Full-system holographic wind turbine generator tower mixing state intelligent identification system based on deep learning

Abstract

The invention discloses a full-system holographic wind turbine generator tower mixing state intelligent identification system based on deep learning. According to the invention, the sensors are arranged at a plurality of key parts of the wind turbine generator mixed tower, so that the wind turbine generator mixed tower can be monitored in all directions, and the intelligent dynamic behavior monitoring of the mixed tower can be realized in combination with the diagnosis strategies of deep learning and online self-learning. The system comprises a data acquisition module, an offline training module, an online monitoring module, a mixed tower state evaluation module and an online learning module. Through the time-frequency double-constraint Attention-BiGRU automatic encoder model, health state evaluation and real-time feedback of all parts and the whole body of the mixed tower are realized, and reliable decision support is provided for safe operation and intelligent operation and maintenance of a wind field.

Description

Technical Field

[0001] This method belongs to the field of wind power generation, and specifically relates to an intelligent identification system for the mixed tower state of a full-system holographic wind turbine based on deep learning Technical Background

[0002] With the development of wind power in China in the past 20 years, onshore wind power has gradually expanded from traditional areas with rich wind resources to low-wind-speed areas. As is well known, the higher the tower barrel, the greater the wind speed and the higher the power generation efficiency. However, low-wind-speed areas pose higher requirements for the tower barrel structure, especially for the design of higher hub heights. Due to its relatively low stiffness, the pure steel tower barrel is prone to resonance with the impeller during the start-up and shutdown of the unit, and there are fatigue problems, high maintenance costs, and transportation restrictions on the bottom diameter. These factors limit its further development in the high tower height direction. In contrast, the steel-concrete tower barrel is gradually favored by owners due to its higher stiffness and economy

[0003] The steel-concrete tower barrel consists of a steel tower section, a transition section, and a concrete section. The lower half of the tower barrel is a concrete structure, and the upper half is a steel structure. Prestressed cables are arranged in the concrete part and are anchored to the transition section and the foundation at the upper and lower ends respectively. Due to the special structure, the concrete structure is composed of multiple segmented tower sections connected together, and each tower section is spliced by concrete segments, using dry connection or wet connection. At the same time, at the steel-concrete transition section, due to the change in the force transmission path, stress concentration is likely to occur at the interface. Therefore, special attention needs to be paid to the stress states of these key parts between tower sections, between the concrete segments of a single tower section, and at the steel-concrete transition section

[0004] However, traditional methods have a single monitoring location and a single monitoring means. They only analyze the operating state of the tower barrel manually by monitoring the changes in vibration amplitude and natural frequency. Due to the complex structure and large scale of the hybrid tower, the distribution of monitoring points is relatively limited. It is difficult to identify damage far from the monitoring points and early minor damage in a timely manner, and it is also vulnerable to human errors caused by insufficient domain knowledge. As a cutting-edge artificial intelligence technology, deep learning has demonstrated its powerful data processing and pattern recognition capabilities in multiple fields. It can effectively process high-dimensional data, thus providing accurate monitoring and diagnosis solutions for complex systems. However, current deep learning methods are often limited by the problem of data imbalance in industrial applications. Especially when there is a lack of fault state data, it is difficult to train an effective model. Secondly, deep learning methods often ignore traditional detection physical information, resulting in low accuracy of existing models. In addition, due to the complex structure and variable operating conditions of the concrete tower barrel, single monitoring is difficult to effectively reflect the true working conditions of the hybrid tower of the wind turbine. Moreover, during the service life of the wind turbine, the overall state of the hybrid tower will change over time, and situations such as inevitable false alarms and missed alarms of the monitoring system need to be considered. Therefore, how to monitor the state of the hybrid tower of the wind turbine in a multi-dimensional, comprehensive, dynamic and intelligent manner has become a key problem that urgently needs to be solved in the intelligent operation and maintenance of the hybrid tower of the wind turbine. Summary of the Invention

[0005] In view of the above problems, the present invention provides an intelligent recognition system for the state of the whole-holographic hybrid tower of a wind turbine based on deep learning. The present invention realizes the full-range monitoring of the hybrid tower structure by arranging sensors at multiple key parts of the hybrid tower (the top of the tower, the transition section of the hybrid tower, the prestressed cable, the tower section, the bolt, the tower foundation), and combines the diagnostic strategies of deep learning and online self-learning to realize the intelligent dynamic behavior monitoring of the hybrid tower.

[0006] As Figure 1 shown, the present invention includes a data acquisition module, an offline training module, an online monitoring module, a hybrid tower state evaluation module, and an online learning module.

[0007] The data acquisition module is responsible for collecting the full-range monitoring data of the hybrid tower of the wind turbine. Sensors are arranged at the top of the tower, the transition section, the prestressed cable, the tower section, the bolt, and the tower foundation of the hybrid tower to collect the sway angle data of the top of the tower, the stress and strain data of the transition section, the vibration data of the prestressed cable, the stress and strain data between the tower sections, the stress and strain data of the bolt, and the sway angle data of the tower foundation under different temperatures, humidities, wind speeds, and operating conditions of the wind turbine.

[0008] The offline training module is responsible for constructing normal behavior models for the top of the hybrid tower, the transition section, the prestressed cable, the tower section, the bolt, and the tower foundation, as Figure 2 shown.

[0009] S1 performs data screening on the collected data, divides the data into a training set and a validation set according to different monitoring parts, and performs normalization processing on the training set and the validation set; the training set only contains data in the normal state.

[0010] S2 constructs an Attention-BiGRU autoencoder model to train the normal behavior models of each part based on the data of different monitoring parts, and selects excellent-performing normal behavior models according to the model evaluation indicators of root mean square error, mean absolute error, and mean absolute percentage error on the validation set for subsequent online monitoring.

[0011] Among them, the Attention-BiGRU autoencoder model is as Figure 3 shown, and it includes an Attention-BiGRU encoder, a feature vector, and an Attention-BiGRU decoder. The Attention-BiGRU encoder consists of an Attention layer and a BiGRU layer with decreasing number of neurons, and the Attention-BiGRU decoder consists of a BiGRU layer with increasing number of neurons and an Attention layer. The feature vector represents the data features after signal dimensionality reduction. When training the model, based on the consideration of the change of the natural frequency of the traditional tower barrel monitoring and the natural frequency and its order of the cable force monitoring, the loss of physical quantities is added to the loss function, that is, the correlation between the vibration signal and the reconstructed signal in the frequency domain.

[0012] The online monitoring module is responsible for evaluating the health levels of various parts of the hybrid tower top, transition section, prestressed cable, tower section, bolt, and tower foundation, as Figure 2 shown.

[0013] S1. Deploy the normal behavior models of each obtained monitoring part to the background server, and collect the data of each key part of the hybrid tower of the wind turbine in real time and transmit it to the background server for data alignment and normalization operations. Data alignment means that the timestamps of the input signals of each model are the same.

[0014] S2. Input the vibration signals of each part into the corresponding normal behavior model to obtain the reconstructed signals.

[0015] S3. Calculate the Wasserstein distance between the input signal and the reconstructed signal to obtain the error between the input signal and the reconstructed signal.

[0016] S4. Input the obtained reconstruction error into the exponentially weighted moving average control chart and combine the judgment criteria to obtain the health level evaluation of each part of the hybrid tower top, transition section, prestressed cable, tower section, bolt, and tower foundation.

[0017] The hybrid tower status evaluation module is responsible for evaluating the overall health level of the hybrid tower of the wind turbine. Based on the health level evaluations of various parts of the hybrid tower, including the tower top, transition section, prestressed cables, tower segments, bolts, and tower foundation, obtained from the online monitoring module, weights are assigned to each model based on the uncertainty of the model, and the weighted results of the model status evaluations are combined to obtain the health level of the hybrid tower.

[0018] The online learning module is responsible for the autonomous iterative learning of the full-system holographic hybrid tower status recognition system of the wind turbine. It compares the health level obtained from the hybrid tower status evaluation module with the on-site maintenance feedback results, and iteratively updates the full-system holographic hybrid tower status recognition system. For the long-term operation of the hybrid tower of the wind turbine, when the state distribution of the operation data changes and there are false negatives in the system, a time-continuous adaptive weighting strategy is used to update the normal behavior models of each monitoring part. For false positives in the system, a strategy of adding penalty terms is used to update the normal behavior models of each monitoring part. At the same time, combined with the performance of each model in the monitoring period, the weights of the models in the multi-voting mechanism are optimized periodically. Description of the Drawings

[0019] Figure 1 Full-system holographic hybrid tower status intelligent recognition system based on deep learning.

[0020] Figure 2 Offline training and online monitoring module.

[0021] Figure 3 Attention-BiGRU autoencoder model. Detailed Implementation Manner

[0022] The data acquisition module is responsible for collecting comprehensive monitoring data of the hybrid tower of the wind turbine. Sensors are arranged at the tower top, transition section, prestressed cables, tower segments, bolts, and tower foundation of the hybrid tower to collect the sway angle data of the tower top, stress and strain data of the transition section of the hybrid tower, vibration data of the prestressed cables, stress and strain data between tower segments, stress and strain data of bolts, sway angle data of the tower foundation, as well as the data timestamps under different temperatures, humidities, wind speeds, and operating conditions of the wind turbine.

[0023] The offline training module is responsible for constructing normal behavior models for the tower top, transition section, prestressed cables, tower segments, bolts, and tower foundation of the hybrid tower, as Figure 2 shown.

[0024] S1. Screen the collected data, divide the data into a training set and a validation set according to different monitoring parts, and perform normalization processing on the training set and the validation set; the training set only contains data in the normal state.

[0025] S2. Then, construct an Attention-BiGRU autoencoder and train the model parameters of the Attention-BiGRU autoencoder with the training set data.

[0026] The input signal can be represented as X ∈ R n×d , where n represents the length of the signal and d represents the dimension of the signal. The Attention layer first performs positional encoding on the input signal using a positional embedding matrix P ∈ R n×d , to obtain the input signal (X + P) ∈ R n×d , and the encoding formula is

[0027] P i,2j = sin(i / 1000 2j / d )

[0028] P i，2j+1 = cos(i / 1000 2j / d )

[0029] where P i,2j represents the element in the i-th row and 2j-th column of the positional embedding matrix; P i,2j+1 represents the element in the i-th row and (2j + 1)-th column of the positional embedding matrix.

[0030] The encoded input signal is input into the attention mechanism and passes through the linear transformation matrices W Q , W K , W V to calculate the query (Q), key (K), and value (V). The calculation formulas are

[0031] Q = (X + P)W q

[0032] K = (X + P)W K

[0033] V = (X + P)W V

[0034] After passing through the softmax function transformation, the attention weights are obtained. These weights reflect the importance of each time step in the signal. These weights are weighted to the Value matrix to obtain the weighted representation of the input signal

[0035]

[0036] where, QK T represents the similarity between the query and the key, d k represents the dimension of the key. To prevent the problem of gradient explosion caused by too large inner product, the outputs of multiple attention mechanisms are concatenated to obtain the output of the attention layer.

[0037] Multihead(Q,K,V) = Concat(head1,…head h )Q o

[0038] where h is the number of attention heads, and W o is the linear transformation matrix of the multi-head attention.

[0039] The weighted signal is input into the BiGRU layer to further learn the temporal information in the signal. The mathematical expression of the gated recurrent unit (GRU) is as follows

[0040] z t = σ(W z · [h t-1 , x t + b z )

[0041] r t = σ(W r · [h r-1 , x t + b r )

[0042]

[0043] where z t represents the update gate, W z represents the weight matrix of the update gate, h t-1 represents the hidden state at time t-1, x t represents the input at time t, b z represents the bias term of the update gate, and σ represents the Sigmoid activation function; r t represents the reset gate, W r represents the weight matrix of the reset gate, b r represents the bias term of the reset gate; represents the candidate hidden state; tanh represents the tanh activation function, W h represents the weight matrix of the candidate hidden state, b represents the bias term of the candidate hidden state, * represents element-wise multiplication, and h t represents the hidden state at time t, that is, the output of the GRU network at time t.

[0044] GRU controls the update and retention of information in the network through the update gate and the reset gate. The input at time t includes x t from the Attention layer, the hidden variable h t-1. The BiGRU works by combining two independent GRU units: one processes the information in the forward direction of the time series, and the other processes the information in the reverse direction. The monitoring signals of the hybrid tower are all periodic. This bidirectional processing method enables the network to simultaneously obtain the context information of the past and the future, enhancing the model's ability to integrate information from vibration signals.

[0045]

[0046] Among them, represents the output of the BiGRU network at time t, and represent the states of the forward hidden layer and the reverse hidden layer at time t respectively, w t1 and w t2 represent the weight matrices of the forward hidden layer and the reverse hidden layer at time t respectively, b t represents the bias term of the reset gate

[0047] The Attention - BiGRU auto - encoder model is as Figure 3 shown, and it includes an Attention - BiGRU encoder, a feature vector, and an Attention - BiGRU decoder. The Attention - BiGRU encoder consists of an Attention layer and a BiGRU layer with decreasing number of neurons, and the Attention - BiGRU decoder consists of a BiGRU layer with increasing number of neurons and an Attention layer. The feature vector space in the middle represents the data features after dimensionality reduction.

[0048] When the Attention - BiGRU network constructs the auto - encoder model and iteratively updates, it is trained using a time - frequency double - constraint loss function.

[0049] L total = L mse + L fre

[0050]

[0051] L fre = 1 - R xy (f)

[0052]

[0053] Among them, L total represents the total loss, L mse represents the traditional mean square error, making the output of the model as close as possible to the numerical reconstruction in the time domain; L fre represents the physical information, making the output of the model reconstruct the number and position of signal resonance peaks in the frequency domain; n represents the number of samples of the signal; xi represents the input signal; y i represents the reconstructed signal; R xy (f) represents the frequency correlation, which is a similarity measure representing the shape of the signal in the frequency domain, and its value is between 0 and 1; S xy (f) represents the cross-spectrum of the input signal and the reconstructed signal; X(f) represents the Fourier transform of the input signal; The trained normal behavior model of the conjugate complex number of the Fourier transform of the reconstructed signal is tested on the validation set. According to the model evaluation metrics of root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), excellent performance tower top normal behavior models, mixed tower transition section normal behavior models, prestressed cable normal behavior models, tower section normal behavior models, bolt normal behavior models, and tower base normal behavior models are selected for subsequent online monitoring.

[0054]

[0055] The online monitoring module is responsible for evaluating the health levels of various parts of the mixed tower, such as the tower top, transition section, prestressed cable, tower section, bolt, and tower base, as Figure 2 shown.

[0056] S1. Deploy the normal behavior models of each monitored part that have been obtained to the background server, and collect the data of each key part of the mixed tower of the wind turbine in real time and transmit it to the background server for data alignment and normalization operations. Data alignment means that the timestamps of the input signals of each model are the same.

[0057] S2. Input the vibration signals of each part into the corresponding normal behavior model to obtain the reconstructed signals of each monitored part.

[0058] S3. Calculate the Wasserstein distance between the input signal and the reconstructed signal to obtain the error between the input signal and the reconstructed signal.

[0059] Since the autoencoder model is trained using the normal data set, what it learns is the data distribution in the normal state of each monitored part. There are differences between the abnormal data and the normal data distribution, and the autoencoder has not learned the mixed tower data in the abnormal state. Therefore, for normal data, the error between the input signal and the reconstructed signal is much smaller than the reconstruction error of abnormal data. Input the input signal and the reconstructed signal into the Wasserstein distance formula to obtain the reconstruction error.

[0060]

[0061] P and Q represent the probability distributions of the input signal and the reconstructed signal, γ ∈ Γ(P,Q) represents the set of joint distributions of the input signal and the reconstructed signal, E (X,Y)～γDenote the expectation calculated according to the joint distribution γ, where X represents the input signal, Y represents the reconstructed signal, and ∥X - Y∥ p represents the distance between signals. When p takes 2, the Euclidean distance is used to measure the distance between signals. The Wasserstein distance regards normal data and abnormal data as two different probability distributions. It not only considers the point-to-point differences but also comprehensively takes into account the differences in the global shape of the signals and has a high tolerance for noise and translation.

[0062] S4. Input the obtained reconstruction error into the exponentially weighted moving average control chart combined with the judgment criterion to obtain the health level assessment of each part of the mixing tower top, transition section, prestressed cable, tower section, bolt, and tower foundation.

[0063] After obtaining the reconstruction errors of a large amount of historical normal data, if a fixed threshold is set subjectively to judge whether the mixing tower is abnormal, it is easily affected by professional knowledge and has a large subjectivity. On a long time scale, the reconstruction error can be regarded as a time series data. Therefore, an exponentially weighted moving average (EWMA) control chart is introduced to monitor the change trend of the error between the input signal and the reconstructed signal of the autoencoder model. Input the obtained reconstruction error into the exponentially weighted moving average control chart to monitor the change trend of the reconstruction error.

[0064] E t = λW t +(1 - λ)E t-1

[0065] where E t is the smoothed value at time t in the EWMA control chart, E0 is the average value of the verification sequence, λ is the weight of the historical residual for the current EWMA value, and W t is the Wasserstein distance between the input signal and the reconstructed signal at time t.

[0066] The upper and lower limits of the EWMA control chart

[0067]

[0068] where cl t represents the upper and lower limits of the EWMA control chart at time t, μ0 is the mean of the Wasserstein distances of the historical data, σ0 is the standard deviation of the Wasserstein distances of the historical data, and K is the threshold coefficient.

[0069] If the EWMA curve of the wind turbine monitoring component exceeds the control limit for multiple consecutive sampling periods, it is considered that the wind turbine component is in an abnormal state.

[0070] The hybrid tower status evaluation module is responsible for evaluating the overall health level of the hybrid tower of a wind turbine. Based on the health levels of various parts of the hybrid tower, including the tower top, transition section, prestressed cables, tower segments, bolts, and tower foundation, obtained from the online monitoring module, and combined with the performance of each normal behavior model within a cycle, the status evaluation results of each part are weighted to obtain the health level of the hybrid tower. The traditional fixed-weight multi-voting mechanism is prone to ignoring the differences in the performance of individual models, which can easily lead to the amplification of incorrect predictions. In response to this, a soft voting mechanism is designed based on the uncertainty of the model, and weights are assigned to the status evaluation results of each model for evaluating the overall health level of the hybrid tower of a wind turbine.

[0071]

[0072] Where: w i is the weight of each model in the multi-voting mechanism, N is the number of models, is the uncertainty of model i, the initial σ is the root mean square error of the model on the validation set, and it is set based on the prediction situation of the model in the previous cycle during the online learning stage. The health levels of each part and the weights of the corresponding models are weighted to obtain the health level of the hybrid tower of a wind turbine.

[0073]

[0074] Where H is the health level of the hybrid tower of a wind turbine, w i is the weight value of model i in the multi-voting mechanism, M i is the status evaluation value of model i. The online learning module is responsible for the autonomous iterative learning of the full-holographic hybrid tower status recognition system of the wind turbine. By comparing the health level obtained from the hybrid tower status evaluation module with the on-site maintenance feedback results, the full-holographic hybrid tower status recognition system of the wind turbine is iteratively updated.

[0075] S1. The initial model is deployed on the back-end server of the wind turbine for the status monitoring of the hybrid tower in the first cycle

[0076] S2. If the status of the tower barrel does not change within the first cycle, the initial model is fine-tuned using the data generated in the first cycle for the monitoring in the second cycle.

[0077] S3. In the second cycle, if the status of the tower barrel does not change either, the initial model is fine-tuned using the data from the first and second cycles for the monitoring in the third cycle.

[0078] S4. If an alarm occurs within the second cycle, the model fine-tuned using the data from the first cycle continues to be used for the monitoring in the third cycle.

[0079] In the long - time scale, with the service of the hybrid tower of wind turbines, the probability distribution of its operation data will change over time, and it is difficult to collect the full - life - cycle data of the hybrid tower of wind turbines and data under various complex working conditions in model training. This inevitably leads to the phenomenon of missed alarms in the model. This indicates that the existing model fails to fully learn and reflect the current operation state of the unit, resulting in limited prediction accuracy and stability. In response, a cost - sensitive learning strategy is adopted during model update, and time - weighted averaging is performed on data in different cycles to highlight the contribution of the latest data.

[0080] α t = e -β(T-t)

[0081]

[0082] Among them, α t is the weight of the t - th cycle; T is the current cycle; β is the decay coefficient (controlling the decay speed); L1 represents the overall loss of the model in multiple cycles; L mse,t represents the time - domain loss of the t - th cycle; L fre,t represents the frequency - domain loss of the t - th cycle. During the cycle monitoring process, false alarms may also occur, which means that the system fails to effectively learn and distinguish the characteristics of false - alarm data. In response, a strategy of increasing the penalty term is adopted to improve the model's recognition ability for such situations.

[0083]

[0084] Among them; L2 represents the overall loss of the model during penalty - term update, L norm,i (x i ,y i ) represents the loss of normal samples, w n represents the weight of normal samples, L false,j (x j ,y j ) represents the loss of false - alarm samples, w n is updated with the cycle; u f represents the weight of false - alarm samples, u f > w n .

[0085] During the model update process, the weights of each model in the multi - voting mechanism are also updated. If no abnormal situation occurs within the cycle, the weights are updated by combining the error between the monitoring signal and the reconstruction signal of each model within the monitoring cycle;

[0086]

[0087] Among them: w iis the weight of each model in the multi-voting mechanism, and N is the number of models. is the uncertainty of model i, and σ is the root mean square error of the model during the monitoring period.

[0088] If there are false alarms or missed alarms in the previous period, the weights are updated according to the F1 score, a performance metric of the model.

[0089]

[0090] where F1 i is the F1 score of model i, Pre is the precision of the model, Rec is the recall of the model, TP is the number of samples predicted as positive and actually positive, FP is the number of samples predicted as positive but actually negative, and FN is the number of samples predicted as negative but actually positive.

Claims

1. A full-system holographic wind turbine mixed tower state intelligent recognition system based on deep learning, characterized in that: include: The data acquisition module is responsible for collecting monitoring data of the wind turbine tower top, transition section, prestressed cable, tower section, bolts, and tower base under different temperatures, humidity, wind speeds, and operating conditions; The offline training module is responsible for building and training normal behavior models for the top, transition section, tower section, bolts, prestressed cables, and tower base of the hybrid tower through data screening and normalization processing; The online monitoring module is responsible for evaluating the health level of the tower top, transition section, prestressed cables, tower sections, bolts, and tower base. The mixed tower status assessment module evaluates the overall health level of the wind turbine mixed tower by combining the health level of each part and the uncertainty of the normal behavior model; The online learning module is responsible for online self-learning, updating and optimizing the normal behavior model and multi-voting mechanism to improve the overall monitoring accuracy of the system.

2. According to the deep learning-based full-system holographic wind turbine mixed tower state intelligent identification system of claim 1, it is characterized in that: The normal behavior model described in the offline training module adopts the Attention-BiGRU autoencoder model, which uses the Attention layer to capture the correlation at different times in the vibration signal and uses the BiGRU network to simultaneously learn the context information in the signal. Among them, the Attention-BiGRU autoencoder model is trained with a time-frequency dual-constraint loss function to learn the data distribution and physical information in the signal. L total =L mse +L fre L fre =1-R xy (f) Where L total represents the total loss, L mse represents the traditional mean square error, which makes the output of the model as close to numerical reconstruction as possible in the time domain; L fre represents physical information, so that the output of the model can reconstruct the number and position of the signal resonance peaks in the frequency domain; n represents the number of signal samples; x i Represents the input signal; y i Represents the reconstructed signal; R xy (f) represents frequency correlation, which represents the similarity measure of the signal frequency domain shape, and its value is between 0 and 1; S xy (f) represents the cross spectrum of the input signal and the reconstructed signal; X(f) represents the Fourier transform of the input signal; represents the complex conjugate of the Fourier transform of the reconstructed signal.

3. The deep learning-based full-system holographic wind turbine mixed tower state intelligent identification system according to claim 1 is characterized in that: The online monitoring module uses the autoencoder model to reconstruct normal data with small errors and abnormal data with large errors to achieve intelligent monitoring and identification of wind turbine towers. The error between the input signal and the reconstructed signal is measured using the Wasserstein distance. P and Q represent the probability distribution of the input signal and the reconstructed signal, γ∈Γ(P,Q) represents the set of joint distributions of the input signal and the reconstructed signal, and E (X,Y)～γ represents the expectation calculated according to the joint distribution γ, where X represents the input signal and Y represents the reconstructed signal, ∥XY∥ p Indicates the distance between signals. p is 2 and the Euclidean distance is used to measure the distance between signals. Exponentially weighted moving average is used to analyze the changes in the error between the input signal and the reconstructed signal. When the error exceeds the control limit continuously, the corresponding model issues an early warning. E t =λW t +(1-λ)E t-1 Where: E t is the smoothed value at time t in the EWMA control chart, E0 is the average value of the verification sequence, λ is the weight of the historical residual to the current EWMA value, W t is the Wasserstein distance between the input signal and the reconstructed signal at time t. The upper and lower limits of the EWMA control chart are as follows where cl t It represents the upper and lower limits of the EWMA control chart at time t, μ0 is the mean of the Wasserstein distance of the historical data, σ0 is the standard deviation of the Wasserstein distance of the historical data, and K is the threshold coefficient. If the EWMA curve of the wind turbine monitoring component exceeds the control limit continuously, it is considered that the wind turbine component is in an abnormal state.

4. The deep learning-based full-system holographic wind turbine mixed tower state intelligent identification system according to claim 1 is characterized in that: The mixed tower state assessment module combines the performance of the normal monitoring model of each part and the state assessment results of the model to obtain the health level of the mixed tower. Among them, the weight of the state assessment result of the model is determined based on the uncertainty of the model output. Where: w i is the weight of each model in the multi-voting mechanism, N is the number of models, is the uncertainty of model i, the initial σ is the root mean square error of the model on the validation set, and the online learning phase is set based on the prediction of the model during the monitoring period.

5. The deep learning-based full-system holographic wind turbine mixed tower state intelligent identification system according to claim 1 is characterized in that: The online learning module continuously optimizes the model and multi-voting mechanism based on the feedback results to adapt to the dynamic changes of actual monitoring data. During the long-term service of wind turbines, their operating data changes over time, and a certain degree of underreporting is inevitable in the early stage of the system. In this regard, a cost-sensitive learning strategy is adopted to perform time-continuous adaptive weighting on data of different periods, highlight the contribution of the latest data, update the system model in a timely manner, and effectively balance the new and old knowledge. α t ＝e -β(T-t) Among them, α t is the weight of the tth cycle; T is the current cycle; β is the decay coefficient (controls the decay speed); L1 represents the overall loss of the model under multiple cycles; L mse,t represents the time domain loss of the tth period; L fre,t represents the frequency domain loss of the tth period During the periodic monitoring process, if the model has false positives, the strategy of adding penalty items is adopted to improve the model's ability to identify similar situations. Among them; L2 represents the overall loss of the model when the penalty term is updated, L norm,i (x i ,y i ) represents the loss of normal samples, w n Represents the weight of normal samples, L false,j (x j ,y j ) represents the loss of false positive samples, w n Updated periodically; f Represents the weight of the false positive sample, u f >w n .

Images