A method for identifying and warning abnormal sound patterns of insulator flashover

Abstract

The present invention discloses a method for identifying and warning abnormal soundprints of insulator flashover, comprising the following steps: sound collection and processing, wherein a sound pickup device collects sound emitted from the top of a high-voltage power tower, and an algorithm is used to process the collected audio to obtain a power spectrum; the power spectrum is converted and processed to obtain a Mel-speech spectrogram; audio data is labeled and classified according to known types, and the audio is repeatedly trained using the algorithm's training function to obtain an optimal audio model containing different sound audio models; the optimal audio model is used to classify, score, and evaluate the audio extracted features, and a threshold is set; the collected audio is input, a collected audio threshold is calculated, and the collected audio threshold is compared with the optimal audio model threshold for similarity; a scoring system is used to indicate the proximity of the thresholds, and the scoring system determines whether to issue an alarm or push information based on the comparison result between the collected audio and the optimal audio model threshold. The method of the present invention has high recognition accuracy, low cost, and high efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring, and in particular to a method for recognizing and warning abnormal sound patterns of insulator flashover. Background Art

[0002] Insulators are specialized insulating components that play a vital role in overhead transmission lines. Initially used on utility poles, they gradually evolved into high-voltage power lines, where multiple disc-shaped insulators are hung on one end of a tower to increase creepage distance. These insulators are typically made of glass or ceramic.

[0003] Insulator flashover abnormalities refer to the occurrence of creepage or arcing on the surface of the insulator. Insulator creepage refers to the polarization of the insulating material surrounding the conductor, resulting in charged areas on the insulating material. Creepage occurs when conductive contaminants such as dust and oil accumulate on the surface of the line insulator. When damp in humid weather, this discharge occurs on the surface of the insulator, generally causing leakage. When the leakage current increases to a certain value, it causes surface breakdown discharge, a condition known as arcing. Insulator creepage is a common phenomenon. If only intermittent and minor creepage occurs, no special treatment is required. However, if continuous and concentrated creepage occurs for a period of time, or even arcing occurs, it can easily lead to a short circuit in the high-voltage line, causing large-scale regional power outages. It can also cause serious accidents such as wildfires due to falling sparks.

[0004] Current monitoring methods for insulator flashover include visual observation, drone inspections, telescope observation, camera recording, and insulation resistance testing during power outages. These methods are either inaccurate, labor-intensive, or difficult to implement. Some also require power outages for monitoring, often resulting in detection of insulators that have already broken down and are no longer viable. Furthermore, these monitoring methods rely solely on manual determination of insulator anomalies such as creepage or arcing, resulting in low accuracy and a significant workload. Summary of the Invention

[0005] In response to the technical problems existing in the background technology, the purpose of the present invention is to provide a method for identifying and warning of abnormal soundprints of insulator flashover. By collecting the sound emitted from the tower head of the high-voltage tower and training and classifying the audio through an algorithm, abnormal alarm information can be pushed in a timely manner with high judgment accuracy, which greatly reduces the workload of monitoring personnel.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying and warning abnormal sound patterns of insulator flashover, characterized by comprising the following steps:

[0008] S1. Sound collection and processing: The sound emitted from the top of the high-voltage tower is collected through a sound pickup device. The collected audio is pre-processed, framed, and windowed using an algorithm. A short-time Fourier transform is then performed to obtain its spectrum. The energy spectrum is obtained by squaring the spectrum, and the energy within each filter band is superimposed to obtain the power spectrum.

[0009] S2. Pass the spectrogram formed by the power spectrum through a Mel filter bank to obtain the Mel Spectrogram audio feature value, and pass the power spectrum through a set of Mel-scale triangular filter banks to obtain the Mel spectrogram;

[0010] S3. Label and classify the audio data according to known types. During machine learning, each audio segment is represented by a corresponding Mel-spectogram. A frequency band corresponding to each frame is a feature. The algorithm's training function is used to repeatedly train the audio to obtain the optimal audio model that includes different sound audio models. The optimal audio model is used to classify and score the feature values ​​extracted from the audio, and a threshold is set. The threshold range is 0 to 1.

[0011] S4. Input the collected audio, obtain the feature data of the collected audio through the prediction function, use this feature data to calculate the diagonal cosine value, and finally obtain the threshold of the collected audio. Compare the similarity between the threshold of the collected audio and the threshold of the optimal audio model, and apply a scoring system to indicate the degree of proximity between the thresholds. The scoring system determines whether to alarm and whether to push information to the staff client based on the comparison result between the collected audio and the threshold of the optimal audio model.

[0012] Furthermore, in step S3, the optimal audio model includes an arcing audio model, a creeping current audio model, and other noise audio models.

[0013] Furthermore, in step S3, the method for modeling the recognition and classification of different sounds includes the following steps:

[0014] P1. Use an algorithm to determine whether the audio frequency exceeds 7kHz. If it does not exceed 7kHz, classify the audio as other noise audio models. If it exceeds 7kHz, proceed to step P2.

[0015] P2. Analyze the audio energy using the spectrum analyzer. If the spectrum analyzer shows a continuous, concentrated, and stable resonance peak in the range of 6.5kHz-8.5kHz, the audio with this characteristic will be classified as a creeping audio model after comparison with the optimal audio model. Otherwise, it will be classified as other noise audio models.

[0016] P3. When the audio is identified and classified as a creeping audio model, it is again determined whether there is a local sudden increase in high-frequency sound intensity within the range of the continuous concentrated stable resonance peak. If a local sudden increase in high-frequency sound intensity of more than 65dB occurs and the duration is greater than the first time threshold, after comparing with the optimal audio model, the audio with this feature is identified and classified as a suspected creeping audio model. Otherwise, it is identified and classified as a creeping audio model.

[0017] P4. When the audio is identified and classified as a suspected arc creeping audio model, an alarm message will be pushed and the corresponding audio will be sent back for secondary manual recognition confirmation.

[0018] Preferably, the first time threshold is 0.5s.

[0019] Furthermore, in step S4, when the threshold of the collected audio triggers the creepage threshold of the optimal audio model and the familiarity exceeds the set percentage, the scoring system immediately issues a creepage alarm and pushes information to the staff client; when the threshold of the collected audio triggers the creepage threshold of the optimal audio model and the familiarity exceeds the set percentage, the collected audio is first continuously analyzed. If the continuous triggering time exceeds the second time threshold, the scoring system issues a creepage alarm and pushes information to the staff client; when the creepage threshold of the optimal audio model is intermittently triggered, no alarm is issued but information is pushed.

[0020] Preferably, the second time threshold is 5 minutes.

[0021] Preferably, the set percentage of the familiarity is 70%.

[0022] Preferably, in order to avoid too frequent alarms, when the scoring system issues a creepage alarm and pushes information to the staff client, the calculation starts from the time of each alarm. No more alarms will be issued within 30 minutes, but information will be pushed.

[0023] Preferably, in order to further ensure the accuracy of the training, in step S3, the classified audio is repeatedly trained for a period of not less than 30 times.

[0024] Preferably, in order to more accurately and completely collect high-definition high-frequency sounds, the sampling rate of the sound pickup device is greater than or equal to 16 kHz, and the quantization bit number is greater than or equal to 16 bits.

[0025] The present invention has the following beneficial effects: the identification and early warning method of the present invention can replace the existing monitoring method, accurately monitor whether arc creepage or creepage abnormalities occur in the insulator through voiceprint recognition, and judge and analyze the severity of the creepage, and can push abnormal alarm information in a timely manner. It has high accuracy, high practicality, low cost, and high efficiency, greatly reducing the workload of power transmission maintenance personnel, with high benefits and suitable for popularization and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is the spectrogram of wind audio.

[0027] Figure 2 This is the spectrogram of the whistle audio.

[0028] Figure 3 This is a spectrogram of the audio of insects and birds singing.

[0029] Figure 4 This is a spectrum analysis diagram of the audio of insects and birds singing.

[0030] Figure 5 This is the spectrum diagram when there is creepage current but no arcing.

[0031] Figure 6 This is the spectrum analysis diagram when there is creepage but no arcing.

[0032] Figure 7 This is the spectrogram when arc creep occurs.

[0033] Figure 8 This is the spectrum analysis diagram when arc creep occurs.

[0034] Figure 9 This is the sound intensity diagram when arc climbing occurs. DETAILED DESCRIPTION

[0035] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0036] Insulators are installed on high-voltage power line connection towers. Since most high-voltage power line connection towers are installed in mountains and forests or in sparsely populated areas, when using a pickup device to collect the sound emitted by the insulators, the interference sounds in the environment can be mainly divided into three types: wind sound, whistle sound (high-voltage power towers near roads) and insect and bird sounds.

[0037] like Figure 1 The following is a spectrogram of wind noise. As can be seen from the figure, there are significant differences in frequency analysis between wind noise and creeping noise. Under normal circumstances, wind noise frequencies are primarily concentrated below 2 kHz. However, inclement weather can affect wind noise frequencies, resulting in higher sound intensity but a duller timbre. Creeping noise, on the other hand, is sharper and exhibits some fluctuation in the high-frequency band. Therefore, during acquisition, wind noise can be initially filtered through physical structures, and then secondary filtering can be performed through post-processing algorithm learning.

[0038] like Figure 2The figure shows a spectrogram of the whistle audio. As can be seen from the figure, the interference of the whistle sound is mainly concentrated in the low and medium frequencies below 5kHz. In special cases, such as the whistle frequency of some large trucks, the frequency may exceed 7kHz. In this case, the algorithm can be used for secondary screening and judgment.

[0039] like Figure 3 The following is a spectrogram of insect chirping and bird calls. As can be seen from the figure, the audio frequency of insect chirping and bird calls is special. The frequency of bird calls is 4-8k, which partially overlaps with the high frequency of creeping electricity. Therefore, in addition to frequency identification, energy identification is also required. Figure 4 The figure shows the spectrum analysis of the chirping of insects and birds. It can be seen from the figure that although the frequency of the chirping of insects and birds exceeds 7kHz, there is no continuous, concentrated and stable resonance peak. Therefore, the use of energy for secondary identification can completely distinguish the chirping of insects and birds from the creeping sound.

[0040] from Figure 1-4 From the above analysis, we can conclude that in the environmental audio, except for the audio of insects and birds, the rest of the audio is concentrated in the medium and low frequencies below 7k.

[0041] The present invention is described using an insulator as an example, but the method of the present invention can also be applied to other insulation controls in a power system.

[0042] The present invention discloses a method for identifying and warning abnormal sound patterns of insulator flashover, which specifically comprises the following steps:

[0043] For sound collection and processing, a sound pickup device captures sound emitted from the top of a high-voltage power tower. Algorithms are used to preprocess the captured audio, including framing and windowing. A short-time Fourier transform is then performed to obtain its frequency spectrum. The energy spectrum is squared to obtain the energy spectrum, and the energy within each filter band is summed to obtain the power spectrum. The spectrogram formed by the power spectrum is passed through a Mel filter bank to obtain the audio eigenvalues. The power spectrum is then passed through a set of Mel-scale triangular filters to obtain the Mel spectrogram. Typically, the number of captured audio inputs exceeds 5,000 to ensure the accuracy of the final data model.

[0044] The first step in machine learning is to extract features. When processing audio signals, we need to convert the signal into a spectrogram, using the data on the spectrogram as the signal's features. The horizontal axis (x) of a spectrogram represents time, and the vertical axis (y) represents frequency. The value corresponding to (x, y) represents the amplitude of frequency y at time x.

[0045] The frequency distribution of a typical spectrogram is linear, but the human ear's perception of frequency is logarithmic, that is, it is sensitive to changes in the low-frequency band and insensitive to changes in the high-frequency band.

[0046] Research has shown that human frequency perception is not linear, and that we are more sensitive to low-frequency signals than high-frequency ones. For example, while people can easily detect the difference between 500 and 1000 Hz, they struggle to discern the difference between 7500 and 8000 Hz. This led to the development of the Mel Scale, a nonlinear transformation of Hz. For signals measured in mel scale units, it allows people to perceive signals of the same frequency difference almost identically.

[0047] Therefore, the linear distribution of the spectrogram will obviously lead to "features that are not useful enough" in feature extraction, so the Mel spectrogram came into being. The vertical axis frequency of the Mel spectrogram and the original frequency are exchanged by the following formula:

[0048]

[0049] f=700(10 m / 2595 -1)

[0050] The Mel-frequency cepstral coefficients (MFCCs) are obtained by performing a cosine transform (DCT, a linear transform similar to Fourier transform) after obtaining the Mel-frequency spectrogram, and then taking a part of the coefficients.

[0051] The specific steps to obtain the Mel spectrogram are as follows:

[0052] (1) Obtaining audio signals

[0053] Python can use the librosa library to read audio files, and can only read wav audio file formats.

[0054] (2) Signal pre-emphasis

[0055] Generally speaking, the high-frequency component of a voice / audio signal is weaker, while the low-frequency component is stronger. Signal pre-emphasis is to pass the signal through a high-pass filter so that the strength of the high-frequency and low-frequency components of the signal does not differ too much.

[0056] (3) Framing

[0057] The original signal is divided into several small blocks based on time, each of which is called a frame. Since the signal covers too long a period of time, performing an FFT on the entire signal only reveals the relationship between the signal's frequency and intensity, but loses the temporal information. To understand how frequency changes over time, we divide the original signal into several frames, perform an FFT on each frame (also called a short-time FFT, as we only take a small period of time), and then concatenate the results in chronological order. This is the principle behind the spectrogram.

[0058] frames variable, each row of which corresponds to each frame;

[0059] frame_size: The length of each frame, usually 20-40ms. If it is too long, the time resolution will be small, while if it is too small, the computation cost will be increased.

[0060] frame_length: the number of samples corresponding to each frame;

[0061] frame_stride: The interval between two adjacent frames. Usually the interval must be less than the length of each frame, that is, there must be overlap between the two frames to prevent the loss of boundary information;

[0062] frame_step: the number of samples between two adjacent frames;

[0063] frame_num: the number of frames required for the entire signal, with zero padding added to ensure that the signal length can be divided into an integer number of frames;

[0064] (4) Window

[0065] After framing, a window function is applied to each frame to achieve a better sidelobe reduction. Usually, the hamming window frames are directly multiplied by the hamming window. Note that this is not a matrix multiplication.

[0066] (The framing process is equivalent to adding a rectangular window to the signal. The spectrum of the rectangular window has large side lobes. Multiplying the window function and the original function in the time domain is equivalent to convolution in the frequency domain. After the convolution of the rectangular window function and the original function, the side lobes are large, which will cause a large difference in the spectrum of the original signal and the corresponding part after the window is added. This is spectral leakage. The hamming window has smaller side lobes, and the spectral leakage caused is also smaller).

[0067] (5) Obtaining the power spectrum

[0068] The frames variable has each row corresponding to each frame. Perform FFT on each row, take the magnitude of the obtained FFT transform, square it, and divide it by the corresponding number of FFT points to get the power spectrum.

[0069] (6) Mel-filter banks

[0070] The last step is to apply the Mel filter to the pow_frames obtained in the previous step.

[0071] A Mel filter bank is a triangular filter bank of equal height, where each filter starts at the midpoint of the previous filter. The corresponding frequencies are linear on the Mel scale, hence the name Mel filter bank. The frequencies corresponding to each filter can be converted to Mel frequencies using the formula mentioned above, linearly divided into several frequency bands on the Mel scale, and then converted back to the actual frequency scale.

[0072] In actual operation, each filter is multiplied by the power spectrum pow_frames, and the result obtained is the energy in that frequency band.

[0073] The result of mel-spectrogram is a matrix where each row is a frame and each column represents the energy of the corresponding mel band.

[0074] (7)Mel-spectogram feature

[0075] During machine learning, each audio segment can be represented by a corresponding mel-spectogram, and a frequency band corresponding to each frame is a feature.

[0076] In practice, each audio file must be of the same length to have the same number of features. Normalization is usually performed, meaning that the mean of each frame is subtracted from the mean of each element to ensure that the mean of each frame is 0.

[0077] The formula for generating a Mel spectrogram is as follows:

[0078]

[0079] Where: |X t (k)| 2 is the power spectrum, k is the subscript of the frequency component; H m (k) is the frequency response corresponding to each filter;

[0080]

[0081] Where m represents the filter number, f(m-1), f(m), and f(m+1) correspond to the starting point, middle point, and end point of the mth filter respectively; Y t (m) is the signal spectrum shape.

[0082] The pickup device is typically installed in the middle of a high-voltage tower, ensuring the distance from each insulator is as even as possible. This reduces installation costs by requiring only one pickup device per high-voltage tower. To more accurately and completely capture high-definition, high-frequency sound, the pickup device has a sampling rate greater than or equal to 16kHz and a quantization bit count greater than or equal to 16 bits.

[0083] Audio data is labeled and classified according to known types for training. During machine learning, each audio segment is represented by a corresponding Mel-spectogram. A frequency band corresponding to each frame is a feature. The audio is divided into creeping audio, electric shock audio, and non-creeping audio. The algorithm's training function is used to repeatedly train the audio at least 30 times to obtain an optimal audio model that includes different sound audio models. The optimal audio model includes a creeping audio model, an electric shock audio model, and other noise audio models. The optimal audio model is used to classify, score, and evaluate the audio extracted feature values, and a threshold is set. The threshold range is 0 to 1. The method for modeling different sound recognition and classification includes the following steps:

[0084] An algorithm determines whether the audio frequency exceeds 7kHz. If not, the audio is classified as other noise audio model. This is because, in environmental sounds, except for the chirping of insects and birds, the frequency of other audio cannot exceed 7kHz. Creepage characteristics are more concentrated in the mid-to-high frequencies above 7kHz, so frequency is chosen to preliminarily diagnose and distinguish creepage.

[0085] If the frequency of the audio is determined to be over 7kHz, the audio energy is analyzed using the spectrum analysis graph. If the spectrum analysis graph shows a continuous, concentrated, and stable resonance peak within the range of 6.5kHz-8.5kHz, the audio with this feature is classified as a creeping audio model after comparison with the optimal audio model, such as Figure 5 、 Figure 6 As shown, otherwise it is identified and classified as other noise audio models; because the frequency of insect and bird sounds in the environmental sound may exceed 7kHz at certain special times, this step is used to perform a secondary analysis of the audio to completely eliminate the interference of insect and bird sounds on the algorithm judgment.

[0086] When the audio is identified and classified as a creeping audio model, it is again determined whether there is a local sudden increase in high-frequency sound intensity within the range of the continuous concentrated stable resonance peak. If a local sudden increase in high-frequency sound intensity of more than 65dB occurs and lasts for more than 0.5s, after comparing with the optimal audio model, the audio with this feature will be classified as a suspected creeping audio model, such as Figure 7-9 Otherwise, it is classified as a creepage audio model. Because arcing is a prerequisite for creepage, and arcing will not occur in the absence of creepage, this step is used to determine whether there is suspected arcing when creepage occurs.

[0087] Some sounds in the environment, such as lightning, the chirping of insects and birds, and the horns of some special vehicles, may also occur during arc creep and may cause certain interference to the algorithm's judgment of arc creep. Therefore, when the audio is identified and classified as a suspected arc creep audio model, an alarm message is pushed and the corresponding audio is sent back for secondary manual recognition confirmation to ensure the accuracy of the arc creep audio model.

[0088] Since creepage is a common phenomenon, if only intermittent and mild creepage occurs, no special treatment is required. However, if continuous and concentrated creepage occurs and the creepage exceeds a certain period of time, it is necessary to push warning information to the staff client to remind them to pay continuous attention, but this situation does not necessarily require special emergency treatment. However, if arc creeping occurs, it can easily cause a short circuit in the high-voltage line, resulting in large-scale regional power outages, and may also cause serious accidents such as wildfires due to falling sparks. Therefore, it is necessary to distinguish between creepage and arc creep. If arc creep occurs, the alarm information needs to be pushed to the staff client to remind them to pay immediate attention and, if necessary, to carry out emergency special treatment such as power outage and maintenance.

[0089] The system takes audio data (continuously collected by a sound pickup device) as input, extracts its feature data through a prediction function, uses this feature data to calculate the diagonal cosine, and ultimately determines the threshold of the audio data. This threshold is then compared with the threshold of the optimal audio model for similarity. A scoring system is used to indicate the proximity of the thresholds. Based on the comparison of the thresholds between the audio data and the optimal audio model, the scoring system determines whether to issue an alarm and whether to push a notification to the staff client. When the audio data threshold triggers the optimal audio model's creepage threshold and the similarity exceeds 70%, the scoring system immediately issues a creepage alarm and pushes a notification to the staff client. When the audio data threshold triggers the optimal audio model's creepage threshold and the similarity exceeds 70%, the system first continuously analyzes the audio data. If the triggering continues for more than 5 minutes, the scoring system issues a creepage alarm and pushes a notification to the staff client. When the optimal audio model's creepage threshold is triggered intermittently, no alarm is issued, but a notification is pushed. To avoid excessively frequent alarms, once the scoring system issues a creepage alarm and pushes information to the staff client, the system will start counting from the time of each alarm. Within 30 minutes, there will be no more alarms but information will be pushed. If creepage occurs after 30 minutes, the system will alarm again, and again within 30 minutes after the alarm, there will be no more alarms, and so on.

[0090] In order to further verify the accuracy of the method of the present invention, 15 suspected creeping arc audios were input, of which 12 were creeping arc audios and 3 were interference audios with lightning sounds, insect chirping, bird chirping and special vehicle horn sounds. Through the algorithm of the present invention, 15 suspected creeping arc audios were finally judged with an accuracy rate of 100%. Through manual screening again, 12 creeping arc audios and 3 creeping electricity audios were screened out, with an accuracy rate of 100%; 35 creeping electricity audios (no creeping arc) were input, and through the algorithm of the present invention, 33 creeping electricity audios were finally judged with an accuracy rate of 94%; 50 noise audios were input, and through the algorithm of the present invention, 48 noise audios were finally judged with an accuracy rate of 96%.

[0091] In summary, the accuracy of the algorithm of the present invention can reach 90%, or even above 94%, which is a high accuracy rate. However, due to the complexity of the types of sounds in nature, and even some sounds that have not been explored, it is inevitable that there will be a small probability of inaccuracy. Even so, replacing manual inspection or judgment with the algorithm of the present invention has greatly improved efficiency, is highly practical, and has low cost, greatly reducing the workload of power transmission maintenance personnel, and is suitable for popularization and use.

[0092] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.

Claims

1. A method for identifying and warning abnormal sound patterns of insulator flashover, characterized in that: The steps include: S1. Sound collection and processing: The sound emitted from the top of the high-voltage tower is collected through a sound pickup device. The collected audio is pre-processed, framed, and windowed using an algorithm. A short-time Fourier transform is then performed to obtain its spectrum. The energy spectrum is obtained by squaring the spectrum, and the energy within each filter band is superimposed to obtain the power spectrum. S2. Pass the spectrogram formed by the power spectrum through a Mel filter bank to obtain the Mel Spectrogram audio feature value, and pass the power spectrum through a set of Mel-scale triangular filter banks to obtain the Mel spectrogram; S3. Label and classify the audio data according to known types. During machine learning, each audio segment is represented by a corresponding Mel-spectogram. A frequency band corresponding to each frame is a feature. The algorithm's training function is used to repeatedly train the audio to obtain the optimal audio model that includes different sound audio models. The optimal audio model is used to classify and score the feature values ​​extracted from the audio, and a threshold is set. The threshold range is 0 to 1. The method for modeling different sound recognition classifications includes the following steps: P1. Use an algorithm to determine whether the audio frequency exceeds 7kHz. If it does not exceed 7kHz, classify the audio as other noise audio models. If it exceeds 7kHz, proceed to step P2. P2. Analyze the audio energy using the spectrum analyzer. If the spectrum analyzer shows a continuous, concentrated, and stable resonance peak in the range of 6.5kHz-8.5kHz, the audio with this characteristic will be classified as a creeping audio model after comparison with the optimal audio model. Otherwise, it will be classified as other noise audio models. P3. When the audio is identified and classified as a creeping audio model, it is again determined whether there is a local sudden increase in high-frequency sound intensity within the range of the continuous concentrated stable resonance peak. If a local sudden increase in high-frequency sound intensity of more than 65dB occurs and the duration is greater than the first time threshold, after comparing with the optimal audio model, the audio with this feature is identified and classified as a suspected creeping audio model. Otherwise, it is identified and classified as a creeping audio model. P4: When the audio is identified and classified as a suspected arc creeping audio model, an alarm message will be pushed and the corresponding audio will be sent back for secondary manual recognition confirmation; S4. Input the collected audio, obtain the feature data of the collected audio through the prediction function, use this feature data to calculate the diagonal cosine value, and finally obtain the threshold of the collected audio. Compare the similarity between the threshold of the collected audio and the threshold of the optimal audio model, and apply a scoring system to indicate the degree of proximity between the thresholds. The scoring system determines whether to alarm and whether to push information to the staff client based on the comparison result between the collected audio and the threshold of the optimal audio model.

2. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 1, characterized in that: In step S3, the optimal audio model includes an arcing audio model, a creeping current audio model, and other noise audio models.

3. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 1, characterized in that: The first time threshold is 0.5s.

4. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 1, characterized in that: In step S4, when the threshold of the collected audio triggers the creepage threshold of the optimal audio model and the familiarity exceeds the set percentage, the scoring system immediately issues a creepage alarm and pushes information to the staff client; when the threshold of the collected audio triggers the creepage threshold of the optimal audio model and the familiarity exceeds the set percentage, the collected audio is first continuously analyzed. If the continuous triggering time exceeds the second time threshold, the scoring system issues a creepage alarm and pushes information to the staff client; when the creepage threshold of the optimal audio model is intermittently triggered, no alarm is issued but information is pushed.

5. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 4, characterized in that: The second time threshold is 5 minutes.

6. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 4, characterized in that: The set percentage of the familiarity is 70%.

7. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 4, characterized in that: When the scoring system issues a creepage alarm and pushes information to the staff client, the calculation starts from the time of each alarm. There will be no more alarms within 30 minutes, but information will be pushed.

8. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 1, characterized in that: In step S3, the classified audio is repeatedly trained for no less than 30 times.

9. The method for identifying and warning abnormal sound patterns of insulator flashover according to claim 1, characterized in that: The sampling rate of the sound pickup device is greater than or equal to 16 kHz, and the number of quantization bits is greater than or equal to 16 bits.

Images