A flow-type detection device for transformer oil quality

Abstract

The present invention provides a flow-type detection device for transformer oil quality, which belongs to the technical field of transformer oil, and comprises: a dark box, in which a flow-type device, a laser detection device and a computing chip are fixedly arranged; the flow-type device comprises an oil inlet pipe, a liquid pump and an oil spray nozzle, the flow-type device is used to spray the transformer oil to be tested in a thin stream, the laser detection device is used to irradiate a laser beam on the sprayed oil thin stream, and collect the laser beam after being blocked and scattered by the oil thin stream to generate an electrical signal, and the computing chip is used to calculate the oil quality of the transformer oil to be tested according to the electrical signal. The oil inlet pipe is used to input the transformer oil to be tested, the oil inlet pipe is connected to the liquid pump, and is used to input the transformer oil to be tested into the liquid pump, and the output end of the liquid pump is fixedly provided with an oil spray nozzle, the oil spray nozzle is conical, and the top of the oil spray nozzle is provided with an oil outlet hole for spraying the oil thin stream.

Description

Technical Field

[0001] The invention belongs to the technical field of transformer oil, and in particular relates to an oil quality flow detection device for transformer oil. Background Art

[0002] Transformer oil is an indispensable and important insulating medium in the power system. The quality of transformer oil is directly related to the normal operation and service life of the transformer. With the long-term operation of the transformer, the transformer oil will gradually age, and various pollutants and aging products will appear, such as colloid, metal particles, fiber flocs, moisture, oxidation products, etc., which seriously affect the insulation performance of the transformer oil and increase safety hazards. Therefore, timely and effective monitoring of transformer oil quality is of vital importance to prevent transformer accidents and extend the service life of equipment.

[0003] At present, the commonly used transformer oil quality detection method is chemical analysis, which refers to the determination of physical and chemical indicators such as acid value, insulation strength, kinematic viscosity, dielectric strength, etc. in transformer oil through chemical tests, and then the quality grade of transformer oil is evaluated based on these indicators. Although this type of method has high detection accuracy, the process is complicated, time-consuming, and requires a large amount of transformer oil samples, and cannot achieve online continuous monitoring. Summary of the invention

[0004] In view of this, the present invention provides a flow-type detection device for transformer oil quality, which can solve the technical problems of the prior art for transformer oil quality detection, namely, the process is complicated, time-consuming, and a large amount of transformer oil samples are required, and online continuous monitoring cannot be achieved.

[0005] The present invention is achieved in that:

[0006] A first aspect of the present invention provides a transformer oil quality flow detection device, which includes a dark box, in which a flow device, a laser detection device and a computing chip are fixedly arranged; the flow device includes an oil inlet pipe, a liquid pump and an oil spray nozzle, the flow device is used to spray the transformer oil to be tested in a thin stream, the laser detection device is used to irradiate a laser beam on the sprayed oil thin stream, and collect the laser beam after being blocked and scattered by the oil thin stream to generate an electrical signal, and the computing chip is used to calculate the oil quality of the transformer oil to be tested according to the electrical signal;

[0007] The oil inlet pipe is used to input the transformer oil to be tested, and the oil inlet pipe is connected to the liquid pump, and is used to input the transformer oil to be tested into the liquid pump. The output end of the liquid pump is fixedly provided with an oil spray nozzle, and the oil spray nozzle is conical, and the top of the oil spray nozzle is provided with an oil outlet hole, which is used to spray a thin stream of oil;

[0008] The laser detection device includes two groups of laser detection structures, and the laser detection structures include a laser emitter and a laser sensor. The laser emitter is used to emit a laser beam to irradiate the oil stream, and the laser sensor is used to receive the laser beam passing through the oil stream and convert it into an electrical signal; wherein the two groups of laser detection structures are respectively a first laser detection structure and a second laser detection structure, the incident direction of the laser beam of the first laser detection structure is perpendicular to the plane where the oil stream is located, and the laser beam of the first laser structure intersects with the oil stream, and the intersection point is recorded as a first intersection point; the laser beam of the second laser detection structure is at an angle of 30° to 60° with the plane where the oil stream is located, and the laser beam of the second laser structure intersects with the oil stream at a second intersection point, and the first intersection point coincides with the second intersection point; the electrical signal collected by the first laser detection structure is recorded as a first electrical signal, and the electrical signal collected by the second laser detection structure is recorded as a second electrical signal.

[0009] The computing chip is provided with an oil quality detection module for performing the following steps:

[0010] S10, continuously acquiring the first electrical signal and the second electrical signal;

[0011] S20, preprocessing the first electrical signal and the second electrical signal to obtain a first signal and a second signal;

[0012] S30, acquiring features of the first signal and the second signal, recording them as a first feature and a second feature respectively, and fusing them into a fused feature;

[0013] S40, using a pre-trained laser-oil quality model, inputting the fusion features, and obtaining the oil quality of the transformer oil to be tested.

[0014] Wherein, the laser beam of the second laser detection structure forms an angle of 45° with the plane where the oil thin stream is located.

[0015] Furthermore, the specific implementation method of the S10 step is: continuously acquiring the first electrical signal and the second electrical signal in real time at a preset acquisition frequency threshold, and temporarily storing the obtained raw electrical signal data in a ring buffer. The acquisition frequency threshold is 1KHz, that is, 1000 data points are collected per second. This frequency can realize real-time detection of the oil flow.

[0016] Furthermore, the specific implementation method of the S20 step is: performing the following preprocessing on the first electrical signal and the second electrical signal to obtain the first signal and the second signal: using a wavelet transform algorithm to perform denoising, selecting different wavelet basis functions to perform wavelet transform on the first electrical signal and the second electrical signal respectively, selecting a suitable high-frequency threshold based on experience, setting the wavelet coefficients above the threshold to zero, and then performing wavelet reconstruction to obtain the denoised signal; performing amplitude normalization calibration on the denoised signal, unifying the signal amplitude to the range of 0-1, thereby eliminating the influence of signal amplitude fluctuations.

[0017] Furthermore, the specific implementation method of step S30 is: adopting a method combining Fourier transform and wavelet packet decomposition to obtain the characteristics of the first signal and the second signal, and the specific steps are: performing Fourier transform and wavelet packet decomposition on the first signal and the second signal respectively, selecting the amplitude of the main frequency component from the spectrum obtained by Fourier transform as the frequency feature, and obtaining the low-frequency approximation coefficients and various levels of detail coefficients from the wavelet packet decomposition as the time scale feature; cascading the feature vectors of the first signal and the second signal into a fused feature vector.

[0018] Furthermore, the laser-oil quality model adopts a support vector machine or an artificial neural network.

[0019] Furthermore, the training steps of the laser-oil quality model are specifically: collecting a large number of transformer oil samples, manually inspecting and rating them to obtain labels, using the present device to perform laser inspection on each sample to obtain the original electrical signal, extracting the fused feature vector and its label of each sample through steps S10-S30 to form a complete training data set, and using the training data set to perform model training.

[0020] Wherein, in step S20, the preprocessing of the first electrical signal and the second electrical signal further includes: detecting the DC component of the signal, and eliminating the DC offset if there is a DC offset.

[0021] Furthermore, in step S30, before acquiring the features of the first signal and the second signal, the first signal and the second signal are first subjected to denoising processing, and the denoising processing adopts a wavelet threshold denoising method, which specifically includes: decomposing the wavelet coefficients of the first signal and the second signal; calculating the difference between the absolute value of the wavelet coefficients at each level and the threshold, and setting the coefficients less than the threshold to zero; and reconstructing the signal to obtain the denoised first signal and the second signal.

[0022] Compared with the prior art, the transformer oil quality flow detection device provided by the present invention has the following beneficial effects:

[0023] 1. Nondestructive rapid detection: Compared with traditional chemical analysis and electrical measurement methods, the present invention adopts non-contact optical principles for nondestructive detection, does not require the consumption of a large amount of samples, has extremely high online real-time detection efficiency, and can meet the plug-and-play requirements of field applications.

[0024] 2. High detection accuracy: By irradiating the liquid flow from two different angles at the same time and collecting optical signals, the structural information inside the liquid can be obtained; combined with signal processing algorithms such as wavelet transform and Fourier transform, the mapping relationship between the transformer oil quality and the optical signal can be accurately extracted, thereby improving the accuracy of the detection results.

[0025] 3. Automatic intelligent detection: By introducing machine learning technology, it is possible to automatically establish a laser-oil quality model without relying on manual experience for judgment, avoiding the influence of subjective factors and improving the objectivity and consistency of detection.

[0026] As mentioned above, the solution of the present invention solves the technical problems of the prior art for detecting the quality of transformer oil, which has the characteristics of complex process, long time consumption, need to consume a large amount of transformer oil samples, and cannot realize online continuous monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0028] Figure 1 A schematic diagram of the structure of the device provided by the present invention;

[0029] Figure 2 A flowchart of the steps performed by the oil quality detection module;

[0030] Figure 3 Schematic diagram of the incident laser beam being scattered;

[0031] In the accompanying drawings, the components represented by the reference numerals are listed as follows:

[0032] 00. Dark box; 1. Oil inlet pipe; 10. Oil trickle; 2. Liquid pump; 3. Fuel injector; 4. Laser transmitter; 5. Laser sensor. DETAILED DESCRIPTION

[0033] 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.

[0034] like Figure 1 The figure is a schematic diagram of the structure of the device provided by the present invention. The device includes a dark box, in which a flow device, a laser detection device and a computing chip are fixedly arranged; the flow device includes an oil inlet pipe, a liquid pump and an oil spray nozzle, the flow device is used to spray the transformer oil to be tested in a thin stream, the laser detection device is used to irradiate the sprayed oil thin stream with a laser beam, and collect the laser beam after being blocked and scattered by the oil thin stream to generate an electrical signal, and the computing chip is used to calculate the oil quality of the transformer oil to be tested according to the electrical signal;

[0035] The oil inlet pipe is used to input the transformer oil to be tested. The oil inlet pipe is connected to the liquid pump and is used to input the transformer oil to be tested into the liquid pump. The output end of the liquid pump is fixedly provided with an oil spray nozzle. The oil spray nozzle is conical, and an oil outlet hole is opened at the top of the cone of the oil spray nozzle for spraying a thin stream of oil.

[0036] The laser detection device includes two groups of laser detection structures, and the laser detection structure includes a laser emitter and a laser sensor. The laser emitter is used to emit a laser beam to irradiate the oil stream, and the laser sensor is used to receive the laser beam passing through the oil stream and convert it into an electrical signal; wherein the two groups of laser detection structures are respectively a first laser detection structure and a second laser detection structure, the incident direction of the laser beam of the first laser detection structure is perpendicular to the plane where the oil stream is located, and the laser beam of the first laser structure intersects with the oil stream, and the intersection point is recorded as the first intersection point; the laser beam of the second laser detection structure is at an angle of 30° to 60° with the plane where the oil stream is located, preferably 45°, and the laser beam of the second laser structure intersects with the oil stream at the second intersection point, and the first intersection point coincides with the second intersection point; the electrical signal collected by the first laser detection structure is recorded as the first electrical signal, and the electrical signal collected by the second laser detection structure is recorded as the second electrical signal.

[0037] like Figure 2 As shown, the computing chip is provided with an oil quality detection module for performing the following steps:

[0038] S10, continuously acquiring the first electrical signal and the second electrical signal;

[0039] S20, preprocessing the first electrical signal and the second electrical signal to obtain a first signal and a second signal;

[0040] S30, acquiring features of the first signal and the second signal, recording them as a first feature and a second feature respectively, and fusing them into a fused feature;

[0041] S40, using a pre-trained laser-oil quality model, inputting the fusion features, and obtaining the oil quality of the transformer oil to be tested.

[0042] The specific implementation of the above scheme is described in detail below:

[0043] First, the oil quality flow detection device of transformer oil of the present invention mainly includes the following structural units: a dark box, a flow device, a laser detection device and a computing chip. The flow device, the laser detection device and the computing chip are fixedly arranged in the dark box. Among them, the flow device includes an oil inlet pipe, a liquid pump and an oil spray nozzle, which is used to spray the transformer oil to be tested in a thin stream; the laser detection device is used to irradiate the sprayed oil thin stream with a laser beam, and collect the laser beam after being blocked and scattered by the oil thin stream to generate an electrical signal; the computing chip is used to calculate the oil quality of the transformer oil to be tested according to the electrical signal.

[0044] Secondly, the present invention provides a specific structure of a flow device. The oil inlet pipe is used to input the transformer oil to be tested, and the oil inlet pipe is connected to the liquid pump, and is used to input the transformer oil to be tested into the liquid pump. The output end of the liquid pump is fixedly provided with an oil spray nozzle, and the oil spray nozzle is conical, and the top of the cone of the oil spray nozzle is provided with an oil outlet hole, which is used to spray a thin stream of oil. The structural design of the flow device enables the transformer oil to be tested to be continuously and stably sprayed out from the oil spray nozzle in the form of a thin liquid stream, providing a stable detection object for the subsequent laser detection device.

[0045] Again, the present invention provides a specific structure of a laser detection device. The laser detection device includes two groups of laser detection structures, the laser detection structures include a laser emitter and a laser sensor, the laser emitter is used to emit a laser beam to irradiate the oil thin stream, and the laser sensor is used to receive the laser beam passing through the oil thin stream and convert it into an electrical signal. Among them, the two groups of laser detection structures are respectively a first laser detection structure and a second laser detection structure, the incident direction of the laser beam of the first laser detection structure is perpendicular to the plane where the oil thin stream is located, and the laser beam of the first laser structure intersects with the oil thin stream, and the intersection point is recorded as the first intersection point; the laser beam of the second laser detection structure is at an angle of 30° to 60° with the plane where the oil thin stream is located, and the laser beam of the second laser structure intersects with the oil thin stream at the second intersection point, and the first intersection point coincides with the second intersection point; the electrical signal collected by the first laser detection structure is recorded as the first electrical signal, and the electrical signal collected by the second laser detection structure is recorded as the second electrical signal. The structural design of the laser detection device enables the laser beam to irradiate the oil flow from different angles, thereby obtaining light scattering information at different angles, thereby improving the comprehensiveness and accuracy of the detection.

[0046] Finally, the present invention provides a specific implementation of an oil quality detection module provided in a computing chip. The oil quality detection module is used to perform the following steps:

[0047] The specific implementation of step S10 is: continuously acquiring the first electrical signal and the second electrical signal. The purpose of this step is to continuously acquire two sets of electrical signal data from the laser detection device as the original input for subsequent processing. In specific implementation, a collection frequency threshold can be set, such as 1KHz, that is, 1000 data points are collected per second, so as to ensure real-time detection of oil flow. The collected raw electrical signal data is temporarily stored in a ring buffer for subsequent real-time processing.

[0048] The specific implementation of step S20 is: pre-processing the first electrical signal and the second electrical signal to obtain the first signal and the second signal. The purpose of this step is to perform pre-processing such as denoising and calibration on the original electrical signal to improve the signal quality. In the specific implementation, the wavelet transform algorithm can be used to denoise the original electrical signal. Wavelet transform is a time-frequency analysis method that can perform localized analysis on the signal in both time and frequency directions, and is suitable for the analysis of non-stationary signals. For the first electrical signal and the second electrical signal, different wavelet basis functions can be used for wavelet transform respectively, and a suitable high-frequency threshold is selected according to experience, and the wavelet coefficients above the threshold are set to zero to obtain new wavelet coefficients, and then wavelet reconstruction is performed to obtain the denoised signal. In addition, it is also necessary to perform amplitude calibration on the signal, and normalize the signal amplitude to the range of 0-1 to eliminate the influence of factors such as laser power fluctuations. The first signal and the second signal obtained after pre-processing have a higher signal-to-noise ratio and a more uniform amplitude, which lays the foundation for feature extraction.

[0049] The specific implementation of step S30 is: obtaining the features of the first signal and the second signal, recording them as the first feature and the second feature respectively, and merging them into a fusion feature. The purpose of this step is to extract the key features that can reflect the quality of the oil product from the preprocessed signal. In the specific implementation, the feature can be extracted by combining Fourier transform and wavelet packet decomposition. Fourier transform can convert the time domain signal to the frequency domain, so as to obtain the frequency characteristics of the signal; wavelet packet decomposition can perform multi-scale decomposition on the signal at different time scales, so as to obtain the details and approximate features of the signal at different time scales. For the first signal and the second signal, Fourier transform and wavelet packet decomposition can be performed respectively, and the amplitude of the main frequency component is selected from the spectrum obtained by Fourier transform as the frequency feature, and the low-frequency approximate coefficient and the detail coefficients at all levels obtained from the wavelet packet decomposition are used as the time scale feature. Finally, the feature vectors of the first signal and the second signal are cascaded to obtain the fusion feature vector. It is worth noting that the parameter setting of feature extraction has a significant impact on the final detection performance, and a large amount of experimental data is needed to optimize the parameters. For example, the number of levels of wavelet packet decomposition, the frequency interval of frequency feature selection and other parameters need to be optimized.

[0050] The specific implementation of step S40 is: using a pre-trained laser-oil model, inputting the fusion feature, and obtaining the oil quality of the transformer oil to be tested. The purpose of this step is to use a machine learning model to infer the specific quality level of the oil from the fusion feature. In specific implementation, a supervised learning model such as a support vector machine (SVM) or an artificial neural network can be used as the laser-oil model.

[0051] The training process of the laser-oil model is as follows:

[0052] 1. Acquisition of training data set: A large number of transformer oil samples need to be collected, and professionals manually test and rate each sample, and use the test results as labels. At the same time, each sample is laser tested using the device of the present invention and the original electrical signal data is recorded. After the feature extraction of the above steps S10-S30, the fused feature vector and its label corresponding to each sample can be obtained to form a complete training data set.

[0053] 2. Model training: Select the appropriate model architecture and related hyperparameters. For example, for SVM, you can choose the Gaussian kernel function, and for neural network, you can choose a 3-layer fully connected network structure. Input the training data set into the model for training, and cross-validation and other methods are required to prevent overfitting. The goal of model training is to enable the model to accurately predict the quality grade label of the oil product from the fused feature vector.

[0054] 3. Model evaluation: Evaluate the performance indicators of the trained model on an independent test data set, such as accuracy, precision, recall, etc., to determine whether the generalization ability of the model meets the requirements. If the evaluation result is not ideal, it is necessary to readjust the model architecture, hyperparameters, training strategies, etc. until a satisfactory model is obtained.

[0055] After obtaining a high-quality laser-oil model, it can be applied to the actual online detection process. The transformer oil to be tested is irradiated with laser by the device of the present invention to obtain the original electrical signal, and the fused feature vector obtained by the processing of the above steps S10-S30 is input into the trained model, so that the oil quality data of the transformer oil can be predicted in real time.

[0056] It should be noted that the specific parameter settings and algorithm selections for the above steps need to be determined based on a large number of experimental comparative analyses according to the actual situation. This embodiment only provides a possible implementation. In addition, in addition to the above method, other feature extraction methods and machine learning models can also be selected to obtain higher detection accuracy and robustness. Other feature extraction methods and machine learning model selections are all within the scope of protection of the present invention.

[0057] The following is a detailed description of a transformer oil quality flow detection device and method of the present invention through a specific embodiment.

[0058] First, the structural part of the device, such as Figure 1 As shown, the device is mainly composed of a dark box 00, a flow device, a laser detection device and a computing chip.

[0059] The dark box 00 is a closed black box without any light source interference inside, and is used to simulate a dark environment. A flow device, a laser detection device and a computing chip are fixedly installed inside the dark box 00.

[0060] The flow device comprises an oil inlet pipe 1, a liquid pump 2 and an oil spray nozzle 3. The oil inlet pipe 1 is a detachable structure and can be replaced according to different pipe diameter specifications to adapt to different types of transformer oil. The present embodiment uses an oil inlet pipe 1 with an inner diameter of 10 mm. One end of the oil inlet pipe 1 is connected to an external oil storage tank (not shown in the figure), and the other end is connected to the oil inlet of the liquid pump 2. The liquid pump 2 selects a micro peristaltic pump with a flow range of 0.01-45 mL / min. The oil spray nozzle 3 adopts a conical structure with a cone angle of 60 degrees, and an oil outlet hole with a diameter of 0.5 mm is opened on the top. The oil outlet of the liquid pump 2 is connected to the conical bottom of the oil spray nozzle 3.

[0061] The laser detection device includes a first laser detection structure and a second laser detection structure. The first laser detection structure consists of a laser emitter 4 and a laser sensor 5. The laser emitter 4 has a wavelength of 650nm and an optical power of 5mW; the corresponding wavelength response range of the laser sensor 5 is 600nm-700nm. The laser beam of the first laser detection structure is perpendicular to the plane where the oil stream 10 ejected from the injector 3 is located, and the laser beam and the oil stream 10 intersect vertically at the first intersection point. The second laser detection structure 2 consists of a laser emitter and a laser sensor, and the specifications and models are the same as those of the first laser detection structure. The laser beam of the second laser detection structure is at an angle of 45 degrees to the plane where the oil stream 10 is located, and the laser beam intersects with the oil stream at the second intersection point, and the first intersection point coincides with the second intersection point.

[0062] The computing chip is an STM32 chip equipped with Linux system. The computing chip integrates software function modules such as oil detection module, data transmission module, fault diagnosis module, etc.

[0063] The second is the specific implementation process of the oil testing method.

[0064] S10, continuously acquiring a first electrical signal and a second electrical signal.

[0065] In this embodiment, the acquisition frequency threshold is set to 2KHz, that is, 2000 data points are collected per second to meet the real-time requirements of continuous detection. The collected original first electrical signal and second electrical signal are first temporarily stored in the ring buffer of the computing chip 4.

[0066] S20, preprocessing the first electrical signal and the second electrical signal to obtain a first signal and a second signal.

[0067] In this embodiment, the preprocessing method adopted is: first, the original electrical signal is subjected to wavelet transform using the Daubechies4 wavelet basis, the wavelet decomposition level is 5, and the high-frequency threshold is selected as 5 times the empirical value, and the wavelet coefficients higher than the threshold are set to zero, thereby filtering out the high-frequency noise components, and then wavelet reconstruction is performed to obtain the denoised first signal and the second signal; the denoised signal is amplitude normalized so that the range of all signal values ​​is unified to the interval of 0 to 1, eliminating the influence of the dimension; finally, the DC component of the denoised signal is detected, and if there is a DC offset, the signal is subjected to moving average filtering, the average value is calculated, and the average value is subtracted from the original signal as the estimated DC component.

[0068] S30: Obtain features of the first signal and the second signal, record them as the first feature and the second feature respectively, and fuse them into a fusion feature.

[0069] In this embodiment, the first signal and the second signal are first subjected to wavelet threshold denoising, and denoised again on the basis of step S20 to further improve the signal quality. Specifically, the Sym5 wavelet is used, the decomposition level is 3, the threshold is selected as the Rigrsure threshold estimation, and the hard threshold processing method is used. Then the first signal and the second signal are respectively subjected to Fourier transform and wavelet packet decomposition, the amplitudes of the first five main frequency components are selected from the spectrum obtained by Fourier transform as the frequency feature, the low-frequency approximate coefficients and the detail coefficients at all levels obtained from the wavelet packet decomposition are used as the time scale feature, and the energy feature and coherence feature of the signal are also extracted, and finally the above feature vectors are cascaded into a fused feature vector with a length of 256.

[0070] S40, using a pre-trained laser-oil model, inputting fusion features, and obtaining the oil quality of the transformer oil to be tested.

[0071] In this embodiment, the laser-oil model adopts a 3-layer fully connected feedforward neural network structure, with an input layer of 256 neurons (the same length as the fusion feature vector), a first hidden layer of 120 neurons, using the ReLU activation function, a second hidden layer of 64 neurons, and an output layer of 5 neurons, corresponding to 5 oil quality grades (non-aged, slightly aged, moderately aged, severely aged, and severely aged), and the output layer uses the Softmax activation function.

[0072] The training process of the model is as follows: first, 100 transformer oil samples with marked quality grades are collected, and the original electrical signal data of each sample is obtained by using the device of the present invention. After the feature extraction of S10-S30, a training data set containing 100 samples is formed. The training data set is randomly divided into a training set and a validation set at a ratio of 8:2. The model parameters are trained on the training set, and the validation set is used to evaluate the model performance. The cross entropy is used as the loss function, and the Adam optimizer is used to iteratively update the model parameters. During the training process, the performance of the current model is evaluated on the validation set after each Epoch is trained. When the loss value on the validation set of 10 consecutive Epochs does not decrease, it is determined that the model has converged and the training is terminated. Finally, the performance indicators of the model evaluated on the test set are: accuracy 88.2%, recall rate 87.1%, precision rate 89.6%, and F1 score 88.3%.

[0073] It can be seen from the above specific embodiments that the transformer oil flow detection device and method of the present invention can realize rapid detection of transformer oil quality with high efficiency.

[0074] Specifically, the principle of the present invention is: Figure 3 As shown in the figure, there are a lot of tiny impurity particles, bubbles, insulating fibers and other scatterers in transformer oil. If the transformer oil is sprayed out in the form of a fine liquid stream and irradiated with a laser beam, the laser beam will inevitably be blocked and scattered by the scatterers. At this time, the light signal passing through the liquid stream is collected and analyzed, which can reflect the subtle differences in the internal structure of the liquid stream, thereby evaluating the quality of the transformer oil.

[0075] In order to realize the technical solution, the present invention designs a special device structure. First, the transformer oil to be tested is sprayed into a fine liquid flow using a liquid pump and a conical nozzle. Secondly, two groups of laser detection devices with different angles are used to irradiate the liquid flow with light, one group of which is at a vertical angle, and the other group is at an inclined angle of 30-60°. The purpose of this dual-angle detection structure is to be able to more comprehensively collect light scattering information at different orientations inside the liquid flow, and to improve the sensitivity and robustness of the detection. Finally, the present invention proposes a complete set of signal processing and model training methods, which can automatically extract features that are highly correlated with oil quality from the collected photoelectric signals, and construct an accurate laser-oil model to complete the intelligent evaluation of oil quality.

[0076] 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 transformer oil quality flow detection device, characterized in that: It comprises a dark box, in which a flow device, a laser detection device and a computing chip are fixedly arranged; the flow device comprises an oil inlet pipe, a liquid pump and an oil spray nozzle, the flow device is used to spray the transformer oil to be tested in a thin stream, the laser detection device is used to irradiate the sprayed oil thin stream with a laser beam, and collect the laser beam after being blocked and scattered by the oil thin stream to generate an electrical signal, and the computing chip is used to calculate the oil quality of the transformer oil to be tested according to the electrical signal; The oil inlet pipe is used to input the transformer oil to be tested, and the oil inlet pipe is connected to the liquid pump, and is used to input the transformer oil to be tested into the liquid pump. The output end of the liquid pump is fixedly provided with an oil spray nozzle, and the oil spray nozzle is conical, and the top of the oil spray nozzle is provided with an oil outlet hole, which is used to spray a thin stream of oil; The laser detection device includes two groups of laser detection structures, and the laser detection structures include a laser emitter and a laser sensor. The laser emitter is used to emit a laser beam to irradiate the oil thin stream, and the laser sensor is used to receive the laser beam passing through the oil thin stream and convert it into an electrical signal; wherein the two groups of laser detection structures are respectively a first laser detection structure and a second laser detection structure, the incident direction of the laser beam of the first laser detection structure is perpendicular to the plane where the oil thin stream is located, and the laser beam of the first laser detection structure intersects with the oil thin stream, and the intersection point is recorded as a first intersection point; the laser beam of the second laser detection structure forms an angle of 30° to 60° with the plane where the oil thin stream is located, and the laser beam of the second laser detection structure intersects with the oil thin stream at a second intersection point, and the first intersection point coincides with the second intersection point; the electrical signal collected by the first laser detection structure is recorded as a first electrical signal, and the electrical signal collected by the second laser detection structure is recorded as a second electrical signal; The computing chip is provided with an oil quality detection module for performing the following steps: S10, continuously acquiring the first electrical signal and the second electrical signal; S20, preprocessing the first electrical signal and the second electrical signal to obtain a first signal and a second signal; S30, acquiring features of the first signal and the second signal, recording them as a first feature and a second feature respectively, and fusing them into a fused feature; S40, using a pre-trained laser-oil quality model, inputting the fusion feature, and obtaining the oil quality of the transformer oil to be tested; Among them, the liquid pump adopts a micro peristaltic pump with a flow range of 0.01-45mL / min; the fuel injector adopts a conical structure with a cone angle of 60 degrees, and a 0.5mm diameter oil outlet hole is opened on the top; the oil outlet of the liquid pump is connected to the conical bottom of the fuel injector; the first laser detection structure is composed of a laser emitter and a laser sensor, the laser emitter wavelength is 650nm, and the optical power is 5mW; the corresponding wavelength response range of the laser sensor is 600nm-700nm; the laser beam of the first laser detection structure is perpendicular to the plane where the oil stream is sprayed from the fuel injector, and the laser beam and the oil stream intersect perpendicularly at the first intersection point; the second laser detection structure is composed of a laser emitter and a laser sensor, and the specifications and models are the same as the first laser detection structure; the laser beam of the second laser detection structure is at an angle of 45 degrees to the plane where the oil stream is, and the laser beam and the oil stream intersect at the second intersection point, and the first intersection point coincides with the second intersection point.

2. The transformer oil quality flow detection device according to claim 1, characterized in that: The laser beam of the second laser detection structure forms an angle of 45° with the plane where the oil thin stream is located.

3. The transformer oil quality flow detection device according to claim 1, characterized in that: The specific implementation method of the S10 step is: continuously acquiring the first electrical signal and the second electrical signal in real time at a preset acquisition frequency threshold, and temporarily storing the obtained raw electrical signal data in a ring buffer. The acquisition frequency threshold is 1KHz, that is, 1000 data points are collected per second. This frequency can realize real-time detection of the oil flow.

4. The transformer oil quality flow detection device according to claim 1, characterized in that: The specific implementation of step S20 is as follows: the first electrical signal and the second electrical signal are preprocessed as follows to obtain the first signal and the second signal: a wavelet transform algorithm is used to perform denoising, different wavelet basis functions are selected to perform wavelet transform on the first electrical signal and the second electrical signal respectively, a suitable high-frequency threshold is selected based on experience, wavelet coefficients above the threshold are set to zero, and then wavelet reconstruction is performed to obtain the denoised signal; The denoised signal is amplitude normalized and calibrated to unify the signal amplitude to the range of 0-1, thereby eliminating the influence of signal amplitude fluctuation.

5. The transformer oil quality flow detection device according to claim 1, characterized in that: The specific implementation method of step S30 is: using a method combining Fourier transform and wavelet packet decomposition to obtain the characteristics of the first signal and the second signal, and the specific steps are: performing Fourier transform and wavelet packet decomposition on the first signal and the second signal respectively, selecting the amplitude of the main frequency component from the spectrum obtained by Fourier transform as the frequency feature, and selecting the low-frequency approximate coefficients and various levels of detail coefficients obtained from the wavelet packet decomposition as the time scale feature; cascading the feature vectors of the first signal and the second signal into a fused feature vector.

6. The transformer oil quality flow detection device according to claim 1, characterized in that: The laser-oil quality model adopts a support vector machine or an artificial neural network.

7. The transformer oil quality flow detection device according to claim 1, characterized in that: The training steps of the laser-oil quality model are specifically: collecting a large number of transformer oil samples, manually inspecting and rating to obtain labels, using the device to perform laser inspection on each sample to obtain the original electrical signal, extracting the fused feature vector and its label of each sample through steps S10-S30 to form a complete training data set, and using the training data set to perform model training.

8. The transformer oil quality flow detection device according to claim 1, characterized in that: In step S20, the preprocessing of the first electrical signal and the second electrical signal further includes: detecting a DC component of the signal, and eliminating the DC offset if there is a DC offset.

9. The transformer oil quality flow detection device according to claim 1, characterized in that: In step S30, before acquiring the features of the first signal and the second signal, the first signal and the second signal are first subjected to denoising processing, wherein the denoising processing adopts a wavelet threshold denoising method, specifically comprising: decomposing the wavelet coefficients of the first signal and the second signal; calculating the difference between the absolute value of the wavelet coefficients at each level and the threshold, and setting the coefficients less than the threshold to zero; and reconstructing the signal to obtain the denoised first signal and the second signal.

Images