Method and system for detecting air leakage sound of fan cabin based on deep learning

Abstract

The invention relates to the technical field of sound detection, and provides a deep learning-based fan cabin air leakage sound detection method and system, and the method comprises the steps: obtaining the audio data of a fan cabin, carrying out the preprocessing of the audio data, generating a Mel language spectrogram of the audio data according to preset spectrogram parameters, and carrying out the processing of the Mel language spectrogram, constructing a Mel language spectrogram image training data set according to the Mel language spectrogram of the audio data; training the optimized MobileNetV3 image classification network by using the image training data set to obtain an image classification model; obtaining a to-be-detected audio signal, performing length alignment on the audio signal, obtaining a Mel-language spectrogram corresponding to the audio signal after length alignment according to a preset spectrogram parameter, inputting the Mel-language spectrogram into the image classification model to obtain a detection result, and determining whether air leakage audio exists in the to-be-detected audio signal according to the detection result, the method has better anti-interference capability, and can detect the air leakage sound of the fan cabin in real time, in a targeted and efficient manner.

Description

Technical Field

[0001] The present invention relates to the technical field of sound detection, and in particular to a method and system for detecting air leakage sound in a fan cabin based on deep learning. Background Art

[0002] Wind power generation, as a highly clean and efficient power generation method, has made great progress in recent years. Among the various systems of wind turbine generator sets, the wind turbine nacelle has the characteristics of high failure rate, long downtime, and difficulty in detection.

[0003] At present, the commonly used methods for detecting wind turbine cabin air leakage in the industry mainly include image method, manual detection method and traditional audio signal processing method. The image method detects the air leakage area by capturing cabin images with a camera, but it is limited by the viewing angle and light source conditions, and it is difficult to directly identify abnormal sounds; the manual detection method relies on the experience and senses of the inspectors, is greatly affected by subjective factors, and is inefficient; the traditional audio signal processing method is easily disturbed by environmental noise, such as talking, rain, thunder, etc., which affects the detection accuracy.

[0004] In view of the limitations of the above methods, there is an urgent need for a detection method that can monitor in real time, resist environmental interference and has high adaptability. Summary of the invention

[0005] In view of the above-mentioned defects, the purpose of the present invention is to propose a method and system for detecting air leakage sound in the fan cabin based on deep learning, which has better anti-interference ability and can detect air leakage sound in the fan cabin in real time, specifically and efficiently.

[0006] To achieve this object, the present invention adopts the following technical solutions:

[0007] A method for detecting air leakage sound in a fan cabin based on deep learning, comprising:

[0008] Acquire audio data of the wind turbine cabin, generate a mel-spectrogram of the audio data according to preset spectrogram parameters after preprocessing the audio data, and construct a mel-spectrogram image training data set according to the mel-spectrogram of the audio data;

[0009] Using the image training data set to train the optimized MobileNetV3 image classification network to obtain an image classification model;

[0010] The audio signal to be detected is obtained, the length of the audio signal is aligned, the Mel-language spectrogram corresponding to the length-aligned audio signal is obtained according to preset spectrogram parameters and input into the image classification model to obtain a detection result, and whether there is leakage audio in the audio signal to be detected is determined according to the detection result.

[0011] Preferably, the step of performing length alignment on the audio signal comprises:

[0012] If the length of the audio signal is greater than a preset standard length, the audio signal is cut into a segment centered at the center position of the audio signal and having a length of the standard length;

[0013] If the length of the audio signal is less than a preset standard length, the length of the audio signal is extended by a number of cycles so that the length of the audio signal is equal to the standard length.

[0014] Preferably, obtaining the Mel-language spectrogram corresponding to the length-aligned audio signal according to preset spectrogram parameters and inputting it into the image classification model includes:

[0015] The audio signal is resampled until a sampling rate of the audio signal is the same as a preset sampling rate.

[0016] Preferably, preprocessing the audio data includes:

[0017] Resampling the audio data until a sampling rate of the audio data is the same as a preset sampling rate;

[0018] The resampled audio data is divided into several short audio data according to a preset time length;

[0019] The normal sound and the air leakage sound in the short audio data are respectively marked.

[0020] Furthermore, the resampling method is a Kaiser window function method.

[0021] Furthermore, constructing a Mel spectrogram image training dataset based on the Mel spectrogram of the audio data includes:

[0022] Performing noise enhancement and volume disturbance enhancement on the short audio data, and generating a Mel-language spectrogram of the enhanced short audio data according to preset spectrogram parameters;

[0023] A mel spectrogram image training dataset is constructed based on the mel spectrogram of the enhanced short audio data and the mel spectrogram of the audio data before enhancement.

[0024] Preferably, the network head of the optimized MobileNetV3 image classification network is obtained by removing several dimensionality-increasing operations from the network head of the existing MobileNetV3 image classification network and retaining only one fully connected layer.

[0025] Preferably, the optimized MobileNetV3 image classification network is obtained by replacing the SE attention mechanism of the existing MobileNetV3 image classification network with the ECA attention mechanism.

[0026] Preferably, the spectrogram parameters include the number of FFT points, the window shift size and the number of filters in the Mel filter bank;

[0027] In the spectrogram parameters, the number of FFT points is 1024, the window shift size is 1 / 2, and the number of filters in the Mel filter bank is 96.

[0028] A deep learning-based detection system for wind turbine nacelle air leakage sound, the system comprising:

[0029] A data set construction module is used to obtain audio data of a wind turbine cabin, generate a Mel spectrogram of the audio data according to preset spectrogram parameters after preprocessing the audio data, and construct a Mel spectrogram image training data set according to the Mel spectrogram of the audio data;

[0030] A model training module, used to train the optimized MobileNetV3 image classification network using the image training data set to obtain an image classification model;

[0031] The sound detection module is used to obtain the audio signal to be detected, perform length alignment on the audio signal, obtain the Mel-language spectrogram corresponding to the length-aligned audio signal according to preset spectrogram parameters and input it into the image classification model to obtain the detection result, and determine whether there is leakage audio in the audio signal to be detected according to the detection result.

[0032] One of the above technical solutions has the following advantages or beneficial effects:

[0033] This method resamples, length-aligns and enhances audio signals, can effectively process audio data of various sources and qualities, and can improve the model's adaptability to different recording conditions. By removing the dimensionality increase operation at the head of the MobileNetV3 network, retaining only one fully connected layer, and replacing the SE attention mechanism with the ECA attention mechanism, the computational complexity of the model is reduced, and the applicability of air leakage sound detection and the anti-interference ability during detection are improved; the Mel-speech spectrogram is generated using preset spectrogram parameters to obtain the optimal Mel-speech spectrogram, which reduces the impact of environmental noise and facilitates training and testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0035] Figure 2 It is a module diagram of a fan nacelle air leakage sound detection system based on deep learning provided by an embodiment of the present application;

[0036] Figure 3It is a Mel spectrogram with zero element completion of a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0037] Figure 4 It is a Mel spectrogram with signal period completion of a method for detecting wind leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0038] Figure 5 It is a Mel spectrogram of audio data of normal volume of a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0039] Figure 6 It is a Mel spectrogram of audio data at 0.1 times the normal volume of a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0040] Figure 7 It is a structural diagram of the network head of an existing MobileNetV3 of a method for detecting wind leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0041] Figure 8 It is a schematic diagram of the structure of the network head of an improved MobileNetV3 of a method for detecting wind leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0042] Fig. 9 It is a structural schematic diagram of a normal sound generated according to preset spectrogram parameters in a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0043] Fig.10 It is a structural schematic diagram of air leakage sound generated according to preset spectrogram parameters in a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application;

[0044] Fig.11 This is a structural diagram of a method for detecting air leakage sound in a fan cabin based on deep learning provided by an embodiment of the present application, in which the SE attention mechanism is replaced by the ECA attention mechanism. DETAILED DESCRIPTION

[0045] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0046] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0047] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0048] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0049] like Figure 1 As shown, a preferred embodiment of this embodiment is a method for detecting air leakage sound in a fan cabin based on deep learning, comprising the steps of:

[0050] S1: Acquire audio data of a wind turbine cabin, generate a Mel spectrogram of the audio data according to preset spectrogram parameters after preprocessing the audio data, and construct a Mel spectrogram image training data set according to the Mel spectrogram of the audio data;

[0051] S2: Using the image training data set to train the optimized MobileNetV3 image classification network to obtain an image classification model;

[0052] S3: Obtain an audio signal to be detected, perform length alignment on the audio signal, obtain a Mel-language spectrogram corresponding to the length-aligned audio signal according to preset spectrogram parameters, and input the spectrogram into the image classification model to obtain a detection result, and determine whether there is leakage audio in the audio signal to be detected according to the detection result.

[0053] Specifically, in step S1, a microphone is used to collect audio data in different wind turbine cabin scenes. These audio data include mechanical noise from normal operation and abnormal conditions such as air leakage, as well as voices, rain and thunder that may affect detection. The newly collected audio data cannot be fully used in the construction of subsequent image training data sets, so the audio data needs to be preprocessed to ensure that subsequent model training can be based on high-quality data. The preprocessing operation can be cropping, resampling, manual labeling or removing background noise of audio data. Among them, in the field of audio and speech signal processing, we can convert the signal into a corresponding Mel spectrogram (the horizontal axis of the Mel spectrogram provided by this method is time, the vertical axis is frequency, the coordinate point value is the energy of speech data, and a two-dimensional plane is used to express three-dimensional information. The size of the energy value is represented by a heat map. The higher the heat value, the stronger the speech energy at the point). The data on the Mel spectrogram is used as the feature of the signal. There are many types of spectrograms, among which the most commonly used is the Mel spectrogram, such as Figure 5 and Figure 6 They are the Mel spectrogram of normal audio data and the Mel spectrogram corresponding to 0.1 times the normal signal volume, respectively, indicating that when converting the Mel spectrogram, the distance between the sound source and the audio collection point has little effect on the recognition effect. The pre-processed audio data is converted into a Mel spectrogram. In the present embodiment, the function in the python version librosa library can be used for processing. The spectrogram parameters of the Mel spectrogram have a greater impact on the accuracy and speed of model prediction, so the optimal spectrogram parameters obtained by experiments in advance are used during conversion to reduce the impact of environmental noise on the Mel spectrogram. Finally, the Mel spectrogram corresponding to the audio data is constructed into a Mel spectrogram image training data set for subsequent training of the model.

[0054] Convolutional neural network (CNN) is one of the representative algorithms of deep learning. It is good at processing images, especially image recognition and other related machine learning problems. It has significant effects in various visual tasks such as image classification, target detection, image segmentation, etc. It is one of the most widely used models. MobileNetV3 is a lightweight convolutional neural network (CNN) architecture, which is specially designed to efficiently perform image classification tasks on mobile devices and other resource-constrained devices. In step S2, the image training data set generated in the previous step is used to train the improved and optimized MobileNetV3 image classification network to obtain an image classification model that can indirectly classify sound. The optimization of the MobileNetV3 image classification network can be the replacement of the attention mechanism or the improvement of the network head structure, so that the trained image classification model is more suitable for the detection of wind leakage sound in the fan cabin.

[0055] In step S3, firstly, it is necessary to obtain the real-time audio signal to be detected. These audio signals may be continuously flowing data and need to be processed by length alignment to ensure the consistency of the data format input to the image classification model. Length alignment ensures that the model can effectively extract features and classify each data block processed. Usually, the audio signal can be adjusted to a fixed length by cropping, padding or other methods. Similarly, the optimal spectrogram parameters obtained in advance through experiments are used to generate the corresponding Mel-language spectrogram, and the Mel-language spectrogram corresponding to the audio signal to be detected is input into the image classification model generated in step S2. The model outputs the classification result of the sound to determine whether there is a sound of air leakage, such as Fig.10 As shown, by observing the Mel spectrogram, it can be found that there is a bright spot on the Mel spectrogram corresponding to the audio signal with leakage audio.

[0056] Preferably, the step of performing length alignment on the audio signal comprises:

[0057] If the length of the audio signal is greater than a preset standard length, the audio signal is cut into a segment centered at the center position of the audio signal and having a length of the standard length;

[0058] If the length of the audio signal is less than a preset standard length, the length of the audio signal is extended by a number of cycles so that the length of the audio signal is equal to the standard length.

[0059] Specifically, the lengths of the audio signals to be detected are different, so the audio signals need to be cropped to the same length. When the length of the audio signal exceeds the preset standard length, the cropping starts from the center of the audio signal to ensure that the cropped segment contains the core part of the original signal as much as possible. Center cropping helps to retain the main information of the audio signal and avoids the information loss that may be caused by cropping from the beginning or end of the signal. Assuming that the length of the audio signal is N and the marked length is L, the starting position of the cropping is start = (NL) / 2, and then the segment from this starting position to start+L is cropped.

[0060] If the length of the audio signal is less than the standard length, a common method is to use 0 elements to pad the signal to the standard length before detection. However, if the proportion of 0 elements is too large and the signal is short, the detection effect of the 0-element padding method is not good. A large proportion of 0 elements will make the signal features unclear and make the black area of ​​the spectrogram account for a large proportion, such as Figure 3 Therefore, in this embodiment, when the length of the audio signal is less than the preset standard length, it is necessary to fill it to the standard length by extending the audio signal. For example, if the standard length is 1 second and an audio signal is 0.4 seconds, the signal is used as a cycle and extended for 1.5 cycles to form a 1-second audio signal. Figure 4 shown.

[0061] In another embodiment, obtaining the Mel-language spectrogram corresponding to the length-aligned audio signal according to preset spectrogram parameters and inputting it into the image classification model includes:

[0062] The audio signal is resampled until a sampling rate of the audio signal is the same as a preset sampling rate.

[0063] Specifically, under the same audio signal, the Mel spectrograms obtained by transformation under different sampling rates are different. In order to make the transformation algorithm more universal and obtain the same Mel spectrogram under the same audio signal, the audio signal needs to be resampled. For example, the audio signal can be resampled using the resampling function in the audio processing library librosa in the Python language, and the preset sampling rate is 48k.

[0064] Further, preprocessing the audio data includes:

[0065] Resampling the audio data until a sampling rate of the audio data is the same as a preset sampling rate;

[0066] The resampled audio data is divided into several short audio data according to a preset time length;

[0067] The normal sound and the air leakage sound in the short audio data are respectively marked.

[0068] Specifically, when obtaining the audio data of the fan cabin, the sampling rate of the audio data is first adjusted to a unified preset value to ensure data consistency. According to engineering experience, it is not advisable to directly detect whether there is air leakage sound in the fan cabin for a longer audio data, but it is advisable to judge a short audio signal, mainly for the following reasons: 1. Short audio signals are relatively stable and will not change too much. 2. Long audio signals can be divided into short audio signals for detection. 3. The computing resources occupied by short audio signal detection are relatively small. Therefore, the resampled audio data is divided according to the preset time length, and the preset time length ranges from 0.3 to 2 seconds. These short audio data are then labeled, and the normal sound of the fan cabin is manually divided into one category, and the air leakage sound of the fan cabin is divided into another category, so that the audio data is suitable for subsequent model training.

[0069] Furthermore, the resampling method is a Kaiser window function method.

[0070] The Kaiser window function method is a resampling technology used in audio and signal processing. It has a smooth transition at the edge of the window function, avoiding the strong edge effect that may be caused by the traditional window function. The smooth transition helps to reduce the ripple phenomenon that occurs during the resampling process, making the signal resampling more natural and maintaining the characteristics of the original audio signal.

[0071] Furthermore, constructing a Mel spectrogram image training dataset based on the Mel spectrogram of the audio data includes:

[0072] Performing noise enhancement and volume disturbance enhancement on the short audio data, and generating a Mel-language spectrogram of the enhanced short audio data according to preset spectrogram parameters;

[0073] A mel spectrogram image training dataset is constructed based on the mel spectrogram of the enhanced short audio data and the mel spectrogram of the audio data before enhancement.

[0074] Specifically, in order to improve the anti-interference ability of the trained image classification model, a certain noise signal, such as speaking or thunder, is collected in advance, and the noise signal is added to the previously generated short audio data in a certain proportion to achieve noise enhancement, and volume changes are applied to the short audio data, including volume increase and volume decrease, which can be achieved by adjusting the audio gain or applying dynamic range compression, and the short audio data after noise enhancement and volume disturbance enhancement is converted into the corresponding Mel-speech spectrogram.

[0075] A mel-spectrogram image training dataset is constructed by combining the mel-spectrogram of the enhanced short audio data and the mel-spectrogram of the audio data before enhancement. By combining the enhanced mel-spectrogram with the original mel-spectrogram, the image classification model can learn how to deal with common audio interferences in practical applications such as noise and volume changes. By combining different types of enhanced data with the original data, the image classification model can learn more diverse features. This diversity helps the model better generalize to unseen data and avoid overfitting to the training data, which helps the model perform more reliably and efficiently when facing various variations in real applications.

[0076] In another embodiment, the network head of the optimized MobileNetV3 image classification network is obtained by removing several dimensionality-increasing operations from the network head of the existing MobileNetV3 image classification network and retaining only one fully connected layer.

[0077] Specifically, the existing MobileNetV3 image classification network is designed to solve the classification of 1000 categories on the ImageNet dataset, while the wind leakage sound detection in the fan cabin has only two categories of classification. In view of the differences caused by the classification categories, the network head structure needs to be improved. In order to simply and efficiently classify the finally extracted feature vector, the redundant and time-consuming multiple dimensionality increase operations in the network head of the existing MobileNetV3 image classification network are removed, and only a simple fully connected layer is used to directly map the feature vector obtained by pooling to the classification category, such as Figure 7 and Figure 8 , which are the existing network head structure and the improved network head structure respectively.

[0078] In another embodiment, the optimized MobileNetV3 image classification network is obtained by replacing the SE attention mechanism of the existing MobileNetV3 image classification network with the ECA attention mechanism.

[0079] Specifically, since the sound of fan leakage is particularly susceptible to noise signals, resulting in little difference between the audio signal and its Mel-level spectrogram, the introduction of the attention module allows the model to selectively pay attention to important areas in the Mel-level spectrogram. By inserting different attention modules after the DW convolution layer of all bottleneck structures of MobileNetV3, replacing the SE attention mechanism in the original MobileNetV3 (such as Fig.11 As shown in the figure, the ECA attention mechanism is most suitable for wind leakage sound detection in the wind turbine cabin. The ECA attention mechanism avoids the use of fully connected layers and instead uses one-dimensional convolution to calculate the dependencies between channels. The computational complexity is significantly lower than that of the SE module. ECA uses fewer parameters to achieve similar functions. Compared with the fully connected layer of the SE module, ECA reduces model parameters while maintaining efficient channel attention calculation.

[0080] Furthermore, the spectrogram parameters include the number of FFT points, the window shift size and the number of filters in the Mel filter bank;

[0081] In the spectrogram parameters, the number of FFT points is 1024, the window shift size is 1 / 2, and the number of filters in the Mel filter bank is 96.

[0082] Specifically, the number of FFT (Fast Fourier Transform) points refers to the number of data points used when calculating the Fourier transform. An FFT point number of 1024 means that each Fourier transform operation will involve 1024 sample points. The number of FFT points determines the frequency domain resolution. The more points, the higher the frequency resolution, but the computational complexity will also increase. In this case, an FFT point number of 1024 can provide a higher frequency resolution, which helps to analyze the frequency components more accurately.

[0083] The window shift size refers to the step size of each sliding when sliding the window in time. A window shift size of 1 / 2 means that the step size of each sliding of the window is 50% of the window length. The window shift size affects the time resolution and the smoothness of the spectrum. A smaller window shift size (such as 1 / 2) means that the window overlap is higher, which can better capture the instantaneous changes of the signal.

[0084] The Mel filter bank is a filter bank designed based on the Mel frequency scale, which is used to convert the spectrum into the Mel frequency domain. The number of filters refers to the number of filters in the Mel filter bank, which is 96 here. The number of filters determines the resolution of the Mel frequency domain. The more filters there are, the higher the frequency resolution on the Mel frequency scale, and the more detailed the characteristics of the spectrum can be captured. 96 filters can usually provide more detailed frequency resolution and are suitable for most audio analysis tasks.

[0085] Applying this spectrogram parameter to the generation of Mel language spectrogram can improve the accuracy and speed of model prediction, such as Fig. 9 and Fig.10 As shown in the figure, the Mel spectrograms corresponding to normal sound and air leakage sound respectively. By observing the Mel spectrogram, it can be found that there is a bright spot on the Mel spectrogram of air leakage sound, which shows that air leakage sound is a high-frequency signal, and the Mel image features of air leakage sound in the fan cabin and normal sound are quite different under less noise interference.

[0086] like Figure 2 As shown, a detection system for wind turbine nacelle air leakage sound based on deep learning, the system comprising:

[0087] The data set construction module 1 is used to obtain audio data of the wind turbine cabin, generate a Mel spectrogram of the audio data according to preset spectrogram parameters after preprocessing the audio data, and construct a Mel spectrogram image training data set according to the Mel spectrogram of the audio data;

[0088] Model training module 2, used to train the optimized MobileNetV3 image classification network using the image training data set to obtain an image classification model;

[0089] The sound detection module 3 is used to obtain the audio signal to be detected, perform length alignment on the audio signal, obtain the Mel-language spectrogram corresponding to the length-aligned audio signal according to preset spectrogram parameters and input it into the image classification model to obtain the detection result, and determine whether there is leakage audio in the audio signal to be detected according to the detection result.

[0090] This embodiment implements a method and implementation process for detecting air leakage sound in a fan cabin based on deep learning. Please refer to the above embodiments and will not go into details here.

[0091] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0092] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for detecting wind leakage sound in a fan cabin based on deep learning, characterized in that: include: Acquire audio data of the wind turbine cabin, generate a mel-spectrogram of the audio data according to preset spectrogram parameters after preprocessing the audio data, and construct a mel-spectrogram image training data set according to the mel-spectrogram of the audio data; Using the image training data set to train the optimized MobileNetV3 image classification network to obtain an image classification model; The audio signal to be detected is obtained, the length of the audio signal is aligned, the Mel-language spectrogram corresponding to the length-aligned audio signal is obtained according to preset spectrogram parameters and input into the image classification model to obtain a detection result, and whether there is leakage audio in the audio signal to be detected is determined according to the detection result.

2. The detection method according to claim 1, characterized in that: The length alignment of the audio signal comprises: If the length of the audio signal is greater than a preset standard length, the audio signal is cut into a segment centered at the center position of the audio signal and having a length of the standard length; If the length of the audio signal is less than a preset standard length, the length of the audio signal is extended by a number of cycles so that the length of the audio signal is equal to the standard length.

3. The detection method according to claim 1, characterized in that: The method of obtaining the Mel language spectrogram corresponding to the length-aligned audio signal according to the preset spectrogram parameters and inputting it into the image classification model includes: The audio signal is resampled until a sampling rate of the audio signal is the same as a preset sampling rate.

4. The detection method according to claim 1, characterized in that: Preprocessing the audio data includes: Resampling the audio data until a sampling rate of the audio data is the same as a preset sampling rate; The resampled audio data is divided into several short audio data according to a preset time length; The normal sound and the air leakage sound in the short audio data are respectively marked.

5. The detection method according to any one of claims 3 or 4, characterized in that: The resampling method is the Kaiser window function method.

6. The detection method according to claim 4, characterized in that: The Mel spectrogram image training dataset is constructed based on the Mel spectrogram of audio data, including: Performing noise enhancement and volume disturbance enhancement on the short audio data, and generating a Mel-language spectrogram of the enhanced short audio data according to preset spectrogram parameters; A mel spectrogram image training dataset is constructed based on the mel spectrogram of the enhanced short audio data and the mel spectrogram of the audio data before enhancement.

7. The detection method according to claim 1, characterized in that: The network head of the optimized MobileNetV3 image classification network is obtained by removing several dimensionality-increasing operations from the network head of the existing MobileNetV3 image classification network and retaining only one fully connected layer.

8. The detection method according to claim 1, characterized in that: The optimized MobileNetV3 image classification network is obtained by replacing the SE attention mechanism of the existing MobileNetV3 image classification network with the ECA attention mechanism.

9. The detection method according to claim 1, characterized in that: The spectrogram parameters include the number of FFT points, the window shift size and the number of filters in the Mel filter bank; In the spectrogram parameters, the number of FFT points is 1024, the window shift size is 1 / 2, and the number of filters in the Mel filter bank is 96.

10. A deep learning-based wind turbine cabin air leakage sound detection system, characterized in that: The system comprises: A data set construction module is used to obtain audio data of a wind turbine cabin, generate a Mel spectrogram of the audio data according to preset spectrogram parameters after preprocessing the audio data, and construct a Mel spectrogram image training data set according to the Mel spectrogram of the audio data; A model training module, used to train the optimized MobileNetV3 image classification network using the image training data set to obtain an image classification model; The sound detection module is used to obtain the audio signal to be detected, perform length alignment on the audio signal, obtain the Mel-language spectrogram corresponding to the length-aligned audio signal according to preset spectrogram parameters and input it into the image classification model to obtain the detection result, and determine whether there is leakage audio in the audio signal to be detected according to the detection result.

Images