Impeller imbalance detection method and system based on multi-source data fusion

Abstract

The invention relates to the technical field of wind power generation. The impeller imbalance detection method based on multi-source data fusion comprises the steps that operation state data of a draught fan unit and environment data of corresponding time are obtained, and a multi-source data set is obtained; preprocessing the multi-source data set to obtain a preprocessed data set; feature parameters in the preprocessed data set are extracted, data fusion is carried out, and a multi-source feature fusion vector is constructed; according to the multi-source feature fusion vector, constructing a time sequence prediction model based on a gating circulation unit, and predicting a theoretical vibration baseline value under the current working condition in real time; and calculating a deviation degree between an actual vibration value and a predicted baseline through residual analysis, dynamically adjusting a detection threshold, and triggering imbalance early warning when the deviation degree exceeds the threshold. The problems that the accuracy of a detection result of a traditional impeller imbalance detection method is affected by various factors, misjudgment and missed judgment are prone to occurring, and the requirement of a modern wind power generation system for high-precision fault detection cannot be met are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and more specifically, to a method and system for detecting impeller imbalance based on multi-source data fusion. Background Art

[0002] In a wind power generation system, the impeller is a core component, and the stability of its operating state directly affects the power generation efficiency and reliability of the entire unit. Impeller imbalance is one of the common causes of fan failures. When the impeller is unbalanced, it will cause abnormal vibration of the unit, which not only reduces the service life of the fan but also may pose serious safety hazards.

[0003] Traditional impeller imbalance detection mainly relies on vibration sensors to judge whether the impeller is balanced by measuring the amplitude and frequency of the unit's vibration signal. However, this detection method has significant limitations. In actual operation, the vibration amplitudes of the x and y signals measured by the vibration sensors are affected by various factors such as wind speed, wind direction, rotational speed, and power generation. For example, different wind speeds and directions will change the aerodynamic force on the impeller, thereby affecting the characteristics of the vibration signal; changes in the rotational speed of the fan and fluctuations in power generation will also interfere with the vibration amplitude. The combined effect of these factors greatly reduces the accuracy of the detection results based on traditional vibration sensors, making it prone to false positives and false negatives, and unable to meet the requirements of modern wind power generation systems for high-precision fault detection. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for detecting impeller imbalance based on multi-source data fusion, aiming to solve the problem that the accuracy of the detection results of traditional impeller imbalance detection methods is affected by various factors, prone to false positives and false negatives, and unable to meet the requirements of modern wind power generation systems for high-precision fault detection.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for detecting impeller imbalance based on multi-source data fusion includes the following steps:

[0007] Obtain the operating state data of the fan unit and the environmental data at the corresponding time to obtain a multi-source data set;

[0008] Preprocess the multi-source data set to obtain a preprocessed data set;

[0009] Extract the characteristic parameters from the preprocessed data set and perform data fusion to construct a multi-source feature fusion vector;

[0010] According to the multi-source feature fusion vector, construct a time series prediction model based on a gated recurrent unit to real-time predict the theoretical vibration baseline value under the current working conditions;

[0011] The deviation between the actual vibration value and the predicted baseline is calculated through residual analysis, and the detection threshold is dynamically adjusted. When the deviation exceeds the threshold, an imbalance warning is triggered.

[0012] Optionally, the specific process of acquiring the operating status data of the fan unit and the environmental data of the corresponding time to obtain the multi-source data set is:

[0013] The vibration amplitude data of the impeller hub in the X-axis and Y-axis directions are synchronously collected through the vibration sensor array;

[0014] The speed encoder is used to obtain the fan main shaft speed data in real time, and the generator output power data is recorded through the power transmitter;

[0015] An ultrasonic anemometer is used to obtain ambient wind speed data, and an electronic wind vane is used to collect wind direction angle data;

[0016] Based on a unified timestamp benchmark, the vibration amplitude data, rotation speed data, power data, ambient wind speed data, and wind direction angle data are aligned in time and space to obtain a multi-source data set.

[0017] Optionally, the specific process of preprocessing the multi-source data set to obtain the preprocessed data set is:

[0018] The cubic spline interpolation method is used to fill in the missing values in the vibration amplitude data, speed data, power data, ambient wind speed data and wind direction angle data, and the outliers in the data of each dimension are eliminated based on the 3σ principle;

[0019] Adaptive median filtering is performed on the vibration amplitude data, and the high-frequency noise and pulse interference are eliminated by combining the wavelet threshold denoising algorithm;

[0020] Standardize the vibration amplitude data, rotation speed data, power data, ambient wind speed data and wind direction angle data respectively;

[0021] The environmental wind speed data and wind direction angle data are smoothed using a sliding window with a window size of 10 sampling points to eliminate instantaneous fluctuations in the environmental data;

[0022] The processed vibration amplitude data, rotation speed data, power data, ambient wind speed data and wind direction angle data are spliced in chronological order to generate a preprocessed data set.

[0023] Optionally, the specific process of extracting feature parameters from the preprocessed data set, performing data fusion, and constructing a multi-source feature fusion vector is as follows:

[0024] Extract statistical features from vibration amplitude data: calculate the mean, variance, peak factor, kurtosis coefficient and waveform factor of X-axis and Y-axis vibration;

[0025] Extract the fluctuation characteristics from the rotational speed data: calculate the rotational speed change rate, the standard deviation of the rotational speed, and the ratio of the standard deviation of the rotational speed to the average rotational speed;

[0026] Extract the energy characteristics from the power data: calculate the mean value of the power spectral density, the power fluctuation entropy, and the power impact factor;

[0027] Extract the dynamic characteristics from the environmental data: calculate the standard deviation of the wind speed, the change rate of the wind direction angle, and the turbulence intensity index;

[0028] Perform short-time Fourier transform on the vibration signal to extract the main frequency amplitude, the frequency band energy ratio, and the sideband frequency components; obtain the distribution entropy of the vibration energy in the frequency domain subspace by using wavelet packet decomposition; perform Hilbert-Huang transform on the rotational speed and power signals to extract the instantaneous frequency modulation characteristics;

[0029] Establish the joint probability density distribution of the wind speed - wind direction angle, and calculate its information gain entropy; extract the coupling dimension characteristics of the vibration signal and the rotational speed fluctuation through the phase space reconstruction technology; construct the cross-correlation matrix of the environmental load vector and the mechanical response vector, and extract its singular value decomposition characteristics;

[0030] Adopt the improved D-S evidence theory to fuse the confidence degrees of the same type of characteristics; use the sparse autoencoder to perform non-linear dimensionality reduction on the high-dimensional characteristics and retain the fault-sensitive principal components; perform dynamic weight allocation on the fused characteristics based on the attention mechanism to generate a multi-source feature fusion vector containing spatio-temporal correlation information; construct a time series feature matrix through a sliding time window as the input sequence of the gated recurrent unit model;

[0031] Use the maximum information coefficient to screen the key characteristics strongly related to the impeller imbalance state; verify the monotonic correlation between the characteristic parameters and the deviation degree of the vibration baseline; establish a feature importance ranking model to ensure that the confidence degree of the multi-source feature fusion vector in characterizing the imbalance fault exceeds 95%;

[0032] Optionally, the specific process of constructing a time series prediction model based on the gated recurrent unit according to the multi-source feature fusion vector to predict the theoretical vibration baseline value under the current working condition in real time is:

[0033] Use the gated recurrent unit to construct a time series prediction model as shown in the following formula (1):

[0034]

[0035] Wherein, represents the multi-source feature fusion vector at the current moment t, with a dimension of D; h t-1 and h t represent the hidden states at the previous moment and the current moment respectively, with a dimension of H; r t represents the reset gate, which is used to control the influence degree of the previous state information on the current candidate hidden state; represents an update gate for controlling the linear interpolation ratio between the previous state information and the current candidate hidden state; represents the candidate hidden state, which is jointly calculated from the previous state filtered by the reset gate and the current input; represents the theoretical vibration baseline value at the current time t; W r 、W z and W h respectively represent the reset gate weight, the update gate weight, and the candidate hidden state weight; b r 、b z and b j respectively represent the bias vectors of the corresponding layers; σ represents the activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication; FC represents the fully connected layer;

[0036] The training of the time series prediction model uses the mean squared error loss function, as shown in Equation (2) below:

[0037]

[0038] where, represents the value of the mean squared error loss function; y t represents the actual vibration amplitude at the current time t; N represents the number of training samples; the model parameters θ = {W r ,W z ,W h ,b r ,b z ,b h} are optimized through the backpropagation algorithm;

[0039] During prediction, the time series feature matrix within the sliding time window is input into the trained time series prediction model, and the theoretical vibration baseline value under the current working condition is output.

[0040] Optionally, the specific process of calculating the deviation degree between the actual vibration value and the predicted baseline through residual analysis and dynamically adjusting the detection threshold to trigger an imbalance warning when the deviation degree exceeds the threshold is as follows:

[0041] Calculate the residual sequence and construct the normalized deviation degree, as shown in Equation (3) below:

[0042]

[0043] where, e t represents the residual at the current time t; δ t represents the normalized deviation degree; M represents the sliding window length;

[0044] Calculate the dynamic threshold using the exponentially weighted moving average, as shown in Equation (4) below:

[0045]

[0046] Among them, S t represents the dynamic threshold; α represents the forgetting factor; δ t-1 Indicates the standardized deviation of the previous moment; μ δ represents the mean of historical residuals;

[0047] When δ t >β·σ δ When , the warning is triggered; where β represents the confidence coefficient; σ δ represents the residual standard deviation;

[0048] K consecutive sampling points satisfy δ t >γ·S t When , it is determined to be an impeller imbalance fault; where γ represents the dynamic adjustment coefficient.

[0049] Optionally, it is characterized in that when an unbalanced fault is detected, the faulty impeller blade is located through vibration modal analysis, and a fault cause analysis model based on a decision tree is established in combination with historical operating data, the fault cause is automatically associated, and a maintenance recommendation report is generated.

[0050] Based on the same inventive concept, the present invention also provides an impeller imbalance detection system based on multi-source data fusion, which is used to implement the impeller imbalance detection method based on multi-source data fusion, including a multi-source data acquisition module, a data preprocessing module, a multi-source feature engineering module, a time series prediction model module and a residual analysis and early warning module connected in sequence;

[0051] The multi-source data acquisition module is used to collect vibration amplitude data in the X-axis and Y-axis directions of the impeller hub, wind turbine main shaft speed data, generator output power data, ambient wind speed data and wind direction angle data through a vibration sensor array, a speed encoder, a power transmitter, an ultrasonic anemometer and an electronic wind vane, and perform spatiotemporal alignment based on a unified timestamp reference to generate a multi-source data set;

[0052] The data preprocessing module is used to fill missing values in multi-source data sets using cubic spline interpolation, remove outliers based on the 3σ principle, perform adaptive median filtering and wavelet threshold denoising on vibration amplitude data, standardize data of each dimension, perform sliding window smoothing on ambient wind speed and wind direction angle data, and splice in chronological order to generate preprocessed data sets;

[0053] The multi-source feature engineering module is used to extract corresponding feature parameters from various types of data in the preprocessed dataset, perform confidence fusion using the improved D-S evidence theory, reduce the dimension using the sparse autoencoder, allocate weights based on the attention mechanism, generate a multi-source feature fusion vector, construct a time-series feature matrix through a sliding time window, screen key features, and establish a feature importance ranking model;

[0054] The time-series prediction model module is used to construct a time-series prediction model using a gated recurrent unit, take the mean squared error loss function as the objective function, optimize the model parameters through the backpropagation algorithm, input the time-series feature matrix within the sliding time window into the trained model, and output the theoretical vibration baseline value under the current working condition;

[0055] The residual analysis and early warning module is used to calculate the residual sequence and standardized deviation degree between the actual vibration value and the predicted baseline, calculate the dynamic threshold using the exponentially weighted moving average, trigger an early warning when the standardized deviation degree meets specific conditions, locate the faulty impeller blade through vibration mode analysis when an imbalance fault is detected, establish a decision tree-based fault cause analysis model in combination with historical operation data, and generate a maintenance recommendation report.

[0056] Based on the same inventive concept, the present invention also provides an electronic device, including a memory and a processor, where the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned impeller imbalance detection method based on multi-source data fusion.

[0057] Based on the same inventive concept, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the above-mentioned impeller imbalance detection method based on multi-source data fusion.

[0058] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0059] Improve detection accuracy: Break through the limitation of traditional detection that only relies on vibration sensors, construct a multi-source dataset by obtaining the operation state data of the fan unit and the environmental data at the corresponding time; comprehensively consider various influencing factors such as wind speed, wind direction, rotational speed, and power generation, and then through data fusion and feature extraction, the detection model can obtain richer and more comprehensive information, avoiding misjudgment and missed judgment caused by interference of multiple factors in a single data source, greatly improving the accuracy of the detection results, and meeting the requirements of modern wind power generation systems for high-precision fault detection.

[0060] Realize real-time monitoring: With the help of a time series prediction model based on gated recurrent units, it is possible to predict the theoretical vibration baseline value under the current working conditions in real time, fully exploit the time series characteristics of the data, timely track the dynamic changes in the operating state of the fan, provide real-time and effective reference for impeller imbalance detection, ensure that the fan is always under monitoring during operation, and lay a foundation for ensuring the safe and stable operation of the unit.

[0061] Dynamically adjust the detection threshold: Calculate the deviation degree between the actual vibration value and the predicted baseline through residual analysis, and dynamically adjust the detection threshold, so that the detection method can adapt to the complex and changeable operating conditions of the fan, avoid false alarms or missed alarms caused by the inadaptability of the fixed threshold under different working conditions, and further improve the reliability and stability of the detection.

[0062] Prolong the service life of the fan: By accurately and timely detecting the impeller imbalance fault, the staff can quickly take corresponding measures for repair and maintenance, effectively reduce the damage caused to the fan by the abnormal vibration caused by impeller imbalance, thereby prolonging the service life of the fan, reducing the equipment maintenance cost, and improving the economic benefits of the wind power generation system.

[0063] Reduce potential safety hazards: Precise detection and timely warning can avoid serious safety accidents caused by impeller imbalance, provide strong guarantee for the safe production of the wind farm, and reduce potential casualties and property losses. Description of the Drawings

[0064] Figure 1 It is a schematic flow chart of an impeller imbalance detection method based on multi-source data fusion according to an embodiment of the present invention;

[0065] Figure 2 It is a schematic structural diagram of an impeller imbalance detection system based on multi-source data fusion according to an embodiment of the present invention. Detailed Embodiments

[0066] The following are the detailed embodiments in combination with the drawings.

[0067] Refer to Figure 1 , An impeller imbalance detection method based on multi-source data fusion includes the following steps:

[0068] Step 1: Obtain the operating state data of the fan unit and the environmental data at the corresponding time to obtain a multi-source data set.

[0069] In some embodiments, the specific process of obtaining the operating state data of the fan unit and the environmental data at the corresponding time to obtain a multi-source data set is as follows:

[0070] Synchronously collect the vibration amplitude data in the X-axis and Y-axis directions at the impeller hub through a vibration sensor array;

[0071] The rotational speed data of the fan main shaft is obtained in real time by using a rotational speed encoder, and the output power data of the generator is recorded by a power transmitter;

[0072] The environmental wind speed data is obtained by using an ultrasonic anemometer, and the wind direction angle data is collected by using an electronic wind vane;

[0073] Based on a unified timestamp benchmark, the vibration amplitude data, rotational speed data, power data, environmental wind speed data, and wind direction angle data are aligned in space and time to obtain a multi-source dataset. The unification of timestamps can be achieved through the time synchronization function of the data acquisition system or an external time synchronization device.

[0074] Step 2: Preprocess the multi-source dataset to obtain a preprocessed dataset.

[0075] In some embodiments, the specific process of preprocessing the multi-source dataset to obtain a preprocessed dataset is as follows:

[0076] The cubic spline interpolation method is used to fill in the missing values in the vibration amplitude data, rotational speed data, power data, environmental wind speed data, and wind direction angle data, and the outliers in the data of each dimension are removed based on the 3σ principle. For the data of a certain dimension, its mean value and standard deviation are calculated, and the data outside the range is regarded as an outlier and removed.

[0077] The vibration amplitude data is processed by adaptive median filtering, and the wavelet threshold denoising algorithm is combined to eliminate high-frequency noise and pulse interference. The wavelet transform is performed on the vibration amplitude data to decompose the data into different frequency sub-bands; the appropriate threshold is selected according to the noise characteristics to process the wavelet coefficients, and the wavelet coefficients smaller than the threshold are set to zero or subjected to shrinkage processing; the data is reconstructed through the inverse wavelet transform to obtain the denoised vibration amplitude data.

[0078] The vibration amplitude data, rotational speed data, power data, environmental wind speed data, and wind direction angle data are respectively standardized. The standardization method that can be adopted can be Z-score standardization, that is, for the data of each dimension, its standardized value is calculated.

[0079] The environmental wind speed data and wind direction angle data are processed by sliding window smoothing with a window size of 10 sampling points to eliminate the instantaneous fluctuations of the environmental data. For each sampling point, the data of 10 sampling points before and after it (including the sampling point itself) can be taken, and the average value of these 10 data is calculated, and this average value is used as the smoothed value of the current sampling point. In this way, the instantaneous fluctuations of the environmental data are eliminated, making the data more stable.

[0080] The processed vibration amplitude data, rotational speed data, power data, environmental wind speed data, and wind direction angle data are concatenated in chronological order to generate a preprocessed dataset.

[0081] Step 3: Extract the feature parameters from the preprocessed dataset, perform data fusion, and construct a multi-source feature fusion vector.

[0082] In some embodiments, the specific process of extracting the feature parameters from the preprocessed dataset, performing data fusion, and constructing a multi-source feature fusion vector is as follows:

[0083] Extract statistical features from the vibration amplitude data: Calculate the mean, variance, peak factor, kurtosis coefficient, and waveform factor of the vibration in the X-axis and Y-axis. The mean reflects the average level of the vibration amplitude; the variance measures the degree of dispersion of the vibration amplitude; the peak factor is used to highlight the peak situation in the vibration; the kurtosis coefficient can reflect the impact characteristics of the vibration signal; and the waveform factor has a certain characterization effect on the waveform shape of the vibration signal.

[0084] Extract the fluctuation features from the rotational speed data: Calculate the rotational speed change rate, which is the ratio of the difference in rotational speed at adjacent times to the time interval, to reflect the speed of change of the rotational speed; calculate the rotational speed standard deviation, which is used to measure the degree of fluctuation of the rotational speed; calculate the ratio of the rotational speed standard deviation to the average rotational speed to eliminate the influence of the average rotational speed on the degree of fluctuation.

[0085] Extract the energy features from the power data: Calculate the mean value of the power spectral density. By averaging the power spectral density, obtain the average energy level of the power in the frequency domain; calculate the power fluctuation entropy, which is used to measure the uncertainty of the power fluctuation; calculate the power impact factor to reflect the sudden change of the power. The power spectral density can be obtained by using the power spectrum estimation method and then perform relevant calculations.

[0086] Extract the dynamic features from the environmental data: Calculate the wind speed standard deviation to reflect the fluctuation of the wind speed; calculate the wind direction angle change rate, which is the ratio of the difference in wind direction angles at adjacent times to the time interval, to reflect the speed of change of the wind direction; calculate the turbulence intensity index, which is used to characterize the degree of atmospheric turbulence.

[0087] Perform short-time Fourier transform on the vibration signal to extract the main frequency amplitude, that is, the amplitude of the main frequency components of the vibration signal; extract the band energy, which reflects the relative distribution of energy in different frequency bands; extract the sideband frequency components, which are related to faults such as impeller imbalance. Use wavelet packet decomposition to obtain the distribution entropy of the vibration energy in the frequency domain subspace to further describe the distribution characteristics of the vibration energy in different frequency bands. Perform Hilbert-Huang transform on the rotational speed and power signals to extract the instantaneous frequency modulation features, which can more accurately reflect the frequency change of the signal at different times.

[0088] Establish the joint probability density distribution of wind speed and wind direction angle, calculate its information gain entropy to capture the joint variation information between wind speed and wind direction angle; extract the coupling dimensionality features of vibration signals and rotational speed fluctuations through phase space reconstruction technology to analyze the internal relationship between the two; construct the cross-correlation matrix of environmental load vectors and mechanical response vectors, and extract its singular value decomposition features to characterize the interaction relationship between environmental loads and mechanical responses.

[0089] Use the improved D-S evidence theory to fuse the confidence of similar features. For similar features such as multiple statistical features of vibration amplitude data, calculate the confidence according to their support degrees for the impeller imbalance fault under different working conditions, and perform fusion to obtain a more reliable feature representation.

[0090] Apply the sparse autoencoder to perform non-linear dimensionality reduction on high-dimensional features. The sparse autoencoder retains the fault-sensitive principal components by learning the compressed representation of input features, removes redundant information, reduces the feature dimension, and improves the efficiency and accuracy of subsequent processing.

[0091] Perform dynamic weight assignment on the fused features based on the attention mechanism. Assign corresponding weights to each feature according to the importance of different features in reflecting the impeller imbalance fault, and generate a multi-source feature fusion vector containing spatio-temporal correlation information. For example, assign a higher weight to features that are more sensitive to the imbalance fault under certain working conditions.

[0092] Construct a time series feature matrix through a sliding time window, arrange the multi-source feature fusion vectors in chronological order, and use them as the input sequence of the gated recurrent unit model. The size of the sliding time window can be adjusted according to the actual situation to capture the feature change information within an appropriate time range.

[0093] Adopt the maximum information coefficient to screen the key features strongly related to the impeller imbalance state. The maximum information coefficient can measure the non-linear correlation between features. By calculating the maximum information coefficient between each feature and the impeller imbalance state, features with higher correlation are screened out, reducing the number of features and improving the generalization ability of the model.

[0094] Verify the monotonic correlation between the feature parameters and the deviation degree of the vibration baseline. Ensure that the selected feature parameters have a monotonic change relationship with the vibration baseline deviation caused by impeller imbalance to a certain extent, thereby improving the ability of the features to characterize faults.

[0095] Establish a feature importance ranking model, rank the features according to factors such as the correlation between the features and the impeller imbalance fault and their impact on the model performance, ensure that the confidence of the multi-source feature fusion vector in characterizing the imbalance fault exceeds 95%, and improve the accuracy and reliability of fault detection.

[0096] Step 4: Based on the multi-source feature fusion vector, construct a time series prediction model based on a gated recurrent unit to predict the theoretical vibration baseline value under the current working condition in real time.

[0097] In some embodiments, the specific process of constructing a time series prediction model based on a gated recurrent unit to predict the theoretical vibration baseline value under the current working condition according to the multi-source feature fusion vector is as follows:

[0098] Use a gated recurrent unit to construct a time series prediction model, as shown in the following formula (1):

[0099]

[0100] where, represents the multi-source feature fusion vector at the current time t, with a dimension of D; h t-1 and h t represent the hidden states at the previous time and the current time respectively, with a dimension of H; r t represents the reset gate, which is used to control the influence degree of the previous state information on the current candidate hidden state; represents the update gate, which is used to control the linear interpolation ratio between the previous state information and the current candidate hidden state; represents the candidate hidden state, which is obtained by jointly calculating the previous state screened by the reset gate and the current input; represents the theoretical vibration baseline value at the current time t; W r 、 and W h represent the reset gate weight, the update gate weight, and the candidate hidden state weight respectively; b r 、 and b h represent the bias vectors of the corresponding layers; σ represents the activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication; FC represents the fully connected layer;

[0101] The time series prediction model is trained using the mean squared error loss function, as shown in the following formula (2):

[0102]

[0103] where, represents the mean squared error loss function value; y t represents the actual vibration amplitude at the current time t; N represents the number of training samples; optimize the model parameters through the backpropagation algorithm

[0104] During prediction, input the time series feature matrix within the sliding time window into the trained time series prediction model, and output the theoretical vibration baseline value under the current working condition.

[0105] Step 5: Calculate the deviation degree between the actual vibration value and the predicted baseline through residual analysis, dynamically adjust the detection threshold, and trigger an imbalance warning when the deviation degree exceeds the threshold.

[0106] In some embodiments, the specific process of calculating the deviation degree between the actual vibration value and the predicted baseline through residual analysis, dynamically adjusting the detection threshold, and triggering an imbalance warning when the deviation degree exceeds the threshold is as follows:

[0107] Calculate the residual sequence and construct the standardized deviation degree as shown in the following formula (3):

[0108]

[0109] where, e t represents the residual at the current time t; δ t represents the standardized deviation degree; M represents the sliding window length;

[0110] Calculate the dynamic threshold using the exponentially weighted moving average as shown in the following formula (4):

[0111]

[0112] where, S t represents the dynamic threshold; α represents the forgetting factor; δ t-1 represents the standardized deviation degree at the previous time; μ δ represents the historical residual mean;

[0113] When δ t > β·σ δ , trigger a warning; where, β represents the confidence coefficient; σ δ represents the residual standard deviation;

[0114] When K consecutive sampling points satisfy δ t > γ·S t , it is determined as an impeller imbalance fault; where, γ represents the dynamic adjustment coefficient.

[0115] In some embodiments, when an imbalance fault is detected, locate the faulty impeller blade through vibration mode analysis, establish a decision tree-based fault cause analysis model in combination with historical operation data, automatically associate the fault cause, and generate a maintenance recommendation report.

[0116] When the system detects an impeller imbalance fault, start the high-precision vibration sensor array to continuously collect the vibration amplitude data in the X-axis and Y-axis directions at the impeller hub with a higher sampling frequency, and ensure that the acquisition duration covers at least one complete impeller rotation cycle. At the same time, obtain the stress-strain data of the blades during operation through strain gauges installed at specific positions on the blades. Use the fast Fourier transform to convert the collected time-domain vibration signals into frequency-domain signals, and identify the natural frequencies and vibration modes of the impeller. Combine the finite element model to simulate and analyze the vibration characteristics of the impeller, providing a theoretical basis for fault location. By comparing the measured modal parameters with the theoretical values, determine the approximate range where the faulty blade is located. Use the modal strain energy method to calculate the strain energy distribution of each blade in different modes. Due to uneven mass distribution or structural damage of the faulty blade, its strain energy will be significantly higher than that of normal blades. By comparing the strain energies of each blade, accurately determine the position of the faulty blade.

[0117] Collect the historical operation data of the fan unit, including but not limited to vibration amplitude data, rotational speed data, power data, ambient wind speed data, wind direction angle data, and detailed records of each fault occurrence, such as fault type, fault time, maintenance measures, etc. Clean the collected data, remove outliers and missing values, and sort them in chronological order.

[0118] Extract the characteristic parameters related to the impeller imbalance fault from the historical operation data, such as the statistical characteristics of vibration signals, the fluctuation characteristics of rotational speed and power, the dynamic characteristics of environmental data, etc. Use methods such as information gain and chi-square test to screen out the key characteristics with strong discrimination ability for the fault cause.

[0119] Adopt classical algorithms such as ID3 and C4.5 to construct a decision tree-based fault cause analysis model. Take the fault cause as the target variable and the screened key characteristics as the input variables, and construct a decision tree structure by recursively dividing the data set. During the construction process, prevent the model from overfitting through pruning strategies and improve the generalization ability of the model.

[0120] Divide the historical operation data into a training set and a test set, use the training set to train the decision tree model, adjust the model parameters to make the model reach the best performance. Use the test set to evaluate the trained model, calculate indicators such as the accuracy rate, recall rate, and F1 value of the model, and verify the effectiveness of the model.

[0121] When an impeller imbalance fault is detected, the current operating data and vibration modal analysis results are input into the trained decision tree model. The model will automatically associate possible fault causes, such as blade wear, mass eccentricity, loose connections, etc., based on preset rules and features. According to the associated fault causes, combined with the maintenance manual and historical maintenance experience of the fan unit, targeted maintenance recommendations are generated. For blade wear failures, it is recommended to repair or replace the blades; for loose connection failures, it is recommended to check and tighten the connection parts. The fault causes, maintenance recommendations, and related operating data are organized into a maintenance recommendation report, which is pushed to maintenance personnel by email, text message, etc. At the same time, the report is stored in the maintenance management system of the fan unit for subsequent query and analysis.

[0122] Based on the same inventive concept, corresponding to any of the above embodiments, refer to Figure 2 , the present invention provides an impeller imbalance detection system based on multi-source data fusion, which is used to implement the aforementioned impeller imbalance detection method based on multi-source data fusion, including a multi-source data acquisition module, a data preprocessing module, a multi-source feature engineering module, a time series prediction model module and a residual analysis and early warning module connected in sequence;

[0123] The multi-source data acquisition module is used to collect vibration amplitude data in the X-axis and Y-axis directions of the impeller hub, wind turbine main shaft speed data, generator output power data, ambient wind speed data and wind direction angle data through a vibration sensor array, a speed encoder, a power transmitter, an ultrasonic anemometer and an electronic wind vane, and perform spatiotemporal alignment based on a unified timestamp reference to generate a multi-source data set;

[0124] The data preprocessing module is used to fill missing values in multi-source data sets using cubic spline interpolation, remove outliers based on the 3σ principle, perform adaptive median filtering and wavelet threshold denoising on vibration amplitude data, standardize data of each dimension, perform sliding window smoothing on ambient wind speed and wind direction angle data, and splice in chronological order to generate preprocessed data sets;

[0125] The multi-source feature engineering module is used to extract corresponding feature parameters from various types of data in the preprocessed data set, use the improved DS evidence theory for confidence fusion, use sparse autoencoders for dimensionality reduction, assign weights based on the attention mechanism, generate multi-source feature fusion vectors, construct a time series feature matrix through a sliding time window, and screen key features to establish a feature importance ranking model;

[0126] The timing prediction model module is used to construct a timing prediction model using a gated recurrent unit, take a mean square error loss function as the objective function, optimize the model parameters through a back propagation algorithm, input the timing feature matrix in the sliding time window into the trained model, and output the theoretical vibration baseline value under the current working condition;

[0127] The residual analysis and early warning module is used to calculate the residual sequence and the standardized deviation degree between the actual vibration value and the predicted baseline, calculate the dynamic threshold by using the exponentially weighted moving average, trigger an early warning when the standardized deviation degree meets specific conditions, and when an imbalance fault is detected, locate the faulty impeller blade through vibration mode analysis, establish a decision tree-based fault cause analysis model in combination with historical operation data, and generate a maintenance advice report.

[0128] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor, where the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the impeller imbalance detection method based on multi-source data fusion in the embodiment.

[0129] Optionally, the above electronic device may be a server.

[0130] In addition, this embodiment also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the impeller imbalance detection method based on multi-source data fusion in the embodiment.

[0131] It can be understood that the processor in the embodiments of the present invention may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0132] The method steps in the embodiments of the present invention can be implemented in a hardware manner or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in a random access memory (RAM), flash memory, read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), register, hard disk, removable hard disk, CD-ROM, or any other form of storage medium well-known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0133] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a storage medium or transmitted through a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)), etc.

Claims

1. An impeller imbalance detection method based on multi-source data fusion, characterized in that, It includes the following steps: Obtain the operation status data of the wind turbine unit and the environmental data at the corresponding time to obtain a multi-source data set; Preprocess the multi-source data set to obtain a preprocessed data set; Extract the characteristic parameters in the preprocessed data set and perform data fusion to construct a multi-source feature fusion vector; According to the multi-source feature fusion vector, construct a time series prediction model based on a gated recurrent unit to predict the theoretical vibration baseline value under the current working condition in real time; Calculate the deviation degree between the actual vibration value and the predicted baseline through residual analysis, dynamically adjust the detection threshold, and trigger an imbalance warning when the deviation degree exceeds the threshold.

2. The impeller imbalance detection method based on multi-source data fusion according to claim 1, characterized in that The specific process of obtaining the operation status data of the wind turbine unit and the environmental data at the corresponding time to obtain a multi-source data set is as follows: Synchronously collect the vibration amplitude data in the X-axis and Y-axis directions at the impeller hub through a vibration sensor array; Use a speed encoder to obtain the wind turbine main shaft speed data in real time, and record the generator output power data through a power transmitter; Use an ultrasonic anemometer to obtain the environmental wind speed data, and use an electronic wind vane to collect the wind direction angle data; Based on a unified timestamp benchmark, perform spatio-temporal alignment on the vibration amplitude data, speed data, power data, environmental wind speed data, and wind direction angle data to obtain a multi-source data set.

3. The impeller imbalance detection method based on multi-source data fusion according to claim 1, wherein, The specific process of preprocessing the multi-source data set to obtain a preprocessed data set is as follows: Use the cubic spline interpolation method to fill in the missing values in the vibration amplitude data, speed data, power data, environmental wind speed data, and wind direction angle data, and eliminate the outliers in the data of each dimension based on the 3σ principle; Perform adaptive median filtering on the vibration amplitude data, and combine the wavelet threshold denoising algorithm to eliminate high-frequency noise and pulse interference; Perform standardization processing on the vibration amplitude data, speed data, power data, environmental wind speed data, and wind direction angle data respectively; Perform sliding window smoothing on the environmental wind speed data and the wind direction angle data, with a window size of 10 sampling points, to eliminate the instantaneous fluctuations of the environmental data; Concatenate the processed vibration amplitude data, speed data, power data, environmental wind speed data, and wind direction angle data in chronological order to generate a preprocessed data set.

4. The impeller imbalance detection method based on multi-source data fusion according to claim 1, wherein, The specific process of extracting the characteristic parameters in the preprocessed data set and performing data fusion to construct a multi-source feature fusion vector is as follows: Extract statistical features from the vibration amplitude data: Calculate the mean value, variance, peak factor, kurtosis coefficient, and waveform factor of the vibration in the X-axis and Y-axis; Extract the fluctuation features from the speed data: calculate the speed change rate, speed standard deviation, and the ratio of the speed standard deviation to the average speed; Extract the energy features from the power data: calculate the mean value of the power spectral density, power fluctuation entropy, and power impact factor; Extract the dynamic features from the environmental data: calculate the wind speed standard deviation, wind direction angle change rate, and turbulence intensity index; Perform short-time Fourier transform on the vibration signal to extract the main frequency amplitude, frequency band energy ratio, and sideband frequency components; use wavelet packet decomposition to obtain the distribution entropy of the vibration energy in the frequency domain subspace; perform Hilbert-Huang transform on the speed and power signals to extract the instantaneous frequency modulation features; Establish the joint probability density distribution of wind speed and wind direction angle, and calculate its information gain entropy; extract the coupling dimension characteristics of vibration signal and speed fluctuation through phase space reconstruction technology; construct the cross-correlation matrix of environmental load vector and mechanical response vector, and extract its singular value decomposition characteristics; The improved DS evidence theory is used to fuse the confidence of similar features; the sparse autoencoder is used to perform nonlinear dimensionality reduction on high-dimensional features and retain the fault-sensitive principal components; the fusion features are dynamically weighted based on the attention mechanism to generate a multi-source feature fusion vector containing spatiotemporal correlation information; the time series feature matrix is constructed through a sliding time window as the input sequence of the gated recurrent unit model; The maximum information coefficient is used to screen key features that are strongly correlated with the impeller imbalance state; the monotonic correlation between the characteristic parameters and the vibration baseline deviation is verified; and a feature importance ranking model is established to ensure that the confidence level of the multi-source feature fusion vector in characterizing the imbalance fault exceeds 95%.

5. The impeller imbalance detection method based on multi-source data fusion according to claim 1, wherein, The specific process of constructing a time series prediction model based on a gated cycle unit according to the multi-source feature fusion vector and predicting the theoretical vibration baseline value under the current working condition in real time is as follows: The gated recurrent unit is used to construct a timing prediction model, as shown in the following formula (1): Among them, represents the multi-source feature fusion vector at the current moment t, with a dimension of D; h t-1 and h t represent the hidden states at the previous moment and the current moment respectively, with a dimension of H; r t represents the reset gate, which is used to control the influence degree of the previous state information on the current candidate hidden state; z t represents the update gate, which is used to control the linear interpolation ratio between the previous state information and the current candidate hidden state; represents the candidate hidden state, which is jointly calculated by the previous state filtered by the reset gate and the current input; represents the theoretical vibration baseline value at the current moment t; W r 、W z and W h represent the reset gate weight, the update gate weight, and the candidate hidden state weight respectively; b r 、b z and b j represent the bias vectors of the corresponding layers respectively; σ represents the activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication; FC represents the fully connected layer; The time series prediction model training adopts the mean square error loss function, as shown in the following formula (2): Among them, represents the mean square error loss function value; y t represents the actual vibration amplitude at the current time t; N represents the number of training samples; the model parameters θ = {W r , W z , W h , b r , b z , b h} are optimized through the backpropagation algorithm; During prediction, the time series feature matrix within the sliding time window is input into the trained time series prediction model, and the theoretical vibration baseline value under the current working condition is output.

6. The impeller imbalance detection method based on multi-source data fusion according to claim 5, characterized in that The specific process of calculating the deviation between the actual vibration value and the predicted baseline through residual analysis, dynamically adjusting the detection threshold, and triggering the imbalance warning when the deviation exceeds the threshold is as follows: Calculate the residual sequence and construct the standardized deviation, as shown in the following formula (3): Among them, e t represents the residual at the current moment t; δ t represents the standardized deviation degree; M represents the sliding window length; The dynamic threshold is calculated using exponentially weighted moving average, as shown in the following formula (4): Among them, S t represents the dynamic threshold; α represents the forgetting factor; δ t-1 represents the normalized deviation degree at the previous moment; μ δ represents the historical residual mean; When δ t > β·σ δ a warning is triggered; where β represents the confidence coefficient; σ δ represents the residual standard deviation. When K consecutive sampling points satisfy δ t > γ·S t it is determined that there is an impeller imbalance fault; where γ represents the dynamic adjustment coefficient.

7. The impeller imbalance detection method based on multi-source data fusion according to claim 1, characterized in that When an unbalanced fault is detected, the faulty impeller blade is located through vibration modal analysis, and a fault cause analysis model based on a decision tree is established in combination with historical operating data. The fault cause is automatically associated and a maintenance recommendation report is generated.

8. An impeller imbalance detection system based on multi-source data fusion is used to implement the impeller imbalance detection method based on multi-source data fusion according to any one of claims 1-7, characterized in that, It includes a multi-source data acquisition module, a data preprocessing module, a multi-source feature engineering module, a time series prediction model module and a residual analysis and early warning module which are connected in sequence; The multi-source data acquisition module is used to collect vibration amplitude data in the X-axis and Y-axis directions of the impeller hub, wind turbine main shaft speed data, generator output power data, ambient wind speed data and wind direction angle data through a vibration sensor array, a speed encoder, a power transmitter, an ultrasonic anemometer and an electronic wind vane, and perform spatiotemporal alignment based on a unified timestamp reference to generate a multi-source data set; The data preprocessing module is used to fill missing values in multi-source data sets using cubic spline interpolation, remove outliers based on the 3σ principle, perform adaptive median filtering and wavelet threshold denoising on vibration amplitude data, standardize data of each dimension, perform sliding window smoothing on ambient wind speed and wind direction angle data, and splice in chronological order to generate preprocessed data sets; The multi-source feature engineering module is used to extract corresponding feature parameters for various types of data in the preprocessed dataset, perform confidence fusion using the improved D-S evidence theory, reduce the dimension using the sparse autoencoder, assign weights based on the attention mechanism, generate a multi-source feature fusion vector, construct a time-series feature matrix through a sliding time window, screen key features, and establish a feature importance ranking model; The time-series prediction model module is used to construct a time-series prediction model using a gated recurrent unit, use the mean squared error loss function as the objective function, optimize the model parameters through the backpropagation algorithm, input the time-series feature matrix within the sliding time window into the trained model, and output the theoretical vibration baseline value under the current working condition; The residual analysis and early warning module is used to calculate the residual sequence and standardized deviation degree between the actual vibration value and the predicted baseline, calculate the dynamic threshold using the exponentially weighted moving average, trigger an early warning when the standardized deviation degree meets specific conditions, and when an imbalance fault is detected, locate the faulty impeller blade through vibration modal analysis, establish a decision tree-based fault cause analysis model in combination with historical operation data, and generate a maintenance recommendation report.

9. An electronic device, characterized in that, It includes a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the impeller imbalance detection method based on multi-source data fusion according to any one of claims 1-7.

10. A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the impeller imbalance detection method based on multi-source data fusion according to any one of claims 1-7.

Images