Offshore wind power blade state identification method based on TDNN model

Abstract

The invention discloses an offshore wind power blade state recognition method based on a TDNN model. The method comprises the following steps: collecting voiceprint data of a blade through a sensor array; noise reduction is carried out on the original voiceprint signal by using an adaptive filtering technology, and background noise is removed; extracting features such as an MFCC coefficient, an LPCC coefficient, short-time energy and a short-time zero-crossing rate of the voiceprint signal; carrying out normalization and standardization on the features by adopting a Z-score method; a TDNN model comprising three hidden layers is used for training, and an Adam optimizer and a dynamic learning rate adjustment strategy are combined, so that the model training efficiency is improved; by adding noise, time stretching, pitch transformation and random cutting and splicing, a training data set is expanded to be three times of the original training data set, and the generalization ability of the model is remarkably improved; through fusion of a TDNN algorithm and a multi-dimensional data enhancement technology, the time sequence dependency relationship of voiceprint signals is effectively captured, and the accuracy, robustness and generalization ability of offshore wind power blade state recognition are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of sound signal processing, and in particular to a method for identifying the state of an offshore wind turbine blade based on a TDNN model. Background Art

[0002] Voiceprint recognition technology was first applied in the fields of speech recognition and identity authentication, but in recent years its potential has gradually expanded to multiple emerging fields. As the core component of the wind power generation system, the health status of offshore wind turbine blades directly affects the efficiency and stability of the wind turbine. Since offshore wind turbines operate in a special environment far away from land, traditional detection methods are often limited by factors such as weather, sea conditions and equipment location. In recent years, combining voiceprint recognition technology to monitor the status of offshore wind turbine blades has become a new research direction. By analyzing the sound signals generated by wind turbine blades during operation and identifying the characteristic patterns therein, early fault warning and status monitoring can be achieved. Voiceprint recognition technology provides an efficient and non-invasive monitoring method for the offshore wind power industry, which can reduce operation and maintenance costs and improve the reliability and economic benefits of wind turbines.

[0003] Traditional monitoring technologies (such as vibration sensors and drone inspections) rely on manual intervention, making it difficult to capture early blade damage in real time and are limited by environmental conditions. Existing voiceprint recognition models have low signal-to-noise ratios and insufficient feature extraction in complex ocean noise environments, resulting in high false alarm rates and weak generalization capabilities. The singleness of data samples leads to insufficient adaptability of the model to changes in blade types and operating conditions. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for identifying the state of an offshore wind turbine blade based on a TDNN model, comprising the following steps: S1. Cover different parts of the fan through the sensor array, synchronously monitor the soundprint signals of different areas of the blade, collect the soundprint data of the blade, and use the specific sound signals during operation as the soundprint data; S2, synchronizing and transmitting the data obtained in the previous step; S3, performing signal preprocessing and noise reduction on the data obtained in the previous step, and retaining the core soundprint information generated by the blade state; S4, preprocessing the core voiceprint information obtained in the previous step to extract multiple features of the voiceprint signal, including Mel frequency cepstral coefficients, linear prediction cepstral coefficients, short-time energy, and short-time zero-crossing rate; S5, normalize the features obtained in the previous step, using a zero mean unit variance normalization method to make each feature value distributed on the same scale and unify the feature parameter scale; S6. The data obtained in the previous step is trained with a TDNN model containing 3 hidden layers, and the Adam optimizer and dynamic learning rate adjustment strategy are used to converge and train the model; S7. For the model obtained in the previous step, introduce multiple data enhancement operations to expand the original data set to three times the original size; S8. Use deep learning technology to train blade sound samples, generate the optimal model, and perform online reasoning and real-time monitoring.

[0005] The technical solution further defined in the present invention is: Furthermore, in step S3, the specific method of noise reduction is: dividing the original sound data into frames, converting them into frequency domain signals, removing noise based on deep learning intelligent algorithm processing, and then converting them back to time domain to obtain noise-reduced sound data; The original sound data is divided into frames, and the short-time energy threshold method is used to calculate the signal energy of each frame, set the dynamic threshold, and remove the silent segments with energy below the threshold, as shown in the following formula: Among them, E n is the short-time energy of the nth frame, x(n+m) is the value of the n+mth sample point, and L is half the frame length; And combined with the short-time zero-crossing rate to distinguish effective signals from background noise: Where sgn[·] is the sign function, x(n) represents the speech signal, and N represents the window length.

[0006] As described above, in the offshore wind turbine blade state recognition method based on the TDNN model, in step S4, short-time Fourier transform and wavelet transform are combined to perform time-frequency analysis on the noise reduction data, and extract the soundprint feature parameters as the TDNN model input; The extraction process of Mel-frequency cepstral coefficients includes the following steps: S4.1, pre-emphasis: enhance high-frequency components through a first-order high-pass filter: Where z represents a complex variable, μ is the pre-emphasis coefficient; S4.2, Frame Windowing: Frame shift 10ms, use Hamming window to smooth the signal, the definition formula is as follows: in, is the weight of the window function, N is the number of sample points, and n is the sample number in the frame; the boundary of the signal is smoothed by multiplying the window function point by point with each frame signal; S4.3. Design Mel filters and calculate the energy output of each filter. After taking the logarithm of the filter group energy, perform discrete cosine transform and retain the first 20 dimensional coefficients as Mel frequency cepstral coefficient features.

[0007] As described above, in a method for identifying the state of an offshore wind turbine blade based on a TDNN model, in step S4, the method for extracting the linear prediction cepstral coefficients is: solving the linear prediction coding coefficients by an autocorrelation method, performing cepstral analysis on the linear prediction coding coefficients, and extracting the first 12-dimensional linear prediction cepstral coefficient features.

[0008] As described above, in the offshore wind turbine blade state recognition method based on the TDNN model, in step S5, the minimum-maximum normalization is used to linearly scale the eigenvalue to the range of [0,1] or [-1,1]; and the Z-score normalization is used to convert it into a standard normal distribution.

[0009] As described above, in a method for identifying the state of an offshore wind turbine blade based on a TDNN model, in step S6, the TDNN model includes three hidden layers, each layer has 128 neurons, and uses a ReLU activation function; the input layer receives the preprocessed feature vector, and the output layer uses a softmax function for multi-classification prediction; the hidden layer includes a series of delayed neurons, and an expansion coefficient is introduced in the hidden layer; each neuron receives not only the input of the current time step, but also the input of several previous time steps; assuming that the hidden layer has N neurons, the output of the i-th neuron at time t is h i (t), then its input is expressed as: Where M is the number of neurons in the input layer, w ij is the weight from the input layer to the hidden layer, L is the number of neurons in the hidden layer, is the delay weight between neurons in the hidden layer, τ represents the number of delay steps, D is the maximum number of delay steps, b i is the bias term.

[0010] As described above, in the offshore wind turbine blade state recognition method based on the TDNN model, in step S6, the output layer of the TDNN model generates the final recognition result according to the output of the hidden layer, and the output layer is set as a softmax classifier to calculate the probability distribution of each sound device; assuming that the output layer has C neurons corresponding to C sound devices, the output of the cth neuron is: Among them, z c (t) is the input of the cth neuron in the output layer and is calculated by the following formula: Among them, vci is the weight from the hidden layer to the output layer, a c is the bias term.

[0011] As described above, in the offshore wind turbine blade state recognition method based on the TDNN model, in step S6, during the training process, the Adam optimizer is used, the learning rate is initially set to 0.001, and is dynamically adjusted according to the loss change during the training process; the batch size is set to 64; and the total number of iterations is set to 5000.

[0012] As described above, in the offshore wind turbine blade state recognition method based on the TDNN model, in step S7, the data enhancement operation includes adding noise, time stretching, pitch shifting, and random cropping and splicing.

[0013] As described above, in a method for identifying the state of an offshore wind turbine blade based on a TDNN model, in step S8, different thresholds are set for different types of fault types according to historical data, and in the online reasoning process, the normal posterior probability is compared with the abnormal posterior probability. When the normal posterior probability is greater than the abnormal posterior probability, the blade is judged to be in a normal state; otherwise, the blade is judged to be in an abnormal state.

[0014] The beneficial effects of the present invention are: In the present invention, by integrating the TDNN algorithm and multi-dimensional data enhancement technology, the temporal dependency of the voiceprint signal is effectively captured, and the accuracy, robustness and generalization ability of offshore wind turbine blade status recognition are significantly improved; the experimental results show that under complex environmental noise, the average recognition accuracy is improved to 95.60%, and the average false alarm rate is reduced to 4.20%, which provides strong technical support for the online monitoring and fault warning of offshore wind turbine blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0016] This embodiment provides a method for identifying the state of an offshore wind turbine blade based on a TDNN model. Figure 1 As shown, the following steps are included: S1. Cover different parts of the fan through the sensor array, synchronously monitor the soundprint signals of different areas of the blades, collect the soundprint data of the blades, and use the specific sound signals during operation as the soundprint data to ensure the comprehensiveness and accuracy of the collected data and provide high-quality data support for subsequent feature extraction.

[0017] At the wind farm, representative wind turbines of different manufacturers, models, installed capacities and years of operation were selected, and multi-protected non-contact industrial microphones were installed 3 to 5 meters from the bottom of the tower to ensure stable performance and high sound pickup quality in harsh environments. Continuous data collection was carried out for one month to record blade sounds and environmental parameters such as unit status, wind speed, wind direction, etc., in order to compare the soundprint characteristics under different conditions and preliminarily reveal how blade sounds change with working conditions.

[0018] The sampling frequency of the sound level meter is set to 2048 Hz, with 4096 sampling points per frame; the sampling frequency of the vibration sensor is 1024 Hz, with 2048 points per frame.

[0019] S2. Synchronize and transmit the data obtained in step S1; transmit data to the data center in real time via PROFIBUS fieldbus and 4G / 5G wireless network, and use AES encryption technology to ensure transmission security; after receiving the data, the data center uses big data processing platform (Hadoop, Spark) and machine learning algorithms (SVM, CNN) to deeply analyze voiceprint and status data, extract valuable information and patterns, and support fan status monitoring and fault diagnosis; data visualization technology (ECharts, Tableau) presents the analysis results intuitively for operation and maintenance personnel to understand.

[0020] S3. Perform signal preprocessing and noise reduction on the data obtained in step S2 to retain the core soundprint information generated by the blade status; this embodiment proposes an adaptive wavelet transform and EMD joint noise reduction algorithm, which integrates the localization characteristics of wavelet transform in the time-frequency domain and the ability of EMD decomposition to process complex non-stationary signals, effectively filters out background noise and improves the quality of picked-up signals; by optimizing the algorithm parameters (EMD decomposition function), high-sensitivity acquisition of long-distance sound waves is achieved to ensure the capture of clear blade soundprint signals in the complex environment of offshore wind farms.

[0021] The environment of offshore wind farms is complex, with great interference from natural and mechanical noise. To reduce the impact of noise, an intelligent noise reduction algorithm is used to process sound data. This algorithm is based on deep learning and learns noise distribution and pure signal characteristics through training samples. It automatically identifies and separates noise and blade soundprints, effectively removes environmental noise, retains soundprint details, and ensures that the data is pure and accurate.

[0022] The noise reduction process is as follows: the original sound data is framed, converted into frequency domain signals, processed by an intelligent algorithm to remove noise, and then converted back to the time domain to obtain the noise-reduced sound data.

[0023] The original sound data is divided into frames, and the short-time energy threshold method is used to calculate the signal energy of each frame, set a dynamic threshold, and remove silent segments with energy below the threshold, as shown in the following formula: Among them, E n is the short-time energy of the nth frame, x(n+m) is the value of the n+mth sample point, and L is half the frame length.

[0024] And combined with the short-time zero-crossing rate to distinguish effective signals from background noise: Here, sgn[·] is the sign function, x(n) represents the speech signal, and N represents the window length. The short-time zero-crossing rate reflects the frequency characteristics of the signal and helps distinguish high-frequency and low-frequency sounds.

[0025] S4. Preprocess the core voiceprint information obtained in step S3 to extract multiple features of the voiceprint signal, including Mel-Frequency Cepstrum Coefficients (MFCC), Linear Predictive Cepstral Coefficients (LPCC), short-time energy, and short-time zero-crossing rate.

[0026] The time-frequency analysis technology is used, combining Short-Time Fourier Transform (STFT) and Wavelet Transform (WT) to simultaneously reflect the time and frequency changes of the sound signal; STFT divides the sound signal into short-time signal segments and performs Fourier transform to reveal the changes of frequency components over time; WT uses wavelet function to decompose the signal to obtain details and approximate information of different frequency bands, which is suitable for non-stationary signal processing; combining the advantages of both, the denoised data is subjected to time-frequency analysis to extract voiceprint feature parameters as the input of the TDNN model.

[0027] The extraction process of Mel-frequency cepstral coefficients (MFCC) includes the following steps: S4.1, pre-emphasis: enhance high-frequency components through a first-order high-pass filter: Wherein, z represents a complex variable, μ is a pre-emphasis coefficient, and its value range is usually between 0.9 and 1. In practical applications, the effect of pre-emphasis can be adjusted by selecting a suitable μ value.

[0028] S4.2, Frame Windowing: Frame shift 10ms, use Hamming window to smooth the signal, the definition formula is as follows: in, is the weight of the window function, N is the number of sample points, and n is the sample number in the frame. By multiplying the window function point by point with each frame of signal, the boundary of the signal can be smoothed and the spectrum leakage effect can be reduced.

[0029] S4.3. Design Mel filters and calculate the energy output of each filter. After taking the logarithm of the filter group energy, perform discrete cosine transform and retain the first 20 dimensional coefficients as Mel frequency cepstral coefficient features.

[0030] The method for extracting linear prediction cepstral coefficients (LPCC) is as follows: solving the linear predictive coding (LPC) coefficients by the autocorrelation method, performing cepstral analysis on the LPC coefficients, and extracting the first 12-dimensional LPCC features.

[0031] S5. Normalize the features obtained in step S4 and use the zero mean unit variance (Z-score) standardization method to make each eigenvalue distributed on the same scale, unify the feature parameter scale, and ensure that the eigenvalues ​​between samples are in the same range, which is conducive to rapid convergence and stable training of the model.

[0032] Normalization processing aims to unify the scale of feature parameters and ensure that the feature values ​​between samples are in the same range, which is crucial for model training. The minimum-maximum normalization and Z-score normalization methods are used. The former linearly scales the feature values ​​to the range of [0,1] or [-1,1], and the latter converts them to a standard normal distribution. The appropriate method is selected according to the feature parameter distribution and model training requirements to ensure that the data input is stable and reliable.

[0033] S6. Train the high-quality data obtained in step S5 with a time delay neural network (TDNN) model containing 3 hidden layers, and use the Adam optimizer and dynamic learning rate adjustment strategy to achieve rapid convergence and stable training of the model; as a neural network architecture designed specifically for time series data analysis, the time delay neural network can significantly improve the accuracy and robustness of blade state recognition by capturing the temporal dependencies in the voiceprint signal.

[0034] In this embodiment, the TDNN model includes 3 hidden layers, each with 128 neurons, and uses the ReLU activation function to increase the nonlinear expression ability; the input layer receives the preprocessed feature vector, and the output layer uses the softmax function for multi-classification prediction; the hidden layer includes a series of delayed neurons, and a dilation coefficient (dilation=2,3,4) is introduced in the hidden layer to expand the temporal perception field; each neuron receives not only the input of the current time step, but also the input of several previous time steps, and captures the temporal dependency of the sound signal through the time delay connection, so that the TDNN can learn the temporal dynamic characteristics of the speech signal.

[0035] Assume that there are N neurons in the hidden layer, and the output of the i-th neuron at time t is h i(t), then its input is expressed as: Where M is the number of neurons in the input layer, w ij is the weight from the input layer to the hidden layer, L is the number of neurons in the hidden layer, is the delay weight between neurons in the hidden layer, τ represents the number of delay steps, D is the maximum number of delay steps, b i is the bias term.

[0036] The output layer is responsible for generating the final recognition result based on the output of the hidden layer. In the voiceprint recognition task, the output layer is usually a softmax classifier, which is used to calculate the probability distribution of each sound device. Assuming that the output layer has C neurons, corresponding to C sound devices, the output of the cth neuron is: Among them, z c (t) is the input of the cth neuron in the output layer and is calculated by the following formula: Among them, v ci is the weight from the hidden layer to the output layer, a c is the bias term.

[0037] During the training process, the Adam optimizer was used, and the learning rate was initially set to 0.001 and dynamically adjusted according to the change in loss during the training process; the batch size was set to 64 to ensure training efficiency and avoid memory overflow problems; after many experiments, it was found that when the number of iterations reached 5000, the accuracy of the model on the validation set tended to be stable, so the total number of iterations was set to 5000 to ensure that the model was fully learned.

[0038] S7. To improve the generalization ability of the model obtained in step S6, a variety of data enhancement operations are introduced, such as adding noise, time stretching, pitch shifting, and random cropping and splicing, so that the original data set is effectively expanded to three times the original size, that is, 30,000 samples; by enriching the training data, the generalization ability of the model is significantly improved, making it more adaptable to various changes in the actual environment.

[0039] Gaussian white noise with different signal-to-noise ratios is superimposed on the original voiceprint signal to simulate the background noise interference in the actual environment and enhance the model's robustness to noise; the time stretching algorithm is used to randomly adjust the speaking speed of the voiceprint signal to between 90% and 110% of the original speed to simulate the voiceprint changes at different speaking speeds and improve the model's adaptability to changes in speaking speed; the frequency conversion technology is used to fine-tune the tone of the voiceprint up and down to simulate the voiceprint characteristics at different pitches and enhance the model's ability to recognize tone changes; fragments of different lengths are randomly cropped from the original voiceprint signal and randomly spliced ​​to generate new training samples. This method increases data diversity and promotes the model's learning of local features; through the above data enhancement technology, the original data set is effectively expanded to three times the original size, that is, 30,000 samples, which greatly enriches the training data and provides more comprehensive learning materials for the TDNN model.

[0040] S8. Use deep learning technology to train blade sound samples, generate accurate recognition models, and perform online reasoning and real-time monitoring.

[0041] The online reasoning process receives sensor signals in real time, performs noise reduction, feature extraction and standardization, inputs the TDNN model, and outputs the blade status classification results (normal / abnormal type), where the reasoning delay meets the real-time requirements.

[0042] Among them, the setting and optimization of warning thresholds are crucial to the realization of fault warning in the online monitoring system based on voiceprint recognition. In the initial stage, thresholds are set for each fault type based on historical data and experience. These thresholds are based on the statistical distribution of voiceprint features, such as mean and standard deviation, and take into account the probability of fault occurrence, severity of consequences and operation and maintenance costs. As the system runs and data accumulates, the warning thresholds need to be dynamically adjusted and optimized. The thresholds are fine-tuned based on the corresponding relationship between real-time monitored voiceprint data and faults to improve the accuracy and timeliness of warnings.

[0043] The present embodiment provides an offshore wind turbine blade state recognition method based on a TDNN model, which aims to solve the problem of blade state recognition under complex environmental noise and long-distance signal acquisition. The recognition method comprises the following steps: collecting blade voiceprint data through a sensor array; using adaptive filtering technology (such as Wiener filtering) to reduce noise on the original voiceprint signal and remove background noise; extracting multiple features of the voiceprint signal, including Mel-frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), short-time energy and short-time zero-crossing rate; using a zero-mean unit variance (Z-score) method to normalize and standardize the features; using a TDNN model containing three hidden layers for training, combining an Adam optimizer and a dynamic learning rate adjustment strategy to improve the model training efficiency; introducing multiple data enhancement techniques, including adding noise, time stretching, pitch shifting, and random cropping and splicing, to expand the training data set to three times the original size, significantly improving the generalization ability of the model.

[0044] This embodiment effectively captures the temporal dependency of voiceprint signals by integrating the TDNN algorithm and multi-dimensional data enhancement technology, and significantly improves the accuracy, robustness and generalization ability of offshore wind turbine blade status recognition. Experimental results show that under complex environmental noise, the average recognition accuracy is improved to 95.60%, and the average false alarm rate is reduced to 4.20%, thus providing strong technical support for online monitoring and fault warning of offshore wind turbine blades.

[0045] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A method for identifying the state of offshore wind turbine blades based on a TDNN model, characterized in that: The following steps are involved: S1. Cover different parts of the fan through the sensor array, synchronously monitor the soundprint signals of different areas of the blade, collect the soundprint data of the blade, and use the specific sound signals during operation as the soundprint data; S2, synchronizing and transmitting the data obtained in the previous step; S3, performing signal preprocessing and noise reduction on the data obtained in the previous step, and retaining the core soundprint information generated by the blade state; S4, preprocessing the core voiceprint information obtained in the previous step to extract multiple features of the voiceprint signal, including Mel frequency cepstral coefficients, linear prediction cepstral coefficients, short-time energy, and short-time zero-crossing rate; S5, normalize the features obtained in the previous step, using a zero mean unit variance normalization method to make each feature value distributed on the same scale and unify the feature parameter scale; S6. The data obtained in the previous step is trained with a TDNN model containing 3 hidden layers, and the Adam optimizer and dynamic learning rate adjustment strategy are used to converge and train the model; S7. For the model obtained in the previous step, introduce multiple data enhancement operations to expand the original data set to three times the original size; S8. Use deep learning technology to train blade sound samples, generate the optimal model, and perform online reasoning and real-time monitoring.

2. The offshore wind turbine blade state recognition method based on the TDNN model according to claim 1 is characterized by: In step S3, the specific method of noise reduction is: dividing the original sound data into frames, converting them into frequency domain signals, removing noise based on the intelligent algorithm of deep learning, and then converting them back to the time domain to obtain the noise-reduced sound data; The original sound data is divided into frames, and the short-time energy threshold method is used to calculate the signal energy of each frame, set the dynamic threshold, and remove the silent segments with energy below the threshold, as shown in the following formula: Among them, E n is the short-time energy of the nth frame, x(n+m) is the value of the n+mth sample point, and L is half the frame length; And combined with the short-time zero-crossing rate to distinguish effective signals from background noise: Where sgn[·] is the sign function, x(n) represents the speech signal, and N represents the window length.

3. The offshore wind turbine blade state recognition method based on the TDNN model according to claim 1 is characterized by: In step S4, short-time Fourier transform and wavelet transform are combined to perform time-frequency analysis on the noise reduction data, and voiceprint feature parameters are extracted as input to the TDNN model; The extraction process of Mel-frequency cepstral coefficients includes the following steps: S4.1, pre-emphasis: enhance high-frequency components through a first-order high-pass filter: Where z represents a complex variable, μ is the pre-emphasis coefficient; S4.2, Frame Windowing: Frame shift 10ms, use Hamming window to smooth the signal, the definition formula is as follows: in, is the weight of the window function, N is the number of sample points, and n is the sample number in the frame; the boundary of the signal is smoothed by multiplying the window function point by point with each frame signal; S4.

3. Design Mel filters and calculate the energy output of each filter. After taking the logarithm of the filter group energy, perform discrete cosine transform and retain the first 20 dimensional coefficients as Mel frequency cepstral coefficient features.

4. The offshore wind turbine blade state recognition method based on the TDNN model according to claim 1 is characterized by: In step S4, the linear prediction cepstral coefficients are extracted by using an autocorrelation method to solve the linear prediction coding coefficients, performing cepstral analysis on the linear prediction coding coefficients, and extracting the first 12-dimensional linear prediction cepstral coefficient features.

5. The offshore wind turbine blade state identification method based on the TDNN model according to claim 3 is characterized by: In step S5, the minimum-maximum normalization is used to linearly scale the eigenvalues ​​to the range of [0, 1] or [-1, 1]; and the Z-score normalization is used to convert them into a standard normal distribution.

6. The offshore wind turbine blade state identification method based on the TDNN model according to claim 1 is characterized by: In step S6, the TDNN model includes 3 hidden layers, each with 128 neurons, and uses a ReLU activation function; the input layer receives the preprocessed feature vector, and the output layer uses a softmax function for multi-classification prediction; the hidden layer includes a series of delayed neurons, and an expansion coefficient is introduced in the hidden layer; each neuron receives not only the input of the current time step, but also the input of several previous time steps; assuming that the hidden layer has N neurons, the output of the i-th neuron at time t is h i (t), then its input is expressed as: Where M is the number of neurons in the input layer, w ij is the weight from the input layer to the hidden layer, L is the number of neurons in the hidden layer, is the delay weight between neurons in the hidden layer, τ represents the number of delay steps, D is the maximum number of delay steps, b i is the bias term.

7. The offshore wind turbine blade state identification method based on the TDNN model according to claim 6 is characterized by: In step S6, the output layer of the TDNN model generates the final recognition result according to the output of the hidden layer. The output layer is set as a softmax classifier to calculate the probability distribution of each sound-emitting device. Assuming that the output layer has C neurons corresponding to C sound-emitting devices, the output of the cth neuron is: Among them, z c (t) is the input of the cth neuron in the output layer and is calculated by the following formula: Among them, v ci is the weight from the hidden layer to the output layer, a c is the bias term.

8. The offshore wind turbine blade state identification method based on the TDNN model according to claim 1 is characterized by: In step S6, during the training process, the Adam optimizer is used, the learning rate is initially set to 0.001, and is dynamically adjusted according to the change in loss during the training process; the batch size is set to 64; and the total number of iterations is set to 5000.

9. The offshore wind turbine blade state identification method based on the TDNN model according to claim 1 is characterized by: In step S7, the data enhancement operations include adding noise, time stretching, pitch shifting, and random cropping and splicing.

10. The offshore wind turbine blade state identification method based on the TDNN model according to claim 1, characterized in that: In step S8, different thresholds are set for different types of faults according to historical data. During the online reasoning process, the normal posterior probability is compared with the abnormal posterior probability. When the normal posterior probability is greater than the abnormal posterior probability, the blade is judged to be in a normal state; otherwise, the blade is judged to be in an abnormal state.

Images