Model training method and fan blade abnormal sound category identification method and system

Abstract

The invention provides a model training method and a fan blade abnormal sound category identification method and system.The training method comprises the steps that training samples are obtained, and the training samples comprise a plurality of sample sound signals generated in the fan blade operation process and real abnormal sound categories corresponding to the sample sound signals; training an abnormal sound category identification model by using the training sample; the abnormal sound category recognition model is used for extracting sound features of the sample sound signals, and calculating a ball domain center and a ball domain radius corresponding to a real abnormal sound category based on the sound features of the sample sound signals corresponding to each real abnormal sound category; and calculating a target distance between the sound feature of the sample sound signal and the ball domain center of each real abnormal sound category, and identifying a predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the ball domain radius of the corresponding real abnormal sound category, thereby reducing the misjudgment condition of the model and improving the accuracy of sound signal identification.

Description

Technical Field

[0001] The present disclosure relates to the technical field of online monitoring and intelligent identification of fan blade damage, and in particular to a model training method and a method and system for identifying abnormal noise categories of fan blades. Background Art

[0002] At present, there are four main online monitoring solutions for blade health status: vibration monitoring, image monitoring, acoustic emission monitoring, and acoustic monitoring.

[0003] Among them, vibration monitoring mainly identifies the natural frequency of the blade from the acceleration vibration spectrum of the blade vibration, and determines whether the blade is damaged by the shift of the natural frequency. Usually, when the blade is damaged, its natural frequency will decrease to a certain extent. However, in wind farm applications, the vibration monitoring solution requires the acceleration sensor to be installed inside the blade at a position 1 / 3 of the blade length from the blade root. This installation method is difficult to implement, and when the wind turbine is running, the energy of the impeller rotation frequency component in the collected vibration signal is much higher than the energy of the natural frequency, making it difficult to locate the natural frequency component, which in turn affects the accuracy of vibration monitoring.

[0004] Image monitoring uses an infrared camera to obtain images inside the blade cavity, and combines image recognition algorithms to monitor the light leakage spots in the blade cavity, and then judge whether the blade has cracks or other damage. However, as the length of the blade increases, its flexibility increases, and the deformation of the blade increases during operation. The range of the camera to obtain images is limited, and it is difficult to observe the tip position of the blade that is prone to damage; in addition, light leakage caused by uneven paint on the blade often interferes with the monitoring results of the image recognition algorithm, resulting in misjudgment and affecting the accuracy of image monitoring.

[0005] Acoustic emission monitoring determines the occurrence of blade cracks by capturing the elastic wave energy generated by the rapid release of strain energy when the blade cracks initiate. In order to collect effective acoustic emission signals, the sensor is required to have high sensitivity and response bandwidth. Therefore, acoustic emission sensors are often expensive, resulting in excessively high costs for acoustic emission monitoring.

[0006] Acoustic monitoring mainly collects the friction noise between the blade and the air, and judges the blade status through the abnormal sound in the sound. Compared with vibration monitoring, acoustic monitoring only requires the installation of the pickup at the blade root baffle, which is simple to construct; compared with image monitoring, the pickup at the blade root baffle in acoustic monitoring can receive the sound of the entire blade, and the monitoring range is wider; compared with acoustic emission monitoring, acoustic monitoring has low cost and is suitable for a wide range of applications.

[0007] However, when identifying acoustic features, existing acoustic monitoring technologies mostly use manual recognition to identify acoustic features in order to determine whether the blades are damaged. With the introduction of artificial intelligence models, existing technologies have adopted models such as BP neural networks, convolutional networks, and support vector machines to replace manual intelligent recognition of acoustic features. However, they generally use octave energy, spectrum center of gravity, or Mel spectrum directly as input to machine learning models such as BP neural networks and convolutional networks to achieve the recognition of different real abnormal sound categories of blades. Although the above method can replace manual abnormal sound recognition, when the input sound signal sample is of an unknown abnormal sound category, the method in the prior art can often only identify the unknown abnormal sound sample as a known category, resulting in misjudgment or missed judgment, affecting the accuracy of sound signal recognition. Summary of the invention

[0008] The technical problem to be solved by the present disclosure is to overcome the defect of low accuracy of sound signal recognition in the prior art, and to provide a model training method and a method and system for identifying the abnormal noise category of fan blades.

[0009] The present invention solves the above technical problems through the following technical solutions:

[0010] A first aspect of the present disclosure provides a training method for an abnormal sound category recognition model, the training method comprising:

[0011] Acquire training samples, wherein the training samples include a plurality of sample sound signals generated during the operation of the fan blades and real abnormal sound categories corresponding to the sample sound signals;

[0012] Using the training samples to train an abnormal sound category recognition model;

[0013] Among them, the abnormal sound category recognition model is used to extract the sound features of the sample sound signal, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signal corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. In response to the absence of a target distance smaller than the spherical radius of each real abnormal sound category, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category.

[0014] Optionally, the abnormal sound category recognition model is used to identify the predicted abnormal sound category corresponding to the sample sound signal as the real abnormal sound category in response to the target distance being smaller than the spherical radius corresponding to the real abnormal sound category.

[0015] Optionally, the step of training the abnormal sound category recognition model using the training samples specifically includes:

[0016] The twin model of the abnormal sound category recognition model is trained using the training samples; wherein the twin model is obtained by replacing the adaptive spherical domain recognition layer in the abnormal sound category recognition model with a fully connected layer;

[0017] The model parameters of the twin model except the fully connected layer are synchronized to the abnormal sound category recognition model.

[0018] Optionally, the abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain a sample time-frequency spectrum, and to extract sound features of the sample time-frequency spectrum.

[0019] Optionally, before the step of training the abnormal sound category recognition model using the training samples, the method further includes:

[0020] Performing amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain a sample time-frequency spectrum;

[0021] The abnormal sound category recognition model is used to extract the sound features of the spectrogram of the sample.

[0022] Optionally, the step of performing amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain the sample time-frequency spectrum specifically includes:

[0023] Performing a short-time Fourier transform on the sample sound signal to obtain a time-frequency spectrum diagram;

[0024] Converting the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel-frequency spectrum;

[0025] The sample Mel-spectrogram is amplitude pooled along the frequency axis to obtain a sample time-frequency spectrum.

[0026] Optionally, before the step of performing short-time Fourier transform on the sample sound signal, the step further includes:

[0027] The sample sound signal is pre-emphasized.

[0028] Optionally, the fan blades are arranged on a wheel hub, and an array-type sound receiving device facing the interior of the blades is provided at the connection between the fan blades and the wheel hub;

[0029] The step of obtaining training samples specifically includes:

[0030] The array-type sound receiving device is used to obtain training samples.

[0031] A second aspect of the present disclosure provides a method for identifying the category of abnormal noise of a fan blade, the method comprising:

[0032] Collect the target sound signals generated during the operation of the fan blades;

[0033] The target sound signal is input into an abnormal sound category recognition model to obtain a predicted abnormal sound category corresponding to the target sound signal; wherein the abnormal sound category recognition model is obtained by using the training method of the abnormal sound category recognition model as described in the first aspect; wherein the abnormal sound category recognition model is used to extract the sound features of the target sound signal, and calculate the target distance between the sound features of the target sound signal and the center of the sphere of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the target sound signal according to each target distance and the sphere radius of the corresponding real abnormal sound category, and in response to the absence of a target distance less than the sphere radius of each real abnormal sound category, identify the predicted abnormal sound category corresponding to the target sound signal as an unknown abnormal sound category.

[0034] Optionally, the abnormal sound category recognition model is used to:

[0035] Performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain a target time-frequency spectrum;

[0036] The sound features of the target spectrogram are extracted.

[0037] Optionally, before the step of inputting the target sound signal into the abnormal sound category recognition model, the method further includes:

[0038] Performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain a target time-frequency spectrum;

[0039] The abnormal sound category recognition model is used to extract the sound features of the target time spectrum graph.

[0040] Optionally, the step of performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain the target time-frequency spectrum specifically includes:

[0041] Performing a short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum;

[0042] Converting the energy amplitude in the time-frequency spectrum into a decibel value to obtain a target Mel-frequency spectrum;

[0043] The target Mel-spectrogram is amplitude pooled along the frequency axis to obtain a target time-frequency spectrum.

[0044] Optionally, before the step of performing short-time Fourier transform on the target sound signal, the step includes:

[0045] The target sound signal is pre-emphasized.

[0046] A third aspect of the present disclosure provides a training system for an abnormal sound category recognition model, the training system comprising:

[0047] An acquisition module, used for acquiring training samples, wherein the training samples include a plurality of sample sound signals generated during the operation of the fan blades and real abnormal sound categories corresponding to the sample sound signals;

[0048] A training module, used to train the abnormal sound category recognition model using the training samples;

[0049] Among them, the abnormal sound category recognition model is used to extract the sound features of the sample sound signal, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signal corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. In response to the absence of a target distance smaller than the spherical radius of each real abnormal sound category, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category.

[0050] Optionally, the abnormal sound category recognition model is used to identify the predicted abnormal sound category corresponding to the sample sound signal as the real abnormal sound category in response to the target distance being smaller than the spherical radius corresponding to the real abnormal sound category.

[0051] Optionally, the training module is specifically used to train the twin model of the abnormal sound category recognition model using the training samples; wherein the twin model is obtained by replacing the adaptive spherical domain recognition layer in the abnormal sound category recognition model with a fully connected layer; and synchronizing the model parameters in the twin model except the fully connected layer to the abnormal sound category recognition model.

[0052] Optionally, the abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain a sample time-frequency spectrum, and to extract sound features of the sample time-frequency spectrum.

[0053] Optionally, the abnormal sound category recognition model is used to perform short-time Fourier transform on the sample sound signal to obtain a time-frequency spectrum; convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel spectrum; and perform amplitude pooling on the sample Mel spectrum along the frequency axis to obtain a sample time-frequency spectrum.

[0054] Optionally, the training system further includes an amplitude pooling module, which is used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain a sample time-frequency spectrum;

[0055] The abnormal sound category recognition model is used to extract the sound features of the spectrogram of the sample.

[0056] Optionally, the amplitude pooling module specifically includes a time-spectrogram unit, a sample Mel-spectrogram unit and a sample time-spectrogram unit:

[0057] The time-spectrogram unit is used to perform short-time Fourier transform on the sample sound signal to obtain a time-spectrogram;

[0058] The sample Mel-spectrogram unit is used to convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel-spectrogram;

[0059] The sample time-frequency spectrum unit is used to perform amplitude pooling on the sample Mel-frequency spectrum along the frequency axis to obtain a sample time-frequency spectrum.

[0060] Optionally, the training system further comprises a pre-emphasis module for performing pre-emphasis processing on the sample sound signal.

[0061] Optionally, the fan blades are arranged on a wheel hub, and an array-type sound receiving device facing the interior of the blades is provided at the connection between the fan blades and the wheel hub;

[0062] The acquisition module is specifically used to acquire training samples using the array sound receiving device.

[0063] A fourth aspect of the present disclosure provides a system for identifying abnormal noise categories of fan blades, the identification system comprising:

[0064] A collection module, used to collect target sound signals generated during the operation of the fan blades;

[0065] A prediction module is used to input the target sound signal into an abnormal sound category recognition model to obtain a predicted abnormal sound category corresponding to the target sound signal; wherein the abnormal sound category recognition model is obtained using the abnormal sound category recognition model training method as described in any one of the first aspects, wherein the abnormal sound category recognition model is used to extract the sound features of the target sound signal, and calculate the target distance between the sound features of the target sound signal and the spherical center of each real abnormal sound category, identify the predicted abnormal sound category corresponding to the target sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category, and in response to the absence of a target distance less than the spherical radius of each real abnormal sound category, identify the predicted abnormal sound category corresponding to the target sound signal as an unknown abnormal sound category.

[0066] Optionally, the abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain a target time-frequency spectrum; and to extract sound features of the target time-frequency spectrum.

[0067] Optionally, the abnormal sound category recognition model is used to perform short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum; convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a target Mel-frequency spectrum; and perform amplitude pooling on the target Mel-frequency spectrum along the frequency axis to obtain a target time-frequency spectrum.

[0068] Optionally, the recognition system further includes a pooling processing module, which is used to perform a short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum; convert the energy amplitude in the time-frequency spectrum to a decibel value to obtain a target Mel-frequency spectrum; and perform amplitude pooling on the target Mel-frequency spectrum along the frequency axis to obtain a target time-frequency spectrum;

[0069] The abnormal sound category recognition model is used to extract the sound features of the target time spectrum graph.

[0070] Optionally, the recognition system further includes a pre-emphasis processing module, configured to perform pre-emphasis processing on the target sound signal.

[0071] A fifth aspect of the present disclosure provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and used to run on the processor, wherein when the processor executes the computer program, the training method of the abnormal sound category recognition model as described in the first aspect or the method for identifying the abnormal sound category of the fan blade as described in the second aspect is implemented.

[0072] The sixth aspect of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the training method for the abnormal sound category recognition model as described in the first aspect or the method for identifying the abnormal sound category of a fan blade as described in the second aspect.

[0073] A seventh aspect of the present disclosure provides a computer program product, including a computer program, which, when executed by a processor, implements the training method of the abnormal sound category recognition model as described in the first aspect or the method for identifying the abnormal sound category of a fan blade as described in the second aspect.

[0074] On the basis of being in accordance with the common sense in the art, the above-mentioned preferred conditions can be arbitrarily combined to obtain the preferred embodiments of the present disclosure.

[0075] The positive and progressive effect of the present disclosure is that the present disclosure uses a plurality of sample sound signals generated during the operation of the fan blades and the real abnormal sound categories corresponding to the sample sound signals as training samples, and uses the training samples to train the abnormal sound category recognition model, wherein the abnormal sound category recognition model is used to extract the sound features of the sample sound signals, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signals corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signals and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. If there is no target distance less than the spherical radius of each real abnormal sound category in response, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category. The abnormal sound category recognition model trained by this training method can identify various abnormal sound categories corresponding to the sound signal, prevent the sample sound signal of the unknown abnormal sound category from being mistakenly identified as other abnormal sound categories, thereby improving the accuracy of the abnormal sound category recognition model in identifying the abnormal sound category. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 A flowchart of a method for training an abnormal sound category recognition model provided in Example 1 of the present disclosure.

[0077] Figure 2 A flowchart for identifying abnormal sound categories of a sample sound signal provided in Embodiment 1 of the present disclosure.

[0078] Figure 3a A time-frequency spectrum diagram corresponding to a sound signal of a periodic whistle type provided in Example 1 of the present disclosure.

[0079] Figure 3b A sample time-frequency spectrum diagram corresponding to a sound signal of a periodic whistle type provided in Example 1 of the present disclosure.

[0080] Figure 3c This is a sound signal waveform diagram corresponding to a sound signal of the periodic whistle type provided in Example 1 of the present disclosure.

[0081] Figure 4 It is a statistical chart of the operating results of Example 1 of the present disclosure and the prior art.

[0082] Figure 5 A schematic diagram of the modules of a training system for an abnormal sound category recognition model provided in Example 1 of the present disclosure.

[0083] Figure 6 A flow chart of a method for identifying abnormal noise categories of fan blades provided in Example 2 of the present disclosure.

[0084] Figure 7A schematic diagram of a module of a system for identifying abnormal noise categories of fan blades provided in Example 2 of the present disclosure.

[0085] Figure 8 A schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present disclosure. DETAILED DESCRIPTION

[0086] The present disclosure is further described below by way of examples, but the present disclosure is not limited to the scope of the examples.

[0087] Prefixes such as "first" and "second" are used in the embodiments of the present disclosure only to distinguish different description objects, and have no limiting effect on the position, order, priority, quantity or content of the described objects. The use of prefixes such as ordinal numbers to distinguish description objects in the embodiments of the present disclosure does not constitute a limitation on the described objects. For the statement of the described objects, please refer to the description in the context of the claims or embodiments, and no unnecessary limitation should be constituted due to the use of such prefixes. In addition, in the description of the present embodiment, unless otherwise specified, the meaning of "plurality" is two or more.

[0088] Example 1

[0089] This embodiment provides a training method for an abnormal sound category recognition model. Figure 1 As shown, the training method comprises the following steps:

[0090] S1. Obtain training samples, wherein the training samples include a plurality of sample sound signals generated during the operation of the fan blades and real abnormal sound categories corresponding to the sample sound signals;

[0091] In this embodiment, when collecting training samples, it is necessary to collect sample sound signals and their corresponding real abnormal sound categories. The collected sample sound signals can be divided into multiple types according to their corresponding real abnormal sound categories, and each sample sound signal includes multiple sample sound signals.

[0092] In a specific example, real sample sound signals can be collected from a wind farm. Based on the sample sound signals collected from the wind farm, the categories of abnormal sounds during the operation of the wind turbine blades are counted, mainly including three categories of abnormal sounds: flexible ring extrusion noise, similar glass fiber breakage noise, and periodic whistle noise. The sound signals of these three categories of abnormal sounds are combined with the sound signals of the normal operation of the wind turbine blades to obtain training samples including four sample sound signals and their corresponding real abnormal sound categories. The specific content of the training samples is shown in Table 1.

[0093] Table 1 Training samples

[0094]

[0095] S2. Training the abnormal sound category recognition model using the training samples;

[0096] Among them, the abnormal sound category recognition model is used to extract the sound features of the sample sound signal, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signal corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. In response to the absence of a target distance smaller than the spherical radius of each real abnormal sound category, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category.

[0097] In this embodiment, during the training process, the loss is calculated based on the predicted abnormal sound category and the actual abnormal sound category corresponding to the sample sound signal, and the parameters of the abnormal sound category recognition model are adjusted according to the loss until the convergence condition is met, thereby obtaining a trained abnormal sound category recognition model.

[0098] In this embodiment, in the process of extracting the sound characteristics of the sample sound signal, because the acoustic characteristics of early blade damage often exist in the mid- and high-frequency range, and the frequency components of the sample sound signal have high energy in the low-frequency components and low energy in the mid- and high-frequency components, the sample sound signal needs to be processed to highlight the mid- and high-frequency acoustic signals.

[0099] In a specific example, the abnormal sound category recognition model may include 2 convolution layers, 2 pooling layers, 2 Dropout layers and an adaptive spherical domain recognition layer, wherein the convolution layers, the pooling layers and the Dropout layers are used to extract the sound features of the sample sound signals; the number of units in the adaptive spherical domain recognition layer is 5, wherein 4 units correspond to 3 types of typical abnormal sounds and healthy sound-free sound, and 1 unit corresponds to an unknown abnormal sound category; the adaptive spherical domain recognition layer is used to calculate the spherical domain center and spherical domain radius of the corresponding real abnormal sound category based on the sound features of the sample sound signals of each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical domain center of each real abnormal sound category, and identify the predicted abnormal sound category of the sample sound signal according to each target distance and the spherical domain radius of the corresponding real abnormal sound category.

[0100] In the specific implementation of step S2, the twin model can be used to train the abnormal sound category recognition model, which specifically includes: first, the twin model of the abnormal sound category recognition model is trained using the training sample; wherein the twin model is obtained by replacing the adaptive spherical domain recognition layer in the abnormal sound category recognition model with a fully connected layer; and then the model parameters of the twin model except the fully connected layer are synchronized to the abnormal sound category recognition model. In one specific implementation example, a fully connected layer including 4 neurons and an activation function of softmax is constructed, and the adaptive spherical domain recognition layer of the abnormal sound category recognition model is replaced to construct a training twin model of the abnormal sound category recognition model, and some samples are randomly selected from the training set as training samples, and the rest are verification samples. The Adam algorithm is used to train the twin model, and the training results of all layers except the fully connected layer in the twin model are synchronized to the abnormal sound category recognition model, and the sample sound signal in the training set is input into the abnormal sound category recognition model to obtain the sound characteristics of the sample sound signal.

[0101] In this embodiment, the abnormal sound category recognition model can calculate the spherical center and spherical radius corresponding to the real abnormal sound category based on the sound features of the sample sound signals of each real abnormal sound category.

[0102] In a specific example, the mean of the sound features of the sample sound signals of each real abnormal sound category is calculated to obtain the spherical center of each real abnormal sound category. The calculation formula is as follows, where Fi represents the spherical center of the i-th type of sample sound signal, k f i p The p-th dimension of the sound signal representing the i-th sample in the i-th category feature, P is the number of dimensions of the sound feature, Q i surface represents the number of sample sound signals of the i-th category, i= The abnormal sound category of the sample sound signal corresponding to 1 is healthy without abnormal sound, the abnormal sound category of the sample sound signal corresponding to i=2 is the abnormal sound of flexible ring extrusion, the abnormal sound category of the sample sound signal corresponding to i=3 is similar to the abnormal sound of glass fiber breakage, and the abnormal sound category of the sample sound signal corresponding to i=4 is a periodic whistle:

[0103]

[0104] In this embodiment, the target distance between sample sound signal A and sample sound signal B is used to characterize the degree of proximity between sample sound signal A and sample sound signal B. The larger the target distance, the smaller the degree of proximity between sample sound signal A and sample sound signal B, which means that the possibility that sample sound signal A and sample sound signal B are the same category of sound signals is smaller. In a specific implementation, the target distance can be Euclidean distance, Minkowski distance, Manhattan distance, etc.

[0105] In a specific example, firstly, the Euclidean distance between the sound feature of the sample sound signal of each real abnormal sound category and the center of the sphere of the corresponding real abnormal sound category is calculated based on the sound feature of the sample sound signal of each real abnormal sound category, and then the mean and standard deviation of the Euclidean distance between the sound feature of the sample sound signal of each real abnormal sound category and the center of the sphere of the corresponding real abnormal sound category are calculated, which are respectively recorded as [μ i ,δ i ], where i = 1, 2, 3, 4, and finally the sphere radius is calculated based on the 3sigma criterion, using R i Represents the radius of the sphere. According to the 3sigma criterion, R i =3δ i .

[0106] In this embodiment, if there is no corresponding target distance less than the spherical radius in all real abnormal sound categories, that is, the sound feature of the sample sound signal does not meet the mean feature of the sound signal of any abnormal sound category, the predicted abnormal sound category of this sound signal is identified as an unknown abnormal sound category. For example, the Euclidean distance is If not present The predicted abnormal sound category of this sound signal is identified as an unknown abnormal sound category.

[0107] In a specific example, as shown in Table 2 below, the target distance between a sound signal and each real abnormal sound category and the spherical radius corresponding to the real abnormal sound category are shown.

[0108] Table 2 The target distance between a sound signal and each real abnormal sound category and the corresponding spherical radius of the real abnormal sound category

[0109]

[0110]

[0111] According to Table 2, the target distance between the sound feature of the sound signal in the healthy and no abnormal sound category and the center of the sphere of the healthy and no abnormal sound category is greater than the sphere radius corresponding to the healthy and no abnormal sound category, so the sound signal is not in the healthy and no abnormal sound category.

[0112] The target distance between the sound feature of the sound signal in the flexible ring extrusion noise and the center of the sphere of the flexible ring extrusion noise is greater than the sphere radius corresponding to the flexible ring extrusion noise, so the sound signal is not the flexible ring extrusion noise.

[0113] In the abnormal sound similar to glass fiber breakage, the target distance between the sound feature of the sound signal and the center of the sphere similar to the abnormal sound of glass fiber breakage is greater than the sphere radius corresponding to the abnormal sound similar to glass fiber breakage, so the sound signal is not the abnormal sound similar to glass fiber breakage.

[0114] In a periodic whistle, the target distance between the sound feature of the sound signal and the center of the sphere of the periodic whistle is smaller than the radius of the sphere corresponding to the periodic whistle, so the sound signal is a periodic whistle.

[0115] In an optional implementation manner, the abnormal sound category recognition model is specifically used to identify the predicted abnormal sound category corresponding to the sample sound signal as the real abnormal sound category in response to the target distance being smaller than the spherical radius corresponding to the real abnormal sound category.

[0116] In this embodiment, if the target distance in a certain real abnormal sound category is smaller than the spherical radius of the corresponding real abnormal sound category, that is, the sound feature of the sample sound signal meets the mean feature of the sound signal of the abnormal sound category, the predicted abnormal sound category of the sound signal is identified as the real abnormal sound category. For example, the Euclidean distance is like The predicted abnormal sound category of the sound signal is identified as the abnormal sound category of the sample sound signal corresponding to i.

[0117] In an optional implementation, the abnormal sound category recognition model is specifically used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis to obtain a sample time-frequency spectrum, and to extract sound features of the sample time-frequency spectrum.

[0118] In this embodiment, when extracting the sound features of the sample sound signal, the existing technology mostly uses frequency domain feature extraction methods such as octave energy and spectrum center of gravity. These extraction methods need to first obtain the spectrum of the sound signal through fast Fourier transform. In this process, the non-stationary local characteristics of the sound signal are ignored, and the obtained features have a weak ability to characterize the laws of the sound signal. In addition, although the Mel spectrum feature can characterize the local characteristics of the sound signal, the operating conditions of the wind turbine are changeable, especially the speed of the wind wheel is changeable, which leads to fluctuations in the characteristic frequencies of similar abnormal sounds during the operation of the blades, resulting in the Mel spectrum feature under similar abnormal sounds. The robustness is not strong. To address this problem, this embodiment constructs a sample time-frequency spectrum by amplitude pooling the time-frequency spectrum of the sample sound signal along the frequency axis. This method allows the characteristics of similar abnormal sounds to fluctuate within a certain frequency range without affecting the final time-frequency spectrum, thereby improving the robustness of acoustic feature extraction.

[0119] Specifically, to perform amplitude pooling of the time-frequency spectrum of the sample sound signal along the frequency axis, it is necessary to first construct the octave interval, and then calculate each octave interval for pooling. The pooling methods include average pooling, maximum pooling, etc. Among them, average pooling can smooth the random noise in the data, thereby making the features more stable, performing better in suppressing noise, and having better effects.

[0120] In a specific example, Figure 2The figure shows a flowchart of abnormal sound category recognition of a sample sound signal. The Euclidean distance is calculated according to the sample sound signal to be identified, and then the relationship between the corresponding Euclidean distance and the spherical radius in each real abnormal sound category is judged respectively. If there is a real abnormal sound category that meets the condition that the Euclidean distance is less than the spherical radius, the sample sound signal is determined to be this real abnormal sound category; if there is no real abnormal sound category that meets the condition that the Euclidean distance is less than the spherical radius, the sample sound signal is determined to be an unknown abnormal sound category.

[0121] In an optional embodiment, after the above step S1 and before the above step S2, the method further includes step S0, performing amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis to obtain a sample time-frequency spectrum. In this embodiment, the abnormal sound category recognition model is used to extract the sound features of the sample time-frequency spectrum.

[0122] In this embodiment, after obtaining the training sample, the sample sound signal needs to be processed, and the time-frequency spectrum of the sample sound signal is amplitude pooled along the frequency axis to obtain the sample time-frequency spectrum, and then the abnormal sound category recognition model is used to extract the sound features of the sample time-frequency spectrum.

[0123] In another optional implementation, the time-frequency spectrum of the sample sound signal is amplitude pooled along the frequency axis using an abnormal sound category recognition model to obtain a sample time-frequency spectrum, and the sound features of the sample time-frequency spectrum are extracted.

[0124] In an optional implementation manner, the step of performing amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction specifically includes:

[0125] S01, performing short-time Fourier transform on the sample sound signal to obtain a time-frequency spectrum diagram;

[0126] S02, converting the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel-frequency spectrum;

[0127] S03. Perform amplitude pooling on the sample Mel-frequency spectrum graph along the frequency axis to obtain a sample time-frequency spectrum graph.

[0128] In this embodiment, a short-time Fourier transform is performed on the sample sound signal to obtain a time-frequency spectrum, such as Figure 3a As shown, it is a time-frequency spectrum diagram corresponding to a sample sound signal of a periodic whistle category.

[0129] Then, the energy amplitude in the two-dimensional time-frequency spectrum is converted into decibel value to obtain the sample Mel spectrum. The conversion formula is as follows, where y i Represents the i-th energy amplitude of the original time-frequency spectrum:

[0130] yi =20·log 10 (y i )

[0131] This conversion can further suppress the low-frequency components in the sample sound signal while enhancing the energy of the high-frequency components.

[0132] Finally, the sample Mel spectrum graph is amplitude pooled along the frequency axis to construct an amplitude pooled sample time spectrum graph, such as Figure 3b As shown, it is a sample time-frequency spectrum diagram corresponding to a sample sound signal of a periodic whistle category.

[0133] In an optional implementation manner, before S01, the step further includes: performing pre-emphasis processing on the sample sound signal.

[0134] In this embodiment, the sample sound signal is pre-emphasized to enhance the high-frequency information in the sample sound signal and compensate for the energy loss of the high-frequency part of the sample sound signal, thereby improving the quality of the sample sound signal. Figure 3c As shown, it is a sound signal waveform corresponding to a sample sound signal of a periodic whistle type, and the sound signal above 4000HZ in the waveform can be pre-enhanced.

[0135] In an optional embodiment, the fan blades are arranged on a hub, and an array type sound receiving device facing the inside of the blades is arranged at the connection between the fan blades and the hub. In this embodiment, the above step S1 specifically includes: obtaining training samples using the array type sound receiving device.

[0136] In this embodiment, when collecting sample sound signals, a microphone or other sound receiving device can be used. A hole is drilled on the root baffle of each fan blade, and an array microphone is fixed with bolts, and the microphone is directed toward the cavity of the fan blade. This is different from the prior art that uses an array microphone as a sound signal collection device and installs it on the outer wall of the tower or a bracket near the fan. In this embodiment, the microphone and the fan blade are in a one-to-one relationship, and the collected sound signal and the fan blade are also in a one-to-one relationship. In addition, because the microphone is at the closable blade root, the collected noise can be reduced while monitoring the entire fan blade. In addition, the construction is simple and the cost is low.

[0137] In a specific example, take the test data samples shown in Table 3 below.

[0138] Table 3 Test data samples

[0139]

[0140] The test data samples in Table 3 are input into the abnormal sound category recognition model to obtain the predicted abnormal sound category. The model is repeated 10 times, and the statistical results show that the recognition accuracy of the test data samples in this embodiment reaches 91.43%, and the standard deviation is 1.56%. The recognition result statistics of the test data samples are shown in Table 4.

[0141] Table 4. Statistics of recognition results of test data samples

[0142]

[0143]

[0144] As can be seen from Table 4, in the embodiment, the recognition accuracy of the sound signal of each abnormal sound category can reach more than about 80%, and the sample sound signal of the unknown category can be identified as an outlier, thereby avoiding misjudgment or omission of the abnormal sound of the unknown category.

[0145] The present embodiment and the prior art were run 10 times, and the accuracy and standard deviation were counted. The statistical results are as follows: Figure 4 As shown by Figure 4 It can be seen that the recognition accuracy of the abnormal sound category recognition model trained in this embodiment is higher than that of the prior art, and the standard deviation of the accuracy of 10 tests is lower than that of the prior art, and the recognition stability is higher.

[0146] Corresponding to the aforementioned training method for the abnormal sound category recognition model, this embodiment also provides a training system for the abnormal sound category recognition model, such as Figure 5 As shown, the training system comprises:

[0147] An acquisition module 41 is used to acquire training samples, wherein the training samples include a plurality of sample sound signals generated during the operation of the fan blades and real abnormal sound categories corresponding to the sample sound signals;

[0148] A training module 42, used to train the abnormal sound category recognition model using the training samples;

[0149] Among them, the abnormal sound category recognition model is used to extract the sound features of the sample sound signal, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signal corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. In response to the absence of a target distance smaller than the spherical radius of each real abnormal sound category, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category.

[0150] In an optional implementation manner, the abnormal sound category recognition model is used to identify the predicted abnormal sound category corresponding to the sample sound signal as the real abnormal sound category in response to the target distance being smaller than the spherical radius corresponding to the real abnormal sound category.

[0151] In an optional embodiment, the training module is specifically used to train the twin model of the abnormal sound category recognition model using the training samples; wherein the twin model is obtained by replacing the adaptive spherical domain recognition layer in the abnormal sound category recognition model with a fully connected layer; and synchronizing the model parameters in the twin model except the fully connected layer to the abnormal sound category recognition model.

[0152] In an optional implementation, the abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis to obtain a sample time-frequency spectrum, and to extract sound features of the sample time-frequency spectrum.

[0153] In an optional embodiment, the abnormal sound category recognition model is used to perform a short-time Fourier transform on the sample sound signal to obtain a time-frequency spectrum; convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel-frequency spectrum; and perform amplitude pooling on the sample Mel-frequency spectrum along the frequency axis to obtain a sample time-frequency spectrum.

[0154] In an optional embodiment, the training system also includes an amplitude pooling module, which is used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis to obtain a sample time-frequency spectrum; the abnormal sound category recognition model is used to extract sound features of the sample time-frequency spectrum.

[0155] In an optional implementation, the amplitude pooling module specifically includes a time-spectrogram unit, a sample Mel-spectrogram unit and a sample time-spectrogram unit:

[0156] The time-spectrogram unit is used to perform short-time Fourier transform on the sample sound signal to obtain a time-spectrogram;

[0157] The sample Mel-spectrogram unit is used to convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel-spectrogram;

[0158] The sample time-frequency spectrum unit is used to perform amplitude pooling on the sample Mel-frequency spectrum along the frequency axis to obtain a sample time-frequency spectrum.

[0159] In an optional implementation, the training system further includes a pre-emphasis module for performing pre-emphasis processing on the sample sound signal.

[0160] In an optional embodiment, the fan blades are arranged on a wheel hub, and an array-type sound receiving device facing the interior of the blades is provided at the connection between the fan blades and the wheel hub;

[0161] The acquisition module is specifically used to acquire training samples using the array sound receiving device.

[0162] As for the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system embodiment described above is only illustrative, wherein the units described as separate components may or may not be physically separated, and the components as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the disclosed solution.

[0163] Example 2

[0164] This embodiment provides a method for identifying the type of abnormal noise of fan blades. Figure 6 As shown, the identification method comprises the following steps:

[0165] S51, collecting target sound signals generated during the operation of the fan blades;

[0166] S52, inputting the target sound signal into an abnormal sound category recognition model to obtain a predicted abnormal sound category corresponding to the target sound signal; wherein the abnormal sound category recognition model is obtained by using the training method of the abnormal sound category recognition model in Example 1. wherein the abnormal sound category recognition model is used to extract the sound features of the target sound signal, and calculate the target distance between the sound features of the target sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the target sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category, and in response to the absence of a target distance less than the spherical radius of each real abnormal sound category, identify the predicted abnormal sound category corresponding to the target sound signal as an unknown abnormal sound category.

[0167] In the identification method provided in this embodiment, the abnormal sound category recognition model can be used to identify the abnormal sound category corresponding to the target sound signal generated during the operation of the fan blade, including the unknown abnormal sound category, to prevent the target sound signal of the unknown abnormal sound category from being mistakenly identified as other abnormal sound categories, thereby improving the accuracy of identifying the abnormal sound category of the fan blade.

[0168] In an optional implementation, the abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis to obtain a target time-frequency spectrum; and to extract sound features of the target time-frequency spectrum.

[0169] In an optional embodiment, after the above step S51 and before the above step S52, step S50 is further included: amplitude pooling of the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain a target time-frequency spectrum. In this embodiment, the abnormal sound category recognition model is used to extract the sound features of the target time-frequency spectrum.

[0170] In this embodiment, after collecting the target sound signal, the target sound signal needs to be processed, and the time-frequency spectrum of the target sound signal is amplitude pooled along the frequency axis to obtain the target time-frequency spectrum, and then the abnormal sound category recognition model is used to extract the sound features of the target time-frequency spectrum.

[0171] In an optional implementation manner, the step of performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain the target time-frequency spectrum specifically includes:

[0172] Performing a short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum;

[0173] Converting the energy amplitude in the time-frequency spectrum into a decibel value to obtain a target Mel-frequency spectrum;

[0174] The target Mel-spectrogram is amplitude pooled along the frequency axis to obtain a target time-frequency spectrum.

[0175] In an optional implementation, the step of performing a short-time Fourier transform on the target sound signal includes: performing a pre-emphasis process on the target sound signal. Corresponding to the aforementioned method for identifying the abnormal sound category of a fan blade, this embodiment further provides a system for identifying the abnormal sound category of a fan blade, such as Figure 7 As shown, the identification system comprises:

[0176] The acquisition module 71 is used to acquire the target sound signal generated during the operation of the fan blades;

[0177] The prediction module 72 is used to input the target sound signal into an abnormal sound category recognition model to obtain a predicted abnormal sound category corresponding to the target sound signal; wherein the abnormal sound category recognition model is obtained by using the training method of the abnormal sound category recognition model in Example 1.

[0178] In an optional implementation, the abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis to obtain a target time-frequency spectrum; and to extract sound features of the target time-frequency spectrum.

[0179] In an optional embodiment, the abnormal sound category recognition model is used to perform a short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum; convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a target Mel-frequency spectrum; and perform amplitude pooling on the target Mel-frequency spectrum along the frequency axis to obtain a target time-frequency spectrum.

[0180] In an optional embodiment, the recognition system further includes a pooling processing module, which is used to perform a short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum; convert the energy amplitude in the time-frequency spectrum into a decibel value to obtain a target Mel-frequency spectrum; and perform amplitude pooling on the target Mel-frequency spectrum along the frequency axis to obtain a target time-frequency spectrum;

[0181] The abnormal sound category recognition model is used to extract the sound features of the target time spectrum graph.

[0182] In an optional implementation, the recognition system further includes a pre-emphasis processing module for performing pre-emphasis processing on the target sound signal.

[0183] As for the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system embodiment described above is only illustrative, wherein the units described as separate components may or may not be physically separated, and the components as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the disclosed solution.

[0184] Example 3

[0185] This embodiment provides an electronic device, including a processor, a memory, and a computer program stored in the memory and used to run on the processor. When the computer program is executed by the processor, it implements any one of the training methods for abnormal sound category recognition models described in Example 1 or the method for identifying abnormal sound categories of fan blades as described in Example 2.

[0186] Figure 8 A schematic diagram of the structure of an electronic device provided in this embodiment. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor, and when the processor executes the program, it implements any one of the training methods for the abnormal sound category recognition model described in Example 1 or the method for recognizing the abnormal sound category of the fan blade described in Example 2. Figure 8 The electronic device 90 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0187] like Figure 8 As shown, the electronic device 90 may be in the form of a general-purpose computing device, for example, it may be a server device. The components of the electronic device 90 may include, but are not limited to: at least one processor 91, at least one memory 92, and a bus 93 connecting different system components (including the memory 92 and the processor 91).

[0188] The bus 93 includes a data bus, an address bus, and a control bus.

[0189] The memory 92 may include a volatile memory, such as a random access memory (RAM) 921 and / or a cache memory 922 , and may further include a read-only memory (ROM) 923 .

[0190] The memory 92 may also include a program tool 925 (or utility) having a set (at least one) of program modules 924, such program modules 924 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include an implementation of a network environment.

[0191] The processor 91 executes various functional applications and data processing by running the computer program stored in the memory 92, such as the training method of any abnormal sound category recognition model described in Example 1 or the method for identifying the abnormal sound category of the fan blade as described in Example 2.

[0192] The electronic device 90 may also communicate with one or more external devices 94 (e.g., keyboards, pointing devices, etc.). Such communication may be performed via an input / output (I / O) interface 95. Furthermore, the electronic device 90 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 96. As shown, the network adapter 96 communicates with other modules of the electronic device 90 via a bus 93. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 90, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems, etc.

[0193] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0194] Example 4

[0195] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the training method of any abnormal sound category recognition model described in Example 1 or the method for identifying the abnormal sound category of a fan blade described in Example 2 is implemented.

[0196] The readable storage medium may include but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device or any suitable combination of the above.

[0197] Example 5

[0198] This embodiment also provides a computer program product, including a computer program, which, when executed by a processor, implements any one of the training methods for abnormal sound category recognition models described in Example 1 or the method for identifying abnormal sound categories of fan blades described in Example 2.

[0199] Among them, the program code for executing the computer program product of the present disclosure can be written in any combination of one or more programming languages, and the program code can be executed completely on the user device, partially on the user device, as an independent software package, partially on the user device and partially on a remote device, or completely on the remote device.

[0200] Although the specific embodiments of the present disclosure are described above, those skilled in the art should understand that this is only an example, and the protection scope of the present disclosure is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present disclosure, but these changes and modifications all fall within the protection scope of the present disclosure.

Claims

1. A training method for an abnormal sound category recognition model, characterized in that: The training method comprises: Acquire training samples, wherein the training samples include a plurality of sample sound signals generated during the operation of the fan blades and real abnormal sound categories corresponding to the sample sound signals; Using the training samples to train an abnormal sound category recognition model; Among them, the abnormal sound category recognition model is used to extract the sound features of the sample sound signal, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signal corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. In response to the absence of a target distance smaller than the spherical radius of each real abnormal sound category, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category.

2. The training method of the abnormal sound category recognition model according to claim 1, characterized in that: The abnormal sound category recognition model is used to identify the predicted abnormal sound category corresponding to the sample sound signal as the real abnormal sound category in response to the target distance being smaller than the spherical radius corresponding to the real abnormal sound category.

3. The training method of the abnormal sound category recognition model according to claim 1, characterized in that: The step of using the training samples to train the abnormal sound category recognition model specifically includes: The twin model of the abnormal sound category recognition model is trained using the training samples; wherein the twin model is obtained by replacing the adaptive spherical domain recognition layer in the abnormal sound category recognition model with a fully connected layer; The model parameters of the twin model except the fully connected layer are synchronized to the abnormal sound category recognition model.

4. The training method of the abnormal sound category recognition model according to claim 1, characterized in that: The abnormal sound category recognition model is used to perform amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain a sample time-frequency spectrum, and to extract sound features of the sample time-frequency spectrum.

5. The training method of the abnormal sound category recognition model according to claim 1, characterized in that: Before the step of using the training samples to train the abnormal sound category recognition model, the method further includes: Performing amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain a sample time-frequency spectrum; The abnormal sound category recognition model is used to extract the sound features of the spectrogram of the sample.

6. The training method of the abnormal sound category recognition model according to claim 4 or 5, characterized in that: The step of performing amplitude pooling on the time-frequency spectrum of the sample sound signal along the frequency axis direction to obtain the sample time-frequency spectrum specifically comprises: Performing a short-time Fourier transform on the sample sound signal to obtain a time-frequency spectrum diagram; Converting the energy amplitude in the time-frequency spectrum into a decibel value to obtain a sample Mel-frequency spectrum; The sample Mel-spectrogram is amplitude pooled along the frequency axis to obtain a sample time-frequency spectrum.

7. The training method of the abnormal sound category recognition model according to claim 6, characterized in that: Before the step of performing short-time Fourier transform on the sample sound signal, the step further includes: The sample sound signal is pre-emphasized.

8. The method for training an abnormal sound category recognition model according to any one of claims 1 to 5, characterized in that: The fan blades are arranged on the hub, and an array type sound receiving device facing the inside of the blades is arranged at the connection between the fan blades and the hub; The step of obtaining training samples specifically includes: The array-type sound receiving device is used to obtain training samples.

9. A method for identifying abnormal noise categories of fan blades, characterized in that: The identification method comprises: Collect the target sound signals generated during the operation of the fan blades; The target sound signal is input into an abnormal sound category recognition model to obtain a predicted abnormal sound category corresponding to the target sound signal; wherein the abnormal sound category recognition model is obtained by using the training method of the abnormal sound category recognition model according to any one of claims 1 to 8; wherein the abnormal sound category recognition model is used to extract the sound features of the target sound signal, and calculate the target distance between the sound features of the target sound signal and the center of the sphere of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the target sound signal according to each target distance and the sphere radius of the corresponding real abnormal sound category, and in response to the absence of a target distance less than the sphere radius of each real abnormal sound category, identify the predicted abnormal sound category corresponding to the target sound signal as an unknown abnormal sound category.

10. The method for identifying abnormal noise types of fan blades according to claim 9, characterized in that: The abnormal sound category recognition model is used for: Performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain a target time-frequency spectrum; The sound features of the target spectrogram are extracted.

11. The method for identifying abnormal noise types of fan blades according to claim 9, characterized in that: Before the step of inputting the target sound signal into the abnormal sound category recognition model, the method further includes: Performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain a target time-frequency spectrum; The abnormal sound category recognition model is used to extract the sound features of the target time spectrum graph.

12. The method for identifying abnormal noise types of fan blades according to claim 10 or 11, characterized in that: The step of performing amplitude pooling on the time-frequency spectrum of the target sound signal along the frequency axis direction to obtain the target time-frequency spectrum specifically includes: Performing a short-time Fourier transform on the target sound signal to obtain a time-frequency spectrum; Converting the energy amplitude in the time-frequency spectrum into a decibel value to obtain a target Mel-frequency spectrum; The target Mel-spectrogram is amplitude pooled along the frequency axis to obtain a target time-frequency spectrum.

13. The method for identifying abnormal noise types of fan blades according to claim 12, characterized in that: The step of performing short-time Fourier transform on the target sound signal comprises: The target sound signal is pre-emphasized.

14. A training system for an abnormal sound category recognition model, characterized in that: The training system comprises: An acquisition module, used for acquiring training samples, wherein the training samples include a plurality of sample sound signals generated during the operation of the fan blades and real abnormal sound categories corresponding to the sample sound signals; A training module, used to train the abnormal sound category recognition model using the training samples; Among them, the abnormal sound category recognition model is used to extract the sound features of the sample sound signal, and calculate the spherical center and spherical radius of the corresponding real abnormal sound category based on the sound features of the sample sound signal corresponding to each real abnormal sound category, and calculate the target distance between the sound features of the sample sound signal and the spherical center of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the sample sound signal according to each target distance and the spherical radius of the corresponding real abnormal sound category. In response to the absence of a target distance smaller than the spherical radius of each real abnormal sound category, the predicted abnormal sound category corresponding to the sample sound signal is identified as an unknown abnormal sound category.

15. A fan blade abnormal noise classification identification system, characterized in that: The identification system comprises: A collection module, used to collect target sound signals generated during the operation of the fan blades; A prediction module, used for inputting the target sound signal into an abnormal sound category recognition model to obtain a predicted abnormal sound category corresponding to the target sound signal; wherein the abnormal sound category recognition model is obtained by using the training method of the abnormal sound category recognition model described in any one of claims 1 to 8, wherein the abnormal sound category recognition model is used to extract the sound features of the target sound signal, and calculate the target distance between the sound features of the target sound signal and the center of the sphere of each real abnormal sound category, and identify the predicted abnormal sound category corresponding to the target sound signal according to each target distance and the sphere radius of the corresponding real abnormal sound category, and in response to the absence of a target distance less than the sphere radius of each real abnormal sound category, identify the predicted abnormal sound category corresponding to the target sound signal as an unknown abnormal sound category.

16. An electronic device comprising a memory, a processor, and a computer program stored in the memory and used to run on the processor, characterized in that: When the processor executes the computer program, the training method of the abnormal sound category recognition model described in any one of claims 1 to 8 or the recognition method of the abnormal sound category of the fan blade described in any one of claims 9 to 13 is implemented.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the training method of the abnormal sound category recognition model described in any one of claims 1 to 8 or the recognition method of the abnormal sound category of a fan blade described in any one of claims 9 to 13 is implemented.

18. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the training method of the abnormal sound category recognition model described in any one of claims 1 to 8 or the identification method of the abnormal sound category of a fan blade described in any one of claims 9 to 13 is implemented.

Images