Substation main transformer fault prediction method based on voiceprint features

Abstract

The invention relates to the technical field of transformer substation equipment fault prediction, in particular to a transformer substation main transformer fault prediction method based on voiceprint features. The method comprises the following steps: acquiring image data, fusion data and beam data of a detection object; wherein the fusion data comprises image data of the detection object and sound intensity distribution data at the position of the detection object; constructing a first input based on the image data and the fusion data; constructing a second input based on the beam data; processing the first input based on a first model to obtain a first feature vector; processing the second input based on a second model to obtain a second feature vector; splicing the first feature vector and the second feature vector to obtain a feature vector; and processing the feature vector based on the third model to obtain fault prediction output. According to the method, the fault prediction of equipment such as a main transformer of a transformer substation can be realized by using the hidden working states of the image data and the beam data.

Description

Technical Field

[0001] The present invention relates to the technical field of substation equipment fault prediction, and more specifically, to a method for predicting main transformer faults in a substation based on voiceprint features. Background Art

[0002] The inspection work of equipment in a substation is of great significance for maintaining the reliable operation of the equipment in the substation and ensuring the reliability of the power system. With the continuous increase in electricity consumption in China, it has brought great pressure to the operation and maintenance work of substations. With the development and upgrading of inspection technologies, most substations have basically realized an intelligent remote inspection system that combines machines, inspection robots, and on-line monitoring systems, and uses technologies such as visible light recognition and infrared thermal imaging to remotely inspect the operating status, equipment defects, and equipment failures of power equipment.

[0003] At present, many traditional industrial equipment operation and maintenance methods, such as the application of infrared and thermosensitive monitoring methods, provide a reference basis for monitoring the operating status of equipment, reduce the workload of manual inspections, and improve the efficiency of equipment status monitoring and the reference basis for evaluation. However, the above monitoring methods have the following disadvantages: (1) The sensors need to be installed in contact; for the existing sensor solutions, the installation positions are not flexible, and they need to be closely attached to the equipment (such as vibration sensors), which will affect the health of the equipment itself; (2) The inspection efficiency is low and it is easy to miss inspections; relying too much on manual regular inspections, the efficiency is low, the cost is high, and it is easy to miss inspections and misjudgments. Currently, applications such as infrared imagers (monitoring temperature, but only some faults will have temperature changes) and vibration sensors will result in missed inspections; (3) There is a lag in fault discovery; only after a fault occurs, it is known that there is a fault, and the response is not timely; currently, applications such as infrared and thermosensitive detection methods are used for fault location after a fault occurs, and they cannot play a role in fault prevention, and their value is limited. Moreover, the transformer physical parameters on which traditional methods are based also have corresponding defects in performing fault detection tasks. Specifically, it is difficult to realize on-line monitoring of insulation performance parameters, the temperature parameters are difficult to be widely applied due to the high price of the equipment, and the vibration parameters are not suitable for detecting high-field-strength positions because they need to measure the vibration source in contact.

[0004] On the other hand, with the development of artificial intelligence technology, it has provided unprecedented opportunities for the intelligent diagnosis of main grid equipment based on acoustic features. The signal recognition mode based on deep learning is more suitable for solving complex non-linear problems in power equipment than traditional information processing modes, and can flexibly and efficiently process massive complex information in terms of expression and modeling capabilities, providing more accurate and intelligent judgment results. However, the existing models are difficult to be applied to the fault prediction of substation equipment due to the lack of training samples. Summary of the Invention

[0005] The present invention provides a method for predicting the faults of main transformers in substations based on voiceprint features, which can overcome certain or some defects of the prior art.

[0006] A method for predicting the faults of main transformers in substations based on voiceprint features according to the present invention includes: Obtaining image data, fusion data and beam data of a detection object; wherein, the fusion data includes the image data of the detection object and the sound intensity distribution data at the detection object; Constructing a first input based on the image data and the fusion data; Constructing a second input based on the beam data; Processing the first input based on a first model to obtain a first feature vector; Processing the second input based on a second model to obtain a second feature vector; Concatenating the first feature vector and the second feature vector to obtain a feature vector; Processing the feature vector based on a third model to obtain a fault prediction output.

[0007] The method of the present invention can utilize the working states implied by the image data and the beam data to realize the fault prediction of equipment such as main transformers in substations. It can make relatively full use of image features and voiceprint features to realize fault prediction with a small sample size, and is particularly suitable for fault prediction under unbalanced sample data.

[0008] Preferably, The constructing of the first input based on the image data and the fusion data includes: Identifying the detection object in the image data to construct an identification frame; wherein, the identification frame is the smallest circumscribed rectangle frame that completely contains the detection object; Obtaining the image of the area covered by the identification frame and converting it into the grayscale image; Reconstructing the grayscale based on the sound intensity distribution data at the image of the area covered by the identification frame to obtain the first input.

[0009] Based on the above, it can preferably achieve the elimination of irrelevant image features, especially retain the image features of the area where the detection object is located. By integrating the sound intensity distribution data into the image data, it can preferably introduce the sound wave distribution information at the detection object, and introduce the difference information of the vibration parts of the detection object in different working states in the overall prediction method, so that it can better realize the prediction of anomalies.

[0010] Preferably, The reconstructing of the grayscale based on the sound intensity distribution data at the image of the area covered by the identification frame to obtain the first input includes: Obtaining the average grayscale value of the grayscale image ; wherein, , N is the total number of pixel points of the grayscale image, is the sum of the grayscale values of all pixel points of the grayscale image; Obtain the grayscale compensation value at each pixel point of the grayscale image ; wherein, , S is the sound intensity at the corresponding pixel point, is the maximum sound intensity of the image in the area covered by the recognition frame, is the minimum sound intensity of the image in the area covered by the recognition frame; Subtract the corresponding grayscale compensation value from the grayscale value at each pixel point in the grayscale image , and obtain the first input.

[0011] Based on the above, it is possible to preferably introduce sound intensity distribution data into the image data.

[0012] Preferably, Constructing the second input based on the beam data includes, Extract the main frequency eigenvalue, acoustic vibration entropy eigenvalue, 50Hz odd and even harmonic ratio eigenvalue, fundamental frequency ratio eigenvalue and high and low frequency eigenvalue in the beam data to construct a voiceprint feature sequence as the second input.

[0013] Based on the above, it is possible to obtain the voiceprint features with a typical normal distribution in the beam data to construct the second input. Based on this method, in particular, it is possible to mine the standard voiceprint feature distribution information from normal data, making the entire prediction method particularly sensitive to abnormal data and achieving an improvement in the prediction accuracy of the prediction model under the imbalance of sample data.

[0014] Preferably, The first model uses a convolutional neural network, which has a convolutional layer, an activation layer and a pooling layer; the second model uses a neural network, which has an input layer and a hidden layer; the third model includes a fully connected layer and an output layer; the first feature vector and the second feature vector are spliced into a feature vector at the fully connected layer and output to the output layer after weighted calculation by the fully connected layer. The output layer is used to obtain the probability of the fault category based on the output of the fully connected layer to complete the fault prediction output. Thus, it is possible to preferably realize the separate processing of different types of data and the fusion processing of data from different sources.

[0015] Preferably, The parameters of the convolutional layer, activation layer and pooling layer in the first model, the parameters of the hidden layer in the second model and the parameters of the fully connected layer in the third model are all obtained by separate training. This enables the use of richer sample data to train different models, which is beneficial to the improvement of model accuracy.

[0016] Preferably, When obtaining the parameters of the convolutional layer, activation layer, and pooling layer in the first model, it includes Constructing a first sample set with the abnormal category as the label; among them, the construction method of each sample data in the first sample set is the same as that of the first input; Using a convolutional neural network as the initial model; Training the initial model, and using the relevant parameters in the trained initial model as the parameters of the convolutional layer, activation layer, and pooling layer in the first model.

[0017] Based on this, it is possible to better obtain the relevant data in the first model.

[0018] Preferably, When obtaining the hidden layer parameters in the second model, it includes Constructing a second sample set with the abnormal category as the label; among them, the construction method of each sample data in the second sample set is the same as that of the second input; Using a neural network as the initial model; Training the initial model, and using the relevant parameters in the trained initial model as the hidden layer parameters in the second model.

[0019] Based on this, it is possible to better obtain the relevant data in the second model.

[0020] Preferably, When obtaining the fully connected layer parameters in the third model, it includes Constructing a third sample set with the abnormal category as the label; among them, each sample data in the third sample set includes the first data obtained based on the construction method of the first input and the second data obtained based on the construction method of the second input; Substituting the obtained parameters of the convolutional layer, activation layer, and pooling layer in the first model into the first model, and substituting the obtained hidden layer parameters in the second model into the second model; Using the outputs of the first model and the second model for the third sample set as the input of the third model, training the third model, and then obtaining the fully connected layer parameters in the third model.

[0021] Based on this, it is possible to better obtain the relevant data in the third model.

[0022] Preferably, the output layer in the third model uses the softmax function. Based on this, it is possible to better realize the prediction of multi-category anomalies. Description of the Drawings

[0023] Figure 1Schematic block diagram of a hand-held mobile remote collaboration imager based on the Internet of Things in Embodiment 1; Figure 2 Schematic diagram of a prediction model adopted by a substation main transformer fault prediction method based on voiceprint features in Embodiment 1. Detailed implementation manners

[0024] To further understand the content of the present invention, the present invention will be described in detail in combination with embodiments. It should be understood that the embodiments are only for explaining the present invention rather than limiting it.

[0025] Embodiment 1 For the inspection of core equipment such as transformers and reactors in a substation, currently it mainly relies on manual "listening to sounds to identify abnormalities". This method has high requirements for personnel experience and low accuracy, especially being restricted by the audible sound frequency range of the human ear.

[0026] Based on this, this embodiment provides a hand-held mobile remote collaboration imager based on the Internet of Things, as Figure 1 shown, which includes a device body, and a reference plane is constructed at the device body; the following are provided at the device body, An image acquisition module for acquiring image data of a detection object; wherein, the imaging plane of the image acquisition module is parallel to the reference plane; A sound acquisition module for acquiring sound data of the detection object; wherein, the sound acquisition module includes a microphone array composed of multiple microphone units, and the microphone array has at least 2 transverse microphone units distributed at equal distances laterally away from the image acquisition module along the reference plane and at least 2 longitudinal microphone units distributed at equal distances longitudinally away from the image acquisition module along the reference plane; A ranging module for acquiring distance data between the detection object and the reference plane; An attitude acquisition module for acquiring the deflection angle data of the reference plane; A processing module, including a data fusion unit and a beam output unit; wherein, the data fusion unit is used to fuse the image data and the sound data to obtain fusion data, and the beam output unit is used to obtain beam data based on the sound data.

[0027] The hand-held mobile remote collaboration imager provided by the present invention can acquire the image data and sound data of the area where the detection object is located by setting the image acquisition module and the sound acquisition module, and can provide the original data for subsequent fault detection based on sound features. Latent faults can be detected earlier by identifying faults based on sound features.

[0028] In this embodiment, the sound data includes the angle between the sound source propagation direction and the lateral direction of the reference plane The included angle between the sound source propagation direction and the longitudinal direction of the reference plane , the deflection angle data of the reference plane includes the included angle between the transverse direction of the reference plane and the horizontal direction and the included angle between the longitudinal direction of the reference plane and the vertical direction ; specifically, the data fusion unit is used to Based on the included angle , the included angle , the included angle , the included angle and the distance data L, obtain the camera coordinates of the sound source in the camera coordinate system ; where ; ; ; Based on the transformation matrix between the camera coordinate system and the image coordinate system, obtain the coordinates of the sound source in the image coordinate system ; Based on the coordinates of the sound source in the image coordinate system and the sound intensity distribution at the sound source, complete the acquisition of the fusion data.

[0029] Based on the above, it is possible to realize the positioning of the sound source in the image data based on the propagation angle and distance information of the sound source during data acquisition, so that the acquisition of the fusion data can be better realized.

[0030] It can be understood that in this embodiment, the alignment of the sound intensity data and the image data can be realized by finding the center coordinates of the sound source , and then the fusion data can be obtained more accurately.

[0031] In this embodiment, the included angle between the sound source propagation direction and the transverse direction of the reference plane is obtained based on the at least 2 transverse microphone units. Specifically ; where m is the total number of transverse microphone units is the moment when the first transverse microphone unit in the direction away from the image acquisition module collects the sound wave data at the detection object is the moment when the i-th transverse microphone unit in the direction away from the image acquisition module collects the sound wave data at the detection object, and d is the array pitch of the at least 2 transverse microphone units.

[0032] Based on the above, through the construction of the at least 2 transverse microphone units, the propagation angle data of the sound source propagation direction relative to the transverse direction of the reference plane can be better realized.

[0033] In this embodiment, the included angle between the sound source propagation direction and the longitudinal direction of the reference plane is obtained based on the at least two longitudinal microphone units. Specifically, ; where m is the total number of longitudinal microphone units, is the time when the first longitudinal microphone unit in the direction away from the image acquisition module collects the acoustic wave data at the detection object, is the time when the jth longitudinal microphone unit in the direction away from the image acquisition module collects the acoustic wave data at the detection object, p is the array pitch of the at least two longitudinal microphone units, and c is the acoustic wave transmission speed.

[0034] Based on the above, through the construction of the at least two longitudinal microphone units, the propagation angle data of the sound source propagation direction relative to the longitudinal direction of the reference plane can be preferably realized.

[0035] In this embodiment, the sound data of the detection object includes the acoustic wave signals collected by each microphone unit. The beam output unit is specifically used for, obtaining the original beam data based on the minimum variance distortionless response algorithm ; compensating the intensity of the original beam data based on the distance data between the detection object and the reference plane, and then obtaining the beam data ; where, ; where, is the reference value obtained by transformation through the transformation matrix between the image coordinate system and the world coordinate system based on the coordinates of the sound source in the image coordinate system .

[0036] Based on the above, the finally obtained beam data can be closer to the acoustic wave distribution at the surface area of the detection object, which is beneficial to subsequent data processing.

[0037] In this embodiment, a communication module is further provided on the device body. The communication module is used to realize the long-distance transmission of the fusion data and the beam data. Thus, the long-distance communication of the collected data can be preferably realized, which is beneficial to the subsequent processing of relevant data based on the upper computer.

[0038] In this embodiment, a power supply module is further provided on the device body. The power supply module is used to supply power to the device body. Therefore, the device body can be used in an outdoor environment.

[0039] In this embodiment, a display module is further provided at the device body. The display module is used to display the fused data, which is beneficial for the user to aim and orient during the data acquisition operation.

[0040] In this embodiment, an environmental monitoring module is further provided at the device body. The environmental monitoring module is used to collect environmental temperature data, environmental humidity data, and environmental air pressure data, which are used to correct the acoustic wave transmission speed, thereby further improving the accuracy of data acquisition.

[0041] Embodiment 2 For equipment such as transformers in a substation, during most of their main life cycle, they are in a normal state, so abnormal data is very rare. This data imbalance makes it difficult to use traditional methods such as voiceprint classification to judge the state of transformers.

[0042] Based on the data collected by the imager in Embodiment 1, this embodiment provides a method for predicting faults in the main transformer of a substation based on voiceprint features, which includes: Obtain the image data, fused data, and beam data of the detection object. Among them, the fused data includes the image data of the detection object and the sound intensity distribution data at the detection object; Construct a first input based on the image data and the fused data; Construct a second input based on the beam data; Process the first input based on the first model to obtain a first feature vector; Process the second input based on the second model to obtain a second feature vector; Concatenate the first feature vector and the second feature vector to obtain a feature vector; Process the feature vector based on the third model to obtain a fault prediction output.

[0043] The method of the present invention can utilize the working states implied by the image data and the beam data to predict faults in equipment such as the main transformer of a substation. It can make relatively full use of image features and voiceprint features to achieve fault prediction with a small sample size, especially suitable for fault prediction under data imbalance.

[0044] The construction of the first input based on the image data and the fused data includes: Identify the detection object in the image data and construct a recognition frame. The recognition frame is the smallest circumscribed rectangle that completely contains the detection object; Obtain the image of the area covered by the recognition frame and convert it into the grayscale image; Based on the sound intensity distribution data at the image in the area covered by the recognition frame, reconstruct the grayscale to obtain the first input.

[0045] Based on the above, it is possible to preferably eliminate irrelevant image features, especially retain the image features in the area where the detection object is located. By integrating the sound intensity distribution data into the image data, it is possible to preferably introduce the sound wave distribution information at the detection object, and introduce the difference information of the vibration parts of the detection object in different working states in the overall prediction method, so that it is possible to better realize the prediction of anomalies.

[0046] The reconstructing the grayscale based on the sound intensity distribution data at the image in the area covered by the recognition frame to obtain the first input includes: Obtain the average grayscale value of the grayscale image ; where , N is the total number of pixel points of the grayscale image, is the sum of the grayscale values of all pixel points of the grayscale image; Obtain the grayscale compensation value at each pixel point of the grayscale image ; where , S is the sound intensity at the corresponding pixel point, is the maximum sound intensity at the image in the area covered by the recognition frame, is the minimum sound intensity at the image in the area covered by the recognition frame; Subtract the corresponding grayscale compensation value from the grayscale value at each pixel point in the grayscale image , to obtain the first input.

[0047] Based on the above, it is possible to preferably introduce the sound intensity distribution data into the image data.

[0048] The constructing the second input based on the beam data includes: Extract the main frequency eigenvalue, acoustic vibration entropy eigenvalue, 50Hz odd and even harmonic ratio eigenvalue, fundamental frequency ratio eigenvalue and high and low frequency eigenvalue in the beam data to construct a voiceprint feature sequence as the second input.

[0049] Based on the above, it is possible to obtain the voiceprint features with a typical normal distribution in the beam data to construct the second input. Based on this method, especially, it is possible to mine the standard voiceprint feature distribution information from the normal data, making the entire prediction method particularly sensitive to abnormal data and achieving an improvement in the prediction accuracy of the prediction model under the imbalance of sample data.

[0050] The first model uses a convolutional neural network, which has a convolutional layer, an activation layer, and a pooling layer; the second model uses a neural network, which has an input layer and a hidden layer; the third model includes a fully connected layer and an output layer; the first feature vector and the second feature vector are concatenated into a feature vector at the fully connected layer and output to the output layer after weighted calculation by the fully connected layer, and the output layer is used to obtain the probability of the fault category based on the output of the fully connected layer to complete the fault prediction output. Thus, it can better achieve the separate processing of different types of data and the fusion processing of data from different sources.

[0051] The parameters of the convolutional layer, activation layer, and pooling layer in the first model, the parameters of the hidden layer in the second model, and the parameters of the fully connected layer in the third model are all obtained through separate training. This enables the use of richer sample data to train different models, which is beneficial to improving the model accuracy.

[0052] When obtaining the parameters of the convolutional layer, activation layer, and pooling layer in the first model, it includes Constructing a first sample set with the abnormal category as the label; among them, the construction method of each sample data in the first sample set is the same as that of the first input; Using a convolutional neural network as the initial model; Training the initial model, and using the relevant parameters in the trained initial model as the parameters of the convolutional layer, activation layer, and pooling layer in the first model.

[0053] Based on this, it can better achieve the acquisition of relevant data in the first model.

[0054] When obtaining the parameters of the hidden layer in the second model, it includes Constructing a second sample set with the abnormal category as the label; among them, the construction method of each sample data in the second sample set is the same as that of the second input; Using a neural network as the initial model; Training the initial model, and using the relevant parameters in the trained initial model as the parameters of the hidden layer in the second model.

[0055] Based on this, it can better achieve the acquisition of relevant data in the second model.

[0056] When obtaining the parameters of the fully connected layer in the third model, it includes Constructing a third sample set with the abnormal category as the label; among them, each sample data in the third sample set includes the first data obtained based on the construction method of the first input and the second data obtained based on the construction method of the second input; Substituting the obtained parameters of the convolutional layer, activation layer, and pooling layer in the first model into the first model, and substituting the obtained parameters of the hidden layer in the second model into the second model; Use the outputs of the first model and the second model for the third sample set as the input for the third model, and train the third model to obtain the fully connected layer parameters in the third model.

[0057] Based on this, it is possible to preferably obtain the relevant data in the third model.

[0058] The output layer in the third model uses the softmax function. Based on this, it is possible to preferably achieve the prediction of multi-class anomalies.

[0059] It is easy to understand that those skilled in the art can combine, split, reorganize, etc. the embodiments of the present application based on one or several embodiments provided by the present application to obtain other embodiments, and these embodiments do not exceed the protection scope of the present application.

[0060] The above schematically describes the present invention and its implementation manners. This description is not restrictive, and what is shown in the embodiments is only part of the implementation manners of the present invention, and the actual structure is not limited thereto. Therefore, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments to this technical solution without creative efforts without departing from the purpose of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A method for predicting the faults of main transformers in a substation based on voiceprint features, which includes obtaining the image data, fusion data, and beam data of the detection object; wherein, the fusion data includes the image data of the detection object and the sound intensity distribution data at the detection object; constructing a first input based on the image data and the fusion data; constructing a second input based on the beam data; processing the first input based on a first model to obtain a first feature vector; processing the second input based on a second model to obtain a second feature vector; concatenating the first feature vector and the second feature vector to obtain a feature vector; processing the feature vector based on a third model to obtain a fault prediction output.

2. The method for predicting the faults of main transformers in a substation based on voiceprint features according to claim 1, wherein: The constructing of the first input based on the image data and the fusion data includes identifying the detection object in the image data to construct a recognition frame; wherein, the recognition frame is the smallest circumscribed rectangle that completely contains the detection object; obtaining the image of the area covered by the recognition frame and converting it into the grayscale image; reconstructing the grayscale based on the sound intensity distribution data at the image of the area covered by the recognition frame to obtain the first input.

3. The method for predicting the faults of main transformers in a substation based on voiceprint features according to claim 2, wherein: The reconstructing of the grayscale based on the sound intensity distribution data at the image of the area covered by the recognition frame to obtain the first input includes Obtain the average grayscale value of the grayscale image ; where , N is the total number of pixel points of the grayscale image, is the sum of the grayscale values of all pixel points of the grayscale image; Obtain the gray compensation value at each pixel of the grayscale image ; where , S is the sound intensity at the corresponding pixel, is the maximum sound intensity in the image of the area covered by the recognition frame, is the minimum sound intensity in the image of the area covered by the recognition frame; Subtract the gray value at each pixel point in the gray image from the corresponding gray compensation value , and obtain a first input.

4. The method for predicting the faults of main transformers in a substation based on voiceprint features according to claim 3, wherein: The constructing of the second input based on the beam data includes extracting the main frequency eigenvalue, acoustic vibration entropy eigenvalue, 50Hz odd and even multiple frequency ratio eigenvalue, fundamental frequency ratio eigenvalue, and high and low frequency eigenvalue in the beam data to construct a voiceprint feature sequence as the second input.

5. The method for predicting the faults of main transformers in a substation based on voiceprint features according to claim 4, wherein: The first model adopts a convolutional neural network, which has a convolutional layer, an activation layer, and a pooling layer; the second model adopts a neural network, which has an input layer and a hidden layer; the third model includes a fully connected layer and an output layer; The first feature vector and the second feature vector are concatenated into a feature vector at the fully connected layer and output to the output layer after weighted calculation by the fully connected layer. The output layer is used to obtain the probability of the fault category based on the output of the fully connected layer to complete the fault prediction output.

6. The method for predicting the faults of main transformers in a substation based on voiceprint features according to claim 5, wherein: The parameters of the convolutional layer, activation layer, and pooling layer in the first model, the parameters of the hidden layer in the second model, and the parameters of the fully connected layer in the third model are all obtained by separate training.

7. The method for predicting the faults of main transformers in a substation based on voiceprint features according to claim 6, wherein: When obtaining the parameters of the convolutional layer, activation layer, and pooling layer in the first model, it includes constructing a first sample set with the abnormal category as the label; wherein, the construction method of each sample data in the first sample set is the same as the construction method of the first input; adopting a convolutional neural network as the initial model; Train the initial model, and use the relevant parameters in the trained initial model as the parameters of the convolutional layer, activation layer, and pooling layer in the first model.

8. A method for predicting main transformer faults in a substation based on voiceprint features according to claim 7, wherein: When obtaining the hidden layer parameters in the second model, it includes: Construct a second sample set with the abnormal category as the label; wherein, the construction method of each sample data in the second sample set is the same as the construction method of the second input; Use a neural network as the initial model; Train the initial model, and use the relevant parameters in the trained initial model as the hidden layer parameters in the second model.

9. A method for predicting main transformer faults in a substation based on voiceprint features according to claim 8, wherein: When obtaining the fully connected layer parameters in the third model, it includes: Construct a third sample set with the abnormal category as the label; wherein, each sample data in the third sample set includes the first data obtained based on the construction method of the first input and the second data obtained based on the construction method of the second input; Substitute the obtained parameters of the convolutional layer, activation layer, and pooling layer in the first model into the first model, and substitute the obtained hidden layer parameters in the second model into the second model; Use the outputs of the first model and the second model for the third sample set as the input of the third model, and train the third model to obtain the fully connected layer parameters in the third model.

10. The method for predicting the main transformer fault in a substation based on voiceprint features according to claim 9, wherein: The output layer in the third model uses the softmax function.

Images