Micro-water content detection system for transformer insulating oil

Abstract

The invention provides a micro-water content detection system for transformer insulating oil, and belongs to the technical field of transformer insulating oil humidity detection.The system comprises a controller, an oil inlet, a detection pipeline and an oil outlet, the oil inlet, the detection pipeline and the oil outlet are sequentially connected, the oil inlet and the oil outlet are used for being connected to an insulating oil external circulation loop of a transformer, and the detection pipeline is a transparent pipe; a plurality of cylindrical hollow channels are arranged in the transparent pipe, the hollow channels are uniformly distributed in the transparent pipe, two electrodes are symmetrically arranged in the middle of each hollow channel, and the two electrodes, the direct-current power supply and the counter form a counting unit. If the insulating oil flowing through the two electrodes enables instantaneous conduction to be formed between the two electrodes, a counter is triggered to count; and a micro-water content calculation module is arranged in the controller and is used for calculating the micro-water content of the insulating oil in the transformer. The problem that the transformer insulating oil cannot be monitored online in real time in the prior art is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of transformer insulating oil humidity detection, and in particular relates to a transformer insulating oil trace water content detection system. Background Art

[0002] Transformers are very important equipment in power systems, and the state of their insulating oil is directly related to the safe operation of transformers. The trace moisture content in transformer insulating oil is one of the key indicators for evaluating the insulation state. Excessive moisture content will accelerate the aging of insulating materials, thereby reducing the insulation strength of the transformer and increasing the risk of equipment failure. Therefore, accurate detection of the trace moisture content in transformer insulating oil is of great significance for preventing insulation failure accidents and extending the service life of transformers. There are two main existing technologies for detecting the trace moisture content of insulating oil: one is to use the capacitance method, which uses the change in the dielectric constant of insulating oil to reflect the moisture content; the other is to use chemical analysis methods, such as Karl Fischer titration. Although the capacitance method can realize online detection, due to the influence of factors such as impurities and temperature in the insulating oil, the detection accuracy is low, and it is difficult to meet the requirements for accurate monitoring of trace moisture content. Although the chemical analysis method has high detection accuracy, it requires regular sampling and laboratory analysis, and cannot realize real-time online detection and cannot timely reflect the dynamic changes of the insulating oil state.

[0003] With the development of intelligent power system, higher requirements are put forward for real-time monitoring and early warning analysis of transformer status. Therefore, it is urgent to develop a new detection technology that can realize high-precision micro-water content detection and real-time online monitoring. Summary of the invention

[0004] In view of this, the present invention provides a system for detecting the trace water content of transformer insulating oil, which can solve the problem that the trace water content in the existing transformer insulating oil cannot be monitored online in real time.

[0005] The present invention is achieved in that:

[0006] The present invention provides a system for detecting the micro-water content of transformer insulating oil, comprising a controller and an oil inlet, a detection pipeline, and an oil outlet connected in sequence, wherein the oil inlet and the oil outlet are used to be connected to an external circulation loop of the insulating oil of the transformer, the detection pipeline is a transparent tube of insulating material whose two ends are respectively connected to the oil inlet and the oil outlet, a plurality of cylindrical hollow channels are arranged in the transparent tube, which are used to connect the oil inlet and the oil outlet, the hollow channels are coaxially and evenly distributed in the transparent tube, and two electrodes in contact with the channel are symmetrically arranged in the middle position of the hollow channel, which are a first electrode and a second electrode, the first electrode is connected to the first pole of a DC power supply, the The second pole of the DC power supply is connected to the first input end of the counter, the second input end of the counter is connected to the second electrode, the output end of the counter is connected to the data input end of the controller, the two electrodes, the DC power supply and the counter form a counting unit, wherein all the counting units share a DC power supply, when the insulating oil flows into the detection pipeline through the oil inlet, if the insulating oil flowing through the two electrodes forms an instantaneous conduction between the two electrodes, the counter is triggered to count; a micro-water content calculation module is provided in the controller, which is used to calculate the micro-water content of the insulating oil in the transformer according to the counting data of all counting units at each collection moment obtained in real time.

[0007] The first electrode is a positive electrode and the second electrode is a negative electrode; the acquisition frequency range of the counter is specifically above 100 Hz; the tip distance of the two electrodes is a sequence distance, and the range is specifically 0.1 mm-10 mm; the step of determining the sequence distance of the tip distance of the two electrodes specifically includes starting from 0.1 mm and increasing in steps of 0.1 mm until 10 mm, forming a sequence distance set.

[0008] Optionally, the two electrodes are in the shape of relative arcs, rectangles or spheres.

[0009] Wherein, the micro-water content calculation module specifically performs the following steps:

[0010] S10, acquiring trigger data collected by a detection system for micro-water content in insulating oil in real time;

[0011] S20, establishing a trigger curve based on the trigger data according to the acquisition time;

[0012] S30, inputting the trigger curve into a pre-trained water content calculation model, inputting the trigger curve, and obtaining a water content detection result of the insulating oil in the transformer.

[0013] Furthermore, the data is collected every 0.02 seconds to 1 second.

[0014] Furthermore, the trigger data is specifically the minimum value of the distance between the tips of two electrodes in the counting unit that generates counts at each acquisition moment.

[0015] Furthermore, the water content calculation model adopts an embedded multi-branch neural network structure, including a fusion Fourier equation group module, a signal processing module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network and a fusion network.

[0016] Furthermore, the fusion Fourier equations module is used to input the trigger curve into the improved Fourier equations to calculate the spectrum analysis parameters, including amplitude spectrum, phase spectrum, and power spectrum density;

[0017] The signal processing module uses a sliding window method to extract features from the spectrum analysis parameters calculated by the fusion Fourier equation group module to obtain long-term features, short-term features and abnormal features;

[0018] The long-term feature branch network is used to obtain first water content data through a deep learning network according to the long-term feature;

[0019] The short-term feature branch network is used to obtain second water content data through a convolutional neural network according to the short-term features;

[0020] The abnormal feature branch network is used to obtain third water content data through a recurrent neural network according to the abnormal feature;

[0021] The fusion network is used to obtain the moisture content data result by fusing and calculating the first moisture content data, the second moisture content data and the third moisture content data using the domain correlation of the long-term characteristics, the short-term characteristics and the abnormal characteristics.

[0022] Furthermore, the fused Fourier equations include Fourier transform equations, inverse Fourier transform equations, short-time Fourier transform equations, wavelet transform equations, and Hilbert transform equations;

[0023] The Fourier transform equation is used to convert the time domain signal into a frequency domain representation, the input is the time domain trigger curve, and the output is the spectrum;

[0024] The inverse Fourier transform equation is used to convert the frequency domain signal back to the time domain representation, the input is the spectrum, and the output is the reconstructed time domain signal;

[0025] The short-time Fourier transform equation is used to analyze the time-frequency characteristics of non-stationary signals, the input is a time-domain trigger curve, and the output is a time-frequency spectrum;

[0026] The wavelet transform equation is used for multi-scale analysis of local features of the signal, the input is the time domain trigger curve, and the output is the wavelet coefficient;

[0027] The Hilbert transform equation is used to calculate the instantaneous frequency and envelope of the signal, the input is the time domain trigger curve, and the output is the analytical signal.

[0028] During the training of the water content calculation model, the step of establishing a training data set specifically includes:

[0029] Collect transformer insulating oil samples under different working conditions;

[0030] The water content of the samples was measured using the standard Karl Fischer titration method;

[0031] Use this detection system to collect trigger curve data of samples;

[0032] Perform improved Fourier transform preprocessing on the trigger curve data;

[0033] Pair the processed spectrum data with standard moisture content results;

[0034] Divide into training set, validation set and test set.

[0035] The step of training the water content calculation model specifically includes:

[0036] Step 1: Initialize network parameters;

[0037] Step 2, forward propagation: Process the input data by improving the Fourier equations, and then calculate and predict the moisture content through the neural network;

[0038] Step 3: Calculate the mean square error loss between the predicted value and the true value;

[0039] Step 4: Back propagation updates the network parameters, including improving the learnable parameters of the Fourier equations;

[0040] Step 5: Repeat steps 2-4 until the loss function converges;

[0041] Step 6: Use the validation set to evaluate model performance and adjust hyperparameters.

[0042] Step 7: Use the test set for final evaluation and optimize parameters.

[0043] The inner diameter of each cylindrical hollow channel is equal.

[0044] The sum of the cross-sectional areas of all cylindrical hollow channels is not less than the cross-sectional area of ​​the oil inlet.

[0045] Furthermore, the sum of the cross-sectional areas of all cylindrical hollow channels is 1.68 times the cross-sectional area of ​​the oil inlet.

[0046] Specifically, the step S10 includes:

[0047] Step 101: Monitor in real time the transient conduction events caused by the insulating oil contacting the electrodes during the flow of the insulating oil through multiple counting units arranged in the detection pipeline. Each counting unit is composed of two electrodes, a DC power supply and a counter. When the insulating oil flows through the two electrodes, it will be instantly turned on, thereby triggering the counter to count.

[0048] Step 102: The controller collects the counting data of each counting unit in real time and stores it as trigger data. The collection frequency can be set between 0.02 seconds and 1 second, for example, 0.1 second can be selected to obtain trigger data with sufficiently high time resolution.

[0049] Step 103: The controller temporarily stores all the collected trigger data to provide raw data support for subsequent trigger curve establishment and water content calculation.

[0050] Specifically, the step S20 includes:

[0051] Step 201: The controller arranges the trigger data at each moment acquired in step S10 into a time series according to the acquisition time sequence, so as to form a trigger curve in the process of detecting the micro-water content of insulating oil.

[0052] Step 202: In order to better describe the time characteristics of the trigger curve, the controller can perform linear interpolation processing on the trigger data so that the trigger curve presents a smooth and continuous state, and better reflects the dynamic characteristics of the micro-water content of the insulating oil changing with time.

[0053] Step 203: The controller saves the established trigger curve data for subsequent water content calculation.

[0054] Specifically, the step S30 includes:

[0055] Step 301: The controller sends the trigger curve established in step S20 as input data to the pre-trained water content calculation model. The model adopts an embedded multi-branch neural network structure, which is composed of a fusion Fourier equation module, a signal processing module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network, and a final fusion network.

[0056] Step 302, the fusion Fourier equations module performs a series of frequency domain analyses such as Fourier transform, inverse Fourier transform, short-time Fourier transform, wavelet transform and Hilbert transform on the input trigger curve to extract frequency spectrum characteristic parameters such as amplitude spectrum, phase spectrum and power spectrum density.

[0057] 5. The method according to claim 1 is characterized in that, in the process of model training, a training data set needs to be established first. Specifically, it includes: collecting transformer insulating oil samples under different working conditions, measuring the water content of the samples using the standard Karl Fischer titration method, and then using the detection system to collect the trigger curve data of these samples, and performing Fourier transform preprocessing on the trigger curve data, pairing the processed spectrum data with the standard water content results, and finally dividing them into training set, verification set and test set.

[0058] In summary, the transformer insulating oil micro-water content detection system proposed by the present invention mainly includes the following features:

[0059] 1. The system structure includes an oil inlet, a detection pipeline and an oil outlet, and the insulating oil circulates through this system. The detection pipeline is made of a transparent tube, using insulating material, and has multiple evenly distributed cylindrical hollow channels inside.

[0060] 2. Two electrodes are symmetrically arranged in the middle of each channel to form a counting unit. When the insulating oil flows through the electrodes, it will be temporarily turned on, triggering the counter to count.

[0061] 3. The controller contains a micro-water content calculation module. According to the real-time counting data of each counting unit, a trigger curve is established and input into the pre-trained moisture content calculation model to obtain the final insulating oil micro-water content detection result.

[0062] 4. The moisture content calculation model adopts an embedded multi-branch neural network structure, including a fusion Fourier equation group module, a signal processing module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network and a fusion network.

[0063] 5. During the model training process, it is necessary to collect insulating oil samples under different working conditions, measure the standard water content, and collect trigger curve data. After improved Fourier transform preprocessing and pairing with the standard results, the training set, validation set, and test set are constructed.

[0064] 6. Model training includes initializing network parameters, calculating predicted values ​​through forward propagation, and updating parameters through back propagation until the loss function converges. Performance evaluation and parameter optimization are then performed through validation sets and test sets.

[0065] Compared with the prior art, the beneficial effects of the transformer insulating oil micro-water content detection system provided by the present invention are:

[0066] 1. Real-time online detection of the micro-water content of insulating oil is realized. This system uses a counting unit to monitor the triggering events in the process of insulating oil flowing in real time, and reflects the dynamic state of insulating oil through trigger curve analysis. It can realize continuous online monitoring and timely detect abnormal changes in the micro-water content of insulating oil.

[0067] 2. Greatly improved detection accuracy. This system uses a deep learning model to analyze the trigger curve and integrates the time-frequency domain features. It has outstanding performance in capturing the slow change trend of the micro-water content of insulating oil, instantaneous fluctuation characteristics, and abnormal changes, and significantly improves the detection accuracy.

[0068] 3. Good adaptability. The deep learning model has strong learning and generalization capabilities, and can adapt to the changes in the characteristics of insulating oil under different working conditions. Through continuous optimization training, the accuracy and reliability of detection can be continuously improved.

[0069] 4. Automation and intelligence are realized. This system adopts embedded design to realize fully automated micro-water content detection and analysis process without manual intervention, providing strong support for intelligent management of transformer status.

[0070] In summary, the present invention solves the problem that the existing water content in transformer insulating oil cannot be monitored online in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of the structure of a device for detecting the trace water content of transformer insulating oil provided by the present invention;

[0072] Figure 2 It is a schematic diagram of a transparent tube having multiple cylindrical hollow channels;

[0073] Figure 3 is a schematic diagram of two electrodes;

[0074] Figure 4 is a schematic diagram of a counting unit;

[0075] Figure 5 Schematic diagram of the steps executed by the micro-water content calculation module;

[0076] The various reference numerals in the accompanying drawings represent as follows: 10, detection pipeline; 11, hollow channel; 111, first electrode; 112, second electrode; 21, oil inlet; 22, oil outlet. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0078] like Figure 1-3As shown, it is a schematic diagram of the device structure of a transformer insulating oil micro-water content detection system provided by the present invention, including a controller, a DC power supply, a counter, and an oil inlet 21, a detection pipeline 10, and an oil outlet 22 connected in sequence. The oil inlet 21 and the oil outlet 22 are used to be connected to the external circulation loop of the transformer insulating oil. The detection pipeline is a transparent tube with two ends respectively connected to the oil inlet 21 and the oil outlet 22. A plurality of cylindrical hollow channels 11 are arranged in the transparent tube for connecting the oil inlet and the oil outlet. The two ends of the transparent tube seal the gaps between the plurality of hollow channels 11. For example, a multi-hole sealing pad can be used, which can support and ensure that more insulating oil flows through the hollow channels 11. The cylindrical hollow channels 11 are evenly distributed in the transparent tube, and can be made of insulating resin or glass. Two electrodes are radially symmetrically arranged in the middle of the hollow channel 11, namely a first electrode 111 and a second electrode 112. The first electrode is connected to the first pole of a DC power supply, the second pole of the DC power supply is connected to the first input end of a counter, the second input end of the counter is connected to the second electrode, and the output end of the counter is connected to the data input end of the controller. The two electrodes, the DC power supply, and the counter form a counting unit. When the insulating oil flows into the detection pipeline through the oil inlet, if the insulating oil flowing through the two electrodes forms an instantaneous conduction between the two electrodes, the counter is triggered to count; a micro-water content calculation module is arranged in the controller, which is used to calculate the micro-water content of the insulating oil in the transformer according to the counting data of all counting units at each acquisition moment obtained in real time.

[0079] Wherein, the micro-water content calculation module specifically performs the following steps:

[0080] S10, acquiring trigger data collected by a detection system for micro-water content in insulating oil in real time;

[0081] S20, establishing a trigger curve based on the trigger data according to the acquisition time;

[0082] S30, inputting the trigger curve into a pre-trained water content calculation model, inputting the trigger curve, and obtaining a water content detection result of the insulating oil in the transformer.

[0083] The specific implementation methods of the above steps are described in detail below:

[0084] First, the specific implementation of step S10. The purpose of this step is to obtain the trigger data collected by the micro-water content detection system in the insulating oil in real time. Specifically, this step includes the following sub-steps:

[0085] (1) By detecting multiple counting units set in the pipeline, the short-term conduction events caused by the insulating oil contacting the electrodes during the flow process are monitored in real time. Each counting unit consists of two electrodes, a DC power supply and a counter. When the insulating oil flows through the two electrodes, it will be instantly turned on, thereby triggering the counter to count.

[0086] (2) The controller collects the counting data of each counting unit in real time and stores it as trigger data. These trigger data reflect the frequency and intensity characteristics of the instantaneous conductive event caused by the trace water content in the insulating oil at a specific collection moment. The collection frequency can be set between 0.02 seconds and 1 second, for example, 0.1 seconds can be selected. In this way, trigger data with sufficiently high time resolution can be obtained.

[0087] (3) The controller temporarily stores all the collected trigger data to provide raw data support for the subsequent trigger curve establishment and water content calculation. These real-time collected trigger data are the basic input for water content calculation.

[0088] In general, step S10 realizes the acquisition of real-time trigger data of the micro-water content detection system in the insulating oil, laying the foundation for subsequent data processing and water content calculation.

[0089] Next is the specific implementation of step S20. The purpose of this step is to establish a trigger curve based on the trigger data according to the acquisition time. It specifically includes the following sub-steps:

[0090] (1) The controller arranges the trigger data at each moment acquired in step S10 into a time series according to the acquisition time sequence. These data points constitute the trigger curve in the process of detecting the micro-water content of insulating oil.

[0091] (2) In order to better describe the time characteristics of the trigger curve, the controller can perform linear interpolation processing on the trigger data so that the trigger curve is smooth and continuous. This can better reflect the dynamic characteristics of the change of the micro-water content of the insulating oil over time.

[0092] (3) The controller saves the established trigger curve data for subsequent water content calculation. The trigger curve contains the time domain characteristics of the micro-water content of the insulating oil, which provides important input for the training and prediction of the deep learning model.

[0093] In general, step S20 realizes the establishment of a trigger curve reflecting the dynamic changes of the micro-water content of the insulating oil based on the trigger data collected in real time, laying a data foundation for the subsequent water content calculation.

[0094] Finally, the specific implementation of step S30 is described. The purpose of this step is to input the trigger curve into the pre-trained water content calculation model to obtain the water content detection result of the insulating oil in the transformer. Specifically, it includes the following sub-steps:

[0095] (1) The controller sends the trigger curve established in step S20 as input data to the pre-trained water content calculation model. The model adopts an embedded multi-branch neural network structure, which consists of a fusion Fourier equation module, a signal processing module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network, and a final fusion network.

[0096] (2) First, the fusion Fourier equations module will perform a series of frequency domain analyses on the input trigger curve, including Fourier transform, inverse Fourier transform, short-time Fourier transform, wavelet transform and Hilbert transform, and extract spectral characteristic parameters such as amplitude spectrum, phase spectrum and power spectrum density.

[0097] (3) The signal processing module uses the sliding window method to further extract the frequency domain characteristic parameters and obtain long-term characteristics, short-term characteristics and abnormal characteristics. These characteristics reflect the different characteristics of the micro-water content of insulating oil in the time and frequency domain.

[0098] (4) Subsequently, the long-term feature branch network, the short-term feature branch network and the abnormal feature branch network respectively use deep learning, convolutional neural network and recurrent neural network technologies to extract the first moisture content data, the second moisture content data and the third moisture content data according to the different input features.

[0099] (5) Finally, the fusion network will combine these three types of moisture content data and use the correlation between long-term characteristics, short-term characteristics and abnormal characteristics to perform the final fusion calculation to obtain the detection results of the micro-water content of the insulating oil.

[0100] In the process of model training, it is necessary to first establish a training data set, which includes: collecting transformer insulating oil samples under different working conditions, measuring the water content of the samples using the standard Karl Fischer titration method, and then using this detection system to collect the trigger curve data of these samples, and pre-processing the trigger curve data by Fourier transform, pairing the processed spectrum data with the standard water content results, and finally dividing them into training set, validation set and test set.

[0101] The steps of model training include: initializing network parameters, processing input data by improving the Fourier equations, then calculating the predicted moisture content through the neural network, calculating the loss function between the predicted value and the true value, back-propagating to update the network parameters until the loss function converges, using the validation set to evaluate the model performance and adjust the hyperparameters, and finally using the test set for final evaluation and parameter optimization.

[0102] Through this multi-branch neural network structure, the time domain and frequency domain characteristics are integrated, which can better capture the dynamic changes of the micro-water content of insulating oil and improve the accuracy and robustness of water content detection.

[0103] In general, step S30 implements inputting the trigger curve into the pre-trained deep learning moisture content calculation model to obtain the detection result of the micro-water content of the transformer insulating oil. The model makes full use of the time-frequency domain features and has strong learning and generalization capabilities.

[0104] In summary, this transformer insulating oil micro-water content detection system includes three main steps: first, the trigger data is collected in real time through the counting unit to establish a trigger curve that reflects the dynamic changes of the micro-water content; then the trigger curve is input into the pre-trained deep learning moisture content calculation model, which integrates the time-frequency domain features and can accurately predict the micro-water content of the insulating oil; the entire detection process realizes automation, real-time and high-precision detection functions.

[0105] Specifically, the principle of the present invention is to use trigger counts to reflect instantaneous conductive events during the flow of insulating oil, and to extract and fusion analyze the trigger curve in the time and frequency domains through a deep learning model, thereby achieving accurate detection of the micro-water content of the insulating oil. The specific technical principles are as follows:

[0106] 1. Trigger counting mechanism: This system sets up multiple counting units in the detection pipeline, each unit consists of two electrodes, a DC power supply and a counter. When the insulating oil flows through the two electrodes, if the micro-water content causes the insulating oil to conduct instantaneously, the counter will be triggered to count. The trigger curve formed by these counting events reflects the conductive properties of the insulating oil at different times, and further reflects the dynamic changes of the micro-water content.

[0107] 2. Time-frequency domain feature extraction: In order to more comprehensively describe the characteristics contained in the trigger curve, this system uses the method of integrating Fourier equations to perform frequency domain analysis on the trigger curve. Including Fourier transform, inverse Fourier transform, short-time Fourier transform, wavelet transform and Hilbert transform, it can extract time-frequency domain feature parameters such as amplitude spectrum, phase spectrum, power spectrum density, time-frequency spectrum, wavelet coefficients, instantaneous frequency and signal envelope. These rich features reflect the changing rules of the micro-water content of insulating oil on different time scales and frequency characteristics.

[0108] 3. Deep learning model: This system adopts an embedded multi-branch neural network structure, takes the extracted time-frequency domain features as input, and learns the slow change trend, instantaneous fluctuation characteristics and sudden change characteristics of the micro-water content of insulating oil through the long-term feature branch network, short-term feature branch network and abnormal feature branch network. Finally, in the fusion network, the correlation between these three types of features is used to comprehensively calculate the final micro-water content detection result.

[0109] This deep learning-based method can fully explore the time-frequency domain information contained in the trigger curve, and has strong learning and generalization capabilities in capturing the dynamic changes in the micro-water content of insulating oil, thereby greatly improving the accuracy and reliability of detection.

[0110] Compared with the prior art, the solution of the present invention has the following outstanding technical advantages:

[0111] 1. Real-time online monitoring of the micro-water content of insulating oil is realized. By triggering the counting mechanism, the dynamic changes of the insulating oil state can be continuously tracked and abnormal conditions can be discovered in time.

[0112] 2. The time-frequency domain features are integrated to improve the accuracy of detection. The deep learning model can effectively extract and integrate the multi-scale features in the trigger curve, and better describe the complex changes in the micro-water content of insulating oil.

[0113] 3. Good adaptability. The deep learning model has strong learning ability and can adapt to the changes in insulating oil performance under different working conditions. Through continuous optimization training, the detection performance can be continuously improved.

[0114] 4. Realized automation and intelligence. The system adopts embedded design, fully realizes automatic detection and analysis, and provides technical support for intelligent management of transformer status.

[0115] In order to better understand and implement the present invention, a specific embodiment 1 of the micro-water content calculation module of the present invention is provided below. The steps included in the embodiment 1 are described in detail as follows: First, the specific implementation of step S10. The purpose of this step is to obtain the trigger data collected by the micro-water content detection system in the insulating oil in real time. Specifically, it includes the following sub-steps:

[0116] Step 101: By detecting multiple counting units set in the pipeline, the short-term conduction events caused by the insulating oil contacting the electrodes during the flow are monitored in real time. Each counting unit consists of two electrodes E1 and E2, a DC power supply V and a counter C. When the insulating oil flows through the two electrodes, it will be instantly turned on, thereby triggering the counter to count. This process can be expressed by the formula:

[0117]

[0118] in, Represents the voltage between the two electrodes E1 and E2. When it is less than the threshold V th When the insulating oil is instantaneously conductive, the counter C increases by 1. Threshold V th It can be adjusted according to actual conditions, for example, it can be set to 0.1V.

[0119] Step 102: The controller collects the counting data C of each counting unit in real time.i (t j ), where i=1,2,...,N represents the i-th counting unit, t j represents the jth sampling moment, the sampling frequency can be set to f s =10Hz, that is, data is collected every 0.1 second. These count data reflect the frequency and intensity characteristics of the instantaneous conductive events caused by the trace water content in the insulating oil at a specific collection moment, and are stored as trigger data.

[0120] Step 103: The controller collects all the trigger data C i (t j ) to provide raw data support for subsequent trigger curve establishment and water content calculation.

[0121] In general, step S10 realizes the acquisition of real-time trigger data of the micro-water content detection system in the insulating oil, laying the foundation for subsequent data processing and water content calculation.

[0122] Next is the specific implementation of step S20. The purpose of this step is to establish a trigger curve based on the trigger data according to the acquisition time. It specifically includes the following sub-steps:

[0123] Step 201: The controller collects the trigger data C at each moment acquired in step S10 according to the acquisition time sequence. i (t j ) are arranged into a time series to form a trigger curve C(t) in the process of detecting the micro-water content of insulating oil, where

[0124] Step 202: In order to better describe the time characteristics of the trigger curve, the controller can i (t j ) to perform linear interpolation processing, so that the trigger curve C(t) is smooth and continuous. The interpolation processing can use the Lagrange interpolation formula:

[0125]

[0126] Among them, M is the total number of sampling points, and t is any time. This can better reflect the dynamic characteristics of the change of the micro-water content of insulating oil over time.

[0127] Step 203: The controller saves the established trigger curve data C(t) for subsequent water content calculation. The trigger curve contains the time domain characteristics of the micro-water content of the insulating oil, which provides important input for the training and prediction of the deep learning model.

[0128] In general, step S20 realizes the establishment of a trigger curve reflecting the dynamic changes of the micro-water content of the insulating oil based on the trigger data collected in real time, laying a data foundation for the subsequent water content calculation.

[0129] Finally, the specific implementation of step S30 is described. The purpose of this step is to input the trigger curve into the pre-trained water content calculation model to obtain the water content detection result of the insulating oil in the transformer. Specifically, it includes the following sub-steps:

[0130] Step 301: The controller sends the trigger curve C(t) established in step S20 as input data to the pre-trained water content calculation model. The model adopts an embedded multi-branch neural network structure, which consists of a fusion Fourier equation module, a signal processing module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network, and a final fusion network.

[0131] First, let's look at the fusion Fourier equation module. Its purpose is to perform frequency domain analysis on the input trigger curve C(t) and extract characteristic parameters such as amplitude spectrum, phase spectrum and power spectrum density. Specifically, it includes the following steps:

[0132] Step 302, Fourier transform: convert the time domain trigger curve C(t) into a frequency domain representation C(ω), where ω is the angular frequency. The Fourier transform formula is:

[0133]

[0134] This step can obtain the spectrum amplitude characteristics.

[0135] Step 303, inverse Fourier transform: convert the frequency domain signal C(ω) back to the time domain representation Used for signal reconstruction. The inverse Fourier transform formula is:

[0136]

[0137] This step helps to analyze the characteristics of the time domain signal.

[0138] Step 304, short-time Fourier transform: Analyze the time-frequency characteristics of the non-stationary trigger curve C(t) to obtain the time-frequency spectrum C(t,ω). The short-time Fourier transform formula is:

[0139]

[0140] Among them, w(t) is a windowing function, which is used to locally analyze non-stationary signals. The time-frequency spectrum can reflect the change of the micro-water content of insulating oil over time.

[0141] Step 305: Wavelet transform: Use multi-scale analysis to extract the local features of the trigger curve C(t). The wavelet transform formula is:

[0142]

[0143] Among them, a is the scale factor, b is the translation factor, and ψ(t) is the mother wavelet function. The wavelet coefficient C(a,b) reflects the characteristics of the signal in different time-frequency regions.

[0144] Step 306: Hilbert transform: Calculate the instantaneous frequency and envelope of the trigger curve C(t) to analyze the nonlinear characteristics of the signal. The Hilbert transform formula is:

[0145]

[0146] in, is the imaginary part of the analytical signal, A(t) is the signal envelope, ω(t) is the instantaneous frequency, and φ(t) is the phase.

[0147] The outputs of the Fusion Fourier Equations module include the amplitude spectrum |C(ω)|, the phase spectrum ∠C(ω), and the power spectrum density S C (ω)=|C(ω)| 2 The frequency domain feature parameters, as well as the time-frequency domain feature parameters such as the time-spectrum C(t,ω), wavelet coefficients C(a,b), instantaneous frequency ω(t) and signal envelope A(t). These rich frequency domain and time-frequency domain features provide important input for subsequent deep learning models.

[0148] Next is the signal processing module, which aims to further extract the frequency domain and time-frequency domain feature parameters calculated by the fusion Fourier equation module to obtain long-term features, short-term features and abnormal features. Specifically, it includes:

[0149] Step 307: Use the sliding window method to divide the time domain trigger curve C(t) into multiple short time windows, perform statistical analysis on the frequency domain and time-frequency domain characteristic parameters in each time window, and extract long-term features, short-term features and abnormal features.

[0150] Long-term characteristics reflect the slow change trend of the micro-water content of insulating oil, which can be expressed by statistics such as mean value and variance. Short-term characteristics reflect the instantaneous fluctuation of the micro-water content of insulating oil, which can be described by features such as peak value and peak interval. Abnormal characteristics can use abnormal detection algorithms to identify abnormal fluctuation points in the trigger curve and analyze sudden changes in the micro-water content of insulating oil.

[0151] These three types of features characterize the dynamic changes of the micro-water content in insulating oil from different time scales and angles, providing rich input for subsequent deep learning models.

[0152] Then there are the long-term feature branch network, short-term feature branch network and abnormal feature branch network, which use different deep learning algorithms to extract the water content prediction results based on the above three types of input features:

[0153] Step 308: The long-term feature branch network uses a deep neural network to use x long As input, output the first moisture content prediction result y long :

[0154] y long =f long (x long θ long );

[0155] Among them, x long is the long-term eigenvector, f long is the deep neural network function, θ long is the network parameter.

[0156] Step 309: The short-term feature branch network uses a convolutional neural network and uses x short As input, output the second moisture content prediction result y short :

[0157] y short =f short (x short θ short );

[0158] Among them, x short is the short-term eigenvector, f short is the convolutional neural network function, θ short is the network parameter.

[0159] Step 310: The abnormal feature branch network uses a recurrent neural network and uses x anom As input, output the third moisture content prediction result y anom :

[0160] y anom =f anom (x anom θ anom );

[0161] Among them, x anom is the abnormal feature vector, f anom is the recurrent neural network function, θ anom is the network parameter.

[0162] Finally, there is the fusion network, which aims to combine the three moisture content prediction results, perform the final fusion calculation, and obtain the detection results of the micro-water content of the insulating oil:

[0163] Step 311, the fusion network uses a weighted average method to combine y long ,y short and anom , considering the correlation between long-term characteristics, short-term characteristics and abnormal characteristics, the final water content detection result y is output:

[0164] y=αy long +βy short +γy anom ;

[0165] Among them, α, β and γ are fusion weight coefficients, which can be optimized through the validation set.

[0166] In the process of model training, we first need to establish a training data set. Specifically, it includes:

[0167] Step 401: Collect transformer insulating oil samples under different working conditions and measure the water content of the samples using the standard Karl Fischer titration method. true .

[0168] Step 402: Use the detection system to collect trigger curve data C(t) of these samples.

[0169] Step 403: Perform Fourier transform preprocessing on the trigger curve data C(t) to obtain the frequency domain characteristic parameter x freq ={|C(ω)|,∠C(ω),S C (ω)}.

[0170] Step 404: The processed frequency domain feature x freq Compared with the standard water content result true Pairing to form training samples (x freq ,y true ).

[0171] Step 405: Divide the training samples into a training set, a validation set, and a test set.

[0172] The steps of model training include:

[0173] Step 501: Initialize network parameters θ long ,θ short ,θ anom And fusion weight coefficients α, β, γ.

[0174] Step 502, forward propagation: Process the input data x by integrating the Fourier equations module and the signal processing module freq , extract the long-term feature x long , short-term characteristics x short and abnormal feature x anom, and then sent them to the long-term feature branch network, short-term feature branch network and abnormal feature branch network respectively to calculate the three water content prediction results y long ,y short and anom Finally, the final water content prediction y is obtained through the fusion network.

[0175] Step 503: Calculate the predicted value y and the true value y true The mean square error loss is:

[0176]

[0177] Among them, N is the number of samples.

[0178] Step 504, back propagation updates network parameters and fusion weight coefficients, including:

[0179]

[0180]

[0181] Among them, η long , η short , η anom , η α , η β and η γ is the corresponding learning rate.

[0182] Step 505: Repeat steps 502-504 until the loss function L converges.

[0183] Step 506: Use the validation set to evaluate the model performance and adjust hyperparameters, such as learning rate, network structure, etc.

[0184] Step 507: Use the test set for final evaluation to further optimize the model parameters.

[0185] Through this multi-branch neural network structure, the time domain and frequency domain characteristics are integrated, which can better capture the dynamic changes of the micro-water content of insulating oil and improve the accuracy and robustness of water content detection.

[0186] The frequency domain and time-frequency domain features extracted by the Fourier equations module are integrated, such as the amplitude spectrum |C(ω)|, the phase spectrum ∠C(ω), the power spectrum density S C (ω), time-frequency spectrum C(t,ω), wavelet coefficients C(a,b), instantaneous frequency ω(t) and signal envelope A(t), etc., reflect the characteristics of the micro-water content in the insulating oil in different time-frequency characteristics.

[0187] The signal processing module further extracts long-term features x long , short-term characteristics x short and abnormal feature xanom , describes the slow changing trend, instantaneous fluctuation and sudden change of the micro-water content of insulating oil on a time scale.

[0188] The three branch networks use deep learning, convolutional and recurrent neural networks respectively to learn the mapping relationship between these features and moisture content from different angles, and obtain three moisture content prediction results.

[0189] The final fusion network makes full use of the correlation between the three prediction results and obtains the final detection result through weighted averaging, which further improves the performance of the model.

[0190] In order to further better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A transformer has been in operation for many years. In order to monitor the micro-water content of its insulating oil, the technical team decided to use the detection system proposed by the present invention for real-time online detection. The capacity of the transformer is 100MVA, and it works at full load all year round. According to the operating conditions of the transformer and the use history of the insulating oil, it is expected that the micro-water content in the insulating oil will fluctuate within the range of 0.1% to 1%.

[0191] The technical team first purchased the insulating oil micro-water content detection system provided by the present invention, and installed and debugged it according to the requirements of the instructions. The specific installation process is as follows:

[0192] 1. Installation of detection system

[0193] like Figure 1 As shown, the main components of the detection system include: oil inlet, detection pipeline, oil outlet and controller. The oil inlet and oil outlet are respectively connected to the external circulation loop of the transformer insulating oil to form an oil circuit. The detection pipeline is made of a transparent tube with an inner diameter of more than 10mm, and the optional length is 5 to 10cm. There are several, such as 10 cylindrical hollow channel tubes evenly distributed in the pipeline, and the diameter of each channel is 2 / 4 / 5mm.

[0194] The two electrodes are installed at radially symmetrical positions in the middle of each channel, and the electrode spacing is at least 1mm. The two electrodes can be two metal patches, symmetrically attached to the radially symmetrical positions in the middle of each channel; the first electrodes of all channels are connected to the positive pole of the DC power supply after being led out of the detection pipeline, and the second electrodes of all channels are connected to the negative input terminal of the counter after being led out of the detection pipeline, and the negative pole of the DC power supply is connected to the positive input terminal of the counter to form a counting unit. The above connections are connected together by wires. For example, the connection can be made in the form of a jumper, and each connection end is provided with a corresponding metal buckle for better contact and fixing the jumper. The data input end of the controller is connected to the output end of the counter. The entire system adopts an embedded design and is installed in a control cabinet next to the transformer.

[0195] Table 1 Main parameters of the detection system

[0196] Parameter name Parameter Value Oil inlet diameter 10mm Detection pipeline length 1m Number of hollow channels 10 Hollow channel diameter 2mm Electrode spacing 1mm DC voltage 5V Counter sampling frequency 200Hz

[0197] Through actual measurement, the total cross-sectional area of ​​the 10 hollow channels in the detection pipeline is π×10×(2mm) 2 / 4=62.8mm 2 , slightly larger than the cross-sectional area of ​​the oil inlet π×(10mm) 2 / 4=78.5mm 2 This ensures that the insulating oil can flow evenly through all hollow channels, fully contacting the electrodes and improving the reliability of detection.

[0198] 2. System debugging

[0199] During installation and commissioning, the initial water content of the transformer insulating oil was measured using the standard Karl Fischer titration method, which was 0.4%. Then the detection system was started and the trigger data was collected for 10 minutes. Figure 2 The trigger curve is shown in Figure 2. It can be seen that the trigger event presents certain randomness and fluctuation characteristics, reflecting the dynamic changes of the micro-water content of the insulating oil in a short period of time.

[0200] Statistical analysis of the collected trigger data revealed that there were 18,240 trigger events in 10 minutes, with an average of 91 triggers per second. This data is basically consistent with the 0.4% water content level of the insulating oil. In order to further verify the detection accuracy of the system, the technical team took 3 insulating oil samples and performed Karl Fischer titration and system detection respectively. The results are shown in Table 2.

[0201] Table 2 Comparison of test results of micro-water content in insulating oil

[0202] Sample No. Karl Fischer titration results The test results of this system error 1 0.35％ 0.37％ -5.71％ 2 0.48％ 0.45％ 6.25％ 3 0.62％ 0.59％ 4.84％

[0203] It can be seen from Table 2 that the detection results of this system are very close to those of the standard titration method, with the maximum relative error not exceeding 7%, which meets the accuracy requirements of the power industry for the detection of trace water content in transformer insulating oil.

[0204] After on-site commissioning and verification, the detection system was found to be stable and reliable. The technical team then officially launched the system and incorporated it into the overall solution for online transformer status monitoring.

[0205] 3. System operation status

[0206] After the transformer was put into operation, the technical team collected the micro-water content data of the transformer insulating oil through this detection system every day. Through the analysis of historical data, it was found that the micro-water content of the insulating oil had obvious daily and seasonal changes: the water content reached a minimum value of about 0.25% at around 6 o'clock in the morning, and then gradually increased with the increase of load, reaching a peak of about 0.45% at around 16 o'clock in the afternoon, and then gradually decreased. This change pattern is highly consistent with the change trend of the daily load of the transformer; the water content is higher in spring and summer, and lower in autumn and winter. This is mainly due to the influence of ambient temperature. When the temperature rises, the solubility of water in the insulating oil increases, which causes the micro-water content to increase.

[0207] Through in-depth analysis of these historical data, the technical team further refined the early warning mechanism for insulating oil status assessment:

[0208] (1) Daily monitoring: If the daily average water content exceeds 0.4%, or there is a significant abnormal fluctuation, a timely warning will be issued.

[0209] (2) Seasonal monitoring: If the average moisture content in spring and summer is higher than 0.5% and in autumn and winter is higher than 0.35%, an abnormal insulating oil condition warning will be issued.

[0210] (3) Emergency monitoring: If the moisture content suddenly increases by more than 0.2% at any time, an immediate warning of possible oil leakage or other faults will be issued.

[0211] Through this refined status monitoring and early warning mechanism, the technical team can promptly detect abnormal changes in the water content in the transformer insulating oil, provide a reliable basis for subsequent maintenance, and significantly reduce the risk of equipment failure.

[0212] 4. Deep learning model training

[0213] In order to further improve the performance of the detection system, the technical team also continued to optimize the deep learning model used in the present invention.

[0214] First, technicians collected 50 insulating oil samples of the transformer under different ambient temperatures, load levels and operating years. The water content of these samples was measured using the standard Karl Fischer titration method, and the results are shown in Table 3.

[0215] Table 3 Water content of insulating oil samples

[0216] Sample No. Micro water content 1 0.25％ 2 0.31％ 3 0.39％ ... ... 50 0.72％

[0217] At the same time, the detection system was used to collect trigger curve data for these 50 samples. Next, the technicians used these sample data to train the deep learning model. First, the fusion Fourier equation group module was used to perform spectrum analysis on the trigger curve data to extract time-frequency domain feature parameters such as amplitude spectrum, phase spectrum, power spectrum density, time-frequency spectrum, wavelet coefficients, instantaneous frequency and signal envelope.

[0218] These characteristic parameters are then paired with the standard trace water content results of the samples and divided into training set, validation set and test set. The following steps are used to train the model:

[0219] Step 1. Initialize the network parameters, including the weights and biases of the long-term feature branch network, the short-term feature branch network, and the abnormal feature branch network, as well as the weight coefficient of the fusion network.

[0220] Step 2. Forward propagation: Input the training sample into the model and calculate the three moisture content prediction results y through each branch network long ,y short and anom ,Then the network is fused to get the final prediction y.

[0221] Step 3. Calculate the predicted value y and the true value y true The mean square error loss L is:

[0222]

[0223] Among them, N is the number of samples.

[0224] Step 4. Back propagation updates network parameters and fusion weights, including:

[0225]

[0226] Among them, η long , η short , η anom , η α , η β and η γ is the corresponding learning rate.

[0227] Step 5. Repeat steps 2-4 until the loss function on the validation set converges.

[0228] Step 6. Use the test set to evaluate the performance of the final model and further optimize the hyperparameters.

[0229] Through such a training process, the deep learning model gradually learned the complex mapping relationship between the micro-water content of insulating oil and the time-frequency domain characteristics of the trigger curve. After multiple rounds of iterative optimization, the average absolute error of the model on the test set was reduced to 3.2%, greatly improving the detection accuracy.

[0230] 5. System operation effect

[0231] Since the adoption of the detection system of the present invention, the technical team has achieved remarkable results in the refined management of the transformer insulating oil status:

[0232] (1) Real-time online monitoring of the micro-water content of insulating oil is realized. Through the trigger counting mechanism, the system can continuously track the dynamic changes of the insulating oil state and collect 91 trigger data per second, providing high-time resolution raw data for subsequent analysis and processing.

[0233] (2) Greatly improved detection accuracy. The deep learning model has a strong learning ability in capturing the slow change trend, instantaneous fluctuation characteristics and abnormal changes of the micro-water content of insulating oil through the effective extraction and fusion of time-frequency domain features. The detection accuracy has reached an average absolute error of 3.2%.

[0234] (3) Good adaptability. Since the deep learning model can continuously optimize training and continuously improve performance, even if the performance of the insulating oil changes with changes in the environment and load, the detection system can adapt in time and maintain high-precision detection.

[0235] (4) Realized automation and intelligent management. The system fully realizes embedded automation design and can complete real-time detection and status analysis of the micro-water content of insulating oil without human intervention, providing strong support for intelligent management of transformer status.

[0236] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A system for detecting the water content of transformer insulating oil, characterized in that: It includes a controller, a DC power supply, a counter, and an oil inlet, a detection pipeline, and an oil outlet connected in sequence. The oil inlet and the oil outlet are used to be connected to the external circulation loop of the insulating oil of the transformer. The detection pipeline is a transparent tube of insulating material with two ends respectively connected to the oil inlet and the oil outlet. A plurality of cylindrical hollow channels are arranged in the transparent tube for connecting the oil inlet and the oil outlet. The hollow channels are coaxially and evenly distributed in the transparent tube. Two electrodes in contact with the channel are symmetrically arranged in the middle position of the hollow channel, which are a first electrode and a second electrode. The first electrode is connected to the first pole of the DC power supply. The second electrode of the DC power supply is connected to the first input end of the counter, the second input end of the counter is connected to the second electrode, the output end of the counter is connected to the data input end of the controller, the two electrodes, the DC power supply and the counter form a counting unit, when the insulating oil flows into the detection pipeline through the oil inlet, if the insulating oil flowing through the two electrodes forms an instantaneous conduction between the two electrodes, the counter is triggered to count; a micro-water content calculation module is provided in the controller, which is used to calculate the micro-water content of the insulating oil in the transformer according to the counting data of all counting units at each collection moment obtained in real time.

2. A transformer insulating oil micro-water content detection system according to claim 1, characterized in that: The micro-water content calculation module specifically performs the following steps: S10, real-time acquisition of trigger data between two electrodes collected by a detection system for micro-water content in insulating oil; S20, establishing a trigger curve with the trigger data according to the acquisition time; S30, inputting the trigger curve into a pre-trained water content calculation model, inputting the trigger curve, and obtaining a water content detection result of the insulating oil in the transformer.

3. A transformer insulating oil micro-water content detection system according to claim 2, characterized in that: The data collection time is specifically collected once every 0.02 seconds to 1 second.

4. A transformer insulating oil micro-water content detection system according to claim 3, characterized in that: The trigger data is specifically the minimum value of the distance between the tips of two electrodes in any hollow channel of the counting unit that generates counts at each acquisition moment.

5. A transformer insulating oil micro-water content detection system according to claim 4, characterized in that: The water content calculation model adopts an embedded multi-branch neural network structure, including a fusion Fourier equation module, a signal processing module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network and a fusion network.

6. A transformer insulating oil micro-water content detection system according to claim 5, characterized in that: The fusion Fourier equations module is used to input the trigger curve into the improved Fourier equations to calculate the spectrum analysis parameters, including amplitude spectrum, phase spectrum, and power spectrum density; The signal processing module uses a sliding window method to extract features from the spectrum analysis parameters calculated by the fusion Fourier equation group module to obtain long-term features, short-term features and abnormal features; The long-term feature branch network is used to obtain first water content data through a deep learning network according to the long-term feature; The short-term feature branch network is used to obtain second water content data through a convolutional neural network according to the short-term features; The abnormal feature branch network is used to obtain third water content data through a recurrent neural network according to the abnormal feature; The fusion network is used to obtain the moisture content data result by fusing and calculating the first moisture content data, the second moisture content data and the third moisture content data using the domain correlation of the long-term characteristics, the short-term characteristics and the abnormal characteristics.

7. A transformer insulating oil micro-water content detection system according to claim 6, characterized in that: The fused Fourier equations include Fourier transform equation, inverse Fourier transform equation, short-time Fourier transform equation, wavelet transform equation and Hilbert transform equation; The Fourier transform equation is used to convert the time domain signal into a frequency domain representation, the input is the time domain trigger curve, and the output is the spectrum; The inverse Fourier transform equation is used to convert the frequency domain signal back to the time domain representation, the input is the spectrum, and the output is the reconstructed time domain signal; The short-time Fourier transform equation is used to analyze the time-frequency characteristics of non-stationary signals, the input is a time-domain trigger curve, and the output is a time-frequency spectrum; The wavelet transform equation is used for multi-scale analysis of local features of the signal, the input is the time domain trigger curve, and the output is the wavelet coefficient; The Hilbert transform equation is used to calculate the instantaneous frequency and envelope of the signal, the input is the time domain trigger curve, and the output is the analytical signal.

8. A system for detecting water content in transformer insulating oil according to claim 1, characterized in that: The inner diameter of each hollow channel is equal.

9. A transformer insulating oil micro-water content detection system according to claim 1, characterized in that: The sum of the cross-sectional areas of all the hollow channels is not less than the cross-sectional area of ​​the oil inlet.

10. A transformer insulating oil micro-water content detection system according to claim 9, characterized in that: The sum of the cross-sectional areas of all the hollow channels is 1.68 times the cross-sectional area of ​​the oil inlet.

Images