and display them visually.

[0006] In one specific implementation of the first aspect, the fault diagnosis model is a deep learning-based multi-class diagnostic model, which includes an input layer, a feature fusion layer, at least one hidden layer, and an output layer; the input layer is used to receive the multi-dimensional feature parameters, the feature fusion layer is used to perform feature-level fusion on the multi-dimensional feature parameters, and the output layer is used to output the probability values ​​corresponding to each fault type; the fault types include at least partial discharge faults, overheating faults, spark discharge faults, and combinations thereof.

[0007] In one specific implementation of the first aspect, the fault diagnosis model adopts a graph neural network architecture. The graph neural network uses the concentration of each gas component in the multi-dimensional feature parameters as node features, and the correlation coefficients calculated by Pearson correlation analysis or Spearman correlation analysis between the concentration sequences of each gas component as edge weights. It extracts the correlation features between each gas component through graph convolution operation, and performs fault classification based on the correlation features.

[0008] In one specific embodiment of the first aspect, the feature extraction unit further includes a spectral decoupling submodule, which uses partial least squares method or independent component analysis to perform mixed gas spectral decoupling and separation on the preprocessed photoacoustic spectral data, eliminates spectral cross-interference between each fault characteristic gas, and obtains independent concentration information of each fault characteristic gas.

[0009] In one specific implementation of the first aspect, the trend prediction module uses a long short-term memory network or a gated recurrent unit network, takes the historical concentration time series of at least one fault characteristic gas among the multi-dimensional feature parameters as input, and outputs the concentration prediction value for the next 72 hours or 168 hours. The trend prediction module automatically updates the model parameters each time new photoacoustic spectral data is obtained.

[0010] In one specific implementation of the first aspect, the early warning decision module includes: The first-level early warning submodule is used to generate an attention-level early warning signal when the concentration value or concentration change rate of any fault characteristic gas exceeds a first preset threshold. The second-level early warning submodule is used to generate an alarm-level early warning signal when the probability value corresponding to the fault type identification result output by the fault diagnosis module exceeds the second preset threshold. The third-level early warning submodule is used to generate an emergency-level early warning signal when the concentration prediction value output by the trend prediction module will reach the third preset threshold within a preset time. And a comprehensive evaluation submodule, used to perform a comprehensive weighted evaluation based on the warning results of the first-level warning submodule, the second-level warning submodule and the third-level warning submodule, and output a comprehensive warning level; The first preset threshold, the second preset threshold, and the third preset threshold are set with reference to the precaution values ​​specified in DL / T 722-2014 "Guidelines for Analysis and Judgment of Dissolved Gases in Transformer Oil".

[0011] In one specific embodiment of the first aspect, the intelligent diagnosis and early warning unit further includes a separation control feedback module; the separation control feedback module compares the concentration value or concentration change rate of at least one fault characteristic gas among the multi-dimensional characteristic parameters with a preset alarm threshold; when the concentration value or concentration change rate exceeds the alarm threshold, the separation control feedback module generates an oil-gas separation control command; the oil-gas separation unit adjusts at least one parameter among gas sampling flow rate, membrane separation pressure difference, or separation duration according to the oil-gas separation control command, with the adjustment range being 10% to 50% of the current parameter value; the photoacoustic spectroscopy detection unit acquires photoacoustic spectral data again after the oil-gas separation unit completes the parameter adjustment, and feeds back the updated photoacoustic spectral data to the intelligent diagnosis and early warning unit.

[0012] In one specific embodiment of the first aspect, the intelligent diagnosis and early warning unit and the photoacoustic spectroscopy detection unit form a detection optimization closed loop; when the preliminary diagnosis result output by the fault diagnosis module indicates the existence of a specific fault type and the probability value corresponding to the fault type is between 0.4 and 0.8, the intelligent diagnosis and early warning unit generates a detection parameter adjustment instruction; the photoacoustic spectroscopy detection unit adjusts at least one parameter among the wavelength output, scanning accuracy, or integration time of the laser source according to the detection parameter adjustment instruction, and performs secondary detection on at least one fault characteristic gas associated with the specific fault type, wherein the scanning accuracy is increased to 1.5 to 3 times the original scanning accuracy, and the integration time is extended to 2 to 5 times the original integration time; the photoacoustic spectral data obtained from the secondary detection is fed back to the fault diagnosis module for diagnostic result verification and correction.

[0013] In one specific embodiment of the first aspect, the system further includes a temperature compensation unit and a pressure compensation unit; the temperature compensation unit is used to monitor the temperature inside the photoacoustic spectral chamber in real time and generate temperature data, and the pressure compensation unit is used to monitor the pressure inside the photoacoustic spectral chamber in real time and generate pressure data; the temperature data and the pressure data are respectively transmitted to the feature extraction unit and the intelligent diagnosis and early warning unit; the feature extraction unit performs dynamic compensation and correction on the photoacoustic spectral signal based on the temperature data and the pressure data; the intelligent diagnosis and early warning unit normalizes the multi-dimensional feature parameters based on the temperature data and the pressure data and then inputs them into the fault diagnosis model; the intelligent diagnosis and early warning unit updates the temperature compensation coefficient and the pressure compensation coefficient online using the least squares method based on the compensated and corrected spectral data and the corresponding temperature data and pressure data.

[0014] Secondly, a method for online monitoring and intelligent diagnosis of transformer insulation oil using an online optical-acoustic spectroscopy monitoring system includes the following steps: Step S1: Dissolved gases are separated from transformer insulating oil by an oil-gas separation unit, and the dissolved gases are subjected to photoacoustic spectral detection by a photoacoustic spectral detection unit to obtain photoacoustic spectral signals that reflect the concentration information of various fault characteristic gases in transformer oil. Step S2: The photoacoustic spectral signal is amplified, filtered, and preprocessed by analog-to-digital conversion through the signal acquisition and preprocessing unit to generate preprocessed photoacoustic spectral data; Step S3: Extract multi-dimensional feature parameters related to fault diagnosis from the preprocessed photoacoustic spectral data through the feature extraction unit. The multi-dimensional feature parameters include the concentration values ​​of each fault characteristic gas, the concentration change rate calculated with a 24-hour time window, the gas component ratio characteristics, and the spectral waveform characteristics. Step S4: The multi-dimensional feature parameters are input into the pre-trained fault diagnosis model through the intelligent diagnosis and early warning unit to identify the fault type and assess the fault severity, and output the diagnosis result, which includes the fault type and its corresponding probability value. Step S5: The intelligent diagnosis and early warning unit uses the time series data of the multi-dimensional feature parameters to predict the concentration change trend of each fault characteristic gas in the next 72 hours or 168 hours through a trend prediction model, and outputs the prediction results. Step S6: The intelligent diagnosis and early warning unit generates a graded early warning signal and outputs it to the remote monitoring terminal based on the diagnosis result and the prediction result, combined with the preset multi-level early warning threshold. Step S7: Based on the fault type and its corresponding probability value in the diagnostic results and / or the concentration value or concentration change rate in the multi-dimensional feature parameters, generate control commands and feed them back to at least one execution unit among the oil-gas separation unit, the photoacoustic spectroscopy detection unit, the temperature compensation unit, or the pressure compensation unit, so as to adjust the working parameters of the corresponding units; Step S8: Obtain the photoacoustic spectral data collected again by the photoacoustic spectral detection unit after the execution unit completes parameter adjustment, and repeat steps S2 to S6 to verify or correct the previous diagnostic results.

[0015] The beneficial effects of this invention are as follows: 1. This invention significantly improves the accuracy and reliability of transformer fault identification by constructing a collaborative architecture of multi-dimensional feature parameter extraction and deep learning diagnostic models. Specifically, the feature extraction unit simultaneously acquires the concentration values, concentration change rates, gas component ratios, and spectral waveforms of each fault characteristic gas, forming a multi-dimensional feature space encompassing gas content information, dynamic change information, and component correlation information. Based on this, the spectral decoupling submodule uses chemometric methods to decouple and separate the spectra of the mixed gas, effectively eliminating spectral cross-interference between fault characteristic gases and ensuring the purity and accuracy of the input features. Furthermore, the fault diagnosis module employs a graph neural network architecture, constructing a correlation graph structure with the concentration of each gas component as node features and the correlation coefficient between gas components as edge weights. Through graph convolution operations, it fully explores the potential correlations between each gas component, achieving accurate identification and severity assessment of transformer internal fault types (including partial discharge faults, overheating faults, spark discharge faults, and their combinations). Simultaneously, the trend prediction module uses a long short-term memory network or a gated recurrent unit network to model the temporal evolution of characteristic gas concentrations, enabling early prediction of gas concentration change trends. The organic combination of the above-mentioned technical means enables this invention not only to accurately diagnose the current equipment status but also to predict the future evolution direction of the status. This forms a complete intelligent analysis chain from "feature extraction, spectral decoupling, graph neural network classification, and time-series trend prediction." Compared to existing technologies that rely on single gas concentration thresholds or traditional shallow machine learning models, this invention significantly reduces the false alarm rate and false negative rate, and improves the reliability and timeliness of diagnostic results. Furthermore, the temperature compensation unit and pressure compensation unit monitor the environmental parameters within the photoacoustic spectral chamber in real time. The feature extraction unit dynamically compensates and corrects the photoacoustic spectral signal based on temperature and pressure data. The intelligent diagnosis and early warning unit normalizes multi-dimensional feature parameters based on temperature and pressure data before inputting them into the fault diagnosis model. This effectively eliminates the negative impact of environmental temperature fluctuations and pressure changes on the accuracy of photoacoustic spectral detection. Simultaneously, the intelligent diagnosis and early warning unit updates the temperature compensation coefficient and pressure compensation coefficient online based on the compensated and corrected spectral data and the corresponding temperature and pressure data, enabling the compensation model to adaptively optimize with environmental conditions. This ensures the long-term stable operation of the system and the consistency of detection results under all-weather, all-season field conditions. 2. This invention achieves coordinated optimization and adaptive adjustment among various system components by constructing a multi-parameter linkage closed-loop control mechanism between the intelligent diagnosis and early warning unit, the oil-gas separation unit, and the photoacoustic spectroscopy detection unit. Specifically, the separation control feedback module dynamically generates oil-gas separation control commands based on the comparison results of the concentration value or concentration change rate of the fault characteristic gas in the multi-dimensional characteristic parameters with the alarm threshold. The oil-gas separation unit adjusts the gas sampling flow rate, membrane separation pressure difference, or separation duration accordingly. When the concentration of the fault characteristic gas abnormally increases, it actively enhances the separation efficiency and shortens the system response time. At the same time, after the separation parameters are adjusted, the photoacoustic spectroscopy detection unit re-acquires photoacoustic spectral data and feeds it back to the intelligent diagnosis and early warning unit, realizing the verification and correction of the previous diagnosis results, forming a closed-loop optimization link of "detection, diagnosis, separation enhancement, re-detection, verification and correction". Simultaneously, when the preliminary diagnostic results output by the fault diagnosis module indicate the existence of a specific fault type and the probability value is within an uncertain range, the intelligent diagnosis and early warning unit generates a detection parameter adjustment command. The photoacoustic spectroscopy detection unit adjusts the wavelength output, scanning accuracy, or integration time of the laser source accordingly, performing a secondary high-precision detection on the fault characteristic gas associated with the specific fault type. The data obtained from the secondary detection is fed back to the fault diagnosis module for diagnostic result verification and correction, achieving proactive information supplementation and improved decision confidence during the suspected diagnostic stage. Furthermore, the early warning decision module integrates the fault type identification results output by the fault diagnosis module and the concentration change trend prediction results output by the trend prediction module, combining them with a three-level progressive early warning threshold for comprehensive weighted evaluation, outputting graded early warning signals. The first-level early warning triggers a attention-level warning based on the current concentration or rate of change exceeding the limit; the second-level warning triggers an alarm-level warning based on the fault diagnosis probability exceeding the limit; and the third-level warning triggers an emergency-level warning based on the trend prediction value reaching the emergency threshold within a preset future time. These three levels of warnings are independent yet comprehensively weighted, achieving comprehensive coverage from "current state abnormality alarm" to "future trend risk warning." The synergistic effect of the aforementioned multi-level early warning mechanism and linked closed-loop control elevates this invention from a traditional one-way open-loop detection mode to an intelligent closed-loop monitoring mode of "perception, diagnosis, decision-making, execution, and re-perception." Each link forms a synergistic relationship of mutual influence and optimization, significantly improving the system's adaptability, response speed, and early warning accuracy. Furthermore, the online model update module uses incremental learning to continuously optimize the fault diagnosis model, enabling it to adapt to new data distributions and fault modes over time without requiring offline retraining of the entire dataset, further enhancing the system's long-term service capability and intelligence level. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0017] Figure 2 This is a schematic diagram of the internal module composition of the intelligent diagnosis and early warning unit of the present invention.

[0018] Figure 3 This is a schematic diagram of the multi-parameter linkage closed-loop control process of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figures 1 to 3 The above describes an online monitoring system for the photoacoustic spectrum of transformer insulating oil.

[0021] I. The present invention provides an online monitoring system for the photoacoustic spectrum of transformer insulating oil, such as... Figure 1 As shown, it mainly includes an oil-gas separation unit, a photoacoustic spectroscopy detection unit, a signal acquisition and preprocessing unit, a feature extraction unit, an intelligent diagnosis and early warning unit, a temperature compensation unit, a pressure compensation unit, a data storage and management unit, and a remote communication and display unit. The intelligent diagnosis and early warning unit forms a bidirectional communication connection with the oil-gas separation unit, the photoacoustic spectroscopy detection unit, the feature extraction unit, the temperature compensation unit, and the pressure compensation unit, forming a closed-loop control system with multi-parameter linkage.

[0022] II. Specific Implementation of the Oil-Gas Separation Unit The oil-gas separation unit is used to separate dissolved gases from transformer insulating oil and deliver the separated gases to the photoacoustic spectroscopy gas chamber.

[0023] In this embodiment, the oil-gas separation unit includes an oil-gas separation membrane module and a micro gas sampling pump. The oil-gas separation membrane module uses a polytetrafluoroethylene AF-2400 hollow fiber membrane. AF-2400 is a copolymer based on perfluorodioxane, possessing extremely high free volume fraction and excellent thermal and chemical stability. After immersion in transformer oil for 10, 20, and 30 days, the mass gain is less than 1%, and the T5% decomposition temperature remains stable at 60°C. The hollow fiber membrane has an inner diameter of 0.5~1.0 mm, a wall thickness of 50~150 μm, and an effective length of 10~30 cm. The membrane module consists of 5~20 hollow fiber membranes connected in parallel to form a membrane bundle, with a total effective membrane area of ​​50~200 cm². 2 .

[0024] The specific process of oil-gas separation is as follows: Transformer insulating oil flows through the outer side of the hollow fiber membrane at a flow rate of 0.5~5 mL / min. Dissolved gas, driven by the concentration gradient, permeates through the membrane wall into the cavity on the inner side of the membrane. A micro gas sampling pump extracts the separated gas at a flow rate of 5~50 mL / min and delivers it to the photoacoustic spectroscopy gas chamber. A shut-off valve and a flow regulating valve are installed at the oil inlet of the membrane module, and a back pressure valve is installed at the oil outlet. By adjusting the oil inlet flow rate and back pressure, the pressure difference across the membrane can be controlled to be 0.05~0.3 MPa. Under normal operating conditions, the T90 response time of the oil-gas separation unit is less than 72.5 minutes.

[0025] The sampling pump speed, membrane pressure difference, and separation duration of the oil-gas separation unit are controlled by the separation control feedback module of the intelligent diagnosis and early warning unit. Specifically, when the intelligent diagnosis and early warning unit detects that the concentration value or concentration change rate of a certain fault characteristic gas exceeds a preset alarm threshold, the separation control feedback module generates an oil-gas separation control command to increase the sampling pump speed (adjustment range of 10%~50% of the current speed), increase the membrane pressure difference (adjustment range of 10%~50% of the current pressure difference), or extend the separation duration (extended to 1.5~3 times the current duration) to enhance the separation efficiency of the target gas.

[0026] III. Specific Implementation of the Photoacoustic Spectroscopy Detection Unit The photoacoustic spectroscopy detection unit is located in the photoacoustic spectroscopy chamber and is used to perform photoacoustic spectroscopy detection on the separated dissolved gas to obtain photoacoustic spectral signals that reflect the concentration information of various fault characteristic gases in transformer oil.

[0027] The photoacoustic spectroscopy cell is a one-dimensional longitudinal resonant photoacoustic cell. The cell body is made of stainless steel or aluminum alloy, and the inner wall is polished to reduce the wall adsorption effect. The resonant cavity length of the photoacoustic cell is 50~150mm, the cross-sectional radius is 5~15mm, the resonant frequency is 500~2000Hz, the quality factor Q is 20~100, and the cell constant (i.e., the photoacoustic signal intensity generated per unit laser power and per unit gas concentration) is 1000~5000 V·cm / W. The inlet of the photoacoustic cell is connected to the outlet of a micro gas sampling pump, and the outlet is connected to the exhaust gas pipeline. The operating pressure of the photoacoustic cell is 0.08~0.12MPa.

[0028] The photoacoustic spectroscopy detection unit also includes a tunable laser source, a microphone or quartz tuning fork sensor, a lock-in amplifier, and a data acquisition card.

[0029] The tunable laser source employs a distributed feedback semiconductor laser or a quantum cascade laser to emit laser beams of multiple wavelengths corresponding to the absorption peaks of each fault-specific gas. In this embodiment, for the seven main fault-specific gases (H2, CH4, C2H6, C2H4, C2H2, CO, CO2) in transformer oil, the wavelength tuning range of the tunable laser source covers the characteristic absorption spectra of each gas in the mid-infrared band (3~12μm) or the near-infrared band (1.5~2.5μm). Specifically, the detection wavelengths are 3.03μm or 1.53μm for C2H2, 3.31μm or 1.65μm for CH4, 4.61μm or 1.56μm for CO, 4.23μm or 1.58μm for CO2, and 2.12μm for H2. The linewidth of the laser source is less than 0.01cm. -1 The output power is 5~50mW, the wavelength tuning accuracy is ±0.01nm, and the wavelength stability is ±0.005nm.

[0030] The laser beam, after collimation and focusing, is incident on the photoacoustic cell, where it is absorbed by the dissolved gas and generates a photoacoustic signal. A microphone or quartz tuning fork sensor is placed in the acoustic resonant cavity of the photoacoustic cell to detect the photoacoustic signal and convert it into an electrical signal. The microphone has a sensitivity of 10~100mV / Pa, a frequency response range of 20~20000Hz, and a noise floor of less than 10nV / √Hz. The lock-in amplifier, with its reference frequency synchronized with the laser modulation frequency (1 / 2 or 1 / 4 of the photoacoustic cell resonant frequency), is used to extract the amplitude and phase information of the photoacoustic signal, with an integration time of 0.1~10s and a dynamic range greater than 100dB. The data acquisition card has a sampling rate of 100kS / s~1MS / s and a resolution of 16~24 bits, used to convert the analog signal output from the lock-in amplifier into a digital signal.

[0031] The detection parameters of the photoacoustic spectroscopy detection unit (including the wavelength output of the laser source, scanning accuracy, and integration time) are controlled by the intelligent diagnosis and early warning unit. When the preliminary diagnosis result output by the fault diagnosis module indicates the existence of a specific fault type and the probability value corresponding to the fault type is between 0.4 and 0.8, the intelligent diagnosis and early warning unit generates a detection parameter adjustment command. The photoacoustic spectroscopy detection unit performs a secondary high-precision detection on the fault characteristic gas associated with the specific fault type according to the command. The scanning accuracy is increased to 1.5 to 3 times the original scanning accuracy (i.e., the wavelength scanning step size is reduced from the conventional 0.1 nm to 0.03 to 0.07 nm), and the integration time is extended to 2 to 5 times the original integration time (i.e., extended from the conventional 0.1 to 1 s to 0.2 to 5 s).

[0032] The system's minimum detection limits meet the following requirements: 0.05-0.1 μL / L for C2H2, 15 μL / L for H2, 0.1-0.5 μL / L for CH4, 0.1-0.5 μL / L for C2H6, 0.1-0.5 μL / L for C2H4, 0.5-2 μL / L for CO, and 1-5 μL / L for CO2. The system's linear correlation coefficient R0... 2 Greater than 0.999.

[0033] IV. Specific Implementation of the Signal Acquisition and Preprocessing Unit The signal acquisition and preprocessing unit is used to acquire the photoacoustic spectral signal output by the photoacoustic spectral detection unit, and to amplify, filter and perform analog-to-digital conversion preprocessing on it.

[0034] Specifically, the signal acquisition and preprocessing unit includes a preamplifier, a bandpass filter, and an analog-to-digital converter (ADC). The preamplifier has a gain of 40–80 dB, an input impedance greater than 1 MΩ, and a noise figure less than 2 dB, and is used for primary amplification of the weak electrical signal output from the microphone. The bandpass filter covers the resonant frequency range of the photoacoustic cell and its harmonic frequencies (typically 500–2000 Hz and its harmonics), and has a stopband attenuation greater than 40 dB / decade, used to filter out noise signals outside the passband. The ADC has a sampling rate of 100 kS / s–1 MS / s and a resolution of 16–24 bits, used to convert the filtered analog signal into a digital signal, generating preprocessed photoacoustic spectral data.

[0035] V. Specific Implementation of the Feature Extraction Unit The feature extraction unit is used to extract multi-dimensional feature parameters related to fault diagnosis from the preprocessed photoacoustic spectral data.

[0036] The feature extraction unit includes a spectral decoupling submodule and a feature calculation submodule.

[0037] The spectral decoupling submodule employs partial least squares or independent component analysis to decouple and separate the mixed-gas spectra of the preprocessed photoacoustic spectral data. Specifically, seven fault characteristic gases with different concentration combinations are pre-prepared using standard gases under laboratory conditions. Photoacoustic spectral data of each standard gas mixture are collected, and a spectral-concentration correction model is constructed. The process of establishing the correction model is as follows: Let C be the concentration matrix of the seven fault characteristic gases (n×7, where n is the number of calibration samples), and S be the corresponding photoacoustic spectral signal matrix (n×m, where m is the number of spectral wavelengths). A partial least squares regression model between S and C is established: C = S × B + E Where B is the regression coefficient matrix (dimension m×7) and E is the residual matrix. Cross-validation determined the number of principal components in the partial least squares method to be 5–10. During actual system operation, the measured photoacoustic spectral signal S... x By inputting the calibration model, the independent concentration values ​​C of each fault-characteristic gas can be obtained. x For spectral cross-interference between multi-component gases, the spectral decoupling submodule eliminates it using the aforementioned multivariate correction method.

[0038] The feature calculation submodule is used to calculate multi-dimensional feature parameters from the decoupled gas concentration data, including: (1) The current concentration values ​​C_i(t) (μL / L) of each fault characteristic gas, i=1~7 correspond to H2, CH4, C2H6, C2H4, C2H2, CO, CO2 respectively; (2) The concentration change rate of each fault characteristic gas R_i(t)=[C_i(t)-C_i(t-Δt)] / Δt, where Δt=24 hours, that is, calculate the concentration change rate (μL / L / day) within 24 hours. (3) Characteristics of gas component ratios, including CH4 / H2 ratio, C2H4 / C2H6 ratio, C2H2 / C2H4 ratio and CO / CO2 ratio; (4) Spectral waveform characteristics, including peak height, peak area, full width at half maximum (FWHM), and peak position shift of each gas characteristic absorption peak.

[0039] The above multi-dimensional feature parameters constitute a 28-dimensional feature vector F=[C1~C7,R1~R7,P1~P7,W1~W7], where P1~P7 are the component ratio features of the seven gases, and W1~W7 are the spectral waveform features of the seven gases.

[0040] VI. Specific Implementation of Temperature Compensation Unit and Pressure Compensation Unit The temperature compensation unit is used to monitor the temperature inside the photoacoustic spectroscopy chamber in real time and generate temperature data T. The temperature compensation unit includes a Pt100 platinum resistance temperature sensor and a temperature signal conditioning circuit installed on the wall of the photoacoustic cell or inside the chamber. The temperature measurement range is -20~80℃, the measurement accuracy is ±0.1℃, and the response time is less than 1s.

[0041] The pressure compensation unit is used to monitor the pressure inside the photoacoustic spectroscopy chamber in real time and generate pressure data P. The pressure compensation unit includes a piezoresistive pressure sensor and a pressure signal conditioning circuit installed at the inlet of the photoacoustic cell or inside the chamber. The pressure measurement range is 0.05~0.2MPa (absolute pressure), the measurement accuracy is ±0.1%FS, and the response time is less than 0.5s.

[0042] Temperature data T and pressure data P are transmitted to the feature extraction unit and the intelligent diagnosis and early warning unit, respectively. The feature extraction unit performs dynamic compensation and correction on the photoacoustic spectral signal based on the temperature data T and pressure data P. The specific method of compensation and correction is as follows: The photoacoustic signal intensity is related to gas concentration, laser power, photoacoustic cell constant, gas temperature, and pressure as follows: S_PAS=K×C×P_laser×Q×(T0 / T)×(P / P0) Where S_PAS is the photoacoustic signal intensity, K is the proportionality constant, C is the gas concentration, P_laser is the laser power, Q is the photoacoustic cell quality factor, T is the actual temperature, T0 is the reference temperature (298.15K), P is the actual pressure, and P0 is the reference pressure (0.101325MPa).

[0043] Based on the above relationship, the feature extraction unit performs temperature-pressure compensation correction on the measured photoacoustic signal: S_corrected=S_measured×(T / T0)×(P0 / P) The intelligent diagnosis and early warning unit normalizes multi-dimensional feature parameters based on temperature data T and pressure data P before inputting them into the fault diagnosis model. The specific normalization method involves subtracting the historical mean from each feature parameter and then dividing by the historical standard deviation (Z-score standardization). Simultaneously, the intelligent diagnosis and early warning unit updates the temperature and pressure compensation coefficients online using the recursive least squares method based on the compensated and corrected spectral data and the corresponding temperature and pressure data, enabling the compensation model to adaptively optimize according to environmental conditions.

[0044] VII. Specific Implementation of the Intelligent Diagnosis and Early Warning Unit The intelligent diagnosis and early warning unit is the core innovation of this invention, which includes a fault diagnosis module, a trend prediction module, an early warning decision module, a separate control feedback module, and a model online update module.

[0045] (a) Fault Diagnosis Module The fault diagnosis module has a built-in pre-trained fault diagnosis model, which is used to identify the fault type and assess the severity of the fault in the transformer based on multi-dimensional feature parameters, and output the diagnosis results (including the fault type and its corresponding probability value).

[0046] The fault diagnosis model is a deep learning-based multi-class diagnostic model, comprising an input layer, a feature fusion layer, at least one hidden layer, and an output layer. In this embodiment, the fault diagnosis model adopts a graph neural network architecture.

[0047] The graph neural network uses the concentrations of seven gas components C1~C7 from the multi-dimensional feature parameters as node features, and the correlation coefficients ρ_ij calculated between the concentration sequences of each gas component by Pearson correlation analysis or Spearman correlation analysis as edge weights to construct a 7-node undirected weighted graph G=(V,E,W), where V={v1,v2,...,v7} is the set of gas nodes, E is the set of edges (any two nodes are connected by an edge), and W={ρ_ij} is the edge weight matrix (7×7).

[0048] The graph convolution operation in a graph neural network is as follows: H^(l+1)=σ(D^(-1 / 2)·A·D^(-1 / 2)·H^(l)·Θ^(l)) Where H^(l) is the node feature matrix of the l-th layer (dimension 7×d_l, where d_l is the feature dimension of the l-th layer), A=W+I is the adjacency matrix with self-loops (I is the identity matrix), D is the degree matrix of A (diagonal matrix, D_ii=Σ_jA_ij), Θ^(l) is the trainable weight matrix of the l-th layer (dimension d_l×d_{l+1}), and σ is the non-linear activation function (using the ReLU function).

[0049] The graph neural network consists of 3-5 graph convolutional layers, with output feature dimensions of 64, 128, 64, and 32 for each layer, respectively. Finally, a global average pooling layer aggregates the features from the seven nodes into a fixed-length feature vector. This vector then passes through 2-3 fully connected layers (each with 64-128 neurons) and a Softmax output layer, outputting the probability values ​​corresponding to each fault type. The fault types include at least: partial discharge faults, low-temperature overheating faults (<300℃), medium-temperature overheating faults (300-700℃), high-temperature overheating faults (>700℃), low-energy discharge faults, high-energy arc discharge faults, and combinations thereof.

[0050] The pre-training process of the fault diagnosis model is as follows: (1) Obtain historical photoacoustic spectral data and corresponding actual fault labels to construct a training dataset. The training dataset contains no less than 5,000 samples, covering all the above-mentioned fault types and normal operating states, with a balanced proportion of samples in each category.

[0051] (2) Preprocess and extract features from the photoacoustic spectral data in the training dataset (according to the method of the feature extraction unit above) to generate a training feature set, with each sample being a 28-dimensional feature vector.

[0052] (3) Divide the training feature set into training set, validation set and test set in a ratio of 7:1.5:1.5.

[0053] (4) The graph neural network is trained in a supervised manner using the training set. The loss function is cross-entropy loss, the optimizer is Adam, the learning rate is 0.0001~0.001, the batch size is 32~128, the training rounds are 100~500, and the patience of the early stopping strategy is 20 rounds.

[0054] (5) Use the validation set to validate the trained fault diagnosis model and perform hyperparameter tuning (including learning rate, batch size, number of network layers, hidden layer dimension, etc.).

[0055] (6) The optimized model is evaluated using a test set, and the diagnostic accuracy is not less than 92%.

[0056] (7) Deploy the optimized fault diagnosis model to the intelligent diagnosis and early warning unit.

[0057] (II) Trend Prediction Module The trend prediction module is used to predict the concentration change trend of each fault characteristic gas in the next 72 hours or 168 hours based on the time series data of multi-dimensional characteristic parameters.

[0058] The trend prediction module employs a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) network. In this embodiment, a two-layer LSTM network is used, with each layer having a hidden state dimension of 64–128. The input to the LSTM network is the historical concentration time series X = [x_{t-τ}, x_{t-τ+1}, ..., x_t] of at least one fault-feature gas, where τ is the historical time window length (72–168 hours, i.e., 3–7 days), and the time step is 1 hour. The output of the LSTM network is the predicted concentration value Y = [y_{t+1}, y_{t+2}, ..., y_{t+T}] for the next 72 or 168 hours, where T is the prediction time window length (72 or 168 hours).

[0059] The computation process of an LSTM network is as follows: f_t=σ(W_f·[h_{t-1},x_t]+b_f) (Forget Gate); i_t=σ(W_i·[h_{t-1},x_t]+b_i)(input gate) o_t=σ(W_o·[h_{t-1},x_t]+b_o) (output gate); c̃_t=tanh(W_c·[h_{t-1},x_t]+b_c)(candidate memory unit) c_t = f_t⊙c_{t-1} + i_t⊙c̃_t (memory unit update); h_t = o_t ⊙ tanh(c_t) (Hidden state output); Where σ is the Sigmoid activation function, ⊙ is the element-wise product, W_f, W_i, W_o, W_c are weight matrices, and b_f, b_i, b_o, b_c are bias terms.

[0060] The LSTM network is trained using historical monitoring data (no fewer than 2000 time-series samples), with the loss function being the mean squared error, the optimizer being Adam, and the learning rate being 0.0001~0.001. The trend prediction module automatically updates the model parameters each time new photoacoustic spectral data is obtained. The prediction accuracy requires the mean absolute percentage error of single-step prediction to be less than 3%, and the mean absolute percentage error of multi-step prediction to be less than 5%.

[0061] (III) Early Warning Decision Module The early warning decision module is used to generate graded early warning signals based on the diagnostic results of the fault diagnosis module and the prediction results of the trend prediction module, combined with preset multi-level early warning thresholds.

[0062] The early warning decision-making module includes a first-level early warning sub-module, a second-level early warning sub-module, a third-level early warning sub-module, and a comprehensive evaluation sub-module.

[0063] The first-level early warning submodule generates a warning signal when the concentration or rate of change of any fault-characteristic gas exceeds a first preset threshold. The first preset threshold is set with reference to the warning values ​​specified in DL / T 722-2014 "Guidelines for Analysis and Judgment of Dissolved Gases in Transformer Oil," as detailed below: ; The second-level early warning submodule is used to generate an alarm-level early warning signal when the probability value corresponding to the fault type identification result output by the fault diagnosis module exceeds the second preset threshold (the second preset threshold is 0.6~0.8).

[0064] The third-level early warning submodule is used to generate an emergency warning signal when the concentration prediction value output by the trend prediction module will reach the third preset threshold (the third preset threshold is 1.2 to 2.0 times the attention value of DL / T 722-2014) within a preset time (72 hours or 168 hours).

[0065] The comprehensive assessment submodule performs a weighted comprehensive assessment based on the warning results from the Level 1, Level 2, and Level 3 warning submodules, and outputs the comprehensive warning level. The weighting rules for the comprehensive assessment are as follows: Comprehensive early warning score S = α1·S1 + α2·S2 + α3·S3 S1, S2, and S3 are the outputs (0 or 1) of the three-level early warning submodule, and α1, α2, and α3 are the corresponding weighting coefficients (α1=0.2, α2=0.3, α3=0.5). When S≥0.7, a red (emergency) warning is output; when 0.4≤S<0.7, an orange (alarm) warning is output; when 0.2≤S<0.4, a yellow (caution) warning is output; and when S<0.2, a green (normal) status is output.

[0066] (iv) Separate control feedback module The separation control feedback module is used to generate oil-gas separation control commands based on the comparison results of the concentration value or concentration change rate of at least one fault characteristic gas in the multi-dimensional characteristic parameters with the preset alarm threshold (the alarm threshold is set to 0.5 to 0.8 times the attention value of DL / T 722-2014).

[0067] When the concentration value or concentration change rate exceeds the alarm threshold, the separation control feedback module generates an oil-gas separation control command. The oil-gas separation unit adjusts at least one parameter—gas sampling flow rate, membrane separation differential pressure, or separation duration—according to this command, with the adjustment range being 10% to 50% of the current parameter value. After the oil-gas separation unit completes the parameter adjustment, the photoacoustic spectroscopy detection unit acquires photoacoustic spectral data again and feeds the updated photoacoustic spectral data back to the intelligent diagnosis and early warning unit, thereby verifying or correcting the previous diagnostic results.

[0068] (v) Online Model Update Module The online model update module is used to update the fault diagnosis model online using an incremental learning approach after obtaining new diagnostic results and corresponding actual fault labels, without the need to retrain the entire dataset.

[0069] The specific method of incremental learning is as follows: When the number of newly acquired labeled samples reaches a preset number (e.g., 100-500), the online model update module triggers the incremental update process—merging the new samples with a portion of the historical training set (e.g., the most recent 1000 samples), and fine-tuning the current model for a small number of rounds (5-20 rounds) with a small learning rate (0.1-0.5 times the initial learning rate) to update the model parameters. During the incremental update process, an elastic weight consolidation method is used to apply additional regularization constraints to important parameters to prevent catastrophic forgetting.

[0070] VIII. Specific Implementation of the Data Storage and Management Unit The data storage and management unit stores historical photoacoustic spectral data, historical diagnostic results, and early warning records. This unit includes local solid-state drive storage and cloud storage interfaces. Local storage uses a time-series database (such as InfluxDB), with the following data retention policy: raw photoacoustic spectral data is retained for 90 days, feature data and diagnostic results are retained for 3 years, and early warning records are permanently retained. The data storage capacity is no less than 1TB, and it supports compressed storage with a compression ratio of 5:1 to 10:1.

[0071] IX. Specific Implementation of the Remote Communication and Display Unit The remote communication and display unit is used to transmit diagnostic results, prediction results, and graded early warning signals to the remote monitoring terminal for visualization. The unit includes a wired communication module (supporting protocols such as Modbus TCP, IEC61850, and DL / T 860) and a wireless communication module (supporting wireless communication methods such as 4G / 5G, WiFi, and LoRa). The remote monitoring terminal is either a local monitoring workstation in the substation or a cloud-based monitoring platform, displaying real-time curves of gas concentration changes, diagnostic results, trend prediction results, and early warning status in graphical form.

[0072] 10. The complete workflow of the system of this invention is as follows: Step S1: The transformer insulating oil enters the oil-gas separation unit through the oil inlet pipeline, and the oil-gas separation is completed in the hollow fiber membrane module. The dissolved gas after separation is transported to the photoacoustic spectroscopy gas chamber by a micro gas sampling pump.

[0073] Step S2: The photoacoustic spectroscopy detection unit performs photoacoustic spectroscopy detection on the dissolved gas in the photoacoustic spectroscopy chamber—the tunable laser source sequentially emits laser beams corresponding to the absorption peaks of the seven fault characteristic gases. After the laser beams are absorbed by the gas, they generate photoacoustic signals. A microphone or quartz tuning fork sensor detects the photoacoustic signals and converts them into electrical signals. A lock-in amplifier extracts the amplitude and phase information of the photoacoustic signals. The data acquisition card converts the analog signals into digital signals and generates photoacoustic spectral data.

[0074] Step S3: The signal acquisition and preprocessing unit amplifies, filters, and performs analog-to-digital conversion preprocessing on the photoacoustic spectral data to generate preprocessed photoacoustic spectral data.

[0075] Step S4: The feature extraction unit extracts multi-dimensional feature parameters (28-dimensional feature vectors) from the preprocessed photoacoustic spectral data. The spectral decoupling submodule uses partial least squares to eliminate spectral cross-interference between gases and obtains the independent concentration information of each gas.

[0076] Step S5: The temperature compensation unit and pressure compensation unit monitor the temperature and pressure inside the photoacoustic spectroscopy chamber in real time, and transmit the temperature data T and pressure data P to the feature extraction unit and the intelligent diagnosis and early warning unit; the feature extraction unit performs dynamic compensation and correction on the photoacoustic spectroscopy signal based on T and P; the intelligent diagnosis and early warning unit performs normalization processing on the multi-dimensional feature parameters based on T and P.

[0077] Step S6: The intelligent diagnosis and early warning unit inputs the normalized multi-dimensional feature parameters into the graph neural network model of the fault diagnosis module to identify the fault type and assess the severity of the fault, and outputs the diagnosis results (fault type and its probability value).

[0078] Step S7: The trend prediction module of the intelligent diagnosis and early warning unit predicts the concentration change trend of each fault characteristic gas in the next 72 hours or 168 hours based on the time series data of multi-dimensional characteristic parameters through an LSTM network, and outputs the prediction results.

[0079] Step S8: The early warning decision module of the intelligent diagnosis and early warning unit generates a graded early warning signal based on the diagnosis results and prediction results, combined with the multi-level early warning thresholds specified in DL / T 722-2014.

[0080] Step S9: The remote communication and display unit transmits the diagnostic results, prediction results, and graded early warning signals to the remote monitoring terminal and displays them visually; the data storage and management unit stores historical data.

[0081] Step S10: The separation control feedback module of the intelligent diagnosis and early warning unit generates control commands and feeds them back to the oil-gas separation unit based on the comparison results of the concentration value or concentration change rate in the multi-dimensional characteristic parameters with the alarm threshold. When the concentration value or concentration change rate exceeds the alarm threshold, the sampling pump speed is increased, the pressure difference across the membrane is increased, or the separation duration is extended (the adjustment range is 10% to 50% of the current parameter value).

[0082] Step S11: After the oil-gas separation unit completes parameter adjustment, the photoacoustic spectroscopy detection unit acquires photoacoustic spectral data again. The system repeats steps S3 to S8 to verify or correct the previous diagnostic results.

[0083] Step S12: After obtaining new labeled samples, the online model update module of the intelligent diagnosis and early warning unit updates the fault diagnosis model online using an incremental learning method.

[0084] The steps S1 to S9 constitute the routine monitoring process, steps S10 to S11 constitute the linkage feedback control process, and step S12 constitutes the model self-evolution process. The three processes are coupled and operate in coordination to form a complete intelligent monitoring closed loop of "perception-diagnosis-decision-execution-re-perception".

[0085] Example 1: This example uses a 110kV oil-immersed power transformer as the monitoring object and employs the online monitoring system for the optical and acoustic spectra of transformer insulation oil provided by this invention to conduct continuous online monitoring for 6 months.

[0086] 1. System configuration: The oil-gas separation unit uses AF-2400 hollow fiber membrane modules (effective membrane area 100cm²). 2 The pressure difference across the membrane was 0.15 MPa, the sampling pump flow rate was 20 mL / min, and the T90 response time was 65 minutes.

[0087] The photoacoustic spectroscopy detection unit adopts a one-dimensional longitudinal resonant photoacoustic cell (resonant cavity length 100mm, cross-sectional radius 10mm, resonant frequency 850Hz, quality factor Q=50), the tunable laser source adopts a quantum cascade laser (wavelength tuning range 3~12μm, output power 20mW), the microphone sensitivity is 50mV / Pa, the lock-in amplifier integration time is 1s, and the data acquisition card has a sampling rate of 500kS / s and a resolution of 24 bits.

[0088] The feature extraction unit uses partial least squares for spectral decoupling, with 7 principal components and a root mean square error of less than 0.05 μL / L for cross-validation.

[0089] The fault diagnosis module uses a 3-layer graph convolutional neural network (each layer has output feature dimensions of 64, 128, and 64, respectively). The pre-training uses 6,000 sets of historical samples (covering seven states: normal, partial discharge, low temperature overheating, medium temperature overheating, high temperature overheating, low energy discharge, and high energy arc discharge). The ratio of training set, validation set, and test set is 7:1.5:1.5, the training rounds are 300, the learning rate is 0.0005, and the diagnostic accuracy on the test set is 94.2%.

[0090] The trend prediction module uses a two-layer LSTM network (with a hidden layer dimension of 128), with a historical time window of 168 hours (7 days) and a prediction time window of 72 hours (3 days). The training uses 2500 sets of time series samples. The mean absolute percentage error for single-step prediction is 2.1%, and the mean absolute percentage error for multi-step prediction is 3.8%.

[0091] The warning thresholds are set according to DL / T 722-2014 standard: C2H2 warning value 5 μL / L (concentration change rate warning value 0.5 μL / L / day), H2 warning value 150 μL / L, and total hydrocarbon warning value 150 μL / L. The second-level warning threshold (probability threshold) is set at 0.7. The third-level warning threshold is set at 1.5 times the warning value of DL / T 722-2014.

[0092] 2. Monitoring process: For the first three months after the system was put into operation, the concentrations of each fault-related gas were within the normal range (H2 concentration 5~25μL / L, CH4 concentration 1~8μL / L, C2H6 concentration 0.5~3μL / L, C2H4 concentration 0.5~2μL / L, C2H2 concentration less than 0.1μL / L, CO concentration 50~150μL / L, CO2 concentration 1000~3000μL / L), and the system output a green (normal) status every day.

[0093] On the 15th day of the 4th month after commissioning, the feature extraction unit detected that the C2H2 concentration increased from 0.08 μL / L to 0.35 μL / L (24-hour change rate of 0.27 μL / L / day). Although it did not exceed the attention value of 5 μL / L specified in DL / T 722-2014, it exceeded the alarm threshold (0.5 times the attention value, i.e., 0.5 times 2.5 μL / L is 1.25 μL / L - here the actual C2H2 concentration was 0.35 μL / L, which did not exceed the alarm threshold; but the C2H2 concentration change rate of 0.27 μL / L / day exceeded the concentration change rate alarm threshold of 0.25 μL / L / day).

[0094] After the separation control feedback module detects that the C2H2 concentration change rate exceeds the alarm threshold, it generates an oil-gas separation control command, increasing the sampling pump flow rate from 20 mL / min to 30 mL / min (adjustment range 50%) and the pressure difference across the membrane from 0.15 MPa to 0.20 MPa (adjustment range 33%). After parameter adjustment, the photoacoustic spectroscopy detection unit acquires photoacoustic spectral data again, and the C2H2 concentration measurement value is corrected to 0.42 μL / L (correcting for the measurement deviation introduced by the change in separation efficiency).

[0095] The fault diagnosis module inputs multi-dimensional feature parameters into the graph neural network model and outputs the diagnosis results: the probability of low-energy discharge fault is 0.73, the probability of partial discharge fault is 0.18, and the probability of normal state is 0.09. Since the probability of low-energy discharge fault (0.73) exceeds the second-level warning threshold (0.7), the second-level warning submodule generates an alarm-level warning signal.

[0096] The trend prediction module predicts that the C2H2 concentration will rise to 1.2~1.8 μL / L in the next 72 hours. Although it has not reached the third-level warning threshold (7.5 μL / L, which is 1.5 times 5 μL / L), the upward trend is obvious.

[0097] The comprehensive evaluation submodule calculates the comprehensive early warning score S=0.2×0+0.3×1+0.5×0=0.3 and outputs a yellow (caution) warning.

[0098] The remote communication and display unit transmitted the diagnostic results (low-energy discharge fault, probability 73%), trend prediction results (C2H2 concentration will rise to 1.2~1.8μL / L within 72 hours), and a yellow warning signal to the substation monitoring center. Based on the warning information, maintenance personnel conducted online partial discharge detection on the transformer, confirmed the presence of a low-energy partial discharge fault, and promptly arranged for maintenance, preventing further escalation of the fault.

[0099] Six months after commissioning, the online model update module collected 320 new labeled samples, triggering an incremental update process. This process merged the new samples with the 1000 most recent samples in the historical training set, and fine-tuned the graph neural network for 10 rounds at a learning rate of 0.0001 (20% of the initial learning rate of 0.0005). After the incremental update, the model's diagnostic accuracy for low-energy discharge faults improved from 91.5% to 94.8%.

[0100] During six months of continuous operation, the system in this embodiment did not experience any false alarms or missed alarms, issuing two yellow alerts and one orange alert, all of which were consistent with the actual equipment status. Compared with offline detection using traditional gas chromatography during the same period, the detection response time of this system was shortened from an average of 8 hours to 1.5 hours (including oil and gas separation time), and the fault detection time was shortened by 5 to 7 days. Compared with traditional photoacoustic spectroscopy monitoring systems, this system, through dynamic temperature-pressure compensation, improved detection accuracy by approximately 35% under all-weather environmental changes (when the temperature changes from -10℃ to 40℃, the detection deviation of the traditional system is ±8%, while the detection deviation of this system is ±3%); through dynamic control of oil and gas separation, the response time (T90) when the characteristic gas concentration is abnormal was shortened from 65 minutes to 42 minutes; through multi-dimensional feature extraction and graph neural network diagnosis, the fault identification accuracy was improved from 82%~87% of the traditional method to over 94%.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online monitoring system for the photoacoustic spectrum of transformer insulating oil, characterized in that, include: An oil-gas separation unit is used to separate dissolved gases from transformer insulating oil and transport the dissolved gases to a photoacoustic spectroscopy gas chamber. A photoacoustic spectroscopy detection unit is installed in the photoacoustic spectroscopy chamber to perform photoacoustic spectroscopy detection on the dissolved gas and obtain photoacoustic spectral signals that reflect the concentration information of various fault characteristic gases in transformer oil. The signal acquisition and preprocessing unit is used to acquire the photoacoustic spectral signal, and to amplify, filter and perform analog-to-digital conversion preprocessing on the photoacoustic spectral signal to generate preprocessed photoacoustic spectral data. The feature extraction unit is used to extract multi-dimensional feature parameters related to fault diagnosis from the preprocessed photoacoustic spectral data. The multi-dimensional feature parameters include the concentration values ​​of each fault characteristic gas, the concentration change rate calculated with a 24-hour time window, the gas component ratio characteristics, and the spectral waveform characteristics. The intelligent diagnostic and early warning unit includes: The fault diagnosis module has a built-in pre-trained fault diagnosis model, which is used to identify the fault type and assess the fault severity of the transformer based on the multi-dimensional feature parameters, and output the diagnosis results, which include the fault type and its corresponding probability value. The trend prediction module is used to predict the concentration change trend of each fault characteristic gas in the next 72 hours or 168 hours based on the time series data of the multi-dimensional characteristic parameters, and output the prediction results. The early warning decision module is used to generate a graded early warning signal based on the diagnostic results and the prediction results, combined with preset multi-level early warning thresholds; The data storage and management unit is used to store historical photoacoustic spectral data, historical diagnostic results, and early warning records; And a remote communication and display unit, used to transmit the diagnostic results, the prediction results and the graded early warning signals to a remote monitoring terminal and display them visually.

2. The online monitoring system according to claim 1, characterized in that, The fault diagnosis model is a deep learning-based multi-class diagnostic model, which includes an input layer, a feature fusion layer, at least one hidden layer, and an output layer. The input layer is used to receive the multi-dimensional feature parameters, the feature fusion layer is used to perform feature-level fusion on the multi-dimensional feature parameters, and the output layer is used to output the probability values ​​corresponding to each fault type. The fault types include at least partial discharge faults, overheating faults, spark discharge faults, and combinations thereof.

3. The online monitoring system according to claim 2, characterized in that, The fault diagnosis model adopts a graph neural network architecture. The graph neural network uses the concentration of each gas component in the multi-dimensional feature parameters as node features and the correlation coefficients calculated by Pearson correlation analysis or Spearman correlation analysis between the concentration sequences of each gas component as edge weights. It extracts the correlation features between each gas component through graph convolution operation and performs fault classification based on the correlation features.

4. The online monitoring system according to claim 1, characterized in that, The feature extraction unit also includes a spectral decoupling submodule, which uses partial least squares or independent component analysis to perform mixed gas spectral decoupling and separation on the preprocessed photoacoustic spectral data, eliminates spectral cross-interference between each fault characteristic gas, and obtains independent concentration information of each fault characteristic gas.

5. The online monitoring system according to claim 1, characterized in that, The trend prediction module uses a long short-term memory network or a gated recurrent unit network. It takes the historical concentration time series of at least one fault characteristic gas in the multi-dimensional feature parameters as input and outputs the concentration prediction value for the next 72 hours or 168 hours. The trend prediction module automatically updates the model parameters each time new photoacoustic spectral data is obtained.

6. The online monitoring system according to claim 1, characterized in that, The early warning decision module includes: The first-level early warning submodule is used to generate an attention-level early warning signal when the concentration value or concentration change rate of any fault characteristic gas exceeds a first preset threshold. The second-level early warning submodule is used to generate an alarm-level early warning signal when the probability value corresponding to the fault type identification result output by the fault diagnosis module exceeds the second preset threshold. The third-level early warning submodule is used to generate an emergency-level early warning signal when the concentration prediction value output by the trend prediction module will reach the third preset threshold within a preset time. And a comprehensive evaluation submodule, used to perform a comprehensive weighted evaluation based on the warning results of the first-level warning submodule, the second-level warning submodule and the third-level warning submodule, and output a comprehensive warning level; The first preset threshold, the second preset threshold, and the third preset threshold are set with reference to the precaution values ​​specified in DL / T 722-2014 "Guidelines for Analysis and Judgment of Dissolved Gases in Transformer Oil".

7. The online monitoring system according to claim 1, characterized in that, The intelligent diagnosis and early warning unit also includes a separation control feedback module; the separation control feedback module compares the concentration value or concentration change rate of at least one fault characteristic gas among the multi-dimensional characteristic parameters with a preset alarm threshold. When the concentration value or concentration change rate exceeds the alarm threshold, the separation control feedback module generates an oil-gas separation control command; the oil-gas separation unit adjusts at least one parameter among gas sampling flow rate, membrane separation pressure difference, or separation duration according to the oil-gas separation control command, with the adjustment range being 10% to 50% of the current parameter value; the photoacoustic spectroscopy detection unit acquires photoacoustic spectral data again after the oil-gas separation unit completes the parameter adjustment, and feeds back the updated photoacoustic spectral data to the intelligent diagnosis and early warning unit.

8. The online monitoring system according to claim 1, characterized in that, The intelligent diagnosis and early warning unit and the photoacoustic spectroscopy detection unit form a detection optimization closed loop. When the preliminary diagnosis result output by the fault diagnosis module indicates the existence of a specific fault type and the probability value corresponding to the fault type is between 0.4 and 0.8, the intelligent diagnosis and early warning unit generates a detection parameter adjustment command. The photoacoustic spectroscopy detection unit adjusts at least one parameter among the wavelength output, scanning accuracy, or integration time of the laser source according to the detection parameter adjustment command, and performs secondary detection on at least one fault characteristic gas associated with the specific fault type, wherein the scanning accuracy is increased to 1.5 to 3 times the original scanning accuracy, and the integration time is extended to 2 to 5 times the original integration time. The photoacoustic spectral data obtained from the secondary detection is fed back to the fault diagnosis module for diagnostic result verification and correction.

9. The online monitoring system according to claim 1, characterized in that, The system further includes a temperature compensation unit and a pressure compensation unit; the temperature compensation unit is used to monitor the temperature inside the photoacoustic spectroscopy chamber in real time and generate temperature data, and the pressure compensation unit is used to monitor the pressure inside the photoacoustic spectroscopy chamber in real time and generate pressure data; the temperature data and the pressure data are respectively transmitted to the feature extraction unit and the intelligent diagnosis and early warning unit. The feature extraction unit performs dynamic compensation and correction on the photoacoustic spectral signal based on the temperature data and the pressure data; the intelligent diagnosis and early warning unit normalizes the multi-dimensional feature parameters based on the temperature data and the pressure data and then inputs them into the fault diagnosis model; the intelligent diagnosis and early warning unit updates the temperature compensation coefficient and pressure compensation coefficient online using the least squares method based on the compensated and corrected spectral data and the corresponding temperature data and pressure data.

10. A method for online monitoring and intelligent diagnosis of transformer insulating oil using the photoacoustic spectrum of claim 1, characterized in that, Includes the following steps: Step S1: Dissolved gases are separated from transformer insulating oil by an oil-gas separation unit, and the dissolved gases are subjected to photoacoustic spectral detection by a photoacoustic spectral detection unit to obtain photoacoustic spectral signals that reflect the concentration information of various fault characteristic gases in transformer oil. Step S2: The photoacoustic spectral signal is amplified, filtered, and preprocessed by analog-to-digital conversion through the signal acquisition and preprocessing unit to generate preprocessed photoacoustic spectral data; Step S3: Extract multi-dimensional feature parameters related to fault diagnosis from the preprocessed photoacoustic spectral data through the feature extraction unit. The multi-dimensional feature parameters include the concentration values ​​of each fault characteristic gas, the concentration change rate calculated with a 24-hour time window, the gas component ratio characteristics, and the spectral waveform characteristics. Step S4: The multi-dimensional feature parameters are input into the pre-trained fault diagnosis model through the intelligent diagnosis and early warning unit to identify the fault type and assess the fault severity, and output the diagnosis result, which includes the fault type and its corresponding probability value. Step S5: The intelligent diagnosis and early warning unit uses the time series data of the multi-dimensional feature parameters to predict the concentration change trend of each fault characteristic gas in the next 72 hours or 168 hours through a trend prediction model, and outputs the prediction results. Step S6: The intelligent diagnosis and early warning unit generates a graded early warning signal and outputs it to the remote monitoring terminal based on the diagnosis result and the prediction result, combined with the preset multi-level early warning threshold. Step S7: Based on the fault type and its corresponding probability value in the diagnostic results and / or the concentration value or concentration change rate in the multi-dimensional feature parameters, generate control commands and feed them back to at least one execution unit among the oil-gas separation unit, the photoacoustic spectroscopy detection unit, the temperature compensation unit, or the pressure compensation unit, so as to adjust the working parameters of the corresponding units; Step S8: Obtain the photoacoustic spectral data collected again by the photoacoustic spectral detection unit after the execution unit completes parameter adjustment, and repeat steps S2 to S6 to verify or correct the previous diagnostic results.

Images