Transformer oil dissolved gas detection system

Abstract

The invention provides a transformer oil dissolved gas multi-sensor detection system. The method comprises the following steps: degassing the transformer oil, blowing the degassed gas into a chromatographic column by using carrier gas to separate each component, feeding the separated gas into a spectrum detection instrument and a chromatographic detection instrument in sequence to obtain independent spectrogram data, and finally deriving and calculating through a Gaussian fusion algorithm or a neural network algorithm to obtain an independent spectrogram of the transformer oil. And comparing and fusing the two data to obtain final dissolved gas concentration data. According to the method, the chromatographic technology and the spectrum technology are combined for use, the defects of respective detection technologies are overcome, data of the chromatographic technology and the spectrum technology are subjected to fusion calculation, and the reliability and accuracy of a final result are improved.

Description

Technical Field

[0001] The invention relates to the technical field of gas detection, and in particular to a transformer oil dissolved gas chromatography + spectrum detection system. Background Art

[0002] With the continuous construction of ultra-high voltage substations in my country, there is a higher demand for transformer detection. The time interval from the occurrence of faults to the detection of abnormalities is getting shorter and shorter, requiring faster detection and resolution of problems. The main transformers and high-voltage plants of large power plants usually use oil-immersed power transformers. Since they are on the key route of power transmission, the reliability of their operation is closely related to the economic efficiency of power plant operation. Whether it is a thermal fault or an electrical fault, the oil-immersed power transformer will cause the insulation medium-transformer oil to crack and produce some characteristic gases, especially the data detection of hydrogen H2 and acetylene C2H2. Its value is used to identify possible transformer operation faults in the early stage, and there is a higher detection limit requirement. At present, the detection limit of C2H2 in the power system has dropped from 0.5ppm to 0.2ppm.

[0003] The gas chromatography method used in the prior art is very sensitive to pressure and temperature. In actual field use, it is easy to produce fault omissions and false alarms due to aging and different working conditions, which affects the judgment of the transformer status and thus creates danger. In addition, the traditional chromatography requires a large dynamic range for CO2 detection. In the case of high concentrations of CO2, the linearity and accuracy of the data are low, and the accuracy of the chromatographic detector will also decrease at high concentrations in order to adapt to a large range. Summary of the invention

[0004] The purpose of the present invention is to provide a transformer oil dissolved gas detection system, which combines chromatography and spectroscopy to make up for the shortcomings of their respective detection technologies, and sends the gas released from the transformer oil to a spectral detection instrument and a chromatographic detection instrument respectively, and cross-compares the data of the two, and uses a fusion algorithm to process the data from multiple sensors to improve the reliability of the detection results.

[0005] In order to achieve the above-mentioned object, the present invention provides a transformer oil dissolved gas detection system, comprising a degassing unit, a chromatographic column, a spectrum detector, a chromatographic detector and a data fusion processing unit, wherein the degassing unit performs degassing on the transformer oil, and after carrier gas purge, the sample gas is separated into gas components through the chromatographic column, and the separated gas enters the spectrum detector and the chromatographic detector to perform qualitative and quantitative analysis on each gas component, and generate corresponding spectrum data and chromatographic data. The data fusion processing unit performs fusion processing on multi-sensor data based on a Gaussian fusion processing algorithm and a neural network fusion model to obtain fusion processing data of gas detection.

[0006] Furthermore, the data fusion processing unit first performs peak height detection on the spectral data and the chromatographic data. When the difference between the two data exceeds a threshold, the sensor is judged to be abnormal and no subsequent algorithm processing is performed, and re-detection is performed using standard gas; if the data difference does not exceed the threshold, further data fusion processing is performed.

[0007] Furthermore, the Gaussian fusion processing algorithm includes the following steps:

[0008] S101, Gaussian distribution fitting, performing Gaussian distribution fitting on the data peaks of the chromatographic data and the spectral data to obtain Gaussian distribution fitting parameters, including the mean and standard deviation of the data peaks, and the amplitude of the Gaussian distribution;

[0009] S102, Gaussian mixture model construction, using the fitted Gaussian distribution parameters as initial parameters to construct the Gaussian mixture model;

[0010] S103, model fitting, optimizing the Gaussian mixture model using a data fitting algorithm to obtain optimal mixture model parameters;

[0011] S104, peak fusion, obtaining peak fusion parameters of each gas by merging the obtained Gaussian mixture model, wherein the peak fusion parameters include the mean value, standard deviation and amplitude of the Gaussian distribution of the peak;

[0012] S105, result verification, verifies whether the fused result can accurately describe the actual gas component peak.

[0013] Furthermore, in step 102, it is assumed that one or more gas peaks have been detected in the data of each sensor. For each peak, a Gaussian distribution can be used to fit it, and the probability density function of the Gaussian distribution is as follows:

[0014]

[0015] Where: μ is the mathematical expectation of the Gaussian distribution, representing the center position of the peak; σ 2 is the variance, representing the width of the peak.

[0016] Furthermore, in step 102, assuming that there are multiple sensors, a certain gas is detected by multiple sensors, that is, the peaks of the same gas are detected at the same time, then the N peaks are fused. The following takes H2 gas as an example:

[0017]

[0018] Where λ is the weight of the ith peak and satisfies And i >0,μ i and σ i 2are the mean and variance of the ith Gaussian distribution.

[0019] Further, in step 103, the expectation maximization algorithm is used to fit the parameters of the Gaussian mixture model, where:

[0020] The Gaussian distribution is simplified as, which means the kth Gaussian distribution in sample x i The probability density function at :

[0021] p(x i )=N(x i ;μ k ,σ k 2 )

[0022] Calculate the posterior probability γ(z ik ), where z ik Represents the probability that the i-th sample belongs to the k-th component:

[0023]

[0024] According to the obtained posterior probability, update the model parameters:

[0025]

[0026]

[0027]

[0028] Repeat the above steps until the parameters converge and obtain the optimal parameters.

[0029] Furthermore, in step 104, the peaks detected by each sensor are fused into one or more more accurate peaks, and the parameters of the fused peaks are obtained by fitting the Gaussian mixture model parameters.

[0030] Furthermore, in the process of Gaussian fusion processing algorithm, when Gaussian normal distribution characteristics cannot be used for fitting, a neural network fusion model is selected to perform fusion processing on multi-sensor data.

[0031] Furthermore, the neural network fusion model extracts and learns features of the spectral data and the chromatographic data respectively, fuses the features extracted in the middle layer or the last layer of the network, and finally outputs the concentration value of each component.

[0032] Furthermore, the neural network fusion model includes:

[0033] An input layer receives the preprocessed spectral and chromatographic data, performs convolution calculations on the spectral data using the trained 1D convolution kernel, compresses the data volume, and extracts the characteristic values ​​of the peaks;

[0034] Hidden layer, used to allow the neural network to further learn the relationship between different gas components and extract the data features of each component;

[0035] Feature fusion, which aggregates and integrates the data features of different detectors;

[0036] The fully connected layer is used to further process and compress the fused data features;

[0037] The output layer outputs classification results according to different task types, including predicted gas concentration and fault type.

[0038] The present invention has the following beneficial effects and technical advances

[0039] 1. Efficient fault detection and monitoring

[0040] The present invention adopts the chromatographic + spectral joint technology to detect the dissolved gas in the transformer oil, and uses the Gaussian fusion algorithm and the neural network fusion algorithm to process the multi-sensor data. Combining the detection technology of chromatography and spectrum, the content of gases such as H2 and C2H2 in the transformer oil can be detected quickly and accurately, which significantly improves the speed and accuracy of fault detection.

[0041] 2. Improve detection sensitivity and accuracy

[0042] The present invention uses the chromatogram + spectrum joint technology and the Gaussian fusion algorithm and the neural network fusion algorithm. The detection limit of gas is reduced from 0.5ppm to 0.2ppm, meeting the higher requirements of the current power system for detection accuracy. This highly sensitive detection technology greatly improves the accuracy of early transformer fault detection and reduces the missed alarm and false alarm rates.

[0043] 3. Strong anti-interference ability and good adaptability

[0044] The present invention can effectively reduce the problems of missed faults and false alarms caused by pressure and temperature changes in the actual field use of gas chromatography. The system is calibrated in real time through an adaptive algorithm to improve the accuracy of transformer status judgment.

[0045] 4. Improved linearity and accuracy

[0046] The present invention uses the combined technology of chromatography and spectroscopy to detect CO2, which solves the problem that the traditional chromatography requires a large dynamic range for CO2 detection. In the case of high concentration CO2, the new detection method ensures the linearity and accuracy of the data.

[0047] 5. Social and economic benefits

[0048] The present invention improves the reliability of transformer operation in an electric power system, reduces accidents caused by inadequate detection, and improves the safety and stability of the electric power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 This is a schematic diagram of a transformer oil dissolved gas detection system according to the present invention.

[0051] Figure 2 This is a logical architecture diagram of the data fusion algorithm part of the present invention.

[0052] Figure 3 Schematic diagram of Gaussian fusion algorithm of the present invention

[0053] Figure 4 This is a network structure diagram of the neural network fusion model of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0056] Figure 1 This is a system framework diagram of a transformer oil dissolved gas detection system of the present invention. Figure 1A transformer oil dissolved gas detection system is shown, comprising: a degassing unit, a chromatographic column, a spectrum detector, a chromatographic detector, a standard gas control unit and a data fusion processing unit. The system sends the gas desorbed from the transformer oil to the spectrum detection instrument and the chromatographic detection instrument respectively, cross-compares the data of the two, uses a fusion algorithm to process the data from multiple sensors, and improves the reliability of the detection results.

[0057] Specifically, the system first uses a degassing unit to degas the transformer oil, and then uses a carrier gas to purge the sample gas through the chromatographic column after quantification. The chromatographic column is filled with a special stationary phase material that can effectively separate different components in the gas sample.

[0058] Preferably, the chromatographic column separates the gas components according to the different characteristics of the gas molecules (such as molecular size, polarity, etc.), ensuring that each component can enter the subsequent sensor independently, and each component gas separated by the chromatographic column is first sent to the spectral detector.

[0059] Preferably, the spectrum detector adopts a non-enclosed detection method, continuously detects during ventilation, and the lower computer (data fusion processing unit) collects and stores data. The spectrum detector uses a light source of a specific wavelength to perform qualitative and quantitative analysis of gas components by absorption or emission spectroscopy.

[0060] Finally, the gas sample is sent to the chromatographic detector, which further separates and detects the gas components according to the distribution behavior of the gas components between the stationary phase and the mobile phase.

[0061] Preferably, the chromatographic detector uses a thermal conductivity detector (TCD) or a flame ionization detector (FID), etc., and obtains its concentration information by detecting the conductivity or ionization degree of the gas component, and the data is also collected by the lower computer (data fusion processing unit).

[0062] In this embodiment, both the spectral data and the chromatographic data include the detection spectra after the sample gas is purged within a period of time, and the time axis is consistent and the peak positions correspond to each other, meeting the basic requirements of the fusion calculation data format.

[0063] Preferably, after obtaining the spectral data and chromatographic data, the data fusion processing unit will first detect whether there is an excessive data difference caused by a sensor failure. When the difference between the spectral data and the chromatographic data is too large, it is determined that one or more sensor failures may exist. The data fusion processing unit sends a signal to the standard gas control unit, and the standard gas control unit uses standard gas to perform a measurement for the degassing part, compares the standard sample result with the previous data, and calibrates the faulty sensor according to the standard gas data (within the calibrable range) or reports the device failure to remind on-site personnel to repair or replace the equipment.

[0064] like Figure 3 As shown, in the data fusion process, Gaussian fusion processing is first performed. In the present embodiment, the data of the spectrum and the chromatogram are approximately normal distribution, which is in line with the application scenario of the Gaussian fusion algorithm. It should be noted that the Gaussian mixture model (Gaussian Mixture Model, GMM) is a parameterized probability model, which assumes that the data is composed of a combination of several Gaussian distributions, and the peak measured by each sensor can be regarded as a Gaussian distribution, and then the Gaussian mixture model is used to merge these Gaussian distributions to obtain a more accurate description of the gas component peak. The Gaussian mixture model is usually an effective method for fusing the same gas peak data of different sensors. In the data fusion process, if two sensors measure the same gas, but the peak shape, peak width and peak intensity are different due to the characteristics of the instrument or the different environmental conditions, the Gaussian mixture model can be used to fuse these peaks.

[0065] Specifically, the Gaussian fusion processing algorithm includes the following steps:

[0066] S101, Gaussian distribution fitting, performing Gaussian distribution fitting on the data peaks of the chromatographic data and the spectral data to obtain Gaussian distribution fitting parameters, including the mean and standard deviation of the data peaks, and the amplitude of the Gaussian distribution;

[0067] S102, Gaussian mixture model construction, using the fitted Gaussian distribution parameters as initial parameters to construct the Gaussian mixture model;

[0068] S103, model fitting, optimizing the Gaussian mixture model using a data fitting algorithm to obtain optimal mixture model parameters;

[0069] S104, peak fusion, obtaining peak fusion parameters of each gas by merging the obtained Gaussian mixture model, wherein the peak fusion parameters include the mean value, standard deviation and amplitude of the Gaussian distribution of the peak;

[0070] S105, result verification, verifies whether the fused result can accurately describe the actual gas component peak.

[0071] The following is a description of the Gaussian fusion algorithm for data fusion in conjunction with a specific embodiment:

[0072] 1. Data Preprocessing and Peak Detection

[0073] First, the data of the chromatograph and spectrometer are preprocessed, including peak detection and peak parameter extraction. Usually, classical methods in one-dimensional signal processing can be used to detect peaks, such as finding local maxima or setting thresholds. For a spectrometer sensitive to a single gas, there will only be one peak during the entire measurement process. The position of the peak of the same gas may be slightly deviated. Before fitting, several data points need to be moved to keep the peak positions of different sensors consistent.

[0074] 2. Gaussian distribution fitting

[0075] Assume that in the data of each sensor, we have detected one or more gas peaks. For each peak, we can fit it with a Gaussian distribution. The probability density function of the Gaussian distribution is as follows:

[0076]

[0077] Where: μ is the mathematical expectation of the Gaussian distribution, representing the center position of the peak; σ 2 is the variance, representing the width of the peak.

[0078] 3. Gaussian mixture model construction

[0079] Assume that we have several sensors, and a certain gas is detected by multiple sensors, that is, the peaks of the same gas are detected at the same time, then the N peaks are fused. Take H2 as an example:

[0080]

[0081] Where λ is the weight of the ith component (peak) and satisfies And i >0,μ i and σ i 2 are the mean and variance of the ith Gaussian distribution.

[0082] 4. Model parameter fitting

[0083] Use the Expectation Maximization (EM) algorithm to fit the parameters of the Gaussian mixture model. The steps of the EM algorithm are as follows:

[0084] First, the Gaussian distribution can be simplified as, which means the k-th Gaussian distribution in sample x i The probability density function at :

[0085] p(x i )=N(x i ;μ k ,σ k 2 )

[0086] E step (Exception Step):

[0087] Calculate the posterior probability γ(z ik ), where z ik Represents the probability that the i-th sample belongs to the k-th component:

[0088]

[0089] M step (Maximization Step):

[0090] Update the model parameters based on the obtained posterior probability

[0091]

[0092]

[0093]

[0094] Repeat the E and M steps until convergence (such as the parameter change is less than a certain threshold or the maximum number of iterations is reached) to obtain the optimal parameters.

[0095] 5. Peak Fusion

[0096] Eventually, the peaks detected by each sensor will be fused into one or more more accurate peaks. The parameters (mean, variance, weight) of the fused peaks can be obtained by fitting the parameters of the Gaussian mixture model.

[0097] 6. Result Verification

[0098] It is verified that the fused peaks can more accurately describe the actual gas component peaks, and the fitting effects of the fitted Gaussian mixture model and the original data are compared.

[0099] In the process of Gaussian fusion processing algorithm, when the original data is interfered and the Gaussian normal distribution characteristics are not obvious, the Gaussian fitting may fail, that is, the parameters cannot converge or the spectrum peak after fusion is too different from the original data. In this case, the Gaussian fusion algorithm will not be applicable, and the fusion neural network will be selected for processing. Through multimodal learning, the neural network is used to extract and learn the features of the chromatographic and spectral data respectively, and the extracted features are fused in the middle or last layer of the network, and finally the concentration value of each component gas is directly output.

[0100] Specifically, the specific implementation steps of the neural network fusion model include:

[0101] S201.Build the network: Design a multimodal neural network to process chromatographic and spectral data separately.

[0102] S202. Feature extraction: Extract the features of the spectrum and spectrum respectively through the first few layers of the network.

[0103] S203. Feature fusion: Feature fusion is performed in the middle layer or the last layer of the network.

[0104] S204. Subsequent layers: Use fully connected layers, etc. for numerical regression tasks.

[0105] Using neural network for data fusion can realize comprehensive analysis and processing of gas component data detected by spectroscopy and chromatography. This method can automatically learn the complex relationship between different sensor data, thereby improving the accuracy and reliability of fault diagnosis. The following is an explanation of the data fusion realized by the neural network fusion model in conjunction with a specific embodiment.

[0106] like Figure 4 As shown in the figure, the construction of the neural network fusion model includes:

[0107] 1. Data preparation

[0108] First, we need to prepare the gas composition data from spectroscopy and chromatography. Assume we have the following data:

[0109] Chromatographic data: Each sample contains time series data of the concentration of multiple gas components (e.g., hydrogen, acetylene, etc.).

[0110] Spectral data: Each sample contains spectral intensity values ​​at multiple wavelengths.

[0111] 2. Data Preprocessing

[0112] Standardize the data so that the data from different sensors have the same scale. Usually, the concentration of each gas component can be standardized to a range of mean 0 and standard deviation 1, which is convenient for network calculation and easier to converge. Perform filtering and smoothing to remove noise and reduce the difficulty of model training.

[0113] 3. Build a neural network model

[0114] a. Input layer

[0115] The input layer of the neural network receives the preprocessed spectral and chromatographic data, and performs convolution calculations on the spectrum data using 1D convolution kernels. Different from the Gaussian fusion algorithm, it is not necessary to separate the peaks of each component, but to extract the characteristic values ​​of the peaks through the calculation of the convolution kernel.

[0116] b. Hidden Layer

[0117] The hidden layer is used to further learn the complex relationship between different gas components, extract the data features of the components, and provide effective data for feature fusion. The one-dimensional convolution kernel is still used for convolution calculation, and the pooling layer is used to reduce the data dimension. The more layers there are, the more complex the model is, the higher the accuracy is, the more difficult it is to converge, and a larger data set is required.

[0118] c. Feature Fusion

[0119] The fusion stage integrates the data features of different modes. Through fusion, the characteristics of multimodal data can be comprehensively utilized to improve the performance of the model. The fused features contain information from data of different modes and can provide a more comprehensive and accurate representation. Commonly used methods include concatenation, weighted averaging, and attention mechanisms.

[0120] d. Fully connected layer

[0121] The fully connected layer is used to further process and compress the fused features for the final decision or output. The fully connected layer can map high-dimensional features to low-dimensional space, aggregate information, and provide input for the output layer. Usually, the number of nodes is large and non-linear activation functions are used.

[0122] f. Output layer

[0123] The output layer can have different settings depending on the task. If the goal is to classify the fault type, the number of nodes in the output layer is the number of fault types, and the activation function can be Softmax. If the goal is to predict the gas concentration, the number of nodes in the output layer is the number of gas components, and the activation function can be a linear function.

[0124] 4. Model Training

[0125] Use historical data to train the neural network. Choose an appropriate loss function and optimization algorithm, such as cross entropy loss (for classification) or mean square error (for regression). The optimization algorithm can be Adam or SGD (stochastic gradient descent).

[0126] 5. Data fusion results

[0127] a. Output fault type

[0128] If the output layer of the neural network is the fault type classification, the fusion result is the probability distribution of each fault type. Through the comprehensive analysis of different gas component data, the neural network can output the probability of each fault type to help fault diagnosis.

[0129] b. Output gas concentration

[0130] If the output layer of the neural network is a gas concentration prediction, the fusion result is the integrated gas component concentration. These concentration values ​​can be further used for decision analysis or fault diagnosis.

[0131] 6. Mobile network deployment

[0132] Existing neural network training generally relies on Python or Tensorflow, that is, training is performed on the Python platform, and after the training results are exported, a suitable embedded platform needs to be selected for deployment. The current technical solutions include directly selecting a board that can run Python; a computing card with an NPU; using FPGA for parallel accelerated computing; and using the exported network parameters to reconstruct the convolutional neural network using C language code.

[0133] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. A person of ordinary skill in the art can change, modify, replace and modify the above embodiments within the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A transformer oil dissolved gas detection system, characterized in that: The invention comprises a degassing unit, a chromatographic column, a spectrum detector, a chromatographic detector and a data fusion processing unit. The degassing unit degasses transformer oil. The sample gas is purged with a carrier gas and passes through the chromatographic column to separate gas components. The separated gas enters the spectrum detector and the chromatographic detector to perform qualitative and quantitative analysis on each gas component, and generate corresponding chromatographic data and spectrum data. The data fusion processing unit performs fusion processing on multi-sensor data based on a Gaussian fusion processing algorithm and a neural network fusion model to obtain fusion processing data of gas detection.

2. A transformer oil dissolved gas detection system as claimed in claim 1, characterized in that: The data fusion processing unit first performs peak height detection on the spectral data and the chromatographic data. When the difference between the two data exceeds a threshold, the sensor is judged to be abnormal and no subsequent algorithm processing is performed. The standard gas is used for re-detection. If the data difference does not exceed the threshold, further data fusion processing is performed.

3. A transformer oil dissolved gas detection system as claimed in claim 1, characterized in that: The Gaussian fusion processing algorithm comprises the following steps: S101, Gaussian distribution fitting, performing Gaussian distribution fitting on the data peaks of the chromatographic data and the spectral data to obtain Gaussian distribution fitting parameters, including the mean and standard deviation of the data peaks, and the amplitude of the Gaussian distribution; S102, Gaussian mixture model construction, using the fitted Gaussian distribution parameters as initial parameters to construct the Gaussian mixture model; S103, model fitting, optimizing the Gaussian mixture model using a data fitting algorithm to obtain optimal mixture model parameters; S104, peak fusion, obtaining peak fusion parameters of each gas by merging the obtained Gaussian mixture model, wherein the peak fusion parameters include the mean value, standard deviation and amplitude of the Gaussian distribution of the peak; S105, result verification, verifies whether the fused result can accurately describe the actual gas component peak.

4. A transformer oil dissolved gas detection system as claimed in claim 3, characterized in that: In step 101, it is assumed that one or more gas peaks have been detected in the data of each sensor. For each peak, a Gaussian distribution can be used to fit it, and the probability density function of the Gaussian distribution is as follows: Where: μ is the mathematical expectation of the Gaussian distribution, representing the center position of the peak; σ 2 is the variance, representing the width of the peak.

5. A transformer oil dissolved gas detection system as claimed in claim 3, characterized in that: In step 102, assuming that there are multiple sensors, a certain gas is detected by multiple sensors, that is, the peaks of the same gas are detected at the same time, then the N peaks are fused. The following takes H2 gas as an example: Where λ is the weight of the ith peak and satisfies And i >0,μ i and σ i 2 is the mathematical expectation and variance of the i-th Gaussian distribution.

6. A transformer oil dissolved gas detection system as claimed in claim 3, characterized in that: In step 103, the parameters of the Gaussian mixture model are fitted using an expectation maximization algorithm, where: The Gaussian distribution is simplified as, which means the kth Gaussian distribution in sample x i The probability density function at : p(x i )=N(x i ;m k ,s k 2 ) Calculate the posterior probability γ(z ik ), where z ik Represents the probability that the i-th sample belongs to the k-th component: According to the obtained posterior probability, update the model parameters: Repeat the above steps until the parameters converge and obtain the optimal parameters.

7. A transformer oil dissolved gas detection system as claimed in claim 3, characterized in that: In step 104, the peaks detected by each sensor are fused into one or more more accurate peaks, and the parameters of the fused peaks are obtained by fitting the Gaussian mixture model parameters.

8. A transformer oil dissolved gas detection system as claimed in claim 3, characterized in that: In the process of Gaussian fusion processing algorithm, when it is impossible to fit the spectrum using Gaussian normal distribution characteristics, the neural network fusion model is selected to fuse the multi-sensor data.

9. A transformer oil dissolved gas detection system as claimed in claim 8, characterized in that: The neural network fusion model extracts and learns features of spectral data and chromatographic data respectively, independently extracts features of spectral data of different sensors in the input layer and hidden layer of the neural network, fuses the extracted features and processes the data in the middle layer of the neural network, and finally summarizes the data in the output layer to output the concentration information of each gas component.

10. A transformer oil dissolved gas detection system as claimed in claim 9, characterized in that: The neural network fusion model includes: An input layer receives the preprocessed spectral and chromatographic data, and uses different 1D convolution kernels to perform convolution calculations on the spectrum data to extract the characteristic values ​​of the peaks; Hidden layer, used to further learn the relationship between different gas components and extract data features of the components; Feature fusion integrates feature data from different modes so that data collected by different sensors can be calculated in the same dimension; The fully connected layer is used to further process and compress the fused data features; The output layer outputs classification results according to different task types, including fault types and predicted gas concentrations.

Images