A wind power equipment quality assessment method and system based on multi-sensor data fusion

Abstract

The present invention proposes a wind power equipment quality assessment method based on multi-sensor data fusion, comprising: acquiring real-time data of multiple sensors on the wind power equipment, and adopting a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data; performing time domain, frequency domain and entropy domain feature extraction on multiple sensor data at the same time in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set; performing feature fusion on feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain an initial multi-modal fusion feature map; constructing a multi-scale deep feature fusion network model based on a ResNet network, inputting the initial multi-modal fusion feature map, and outputting a multi-scale fusion feature vector; through multi-sensor data fusion and deep feature extraction, the data diversity and complexity problems in the quality assessment of wind power equipment are solved, and the accuracy and robustness of the assessment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of quality assessment, and in particular to a wind power equipment quality assessment method and system based on multi-sensor data fusion. Background Art

[0002] Wind power equipment is a complex electromechanical system with highly coupled functions and structures, coordinated operation of multiple components, and powerful operation control capabilities. Any abnormality in any component may affect the operating status of other components through interactions between components, thus causing system failure and even huge economic losses and safety issues.

[0003] A wind power quality assessment method and device with publication number CN117494591B, wherein a wind power quality assessment method comprises: dividing a wind farm into multiple wind speed intervals according to the wind speed, building a steady-state model and a transient model through a Simulink simulation platform, and obtaining power signals corresponding to the multiple wind speed intervals; reconstructing and clustering the power signals corresponding to the multiple wind speed intervals to obtain power feature vectors; wherein the reconstruction comprises: adding Gaussian white noise to the power signals corresponding to the multiple wind speed intervals respectively, and compressing them, and converting the compressed signals into one-dimensional signals; iteratively training a quality assessment network using the power feature vectors, and judging whether the training is completed; if the training is completed, inputting any power signal into the quality assessment network, and outputting the assessment result.

[0004] However, as a large and complex system, the operating status and quality of wind power equipment are affected by many factors. The current traditional quality assessment methods often rely on data from a single sensor, which is difficult to fully reflect the true status of the equipment and is easily affected by noise, resulting in inaccurate assessment results. Summary of the invention

[0005] In view of this, the present invention proposes a wind power equipment quality assessment method and system based on multi-sensor data fusion. Through multi-sensor data fusion and deep feature extraction, the data diversity and complexity problems in wind power equipment quality assessment are solved, and the accuracy and robustness of the assessment are improved.

[0006] The technical solution of the present invention is implemented as follows: In a first aspect, the present invention provides a wind power equipment quality assessment method based on multi-sensor data fusion, the method comprising the following steps:

[0007] S1, obtains real-time data from multiple sensors on wind power equipment, and uses a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data;

[0008] S2, extracting features in the time domain, frequency domain and entropy domain from multiple sensor data at the same time in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set;

[0009] S3, feature fusion is performed on the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain an initial multi-modal fusion feature map;

[0010] S4, constructing a multi-scale deep feature fusion network model based on the ResNet network, inputting the initial multi-modal fusion feature map into the multi-scale deep feature fusion network model for deep feature extraction and multi-scale deep fusion, and outputting a multi-scale fusion feature vector;

[0011] S5, builds a wind power equipment quality assessment model based on the BiLSTM network, inputs the multi-scale fusion feature vector into the wind power equipment quality assessment model for training, and outputs the quality assessment results of the wind power equipment.

[0012] On the basis of the above technical solution, preferably, the five-point cubic smoothing algorithm is used in step S1 to denoise the real-time data to obtain standard sample data, including the following sub-steps:

[0013] S11, selecting five continuous sampling points from the acquired real-time data, and using a polynomial least square method to approximate the continuous sampling points, and fitting to obtain a cubic polynomial curve equation;

[0014] S12, smoothing each data point in the real-time data according to the fitted cubic polynomial curve equation, calculating the smoothed data point corresponding to each data point, and obtaining standard sample data.

[0015] On the basis of the above technical solution, preferably, step S1 further includes evaluating the quality of the denoised standard sample data, which includes the following sub-steps:

[0016] The signal-to-noise ratio is calculated based on the real-time data and the corresponding denoised data. The expression is:

[0017]

[0018] Where SNR is the signal-to-noise ratio, P s is the power of real-time data, P n is the power of the noise in the signal, x i To run data in real time, is the standard sample data, and n is the total amount of real-time data;

[0019] The correlation index is used to calculate the local feature correlation coefficient between real-time data and standard sample data. The expression is:

[0020]

[0021] In the formula, R is the correlation coefficient, is the covariance between real-time data and standard sample data, It is the product of the standard deviation of real-time data and standard sample data;

[0022] The noise threshold is preset, and the denoising effect is judged based on the results of the signal-to-noise ratio and the correlation coefficient. If the signal-to-noise ratio is greater than the noise threshold and the correlation coefficient is close to 1, the denoising effect is satisfied and the standard sample data set is retained. Otherwise, the parameters of the five-point cubic smoothing algorithm are readjusted.

[0023] On the basis of the above technical solution, preferably, the multi-sensor high-dimensional feature data set in step S2 includes a time domain feature data set, a frequency domain feature data set and an entropy domain feature data set, wherein:

[0024] Time domain feature extraction: extracting dimensional time domain features and dimensionless time domain features of standard sample data, wherein the dimensional time domain features include waveform index, peak index, pulse index and margin index, and adding the extracted feature values ​​to the time domain feature data set;

[0025] Frequency domain feature extraction: Use the fast Fourier transform method to obtain the real-time data spectrum of the standard sample data, extract the center of gravity frequency, root mean square frequency and frequency standard deviation, and add the extracted eigenvalues ​​to the frequency domain feature data set;

[0026] Entropy domain feature extraction: extract entropy domain features from standard sample data, the entropy domain features include energy entropy, information entropy and sample entropy, and add the extracted feature values ​​to the entropy domain feature data set.

[0027] On the basis of the above technical solution, preferably, the step S3 of fusing the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain an initial multimodal fusion feature map includes the following sub-steps:

[0028] S31, combining the feature values ​​extracted from each sensor data to form a feature vector according to the time domain feature data set, the frequency domain feature data set and the entropy domain feature data set;

[0029] S32, expand the dimension of the feature vector of each sensor at the same time and calculate the tensor outer product

[0030] Calculate and obtain the corresponding multi-sensor feature subgraph;

[0031] S33, fusing and splicing the multi-sensor feature sub-graphs corresponding to the same moment to obtain an initial multi-modal fusion feature graph.

[0032] On the basis of the above technical solution, preferably, step S32 includes the following sub-steps:

[0033] Assume there are m sensors, whose feature vector V i The expression is:

[0034]

[0035] Where, d i is the characteristic dimension of the i-th sensor, R y is the correlation coefficient between two eigenvectors, i∈[1,m];

[0036] The bias term is introduced to expand the dimension of the feature vector of each sensor to obtain the feature extension vector, which is expressed as:

[0037]

[0038] The tensor outer product calculation is performed on the feature extension vectors of different sensor data at the same time to obtain the multi-sensor feature subgraph, which is expressed as:

[0039]

[0040] On the basis of the above technical solution, preferably, the multi-scale deep feature fusion network model is constructed based on the ResNet network in step S4, wherein the multi-scale deep feature fusion network model includes an initial convolution layer and several pairs of residual blocks, the several pairs of residual blocks are connected in sequence head to tail, the output end of the initial convolution layer is connected to the input end of the residual block located at the head end, each pair of residual blocks includes a first residual unit and a second residual unit, the output end of the first residual unit is connected to the input end of the second residual unit, the input and output residuals of each residual block are connected, and the outputs of the first residual unit and the second residual unit are input between the first residual unit and the second residual unit in adjacent residual blocks through a splicing and fusion operation.

[0041] On the basis of the above technical solution, preferably, step S5 includes the following sub-steps:

[0042] S51, re-dimensionalizing the output multi-scale fusion feature vector, expanding the feature vector of the time dimension on the basis of information unchanged, and obtaining a training time series data set;

[0043] S52, building a wind power equipment quality assessment model based on the BiLSTM network, and inputting the training time series data set into the wind power equipment quality assessment model for training;

[0044] S53, using the forward LSTM and the reverse LSTM to capture the time series characteristics of the wind power equipment in the time dimension, and combining the outputs of the forward LSTM and the reverse LSTM to integrate all the time series characteristic information of the wind power equipment in the entire time dimension, to obtain the quality status characteristic information of the wind power equipment at each moment;

[0045] S54, using the splicing layer to map the quality status characteristic information of the wind power equipment at each moment into the output sample space, and output the comprehensive evaluation result of the quality status of the wind power equipment.

[0046] Based on the above technical solution, preferably, step S5 further includes optimizing the wind power equipment quality assessment model based on the Adam optimizer and evaluating the model prediction performance, including the following sub-steps:

[0047] Randomly initialize the parameters of the wind power equipment quality assessment model, configure the Adam optimizer, and define the loss function. The expression of the loss function is:

[0048] LOSS = a·MSE + (1-a)·MAE

[0049] Where a is a hyperparameter used to adjust the weight ratio between MSE and MAE. MSE is the mean square error and MAE is the mean absolute error.

[0050] The Adam optimizer is used on the training time series data set to iteratively train the wind power equipment quality assessment model, and the model parameters are adjusted through back propagation to achieve the minimum loss function;

[0051] The RMSE, R-Square and MAE indicators were used to evaluate the model prediction performance.

[0052] In a second aspect, the present invention further provides a comprehensive wind power equipment quality assessment system based on multi-sensor data fusion, characterized in that: the system comprises:

[0053] The acquisition module is used to obtain real-time data from multiple sensors on the wind power equipment, and use a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data;

[0054] The feature extraction module is used to extract features in the time domain, frequency domain and entropy domain from multiple sensor data at the same time in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set;

[0055] The feature fusion module is used to fuse the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain the initial multi-modal fusion feature map;

[0056] The deep extraction and fusion module is used to build a multi-scale deep feature fusion network model based on the ResNet network, input the initial multi-modal fusion feature map into the multi-scale deep feature fusion network model for deep feature extraction and multi-scale deep fusion, and output a multi-scale fusion feature vector;

[0057] The evaluation and prediction module is used to build a wind power equipment quality evaluation model based on the BiLSTM network, input the multi-scale fusion feature vector into the wind power equipment quality evaluation model for training, and output the quality evaluation results of the wind power equipment.

[0058] The wind power equipment quality assessment method and system based on multi-sensor data fusion of the present invention have the following beneficial effects compared with the prior art:

[0059] (1) By collecting real-time data from multiple sensors and performing five-point cubic smoothing and denoising, noise interference is effectively reduced, and the accuracy and reliability of the data are improved. Combined with multi-dimensional feature extraction in the time domain, frequency domain, and entropy domain, as well as deep feature extraction and multi-scale deep fusion of multi-sensor data, the feature representation is more comprehensive and accurate, which improves the accuracy of wind power equipment quality assessment. In addition, the multi-sensor data fusion method can make full use of the complementary information between different sensors, reduce the impact of single sensor failure or data anomaly on the assessment results, and thus enhance the robustness and stability of the assessment;

[0060] (2) By extracting features in the time domain, frequency domain, and entropy domain from multiple sensor data at the same time in the standard sample data, a multi-sensor high-dimensional feature dataset was constructed; this not only enriched the dimension of the dataset, but also improved the accuracy and depth of data analysis and processing;

[0061] (3) Through the improved ResNet residual network structure, deep analysis and efficient fusion of the initial multimodal fusion feature map are achieved, which not only enhances the model's adaptability to complex scenes, but also improves the model's robustness and generalization ability through rich feature representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0063] Figure 1 It is a flowchart of the wind power equipment quality assessment method based on multi-sensor data fusion of the present invention;

[0064] Figure 2It is a schematic diagram of the noise reduction effect of the five-point cubic algorithm of the wind power equipment quality assessment method based on multi-sensor data fusion of the present invention;

[0065] Figure 3 It is a schematic diagram of a multi-sensor characteristic sub-graph splicing method of a wind power equipment quality assessment method based on multi-sensor data fusion of the present invention;

[0066] Figure 4 It is a schematic diagram of a multi-scale deep feature fusion network structure of a wind power equipment quality assessment method based on multi-sensor data fusion of the present invention;

[0067] Figure 5 A schematic diagram of the BiLSTM network structure of the wind power equipment quality assessment method based on multi-sensor data fusion of the present invention;

[0068] Figure 6 It is a schematic diagram of the implementation flow of the wind power equipment quality assessment method based on multi-sensor data fusion of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] like Figure 1 and Figure 6 As shown, a wind power equipment quality assessment method based on multi-sensor data fusion of the present invention comprises the following steps:

[0071] S1, obtains real-time data from multiple sensors on wind power equipment, and uses a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data.

[0072] Step S1 includes the following sub-steps:

[0073] S11, selecting five continuous sampling points from the acquired real-time data, and using a polynomial least square method to approximate the continuous sampling points, and fitting to obtain a cubic polynomial curve equation;

[0074] S12, smoothing each data point in the real-time data according to the fitted cubic polynomial curve equation, calculating the smoothed data point corresponding to each data point, and obtaining standard sample data.

[0075] It should be noted that the real sampling data is approximated by polynomial least squares, and the cubic polynomial curve equation is established using the real sampling points. On the basis of ensuring that the original signal curve characteristics remain unchanged, the interference components are effectively removed. The noise reduction effect is as follows: Figure 2 As shown, the denoising formula is as follows:

[0076]

[0077] Where Y -1 , Y -2 , Y0, Y1, and Y2 are the data values ​​of five consecutive sampling points. The data value after noise reduction for five consecutive sampling points.

[0078] Step S1 also includes the following sub-steps:

[0079] S13, calculate the signal-to-noise ratio based on the real-time data and the corresponding denoised data, and the expression is:

[0080]

[0081] Where SNR is the signal-to-noise ratio, P s is the power of real-time data, P n is the power of the noise in the signal, x i To run data in real time, is the standard sample data, and n is the total amount of real-time data;

[0082] S14, using the correlation index to calculate the local feature correlation coefficient between the real-time data and the standard sample data, the expression is:

[0083]

[0084] In the formula, R is the correlation coefficient, is the covariance between real-time data and standard sample data, It is the product of the standard deviation of real-time data and standard sample data;

[0085] S15, preset the noise threshold, and judge whether the denoising effect is satisfactory according to the results of the signal-to-noise ratio and the correlation coefficient. If the signal-to-noise ratio is greater than the noise threshold and the correlation coefficient is close to 1, the denoising effect is satisfactory and the standard sample data set is retained. Otherwise, readjust the parameters of the five-point cubic smoothing algorithm.

[0086] It should be noted that the signal-to-noise ratio formula is used to calculate the average approximation degree of the denoised signal to the true signal. The larger the value, the higher the average approximation degree of the denoised signal to the true signal, indicating that the denoising algorithm can retain more effective features of the original signal on the basis of denoising. The correlation index is used to calculate the local feature correlation between the denoised signal and the true signal, and the denoising effect of the five-point cubic smoothing algorithm is quantitatively evaluated. R is the correlation coefficient, and its value range is [-1,1]. When R∈[0,1], it means that the estimated signal is positively correlated with the true signal. The closer it is to 1, the greater the local feature correlation between the estimated signal and the true signal, and the better the denoising effect; when R∈[-1,0], it means that the estimated signal is negatively correlated with the true signal. The closer it is to -1, the greater the difference between the signals.

[0087] It can be understood that the real-time data obtained by the wind power equipment sensor is effectively denoised through the five-point cubic smoothing algorithm to generate standard sample data, which not only removes the random noise in the data, but also retains the curve characteristics and effective features of the original signal, realizes the smoothing and denoising of the original data, and provides a more reliable and accurate data basis for subsequent data analysis and processing.

[0088] S2, extracting features in the time domain, frequency domain and entropy domain from multiple sensor data at the same moment in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set.

[0089] The multi-sensor high-dimensional feature data set in this embodiment includes a time domain feature data set, a frequency domain feature data set, and an entropy domain feature data set, wherein:

[0090] Time domain feature extraction: extracting dimensional time domain features and dimensionless time domain features of standard sample data, wherein the dimensional time domain features include waveform index, peak index, pulse index and margin index, and adding the extracted feature values ​​to the time domain feature data set;

[0091] It should be noted that the waveform index is the ratio of the root mean square value of a set of data to the average value, that is, the ratio of the effective value to the average value. It reflects the data waveform and characterizes the degree of fluctuation of the physical variable over time. The formula is as follows:

[0092]

[0093] Where W is the waveform index, δ is the root mean square value, is the average value.

[0094] The peak index is the ratio of the maximum value of the interval data to the root mean square value. It represents the extreme degree of the peak value in the data waveform and reflects the impact information in the data. The formula is as follows:

[0095]

[0096] Where C is the peak index, x max is the maximum value of the data in this interval, x rms is the root mean square value.

[0097] The pulse index is the ratio of the maximum value to the average value of the data set. The difference from the peak index is that the denominator is changed from the effective value to the average value. For the same set of data, the average value is often smaller than the effective value. Therefore, the pulse index is more sensitive to impact information than the peak index. The formula is as follows:

[0098]

[0099] Where I is the pulse index.

[0100] The margin index is the ratio of the peak value to the square root amplitude. It is similar to the peak index, except that the denominator of the margin index is the square root amplitude. Therefore, it is more sensitive than the peak index and reflects the aging, wear and tear of the system. The formula is as follows:

[0101]

[0102] Where L is the pulse index, x r is the square root amplitude.

[0103] Frequency domain feature extraction: Use the fast Fourier transform method to obtain the real-time data spectrum of the standard sample data, extract the center of gravity frequency, root mean square frequency and frequency standard deviation, and add the extracted eigenvalues ​​to the frequency domain feature data set;

[0104] It should be noted that, assuming a set of data x(n) with a length of N, the fast Fourier transform of x(n) can be expressed as:

[0105]

[0106] Where a0 is the static component, a k is the cosine term coefficient, b k is the sine term coefficient, c k is the harmonic amplitude,

[0107] The center frequency can represent the frequency value of the signal component that occupies a larger proportion in the spectrum, reflecting the distribution of the signal power spectrum. If the low-frequency amplitude is large, the center frequency is closer to the origin, otherwise, it is far away from the origin. The formula is as follows:

[0108]

[0109] Where F2 is the center of gravity frequency index, f nis the frequency value of the nth spectral line.

[0110] The root mean square frequency is the algorithmic square root of the mean square frequency. It is similar to the centroid frequency, but the difference is that the root mean square frequency is the weighted average of the square of the signal frequency, which describes the position distribution of the main frequency band of the power spectrum. The formula is as follows:

[0111]

[0112] Where F3 is the root mean square frequency index.

[0113] The frequency standard deviation is the inertia radius centered on the center frequency. If the frequency amplitude near the center frequency is large, the value of the frequency standard deviation will be small, otherwise, the value of the frequency standard deviation will be large. Therefore, the frequency standard deviation mainly describes the dispersion degree of the signal power spectrum energy distribution, and the formula is as follows:

[0114]

[0115] Where F4 is the frequency standard deviation indicator.

[0116] Entropy domain feature extraction: extract entropy domain features from standard sample data, the entropy domain features include energy entropy, information entropy and sample entropy, and add the extracted feature values ​​to the entropy domain feature data set.

[0117] It should be noted that information entropy is used to measure the uncertainty and complexity of information sources. The larger the value, the greater the amount of information and the degree of confusion contained in the information source. The formula is as follows:

[0118]

[0119] In the formula, x is the signal value, X is the set of all signals in the target interval, and P(x) represents the probability of the variable x appearing;

[0120] Energy entropy measures the amount of information in a signal from the perspective of the energy it possesses. The calculation of signal energy entropy requires decomposing the original signal into several components, calculating the energy of each component and its ratio to the total energy of the signal. The formula is as follows:

[0121]

[0122] In the formula, H E is the energy entropy, E i is the energy of each component and its ratio to the total energy of the signal.

[0123] Sample entropy is a parameter that characterizes the complexity of a time series. It is used to measure the probability of generating new patterns in a signal. Sample entropy does not include comparisons of its own data segments, which not only reduces the amount of calculation and improves calculation accuracy, but also makes the sample entropy not dependent on the length of the original data segment.

[0124] S3, feature fusion is performed on the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain the initial multimodal fusion feature map.

[0125] Step S3 includes the following sub-steps:

[0126] S31, combining the feature values ​​extracted from each sensor data to form a feature vector according to the time domain feature data set, the frequency domain feature data set and the entropy domain feature data set;

[0127] S32, performing dimension expansion and tensor outer product calculation on the feature vector of each sensor at the same time to obtain a corresponding multi-sensor feature subgraph;

[0128] Step S32 includes the following sub-steps:

[0129] Assume there are m sensors, whose feature vector V i The expression is:

[0130]

[0131] Where, d i is the characteristic dimension of the i-th sensor, R y is the correlation coefficient between two eigenvectors, i∈[1,m];

[0132] The bias term is introduced to expand the dimension of the feature vector of each sensor to obtain the feature extension vector, which is expressed as:

[0133]

[0134] The tensor outer product calculation is performed on the feature extension vectors of different sensor data at the same time to obtain the multi-sensor feature subgraph, which is expressed as:

[0135]

[0136] Specifically, in this implementation, m=4, then the expression is:

[0137]

[0138] S33, such as Figure 3 As shown in FIG. 1 , the multi-sensor feature sub-graphs corresponding to the same moment are fused and spliced ​​to obtain the initial multi-modal fusion feature graph.

[0139] In this embodiment, a multi-sensor high-dimensional feature data set is constructed by extracting time domain, frequency domain and entropy domain features of multiple sensor data at the same time in the standard sample data; this not only enriches the dimension of the data set, but also improves the accuracy and depth of data analysis and processing.

[0140] S4, builds a multi-scale deep feature fusion network model based on the ResNet network, inputs the initial multimodal fusion feature map into the multi-scale deep feature fusion network model for deep feature extraction and multi-scale deep fusion, and outputs a multi-scale fusion feature vector.

[0141] Among them, the multi-scale deep feature fusion network model includes an initial convolution layer and several pairs of residual blocks, and the several pairs of residual blocks are connected in sequence head to tail. The output end of the initial convolution layer is connected to the input end of the residual block located at the head end. Each pair of residual blocks includes a first residual unit and a second residual unit. The output end of the first residual unit is connected to the input end of the second residual unit. The input and output residuals of each residual block are connected, and the outputs of the first residual unit and the second residual unit are input between the first residual unit and the second residual unit in the adjacent residual block through a splicing and fusion operation.

[0142] like Figure 4 As shown in the figure, the initial multimodal fusion feature map of the input is encoded and embedded into the new vector space using the initial convolution layer with a convolution kernel size parameter of 7*7, so that each feature information extracted by the initial convolution layer has rich feature information of different modalities. The convolution kernel with a size parameter of 1*1 and a step length of 2 is used to adjust the format of the flowing feature map to ensure that the format of the converged feature information is the same. The feature map is reduced in dimension and formatted by a convolution kernel with a step length of 1x1 and a step length of 2 to ensure that the feature maps of different levels or branches are consistent in size and number of channels during the feature fusion process, so as to facilitate subsequent operations. This helps to reduce the amount of calculation while keeping the important information of the feature map from being lost.

[0143] The multi-scale deep feature fusion network model in this embodiment includes four pairs of residual blocks, which are used to extract high-level semantic feature information. All convolution blocks are stacked in sequence to form the main part of the network, ensuring the feature extraction capability of the network, and constructing process feature information of different scales for multi-scale feature deep fusion. The feature maps output by two convolution blocks are cross-layer connected and stacked, aligned with the channel number of the next level feature map and pixel points are added to achieve the fusion of three-level feature information.

[0144] Between residual blocks, the outputs of the first residual unit and the second residual unit are combined through splicing and fusion operations. This cross-scale feature fusion method can capture feature information at different levels, making the final feature vector richer and more comprehensive. In addition, cross-layer connections further promote the fusion of multi-level feature information, realize feature complementarity from low to high levels, and improve the representation ability of the model.

[0145] In this embodiment, based on the improved ResNet residual network structure, deep analysis and efficient fusion of the initial multimodal fusion feature map are achieved, which not only enhances the model's adaptability to complex scenes, but also improves the model's robustness and generalization ability through rich feature representation.

[0146] S5, builds a wind power equipment quality assessment model based on the BiLSTM network, inputs the multi-scale fusion feature vector into the wind power equipment quality assessment model for training, and outputs the quality assessment results of the wind power equipment.

[0147] like Figure 5 As shown, step S5 includes the following sub-steps:

[0148] S51, re-dimensionalizing the output multi-scale fusion feature vector, expanding the feature vector of the time dimension on the basis of information unchanged, and obtaining a training time series data set;

[0149] S52, building a wind power equipment quality assessment model based on the BiLSTM network, and inputting the training time series data set into the wind power equipment quality assessment model for training;

[0150] S53, using the forward LSTM and the reverse LSTM to capture the time series characteristics of the wind power equipment in the time dimension, and combining the outputs of the forward LSTM and the reverse LSTM to integrate all the time series characteristic information of the wind power equipment in the entire time dimension, to obtain the quality status characteristic information of the wind power equipment at each moment;

[0151] S54, using the splicing layer to map the quality status characteristic information of the wind power equipment at each moment into the output sample space, and output the comprehensive evaluation result of the quality status of the wind power equipment.

[0152] It should be noted that through the combination of forward LSTM and backward LSTM, the BiLSTM network can fully capture the bidirectional dependency of wind power equipment data in the time dimension. This bidirectional capture mechanism not only considers the impact of historical data on the current state, but also incorporates the potential impact of future data on the current state, thereby more accurately reflecting the dynamic changes in the quality of wind power equipment and improving the accuracy of the comprehensive assessment of the quality status of wind power equipment.

[0153] In this embodiment, step S5 further includes optimizing the wind power equipment quality assessment model based on the Adam optimizer and evaluating the model prediction performance, including the following sub-steps:

[0154] Randomly initialize the parameters of the wind power equipment quality assessment model, configure the Adam optimizer, and define the loss function. The expression of the loss function is:

[0155] LOSS=w·MSE+(1-w)·MAE

[0156] Where w is a hyperparameter used to adjust the weight ratio between MSE and MAE. MSE is the mean square error and MAE is the mean absolute error.

[0157] The Adam optimizer is used on the training time series data set to iteratively train the wind power equipment quality assessment model, and the model parameters are adjusted through back propagation to achieve the minimum loss function;

[0158] The RMSE, R-Square and MAE indicators were used to evaluate the model prediction performance.

[0159] It should be noted that the random initialization method is used to initialize the network weight parameters, breaking the inherent symmetry of the parameters, improving the expressiveness of the network, and to a certain extent reducing the probability of gradient explosion during network training.

[0160] In addition, in the process of defining the loss function, a suitable loss function is selected to measure the error between the model's predicted value and the actual value, and to provide an optimization direction for the model so that the model can complete the best mapping at the lowest cost.

[0161] MSE is a commonly used loss function in regression models. It represents the sum of the squares of the distances between the model prediction value and the true value. It is sensitive to model prediction errors. If there are outliers in the data, the model will give higher weights to adjust the model performance of the outliers, at the cost of reducing the overall performance of the model. For the prediction of the quality status of wind power equipment by the model proposed in this paper, assuming that the true quality status of the wind power equipment is y i , the model output predicts its quality status as The sample size is n, then the MSE loss function calculation formula is:

[0162]

[0163] The MAE loss function calculates the average of a set of predictions of the model without considering the direction of the error. It is more robust to outliers in the data, and its gradient remains the same throughout the training process. The factor that affects the training speed and quality of the model is the sum of the absolute values ​​of the differences between the predicted values ​​of the metric model and the true values. The calculation formula is:

[0164]

[0165] Taking into account that MSE is easy to solve and the model is easy to converge, while MAE is more robust, the characteristics of the two loss functions are combined and the MSE and MAE joint loss function is adopted. This ensures that the model has high robustness while ensuring the convergence speed and model training performance.

[0166] Configure the Adam optimizer. Adam is a parameter optimization algorithm that performs first-order gradient on a random objective function. It uses the first-order moment and second-order moment of the gradient to estimate the learning rate of the dynamic adjustment parameters. The learning rate after bias correction is within a certain range, making the parameter update relatively stable. At the same time, the Adam optimizer is applicable to a wide range of network models, and is simple to implement and has low memory requirements, which greatly reduces the difficulty of model training. In this embodiment, the Adam algorithm is used to dynamically adjust parameters such as the model learning rate to optimize the wind power equipment quality assessment model. The calculation formula is:

[0167] m t =β1m t-1 +(1-β1)g t

[0168]

[0169] Among them, m t is the first-order momentum; n t is the second-order momentum; β1 and β2 are exponential decay rates; and are the deviation corrections for the first-order momentum and the second-order momentum, respectively; ∈ is a small constant to ensure numerical stability during training; a is the learning rate; θ is the model parameter;

[0170] In this embodiment, RMSE, R-Square and MAE indicators are used to evaluate the model prediction performance, where:

[0171] RMSE is used to estimate the degree of deviation between the model prediction value and the true value. The smaller the value, the higher the prediction accuracy of the model. The calculation formula is:

[0172]

[0173] The R-Square indicator is used to evaluate the model prediction performance. The value range is 0-1. The larger the value, the better the performance of the model. The calculation formula is:

[0174]

[0175] The MAE indicator is used to represent the non-directional error of the model. The smaller the value, the higher the prediction accuracy of the model.

[0176] This embodiment is based on multi-sensor data fusion, through five-point cubic smoothing denoising, multi-domain feature extraction and fusion, ResNet deep feature extraction and multi-scale fusion, and combined with BiLSTM network for time series analysis, to achieve high precision and real-time quality assessment of wind power equipment.

[0177] In a second aspect, a comprehensive wind power equipment quality assessment system based on multi-sensor data fusion in this embodiment includes:

[0178] The acquisition module is used to obtain real-time data from multiple sensors on the wind power equipment, and use a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data;

[0179] The feature extraction module is used to extract features in the time domain, frequency domain and entropy domain from multiple sensor data at the same time in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set;

[0180] The feature fusion module is used to fuse the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain the initial multi-modal fusion feature map;

[0181] The deep extraction and fusion module is used to build a multi-scale deep feature fusion network model based on the ResNet network, input the initial multi-modal fusion feature map into the multi-scale deep feature fusion network model for deep feature extraction and multi-scale deep fusion, and output a multi-scale fusion feature vector;

[0182] The evaluation and prediction module is used to build a wind power equipment quality evaluation model based on the BiLSTM network, input the multi-scale fusion feature vector into the wind power equipment quality evaluation model for training, and output the quality evaluation results of the wind power equipment.

[0183] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind power equipment quality assessment method based on multi-sensor data fusion, characterized by: The method comprises the following steps: S1, obtains real-time data from multiple sensors on wind power equipment, and uses a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data; S2, extracting features in the time domain, frequency domain and entropy domain from multiple sensor data at the same time in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set; S3, feature fusion is performed on the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain an initial multi-modal fusion feature map; S4, construct a multi-scale deep feature fusion network model based on the ResNet network, input the initial multi-modal fusion feature map into the multi-scale deep feature fusion network model for deep feature extraction and multi-scale deep fusion, and output a multi-scale fusion feature vector; S5, builds a wind power equipment quality assessment model based on the BiLSTM network, inputs the multi-scale fusion feature vector into the wind power equipment quality assessment model for training, and outputs the quality assessment results of the wind power equipment.

2. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 1, characterized in that: The five-point cubic smoothing algorithm is used to denoise the real-time data in step S1 to obtain standard sample data, including the following sub-steps: S11, selecting five continuous sampling points from the acquired real-time data, and using a polynomial least square method to approximate the continuous sampling points, and fitting to obtain a cubic polynomial curve equation; S12, smoothing each data point in the real-time data according to the fitted cubic polynomial curve equation, calculating the smoothed data point corresponding to each data point, and obtaining standard sample data.

3. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 2, characterized in that: Step S1 also includes evaluating the quality of the denoised standard sample data, which includes the following sub-steps: The signal-to-noise ratio is calculated based on the real-time data and the corresponding denoised data. The expression is: Where SNR is the signal-to-noise ratio, P s is the power of real-time data, P n is the power of the noise in the signal, x i To run data in real time, is the standard sample data, and n is the total amount of real-time data; The correlation index is used to calculate the local feature correlation coefficient between real-time data and standard sample data. The expression is: In the formula, R is the correlation coefficient, is the covariance between real-time data and standard sample data, It is the product of the standard deviation of real-time data and standard sample data; The noise threshold is preset, and the denoising effect is judged based on the results of the signal-to-noise ratio and the correlation coefficient. If the signal-to-noise ratio is greater than the noise threshold and the correlation coefficient is close to 1, the denoising effect is satisfied and the standard sample data set is retained. Otherwise, the parameters of the five-point cubic smoothing algorithm are readjusted.

4. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 3, characterized in that: The multi-sensor high-dimensional feature data set in step S2 includes a time domain feature data set, a frequency domain feature data set and an entropy domain feature data set. in, Time domain feature extraction: extracting dimensional time domain features and dimensionless time domain features of standard sample data, wherein the dimensional time domain features include waveform index, peak index, pulse index and margin index, and adding the extracted feature values ​​to the time domain feature data set; Frequency domain feature extraction: Use the fast Fourier transform method to obtain the real-time data spectrum of the standard sample data, extract the center of gravity frequency, root mean square frequency and frequency standard deviation, and add the extracted eigenvalues ​​to the frequency domain feature data set; Entropy domain feature extraction: extract entropy domain features from standard sample data, the entropy domain features include energy entropy, information entropy and sample entropy, and add the extracted feature values ​​to the entropy domain feature data set.

5. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 4, characterized in that: Step S3, in which the feature vectors in the multi-sensor high-dimensional feature data set at the same time are subjected to feature fusion to obtain an initial multimodal fusion feature map, includes the following sub-steps: S31, combining the feature values ​​extracted from each sensor data to form a feature vector according to the time domain feature data set, the frequency domain feature data set and the entropy domain feature data set; S32, performing dimension expansion and tensor outer product calculation on the feature vector of each sensor at the same time to obtain a corresponding multi-sensor feature subgraph; S33, fusing and splicing the multi-sensor feature sub-graphs corresponding to the same moment to obtain an initial multi-modal fusion feature graph.

6. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 5, characterized in that: Step S32 includes the following sub-steps: Assume there are m sensors, whose feature vector V i The expression is: Where, d i is the characteristic dimension of the i-th sensor, R y is the correlation coefficient between two eigenvectors, i∈[1,m]; The bias term is introduced to expand the dimension of the feature vector of each sensor to obtain the feature extension vector, which is expressed as: The tensor outer product calculation is performed on the feature extension vectors of different sensor data at the same time to obtain the multi-sensor feature subgraph, which is expressed as:

7. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 1, characterized in that: In step S4, a multi-scale deep feature fusion network model is constructed based on the ResNet network, wherein the multi-scale deep feature fusion network model includes an initial convolution layer and several pairs of residual blocks, the several pairs of residual blocks are connected in sequence head to tail, the output end of the initial convolution layer is connected to the input end of the residual block located at the head end, each pair of residual blocks includes a first residual unit and a second residual unit, the output end of the first residual unit is connected to the input end of the second residual unit, the input and output residuals of each residual block are connected, and the outputs of the first residual unit and the second residual unit are input between the first residual unit and the second residual unit in the adjacent residual block through a splicing and fusion operation.

8. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 7, characterized in that: Step S5 includes the following sub-steps: S51, re-dimensionalizing the output multi-scale fusion feature vector, expanding the feature vector of the time dimension on the basis of information unchanged, and obtaining a training time series data set; S52, building a wind power equipment quality assessment model based on the BiLSTM network, and inputting the training time series data set into the wind power equipment quality assessment model for training; S53, using the forward LSTM and the reverse LSTM to capture the time series characteristics of the wind power equipment in the time dimension, and combining the outputs of the forward LSTM and the reverse LSTM to integrate all the time series characteristic information of the wind power equipment in the entire time dimension, to obtain the quality status characteristic information of the wind power equipment at each moment; S54, using the splicing layer to map the quality status characteristic information of the wind power equipment at each moment into the output sample space, and output the comprehensive evaluation result of the quality status of the wind power equipment.

9. The wind power equipment quality assessment method based on multi-sensor data fusion according to claim 8, characterized in that: The step S5 further includes optimizing the wind power equipment quality assessment model based on the Adam optimizer and evaluating the model prediction performance, including the following sub-steps: Randomly initialize the parameters of the wind power equipment quality assessment model, configure the Adam optimizer, and define the loss function. The expression of the loss function is: LOSS = a·MSE + (1-a)·MAE Where a is a hyperparameter used to adjust the weight ratio between MSE and MAE. MSE is the mean square error and MAE is the mean absolute error. The Adam optimizer is used on the training time series data set to iteratively train the wind power equipment quality assessment model, and the model parameters are adjusted through back propagation to achieve the minimum loss function; The RMSE, R-Square and MAE indicators were used to evaluate the model prediction performance.

10. A comprehensive wind power equipment quality assessment system based on multi-sensor data fusion, characterized by: The system comprises: The acquisition module is used to obtain real-time data from multiple sensors on the wind power equipment, and use a five-point cubic smoothing algorithm to denoise the real-time data to obtain standard sample data; The feature extraction module is used to extract features in the time domain, frequency domain and entropy domain from multiple sensor data at the same time in the standard sample data in time sequence to obtain a multi-sensor high-dimensional feature data set; The feature fusion module is used to fuse the feature vectors in the multi-sensor high-dimensional feature data set at the same time to obtain the initial multi-modal fusion feature map; The deep extraction and fusion module is used to build a multi-scale deep feature fusion network model based on the ResNet network, input the initial multi-modal fusion feature map into the multi-scale deep feature fusion network model for deep feature extraction and multi-scale deep fusion, and output a multi-scale fusion feature vector; The evaluation and prediction module is used to build a wind power equipment quality evaluation model based on the BiLSTM network, input the multi-scale fusion feature vector into the wind power equipment quality evaluation model for training, and output the quality evaluation results of the wind power equipment.

Images