A method for judging an abnormal state of a fan based on unsupervised learning

Abstract

The application relates to the technical field of fan abnormality judgment, and discloses a method for judging abnormal states of a fan based on unsupervised learning, which comprises the following steps: generating standardized operation data corresponding to each structure category; when the fan is in real-time operation, if an initial abnormality alarm event is triggered by the built-in normality judgment logic of the fan, the structure category corresponding to the initial abnormality alarm event is identified, and the standardized operation data of the corresponding fan structure category is subjected to dimension reduction processing by using an unsupervised deep learning technology to obtain low-dimensional feature data after dimension reduction; the low-dimensional feature data is subjected to secondary confirmation by using a preset abnormality judgment logic; if the secondary confirmation result determines that the fan has a real abnormality, the initial abnormality alarm event information is reported, otherwise, the initial abnormality alarm event is determined to be a false alarm and is automatically filtered. The application has high universality, low use threshold, and the dimension reduction conversion matrix can be obtained only by pre-training a large amount of data.

Description

Technical Field

[0001] This invention relates to the field of wind turbine anomaly detection technology, and more specifically, to a method for detecting wind turbine anomalies based on unsupervised learning. Background Technology

[0002] Currently, in the main control PLC of wind turbines, after receiving all the data from the wind turbine via the master station, only simple judgments are typically made regarding abnormal situations, such as whether certain values ​​exceed or fall below a certain threshold within a certain period. However, the introduction of machine learning into the PLC main control program to assist in the diagnosis of abnormalities is gradually gaining importance.

[0003] Chinese patent application CN202510725900.0 discloses a deep learning-based operation and maintenance scheduling method for offshore wind power equipment. This patent uses a variational autoencoder to extract dimensionality reduction features from preprocessed data. However, the variational autoencoder used has high parameter tuning requirements, a complex calculation process, and requires a large amount of computing resources. Chinese patent application CN202111560423.5 discloses a method for detecting and cleaning abnormal data of wind turbine units based on an LSTM-AE integrated sharing framework. It preprocesses the SCADA data of multiple neighboring wind turbine units in the same wind farm to obtain multiple original time series with the same timestamp. It constructs an LSTM-AE integrated sharing framework to introduce the correlation between data of neighboring units. However, this method relies on a large amount of comprehensive data across units, has high memory resource requirements, and is only suitable for SCADA system deployment, and cannot be directly integrated into the local PLC of the wind turbine.

[0004] In summary, existing technologies mostly employ deep learning models such as variational autoencoders and LSTM-AE, which contain nonlinear activation functions and complex network structures. Not only does the training process have extremely high requirements for parameter tuning, but the inference stage also involves a large number of nonlinear operations and matrix transformations. These technologies cannot be run on wind turbine main control PLCs with extremely limited computing resources and can only be deployed in the cloud or wind farm SCADA systems.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] In response to the problems in related technologies, this invention proposes a method for judging abnormal states of wind turbines based on unsupervised learning, so as to overcome the above-mentioned technical problems existing in the existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] According to a first aspect of the present invention, a method for determining abnormal states of wind turbines based on unsupervised learning is provided, the method comprising:

[0009] Acquire sample data under normal operating conditions of the wind turbine, and classify and standardize the sample dataset based on the structural characteristics of the wind turbine to generate standardized operating data corresponding to each structural category;

[0010] When the wind turbine is running in real time, if the wind turbine's built-in normal judgment logic triggers an initial abnormal alarm event, the wind turbine structure category corresponding to the initial abnormal alarm event is identified, and unsupervised deep learning technology is used to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure category to obtain dimensionality-reduced low-dimensional feature data.

[0011] The low-dimensional feature data is reconfirmed by a preset anomaly judgment logic; if the reconfirmation result determines that the wind turbine has a real anomaly, the initial anomaly alarm event information is reported; otherwise, the initial anomaly alarm event is determined to be a false alarm and is automatically filtered; wherein, the anomaly judgment logic adopts a rate of change fluctuation judgment method or a normal weighted neural network judgment method.

[0012] Preferably, the step of using unsupervised deep learning technology to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure category to obtain the dimensionality-reduced low-dimensional feature data includes:

[0013] Standardized operating data for the corresponding wind turbine structure category is extracted, and the operating parameters at each data acquisition time are combined into a high-dimensional operating feature vector. The high-dimensional operating feature vectors at all times are used to form a high-dimensional feature vector set for unsupervised deep learning training.

[0014] The high-dimensional feature vector set is input into a linear autoencoder network to train and obtain the dimensionality reduction transformation matrix and inverse transformation matrix corresponding to the current dimension. The dimensionality reduction transformation matrix and inverse transformation matrix are combined to generate a low-dimensional feature vector set and verify the effectiveness of dimensionality reduction.

[0015] The validated low-dimensional feature vector set is used as the high-dimensional input feature vector set for the new round of training. The single-level training and validation process is repeated to obtain the intermediate dimensionality reduction transformation matrix corresponding to each dimension level. The iteration process is terminated when the reconstruction error can no longer meet the preset engineering threshold.

[0016] The intermediate dimensionality reduction transformation matrices corresponding to each dimension level are multiplied sequentially according to the training order to obtain the final dimensionality reduction transformation matrix of the corresponding wind turbine structure category, and at the same time, the low-dimensional feature data of the corresponding wind turbine structure category are obtained.

[0017] Preferably, the step of inputting the high-dimensional feature vector set into a linear autoencoder network, training to obtain the dimensionality reduction transformation matrix and inverse transformation matrix corresponding to the current dimension, and combining the dimensionality reduction transformation matrix and inverse transformation matrix to generate a low-dimensional feature vector set and verify the effectiveness of dimensionality reduction includes:

[0018] Define a linear autoencoder network that includes an encoder and a decoder, and input the high-dimensional feature vector set of the current dimension into the linear autoencoder network to perform an unsupervised training process;

[0019] In each round of training, the dimensionality reduction transformation matrix and the inverse transformation matrix corresponding to the current dimension are obtained; the encoder compresses the input vector set of the current dimension using the dimensionality reduction transformation matrix and outputs the dimensionality-reduced low-dimensional feature vector set; the decoder reconstructs the dimensionality-reduced low-dimensional feature vector set back to the original dimension using the inverse transformation matrix and calculates the reconstruction error.

[0020] The reconstruction error is calculated to verify the effectiveness of dimensionality reduction. If the reconstruction error is less than the preset engineering threshold, the dimensionality reduction is considered successful, the current low-dimensional feature vector set is retained, and the current low-dimensional feature vector set is used as the input for the next round of training. If the reconstruction error is greater than or equal to the preset engineering threshold, the dimensionality reduction is considered unsuccessful, and the dimensionality reduction iteration process is terminated.

[0021] Preferably, the step of performing secondary confirmation on low-dimensional feature data through a preset anomaly judgment logic; if the secondary confirmation result determines that the wind turbine has a real anomaly, then the initial anomaly alarm event information is reported; otherwise, the initial anomaly alarm event is determined to be a false alarm and automatically filtered, including:

[0022] Define anomaly detection logic, wherein the anomaly detection logic adopts a rate of change fluctuation detection method or a normal weighted neural network detection method;

[0023] Calculate the instantaneous information entropy of the low-dimensional feature data at the current moment, and determine the anomaly type corresponding to the initial anomaly alarm event based on the numerical characteristics of the instantaneous information entropy;

[0024] Based on the error cost parameters of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, the expected error loss of the two judgment methods is calculated respectively.

[0025] If the expected misjudgment loss of the rate of change fluctuation judgment method is greater than the expected misjudgment loss of the normal weighted neural network judgment method, then the normal weighted neural network judgment method will be used as the current anomaly judgment method, and the low-dimensional feature data will be reconfirmed through the normal weighted neural network judgment method.

[0026] If the expected misjudgment loss of the normal weighted neural network judgment method is greater than the expected misjudgment loss of the rate of change fluctuation judgment method, then the rate of change fluctuation judgment method will be used as the current anomaly judgment method, and the low-dimensional feature data will be reconfirmed through the rate of change fluctuation judgment method.

[0027] If the expected misjudgment loss of the normal weighted neural network judgment method is equal to the expected misjudgment loss of the rate of change fluctuation judgment method, then any judgment method can be selected as the current anomaly judgment method.

[0028] If the secondary confirmation results indicate that the fan has a real abnormality, the initial abnormality alarm event information will be reported; otherwise, the initial abnormality alarm event will be determined to be a false alarm and automatically filtered out.

[0029] Preferably, the misjudgment cost parameters for the rate of change fluctuation-based judgment method and the normal weighted neural network judgment method under the current anomaly type are calculated separately for the expected misjudgment loss of the two judgment methods, including:

[0030] Obtain the historical false alarm rate statistics of the current anomaly type in the same type of wind turbine structure from the target database, and use it as the prior probability;

[0031] The performance differences between the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type are analyzed, and a set of misjudgment cost parameters for the two judgment methods are set according to the performance differences.

[0032] Using prior probability and scenario-specific misjudgment cost parameter sets as input, the minimum risk decision algorithm is used to calculate the expected false alarm loss for choosing either the rate of change fluctuation judgment method or the normal weighted neural network judgment method under the current anomaly type.

[0033] Preferably, the analysis of the change rate fluctuation judgment method and the normal weighted neural network judgment method differs in response performance under the current anomaly type. Based on the response performance difference, false negative cost parameters and false positive cost parameters are set for the two judgment methods, and the integrated set of scenario-specific false positive cost parameters includes:

[0034] The false negative rate, false positive rate and response delay of the change rate fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type are statistically analyzed to form a set of performance indicators for different scenarios.

[0035] Based on the actual loss data caused by the current anomaly type in the historical operation and maintenance records, a unit loss benchmark is defined and a unit loss benchmark table is formed. The unit loss mechanism includes unit loss due to equipment damage, unit loss due to downtime operation and maintenance, unit loss due to safety risks, and unit loss due to invalid inspections.

[0036] The time series features of low-dimensional feature data at the current moment are obtained. Based on the performance index set of different scenarios and the unit loss benchmark table, and combined with the fault evolution results reflected by the time series features, the correlation between fault detection timing and loss degree is dynamically quantified.

[0037] Based on the correlation obtained by dynamic quantization, each performance indicator in the scenario-specific performance indicator set is multiplied by the corresponding unit loss, and the calculation results are weighted and corrected according to the fault evolution results to obtain the missed detection cost parameters and false detection cost parameters of the two judgment methods under the current anomaly type. The missed detection cost parameters and false detection cost parameters are integrated to obtain the scenario-specific misjudgment cost parameter set of the two judgment methods.

[0038] Preferably, the step of using prior probability and scenario-specific misjudgment cost parameter sets as input, and employing a minimum risk decision algorithm to calculate the expected false alarm loss for selecting either the rate of change fluctuation judgment method or the normal weighted neural network judgment method under the current anomaly type, includes:

[0039] The prior probabilities are mapped to the initial class weights of the decision tree nodes, and the misjudgment cost parameter set for each scenario is used as the misjudgment cost coefficient of the decision tree loss function; an initial decision tree structure is constructed that includes two decision branches: judgment of the rate of change fluctuation and judgment of the normal weighted neural network.

[0040] The initial decision tree structure is used as the optimization object, with the objective function being to minimize the expected misjudgment loss. An adaptive optimization algorithm is used to perform global pruning optimization on the initial decision tree structure. The pruning process starts from the leaf nodes of the initial decision tree and traverses each leaf node layer by layer from bottom to top. At each layer, the subtree branches under the current leaf node are attempted to be pruned. After each pruning attempt, the total loss function value of the decision tree is recalculated. If the total loss function value decreases, the pruning operation is retained and the next leaf node is traversed until the root node is reached. The final decision tree structure is used as the optimized hybrid decision tree model.

[0041] The performance index set of the two judgment methods under the current anomaly type is used as the input feature of the hybrid decision tree model. Through node splitting and classification reasoning of the hybrid decision tree model, the comprehensive misjudgment probability of choosing the rate of change fluctuation judgment method and the comprehensive misjudgment probability of choosing the normal weighted neural network judgment method are output respectively.

[0042] By combining the combined false positive probability of the two judgment methods with the prior probability of the current anomaly type, the expected loss of underreporting and the expected loss of false positive for the two judgment methods are calculated and summed to obtain the total expected false positive loss of the rate of change fluctuation judgment method and the total expected false positive loss of the normal weighted neural network judgment method.

[0043] Preferably, the step of outputting the comprehensive misjudgment probability of selecting the rate of change fluctuation judgment method and the comprehensive misjudgment probability of selecting the normal weighted neural network judgment method through node splitting and classification reasoning of the hybrid decision tree model includes:

[0044] Obtain the false negative rate and false positive rate of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, combine the false negative rate and false positive rate into an input feature vector, and input it into the root node of the hybrid decision tree model;

[0045] The hybrid decision tree model performs layer-by-layer classification reasoning on the input feature vector according to the splitting rules of each node. At each non-leaf node, it selects a branch path based on the comparison between the false negative rate or false positive rate and the corresponding splitting threshold until it reaches the leaf node.

[0046] Extract the pre-stored comprehensive misjudgment probability value from the leaf nodes, and output the first comprehensive misjudgment probability corresponding to the selection of the rate of change fluctuation judgment method and the second comprehensive misjudgment probability corresponding to the selection of the normal weighted neural network judgment method.

[0047] Preferably, the secondary confirmation of low-dimensional feature data through the rate of change fluctuation judgment method includes:

[0048] Based on the low-dimensional feature data of the corresponding wind turbine structure category, low-dimensional feature vectors are extracted for the current time, the previous time, and the second time before that.

[0049] Based on the low-dimensional feature vector at the current moment and the low-dimensional feature vector at the previous moment, the relative change of the current moment relative to the previous moment is calculated in each dimension, and all relative changes are summed to obtain the first rate of change; using the same calculation method, based on the low-dimensional feature vector at the previous moment and the low-dimensional feature vector at the second preceding moment, the second rate of change is calculated.

[0050] Calculate the absolute value of the difference between the first rate of change and the second rate of change, and use it as the rate of change fluctuation value; compare the rate of change fluctuation value with a preset fluctuation threshold.

[0051] If the rate of change fluctuation is less than the preset fluctuation threshold, the fan status is determined to be normal and the initial abnormal alarm event is a false alarm; if the rate of change fluctuation is greater than or equal to the preset fluctuation threshold, the fan is determined to have a real abnormality.

[0052] Preferably, the secondary confirmation of low-dimensional feature data using a normal weighted neural network includes:

[0053] Define a standard normal function for calculating time decay weights;

[0054] Based on the preset time window duration and the data collection time interval for a single set, determine the historical time window range corresponding to the current moment, as well as the historical time sequence within that window that is used for calculation.

[0055] Based on the low-dimensional feature data of the corresponding wind turbine structure category, extract the low-dimensional feature data of the current moment and the low-dimensional feature data corresponding to each historical moment in the historical time sequence;

[0056] The time decay weight of each historical moment relative to the current moment is calculated using the standard normal function, and the time decay weight of all historical moments is normalized to obtain the normalized weighted weight sequence.

[0057] Using the normalized weighted sequence, the low-dimensional feature data corresponding to each historical moment in the historical time sequence are weighted and summed to obtain the time-series weighted feature vector of the current moment.

[0058] The time-weighted feature vector is input into a pre-trained neural network model, and the neural network model outputs the anomaly detection result at the current time.

[0059] If the anomaly detection result indicates that the fan is in normal condition, the initial anomaly alarm event is determined to be a false alarm; if the anomaly detection result indicates that the fan is in abnormal condition, the fan is determined to have a real anomaly.

[0060] According to a second aspect of the present invention, a system for determining abnormal states of wind turbines based on unsupervised learning is provided, the system comprising:

[0061] The data classification and processing module is used to acquire sample data under normal operating conditions of the wind turbine, and to classify and standardize the sample dataset based on the structural characteristics of the wind turbine to generate standardized operating data corresponding to each structural category.

[0062] The data dimensionality reduction module is used to identify the wind turbine structure category corresponding to the initial abnormal alarm event when the wind turbine is running in real time and the wind turbine's built-in normal judgment logic triggers the initial abnormal alarm event. It then uses unsupervised deep learning technology to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure category to obtain the dimensionality-reduced low-dimensional feature data.

[0063] The wind turbine anomaly secondary confirmation module is used to perform secondary confirmation on low-dimensional feature data through preset anomaly judgment logic. If the secondary confirmation result determines that there is a real anomaly in the wind turbine, the initial anomaly alarm event information is reported. Otherwise, the initial anomaly alarm event is determined to be a false alarm and is automatically filtered. The anomaly judgment logic adopts a rate of change fluctuation judgment method or a normal weighted neural network judgment method.

[0064] The beneficial effects of this invention are as follows:

[0065] 1. This invention targets wind turbine application scenarios, classifies data according to wind turbine characteristics, processes the classified data through deep learning to obtain a dimensionality reduction transformation matrix, and defines a rate of change based on the dimensionality reduction data. The rate of change over time is used to assist in judging the abnormal state of the wind turbine. This invention has strong versatility and low barriers to entry. The dimensionality reduction transformation matrix can be obtained by pre-training with a large amount of data, without requiring engineers to have rich engineering experience and mathematical foundation. It also requires less computing resources and can be adapted to be deployed in the PLC of wind turbines. Its anomaly judgment process only requires one matrix multiplication and two rate of change calculations, which has the technical advantages of high efficiency, lightweight and easy implementation.

[0066] 2. This invention has strong versatility. Since the deep learning process only requires data for training and does not require professional personnel to perform complex parameter tuning, an effective dimensionality reduction model can be generated as long as there is sufficient wind turbine operating data. At the same time, the dimensionality reduction process of this invention involves simple and limited matrix calculations, which is more suitable for industrial environments with limited computing resources such as PLCs compared to traditional deep learning methods. Moreover, this invention only requires the operating data of a single wind turbine to complete the anomaly judgment, with low requirements for memory resources. It can be directly integrated and deployed in the PLC main control program of the wind turbine, which greatly improves the engineering practicality and ease of implementation of the solution. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart of a method for determining abnormal wind turbine states based on unsupervised learning according to an embodiment of the present invention;

[0069] Figure 2 This is one of the schematic diagrams of a neural network in a method for judging abnormal wind turbine states based on unsupervised learning according to an embodiment of the present invention;

[0070] Figure 3 This is the second schematic diagram of the neural network in a method for judging abnormal wind turbine states based on unsupervised learning according to an embodiment of the present invention.

[0071] Figure 4 This is the third schematic diagram of the neural network in a method for judging abnormal wind turbine states based on unsupervised learning according to an embodiment of the present invention.

[0072] Figure 5This is a flowchart of the normal weighted neural network judgment method in a method for judging abnormal wind turbine states based on unsupervised learning according to an embodiment of the present invention. Detailed Implementation

[0073] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0074] According to an embodiment of the present invention, a method for judging abnormal states of wind turbines based on unsupervised learning is provided.

[0075] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for determining abnormal wind turbine states based on unsupervised learning includes:

[0076] S1. Obtain sample data under normal operating conditions of the wind turbine, and perform classification and standardization preprocessing on the sample dataset based on the structural characteristics of the wind turbine to generate standardized operating data corresponding to each structural category.

[0077] Specifically, based on the structural characteristics of the wind turbine, including yaw, pitch, generator, gearbox, converter, power grid, main shaft hydraulics, and nacelle / tower base, the wind turbine sample data is categorized into yaw data, pitch data, generator data, gearbox data, converter data, power grid data, main shaft hydraulics data, and nacelle / tower base data. This classification method is purely based on engineering experience, primarily distinguishing based on the structural components of the wind turbine. For each category of data, related data are combined based on engineering experience. For example, generator data has a strong correlation with yaw and pitch data. The combination method refers to the specific structure of the wind turbine, mainly combining data from related components. Combining unrelated data only affects the training process and slightly increases the computational load, without affecting the actual results.

[0078] S2. When the wind turbine is running in real time, if the wind turbine's built-in normal judgment logic triggers an initial abnormal alarm event, the wind turbine structure category corresponding to the initial abnormal alarm event is identified, and unsupervised deep learning technology is used to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure category to obtain dimensionality-reduced low-dimensional feature data.

[0079] Among them, unsupervised deep learning technology is used to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure categories, and the resulting low-dimensional feature data includes:

[0080] Standardized operating data for the corresponding wind turbine structure category is extracted, and the operating parameters at each data acquisition time are combined into a high-dimensional operating feature vector. The high-dimensional operating feature vectors at all times are used to form a high-dimensional feature vector set for unsupervised deep learning training.

[0081] The high-dimensional feature vector set is input into a linear autoencoder network to train and obtain the dimensionality reduction transformation matrix and inverse transformation matrix corresponding to the current dimension. The dimensionality reduction transformation matrix and inverse transformation matrix are combined to generate a low-dimensional feature vector set and verify the effectiveness of dimensionality reduction.

[0082] The process involves inputting a high-dimensional feature vector set into a linear autoencoder network, training it to obtain the dimensionality reduction transformation matrix and inverse transformation matrix corresponding to the current dimension, combining the dimensionality reduction transformation matrix and inverse transformation matrix to generate a low-dimensional feature vector set, and verifying the effectiveness of dimensionality reduction.

[0083] Define a linear autoencoder network that includes an encoder and a decoder, and input the high-dimensional feature vector set of the current dimension into the linear autoencoder network to perform an unsupervised training process;

[0084] In each round of training, the dimensionality reduction transformation matrix and the inverse transformation matrix corresponding to the current dimension are obtained; the encoder compresses the input vector set of the current dimension using the dimensionality reduction transformation matrix and outputs the dimensionality-reduced low-dimensional feature vector set; the decoder reconstructs the dimensionality-reduced low-dimensional feature vector set back to the original dimension using the inverse transformation matrix and calculates the reconstruction error.

[0085] The reconstruction error is calculated to verify the effectiveness of dimensionality reduction. If the reconstruction error is less than the preset engineering threshold, the dimensionality reduction is considered successful, the current low-dimensional feature vector set is retained, and the current low-dimensional feature vector set is used as the input for the next round of training. If the reconstruction error is greater than or equal to the preset engineering threshold, the dimensionality reduction is considered unsuccessful, and the dimensionality reduction iteration process is terminated.

[0086] The validated low-dimensional feature vector set is used as the high-dimensional input feature vector set for the new round of training. The single-level training and validation process is repeated to obtain the intermediate dimensionality reduction transformation matrix corresponding to each dimension level. The iteration process is terminated when the reconstruction error can no longer meet the preset engineering threshold.

[0087] The intermediate dimensionality reduction transformation matrices corresponding to each dimension level are multiplied sequentially according to the training order to obtain the final dimensionality reduction transformation matrix of the corresponding wind turbine structure category, and at the same time, the low-dimensional feature data of the corresponding wind turbine structure category are obtained.

[0088] Specifically, the present invention is an auxiliary method. After a wind turbine generates an alarm according to the normal judgment logic, this method performs secondary judgment for confirmation. The present invention is regarded as an auxiliary method rather than a main judgment means, which mainly takes into account the reduction of computational complexity, so that this method can be directly applied to the main control program of the wind turbine. The PLC program is a cyclic task with a relatively short cycle, and the single-cycle loop of the abnormal alarm task of a general wind turbine needs to be controlled within 20ms.

[0089] Since traditional deep learning and neural networks need to be processed through a nonlinear transfer function after weighting each neuron node, the present invention, which is positioned as an auxiliary judgment method, can tolerate a slightly larger error than conventional methods. Therefore, the nonlinear transfer function is removed.

[0090] After the nonlinear transfer function is removed, the dimensionality reduction process of deep learning is only a matrix calculation. For example, if 100 pieces of data are finally reduced to 80 pieces of data, the dimensionality reduction process of deep learning will generate 20 dimensionality reduction matrices, which are respectively a 100*99 matrix for reducing the data to 99 dimensions, a 99*98 matrix for reducing the data to 98 dimensions, and so on. Since each dimensionality reduction needs to go through a nonlinear transfer function, the traditional deep learning method cannot simplify this calculation process, and 20 matrix operations and 20 nonlinear transfer functions are required. After the nonlinear transfer function is removed, the dimensionality reduction process becomes a simple matrix multiplication, so the 20 matrices can be directly multiplied to form a 100*80 matrix. The actual dimensionality reduction only needs to multiply the original 100 pieces of data with this 100*80 matrix. The dimensionality reduction process of deep learning specifically includes:

[0091] Considering a set of data of a wind turbine, such as wind speed, wind direction, active power, reactive power, etc., such a set of data generated at each moment is regarded as an n-dimensional vector, that is, (wind speed, wind direction, active power, reactive power, ...), where n represents the number of this set of data. There is a correlation between these data, that is, n pieces of data can be obtained by calculation from m pieces of data, where m<n, and this calculation process can be obtained through an m*n transformation matrix, that is:

[0092] (A1, A2, ..., An)*T=(B1, B2, B3, ..., Bm);

[0093] (B1, B2, B3, ..., Bm)*T'=(A1, A2, ..., An);

[0094] Wherein, T is an m*n matrix, T' is an n*m matrix, and T*T' is an identity matrix.

[0095] However, in practical engineering applications, due to errors such as data jitter, it is difficult to calculate matrix T using traditional mathematical methods. The ultimate goal of deep learning is to calculate matrix T using a large amount of data. The process of deep learning is to reduce the dimensionality step by step. The following describes this process with specific examples.

[0096] Suppose we have a set of wind turbine data with 100 elements, i.e., n=100. First, we attempt to reduce this 100-dimensional data to 99 dimensions, i.e., find a 99*100 matrix T1 such that (A1, A2, ..., A100)*T1 = (B1, B2, B3, ..., B99). Simultaneously, we find a 100*99 matrix T1' such that (B1, B2, B3, ..., B99)*T1' = (A1, A2, ..., A100). This constructs a matrix as follows: Figure 2 The neural network shown.

[0097] If T1 can be found, then this 100-dimensional vector (A1, A2, ..., A100) can be represented by a 99-dimensional vector (B1, B2, B3, ..., B99). This process is then repeated to attempt to represent the 99-dimensional vector (B1, B2, B3, ..., B99) using a 98-dimensional vector, thus constructing a representation as follows: Figure 3 The neural network shown.

[0098] Repeat the above process until no Ts matrix can be found. Multiply the T1, T2, ..., Ts matrices found in the above process to obtain the final transformation matrix T.

[0099] The methods for finding Ts include:

[0100] A large amount of normal wind turbine data (samples) needs to be prepared, and intermediate values ​​(T1, T2, ... Ts) are obtained through neural network training. Due to the presence of errors and fluctuations in the data, it is impossible to find a perfectly matching (T1, T2, ... Ts), so an allowable error needs to be set during training.

[0101] like Figure 4 As shown, let vector A be (A1, A2, ..., An), and vector A' be (A1', A2', ..., An'). Define the error Diff = |A1 - A1'| / A1 + |A2 - A2'| / A2 + ... + |An - An'| / An, and define a threshold (because normalization has been performed, based on engineering experience, the error for each data point can be defined as 0.01, therefore the total error is 0.01*n). If a transformation matrix is ​​found such that all data in the sample satisfy the error Diff being less than the threshold, then dimensionality reduction is possible; otherwise, dimensionality reduction is not possible.

[0102] S3. The low-dimensional feature data is reconfirmed through the preset anomaly judgment logic. If the reconfirmation result determines that there is a real anomaly in the wind turbine, the initial anomaly alarm event information is reported. Otherwise, the initial anomaly alarm event is determined to be a false alarm and is automatically filtered. The anomaly judgment logic adopts the rate of change fluctuation judgment method or the normal weighted neural network judgment method.

[0103] By dynamically selecting either the rate of change fluctuation judgment method or the normal weighted neural network judgment method for secondary confirmation of wind turbine anomalies, this invention significantly reduces the risk of false alarms and missed alarms caused by a single judgment logic, minimizing the cost of misjudgment. This method first identifies the anomaly type based on the instantaneous information entropy of low-dimensional feature data, and combines this with the historical false alarm rate as a priori probability. Simultaneously, it analyzes the performance differences between the two judgment methods under the current anomaly type, such as the missed alarm rate, false alarm rate, and response delay, constructing a set of misjudgment cost parameters for different scenarios. Then, it calculates the expected misjudgment loss for each method and selects the optimal method for secondary confirmation. This adaptive decision-making mechanism not only improves the accuracy and robustness of anomaly judgment but also dynamically adjusts the judgment strategy based on the actual fault evolution results, effectively avoiding cascading losses such as ineffective inspections and equipment damage, significantly optimizing the intelligence level and economic benefits of wind farm operation and maintenance.

[0104] Specifically, a pre-defined anomaly detection logic is used to perform secondary verification on low-dimensional feature data. If the secondary verification determines that the wind turbine has a genuine anomaly, an initial anomaly alarm event is reported; otherwise, the initial anomaly alarm event is determined to be a false alarm and is automatically filtered, including:

[0105] Define anomaly detection logic, which adopts either the rate of change fluctuation judgment method or the normal weighted neural network judgment method.

[0106] It should be noted that the anomaly detection logic completes all construction and verification work in the offline stage. The fluctuation threshold of the rate of change fluctuation detection method and the network structure and weight parameters of the normal weighted neural network detection method are all pre-calibrated and stored. In real-time operation, only the pre-stored logic modules need to be called.

[0107] Calculate the instantaneous information entropy of the low-dimensional feature data at the current moment, and determine the anomaly type corresponding to the initial anomaly alarm event based on the numerical characteristics of the instantaneous information entropy.

[0108] Specifically, the calculation process of instantaneous information entropy is as follows: First, the low-dimensional feature data (e.g., values ​​in multiple dimensions) at the current moment is divided into several equally wide intervals according to their value range. The frequency of the values ​​in each interval of each dimension is counted, and the information entropy of the entire feature vector is calculated using these frequencies. For example, for each dimension of data, its value fluctuation range is divided into 10 intervals, and the interval in which the data of that dimension falls at the current moment is recorded. Then, a set of interval distributions is obtained by combining all dimensions, and the entropy value is calculated according to the basic definition of information entropy (i.e., the sum of negative frequencies multiplied by logarithmic frequencies).

[0109] Based on common wind turbine anomaly patterns, two empirical thresholds can be preset: when the entropy value is too low, it indicates that the data changes very little or even stagnates, corresponding to sensor stagnation or data freezing anomalies; when the entropy value is in the middle range, it corresponds to normal fluctuations or sudden wind speed changes; when the entropy value is too high, it corresponds to random interference or signal noise anomalies. The low threshold is set to 95% of the normal data entropy values ​​in the normal operating entropy value distribution being higher than this value; values ​​below this value are judged as data freezing or saturation. The high threshold is set to 95% of the normal data entropy values ​​in the normal operating entropy value distribution being lower than this value; values ​​above this value are judged as random disturbances.

[0110] Based on the error cost parameters of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, the expected error loss of the two judgment methods is calculated respectively.

[0111] Among them, the misjudgment cost parameters of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type are calculated respectively, and the expected misjudgment loss of the two judgment methods includes:

[0112] Obtain the historical false alarm rate statistics of the current anomaly type in the same type of wind turbine structure from the target database, and use it as the prior probability;

[0113] Analyze the performance differences between the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, and set scenario-specific misjudgment cost parameter sets for the two judgment methods based on the performance differences, specifically including:

[0114] The false negative rate, false positive rate, and response delay of the change rate fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type are statistically analyzed to form a set of performance indicators for different scenarios.

[0115] Specifically, sample segments clearly marked as real anomalies and normal conditions are extracted from historical operational data of wind farms. Real anomaly samples require at least 50 independent events (at different times and on different wind turbines), while normal samples require at least 200 consecutive periods without anomalies. Each sample segment contains low-dimensional feature data for 30 seconds before and after triggering the initial alarm.

[0116] Based on the operating conditions of the wind turbine and the external environment, the following three typical scenarios are identified:

[0117] Scenario A: Stable operation near rated power, for example, power fluctuation less than 5% of the rated value, and wind speed change rate less than 0.5 m / s. 2 .

[0118] Scenario B: Variable power adjustment process, for example, the active power command changes by more than 10% of the rated value, and the change lasts for more than 5 seconds.

[0119] Scenario C: Drastic wind speed fluctuations, for example, the average wind speed changes by more than 3 m / s within 10 seconds.

[0120] Statistical methods include:

[0121] False negative rate: The proportion of normal samples among real abnormal samples.

[0122] False alarm rate: The proportion of abnormal outputs in normal samples.

[0123] Response latency: The average time taken from inputting low-dimensional feature data to outputting the judgment result. Each judgment method is measured 100 times and the average value is taken.

[0124] Results Record: Form a table. For example, in scenario A, the false alarm rate is 5% and the false alarm rate is 8% for the rate of change fluctuation method, with a delay of 2ms; the false alarm rate is 2% and the false alarm rate is 3% for the normal weighted neural network method, with a delay of 15ms, and so on.

[0125] Based on the actual loss data caused by the current anomaly type in the historical operation and maintenance records, a unit loss benchmark is defined and a unit loss benchmark table is formed. The unit loss mechanism includes unit loss due to equipment damage, unit loss due to downtime operation and maintenance, unit loss due to safety risks, and unit loss due to ineffective inspections.

[0126] Specifically, the unit loss for equipment damage is calculated as follows: Based on statistics from the past three years, cases where equipment damage ultimately occurred due to this type of anomaly, such as unreported abnormal gearbox oil temperature, are analyzed. The average repair or replacement cost is calculated, including spare parts, labor, and hoisting costs. If no damage cases exist, the loss is calculated based on the estimated lifespan of the equipment. For example, the total lifespan cost of the gearbox is divided by the expected lifespan in days and then multiplied by the average number of days before premature failure.

[0127] Losses for maintenance units due to downtime: Calculate the average downtime (in hours) caused by this type of anomaly, multiply by the rated power of the wind turbine, and then multiply by the grid electricity price.

[0128] Unit loss due to safety risk: Preset reference value based on accident level.

[0129] Losses incurred by units due to invalid inspections: The manpower, vehicle, and time costs of a single invalid on-site inspection, calculated based on the local average maintenance cost.

[0130] For each type of anomaly, such as abnormal generator bearing temperature change rate, a baseline loss table should be established. For example: equipment damage loss is 20,000 yuan / time, downtime maintenance loss is 4,000 yuan / hour (indirect losses, but underreporting may result in multiple hours of downtime, which needs to be considered in conjunction with the duration), safety risk loss is 0, and ineffective inspection loss is 1,000 yuan / time. Note: Downtime maintenance losses are usually calculated by multiplying the hourly rate by the unit power generation loss, rather than being a fixed value.

[0131] The time series features of low-dimensional feature data at the current moment are obtained. Based on the performance index set of different scenarios and the unit loss benchmark table, and combined with the fault evolution results reflected by the time series features, the correlation between fault detection timing and loss degree is dynamically quantified.

[0132] Specifically, the system acquires low-dimensional feature data from a preset number of sampling points prior to the current moment (the number of sampling points is determined based on the data acquisition interval and the required time window, such as backtracking at a fixed duration), and extracts three types of time series features from them: trend features, used to determine whether the value has been continuously rising, continuously falling, or oscillating in recent moments, which can be obtained by comparing the direction of the mean change between the current moment and the previous few moments); fluctuation amplitude features, which is the ratio of the difference between the maximum and minimum values ​​of the feature values ​​in the previous short period to the average fluctuation amplitude under normal operating conditions; and abrupt change features, used to detect whether the rate of change of adjacent sampling points significantly exceeds the normal fluctuation range, such as exceeding a set multiple of the upper limit of normal fluctuation.

[0133] Based on the above characteristics, the current abnormal development stage is determined as an early stage, an acceleration stage, or a critical stage. The specific determination rules are as follows: if the characteristic value has just deviated, the fluctuation amplitude is small, and the trend is not obvious, it is classified as an early stage; if the characteristic value continues to change in the same direction and the rate of change gradually increases (for example, more than three of the last five sampling points have changed in the same direction continuously and the amount of change is increasing), it is classified as an acceleration stage; if the current characteristic value is close to or exceeds the pre-statistical historical fault threshold (for example, exceeding the set multiple of the normal operation upper limit), and the fluctuation is violent, it is classified as a critical stage.

[0134] Establish a correlation between the timing of fault detection and the degree of loss. The basic principle is: the earlier the fault is detected (e.g., intervention in the early stages), the smaller the loss; the later the detection (e.g., intervention in the critical stage), the greater the loss. Specifically, in quantification, the combination of the delay time output by the judgment method and the fault evolution stage is used as the metric. For the same abnormal event, let the actual fault initiation time be... The time when the judgment method outputs an abnormality is Then a delay was found. However, it is difficult to obtain the exact value in practice. Therefore, the failure evolution stage (early / accelerated / critical) is used as a proxy indicator for the detection timing, where the early stage corresponds to high timeliness and the critical stage corresponds to low timeliness.

[0135] Based on historical operation and maintenance records, the average total loss generated when the same anomaly type was discovered at different evolution stages is statistically analyzed, covering equipment damage losses, downtime operation and maintenance losses, and security risk losses, thereby obtaining a quantitative relationship between loss and discovery delay. Simultaneously, by combining data from the aforementioned performance indicator set for different scenarios and the unit loss benchmark table, this quantitative relationship is adapted to different scenarios, ultimately achieving dynamic quantification of the correlation between the timing of fault discovery and the degree of loss.

[0136] Based on the correlation obtained by dynamic quantization, each performance indicator in the scenario-specific performance index set is multiplied by its corresponding unit loss, and the calculation results are weighted and corrected according to the fault evolution results to obtain the false negative cost parameters and false positive cost parameters of the two judgment methods under the current anomaly type. The false negative cost parameters and false positive cost parameters are integrated to obtain the scenario-specific misjudgment cost parameter set of the two judgment methods, specifically including:

[0137] Step 1: Obtain performance metrics and unit loss benchmarks;

[0138] For the current anomaly type and current operating scenario, extract the false negative rate and false positive rate of each of the two judgment methods from the established performance indicators for each scenario. At the same time, find the corresponding unit loss from the unit loss benchmark table, including equipment damage loss, downtime maintenance loss, security risk loss used for calculating the cost of false negatives, and invalid inspection loss used for calculating the cost of false positives.

[0139] Step 2: Determine the fault evolution stage and weighted correction coefficients;

[0140] Based on the low-dimensional feature data at the current moment, the evolution stage of the fault (early stage, acceleration stage, or critical stage) is determined according to the time series characteristics. A corresponding weighted correction coefficient is determined based on the stage: the loss amplification is relatively low in the early stage, intermediate in the acceleration stage, and highest in the critical stage. This weighted correction coefficient is used to scale the cost of missed detections.

[0141] Step 3: Calculate the cost parameters for underreporting;

[0142] For each judgment method, perform the following operations: multiply the false negative rate of the judgment method by the equipment damage loss, downtime maintenance loss, and safety risk loss to obtain three preliminary false negative costs; add the three preliminary false negative costs to obtain the preliminary total false negative cost; multiply the preliminary total false negative cost by the weighted correction coefficient corresponding to the fault evolution stage to obtain the false negative cost parameter of the judgment method under the current anomaly type.

[0143] Step 4: Calculate the false alarm cost parameters;

[0144] The calculation of the false alarm cost parameter does not consider the fault evolution stage. Instead, the false alarm rate of the judgment method is directly multiplied by the invalid inspection loss to obtain the false alarm cost parameter.

[0145] Step 5: Integrate the set of parameters for misjudgment costs in different scenarios;

[0146] Combining the false positive cost parameters and the missed false positive cost parameters calculated under the same judgment method constitutes the scenario-specific false positive cost parameter set for this judgment method under the current anomaly type and current scenario. Steps three to five are executed for the rate of change fluctuation judgment method and the normal weighted neural network judgment method respectively, resulting in two complete parameter sets for subsequent minimum risk decision selection. Using the prior probability and the scenario-specific false positive cost parameter set as input, the minimum risk decision algorithm is used to calculate the expected false positive loss for selecting the rate of change fluctuation judgment method or the normal weighted neural network judgment method under the current anomaly type.

[0147] Using prior probabilities and scenario-specific misjudgment cost parameter sets as input, the minimum risk decision algorithm is used to calculate the expected false alarm loss for choosing either the rate of change fluctuation judgment method or the normal weighted neural network judgment method under the current anomaly type.

[0148] The prior probabilities are mapped to the initial class weights of the decision tree nodes, and the misjudgment cost parameter set for each scenario is used as the misjudgment cost coefficient of the decision tree loss function. An initial decision tree structure is constructed that includes two decision branches: judgment of the rate of change fluctuation and judgment of the normal weighted neural network.

[0149] Specifically, during training, the decision tree calculates the difference between the predicted class and the true class based on the splitting result of each node. If the true class is an anomaly and the prediction is normal (i.e., a false negative occurs), the loss is calculated using the false negative cost parameter. Conversely, if the true class is normal and the prediction is an anomaly (i.e., a false positive occurs), the loss is calculated using the false positive cost parameter. Therefore, during the training phase, the prediction result is a simulation output of the decision tree based on historical samples.

[0150] Mapping prior probabilities to initial class weights for decision tree nodes includes: the prior probability, which is the historical false positive rate of the current anomaly type (between 0 and 1). When mapping this to the initial class weights of the decision tree root node, the following method is used:

[0151] Let the prior probability be... Then the initial weight of the true anomaly category in the root node is set to (Or set according to the actual category distribution), while the initial weight of the "false alarm" category is set to .

[0152] If a decision tree uses a weighted loss function, the weight of each sample is initialized according to the reciprocal of the prior probability of its class, giving higher weights to the minority class. In practice, class weight can be expressed as: total number of samples / (number of classes × number of samples in that class) × normalization factor. However, for simplicity, we directly use... As a weighting coefficient for false alarm categories.

[0153] The misjudgment cost coefficients, which are used as the set of misjudgment cost parameters for different scenarios in the decision tree loss function, include:

[0154] The scenario-specific misclassification cost parameter set includes both false negative and false positive cost parameters. During decision tree training, the loss function must consider not only the number of classification errors but also the different costs associated with different error types. Specific implementation details include:

[0155] The cost of a true anomaly but a predicted normal result is set as the false negative cost parameter; the cost of a true normal result but a predicted anomaly is set as the false positive cost parameter. When calculating the reduction in impurity after a split, each node in the decision tree does not directly use the classification error rate, but instead uses a weighted cost. For example, cost-sensitive entropy is used, with the error cost of each sample included as a weight in the split evaluation metric.

[0156] The initial decision tree structure is used as the optimization object, with the objective function being to minimize the expected misjudgment loss. An adaptive optimization algorithm is used to perform global pruning optimization on the initial decision tree structure. The pruning process starts from the leaf nodes of the initial decision tree and traverses each leaf node layer by layer from bottom to top. At each layer, an attempt is made to prune the subtree branches under the current leaf node. After each pruning attempt, the total loss function value of the decision tree is recalculated. If the total loss function value decreases, the pruning operation is retained and the next leaf node is traversed until the root node is reached. The final decision tree structure is used as the optimized hybrid decision tree model.

[0157] Specifically, the initial decision tree is a simple binary tree, with the root node containing historical performance samples for both decision methods. Each sample consists of a feature vector (false negative rate, false positive rate, response latency) and a label (the probability that the sample is actually misclassified). The node splitting rule is as follows:

[0158] At each node, all features are traversed, including false negative rate, false positive rate, response latency, and candidate split thresholds, such as false negative rate <0.05, <0.10, etc. Split features and thresholds that minimize the cost-sensitive loss function are selected. To avoid overfitting, no depth limit is set initially for tree growth, continuing until the number of samples within a node is less than the minimum split number or all samples belong to the same cost category. The final result is a complete decision tree containing two decision branches: one for selecting fluctuations in the rate of change and the other for selecting a normally weighted neural network.

[0159] The specific implementation steps for global pruning optimization of the initial decision tree structure using an adaptive optimization algorithm include:

[0160] Step 1: Calculate the total loss of the initial decision tree;

[0161] Let the initial decision tree be For each leaf node Calculate the sum of the false negative costs and the sum of the false positive costs for samples at this node (using pre-defined cost parameters). The total loss equals the sum of the costs of all leaf nodes, and is recorded as follows: .

[0162] Step 2: Start from the bottommost leaf node and process upwards layer by layer.

[0163] Step 3: Prune all subtrees under the current internal node, turning it into a leaf node. After pruning, recalculate the cost of the new leaf node: For all training samples falling into this node, count the number of missed and false positives, multiply them by the corresponding missed and false positive cost parameters respectively, and sum them to obtain the cost of the node. Calculate the total loss of the entire decision tree after pruning. .

[0164] Step 4: Evaluate the benefits of pruning;

[0165] if Then accept this pruning and update. And continue traversing upwards through the parent nodes.

[0166] if If the pruning is abandoned, the traversal continues to the next internal node on the same level.

[0167] Step 5: Since pruning may affect the pruning effect of higher-level nodes, after completing one bottom-up traversal, repeat steps 2 to 4 until no pruning operation is accepted in two consecutive traversals, or the preset maximum number of iterations is reached. The maximum number of iterations is preferably 5. The final decision tree is the optimal structure and is used for subsequent reasoning.

[0168] The performance index set of the two judgment methods under the current anomaly type is used as the input feature of the hybrid decision tree model. Through node splitting and classification reasoning of the hybrid decision tree model, the comprehensive misjudgment probability of choosing the rate of change fluctuation judgment method and the comprehensive misjudgment probability of choosing the normal weighted neural network judgment method are output respectively.

[0169] Specifically, through node splitting and classification reasoning in the hybrid decision tree model, the comprehensive misjudgment probability of choosing the rate of change fluctuation judgment method and the comprehensive misjudgment probability of choosing the normal weighted neural network judgment method are output respectively, including:

[0170] Obtain the false negative rate and false positive rate of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, combine the false negative rate and false positive rate into an input feature vector, and input it into the root node of the hybrid decision tree model;

[0171] The hybrid decision tree model performs layer-by-layer classification reasoning on the input feature vector according to the splitting rules of each node. At each non-leaf node, it selects a branch path based on the comparison between the false negative rate or false positive rate and the corresponding splitting threshold until it reaches the leaf node.

[0172] Extract the pre-stored comprehensive misjudgment probability value from the leaf nodes, and output the first comprehensive misjudgment probability (the probability of false alarm and the probability of false alarm when selecting the rate of change fluctuation judgment method) and the second comprehensive misjudgment probability (the probability of false alarm and the probability of false alarm when selecting the normal weighted neural network judgment method) corresponding to the selection of the rate of change fluctuation judgment method.

[0173] By combining the combined false positive probability of the two judgment methods with the prior probability of the current anomaly type, the expected loss of underreporting and the expected loss of false positive for the two judgment methods are calculated and summed to obtain the total expected false positive loss of the rate of change fluctuation judgment method and the total expected false positive loss of the normal weighted neural network judgment method.

[0174] It should be noted that the total expected false alarm loss of the rate of change fluctuation judgment method and the total expected false alarm loss of the normal weighted neural network judgment method include:

[0175] Step 1: For the current anomaly type and the current scenario, both the rate of change fluctuation judgment method and the normal weighted neural network judgment method have a set of performance indicators, including false negative rate, false positive rate, and response latency. These indicators are merged into a fixed-length feature vector, for example: [False Negative Rate_Rate of Change, False Positive Rate_Rate of Change, Latency_Rate of Change, False Negative Rate_Neural Network, False Positive Rate_Neural Network, Latency_Neural Network]. This feature vector is then input into the optimized hybrid decision tree model.

[0176] Step 2: Input the above feature vector into the root node of the hybrid decision tree model; according to the pre-stored splitting rules (split features, splitting threshold, comparison direction) of each non-leaf node, compare the value of the corresponding feature in the current feature vector with the threshold, and select the left branch or the right branch; repeat the above process until the leaf node is reached.

[0177] Step 3: The leaf nodes pre-store not a single overall probability, but four probability values:

[0178] : The probability of missed reports when selecting the method for judging rate of change fluctuation;

[0179] : The probability of false alarms when selecting a method to judge rate of change fluctuation;

[0180] : The probability of false negatives when choosing the normal weighted neural network as the judgment method;

[0181] : The false alarm probability when choosing the normal weighted neural network judgment method;

[0182] Output respectively:

[0183] Group 1 (corresponding to the method for judging rate of change fluctuation): Probability of underreporting False alarm probability ;

[0184] The second group (corresponding to the judgment method of normal weighted neural network): false negative probability False alarm probability ;

[0185] Step 4: Obtain the prior probability of the current exception type Parameters for the underreporting cost of obtaining the rate of change fluctuation judgment method The cost parameter of false negatives in the normal weighted neural network judgment method ; False alarm cost parameters of the rate of change fluctuation judgment method and the normal weighted neural network judgment method and Calculate the total expected loss for both judgment methods:

[0186] Methods for judging rate of change fluctuation:

[0187] ;

[0188] Normal weighted neural network judgment method:

[0189] ;

[0190] Step 5: Select the current anomaly detection method:

[0191] like If so, then the normal weighted neural network judgment method is selected.

[0192] like If so, then choose the rate of change fluctuation judgment method.

[0193] like You can choose any of them.

[0194] If the expected misjudgment loss of the rate of change fluctuation judgment method is greater than the expected misjudgment loss of the normal weighted neural network judgment method, then the normal weighted neural network judgment method will be used as the current anomaly judgment method, and the low-dimensional feature data will be reconfirmed through the normal weighted neural network judgment method.

[0195] Among these, the secondary verification of low-dimensional feature data using a normal weighted neural network includes:

[0196] Define a standard normal function for calculating time decay weights;

[0197] Based on the preset time window duration and the data collection time interval for a single set, determine the historical time window range corresponding to the current moment, as well as the historical time sequence within that window that is used for calculation.

[0198] Based on the low-dimensional feature data of the corresponding wind turbine structure category, extract the low-dimensional feature data of the current moment and the low-dimensional feature data corresponding to each historical moment in the historical time sequence;

[0199] The time decay weight of each historical moment relative to the current moment is calculated using the standard normal function, and the time decay weight of all historical moments is normalized to obtain the normalized weighted weight sequence.

[0200] Using the normalized weighted sequence, the low-dimensional feature data corresponding to each historical moment in the historical time sequence are weighted and summed to obtain the time-series weighted feature vector of the current moment.

[0201] The time-weighted feature vector is input into a pre-trained neural network model, and the neural network model outputs the anomaly detection result at the current time.

[0202] If the anomaly detection result indicates that the fan is in normal condition, the initial anomaly alarm event is determined to be a false alarm; if the anomaly detection result indicates that the fan is in abnormal condition, the fan is determined to have a real anomaly.

[0203] If the expected misjudgment loss of the normal weighted neural network judgment method is greater than the expected misjudgment loss of the rate of change fluctuation judgment method, then the rate of change fluctuation judgment method will be used as the current anomaly judgment method, and the low-dimensional feature data will be reconfirmed through the rate of change fluctuation judgment method.

[0204] Specifically, when the master station or slave station reports an anomaly through general means, it is necessary to perform secondary confirmation according to the rate of change fluctuation judgment method or the normal weighted neural network judgment method.

[0205] Among them, such as Figure 5 As shown, the judgment methods of the normal weighted neural network include:

[0206] Let the data after dimensionality reduction via deep learning at each time t be (Z(t)1, Z(t)2, ..., Z(t)i). We take the current time t0, the previous time t1, and the time before that t2, for a total of three sets of data. We define the change between the previous time and the current time as follows:

[0207] C=|Z(t0)1-Z(t1)1| / Z(t0)1+|Z(t0)2-Z(t1)2| / Z(t0)2+...|Z(t0)iZ(t1)i| / Z(t0)i;

[0208] Similarly, the rate of change between the previous time t1 and the time before that t2 is C1. If |C-C1| is less than the threshold, then there is no abnormality; otherwise, there is an abnormality.

[0209] Define the standard normal function as: .

[0210] Take x historical time points, i.e., (t1, t2, ..., tx). The value of x is obtained by combining the time interval and duration of each data set: x = duration / time interval of each data set, with the duration typically being 1-5 seconds. Obtain the weight of each time point (...). (t0-t1), (t0-t2), ..., (t0-tx)), and normalize the weights so that the sum of all weights is 1. The normalized weights are ( '(t0-t1), '(t0-t2),..., (t0-tx)). Based on the weighted sum, the final input of the model (C1t, C2t...Cmt) is obtained, for example:

[0211] C1t= (t0-t1)*B1t1+ (t0-t2)*B1t2+...+ (t0-tx)*B1tx;

[0212] For each wind turbine, the data is generally continuous. However, when the energy management platform controls the active and reactive power of the turbines, the data often changes abruptly. Therefore, when the active and reactive power of each turbine changes, the historical data needs to be truncated. For example, if the duration is 5 seconds, and the active power changes 1 second ago, only data from that 1 second should be used for the historical moment until the 5 seconds have elapsed and the data returns to normal. If only data from the 1 second is used, the time period x will be smaller than normal, but after weighted normalization, it will not affect the input calculation.

[0213] The model training process is a supervised learning process, as follows:

[0214] (1) Prepare a large amount of normal data (including data after controlling active power and reactive power).

[0215] (2) Dimensionality reduction is performed using deep learning methods based on normal data.

[0216] (3) For each set of data, determine x for each set of data according to the preset duration (e.g., 1 second) and the time interval between the generation of the original data, and then use the standard normal function. (z) After normalization, weighting is performed to obtain the input of the neural network model.

[0217] The neural network model is trained by inputting the neural network model, and its output is compared with the data derived from t0. The wind turbine is judged to be abnormal based on the comparison result. The specific comparison method is to compare using the Diff calculation formula, and the error is defined as 0.01*n.

[0218] It should be noted that the neural network used in this invention is consistent with the neural network structures shown in Figures 2 to 4, all of which are neural networks containing only one hidden layer. This invention uses two different neural network architectures simultaneously: one is as follows... Figures 2-4 The diagram shows a deep learning network formed by the fusion of neural networks. The neural network model here is a normal neural network, a single-hidden-layer ordinary neural network. The number of nodes in its hidden layer is typically less than the number of nodes in the input and output layers. The specific value needs to be determined based on the actual training effect. The training process uses the formula Diff=|A1-A1'| / A1+|A2-A2'| / A2+...|An-An'| / An to define the error between the actual t0 data and the t0 derived data output by the neural network.

[0219] The secondary confirmation of low-dimensional feature data through the method of judging the rate of change fluctuation includes:

[0220] Based on the low-dimensional feature data of the corresponding wind turbine structure category, low-dimensional feature vectors are extracted for the current time, the previous time, and the second time before that.

[0221] Based on the low-dimensional feature vector at the current moment and the low-dimensional feature vector at the previous moment, the relative change of the current moment relative to the previous moment is calculated in each dimension, and all relative changes are summed to obtain the first rate of change; using the same calculation method, based on the low-dimensional feature vector at the previous moment and the low-dimensional feature vector at the second preceding moment, the second rate of change is calculated.

[0222] Calculate the absolute value of the difference between the first rate of change and the second rate of change, and use it as the rate of change fluctuation value; compare the rate of change fluctuation value with a preset fluctuation threshold.

[0223] If the rate of change fluctuation is less than the preset fluctuation threshold, the fan status is determined to be normal and the initial abnormal alarm event is a false alarm; if the rate of change fluctuation is greater than or equal to the preset fluctuation threshold, the fan is determined to have a real abnormality.

[0224] If the expected misjudgment loss of the normal weighted neural network judgment method is equal to the expected misjudgment loss of the rate of change fluctuation judgment method, then any judgment method can be selected as the current anomaly judgment method.

[0225] If the secondary confirmation results indicate that the fan has a real abnormality, the initial abnormality alarm event information will be reported; otherwise, the initial abnormality alarm event will be determined to be a false alarm and automatically filtered out.

[0226] Specifically, when the main station or slave station reports an anomaly through general means, it is necessary to make a secondary confirmation based on the rate of change fluctuation judgment method or the normal weighted neural network judgment method.

[0227] The method for judging the fluctuation of the rate of change includes: Let the data after dimensionality reduction via deep learning at each time t be (Z(t)1, Z(t)2, ..., Z(t)i). Take the current time t0, the previous time t1, and the time before that t2, for a total of 3 sets of data. Define the rate of change between the previous time and the current time as C = |Z(t0)1 - Z(t1)1| / Z(t0)1 + |Z(t0)2 - Z(t1)2| / Z(t0)2 + ... + |Z(t0)iZ(t1)i| / Z(t0)i. Similarly, the rate of change between the previous time t1 and the time before that t2 is C1. If |C - C1| is less than a threshold, it is considered that there is no anomaly; otherwise, it is considered that there is an anomaly.

[0228] According to another embodiment of the present invention, a system for determining abnormal wind turbine states based on unsupervised learning is also provided, the system comprising:

[0229] The data classification and processing module is used to acquire sample data under normal operating conditions of the wind turbine, and to classify and standardize the sample dataset based on the structural characteristics of the wind turbine to generate standardized operating data corresponding to each structural category.

[0230] The data dimensionality reduction module is used to identify the wind turbine structure category corresponding to the initial abnormal alarm event when the wind turbine is running in real time and the wind turbine's built-in normal judgment logic triggers the initial abnormal alarm event. It then uses unsupervised deep learning technology to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure category to obtain the dimensionality-reduced low-dimensional feature data.

[0231] The wind turbine anomaly secondary confirmation module is used to perform secondary confirmation on low-dimensional feature data through preset anomaly judgment logic. If the secondary confirmation result determines that there is a real anomaly in the wind turbine, the initial anomaly alarm event information is reported. Otherwise, the initial anomaly alarm event is determined to be a false alarm and is automatically filtered. The anomaly judgment logic adopts the rate of change fluctuation judgment method or the normal weighted neural network judgment method.

[0232] 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 within the protection scope of the present invention.

Claims

1. A method for judging abnormal states of wind turbines based on unsupervised learning, characterized in that, The method includes: Acquire sample data under normal operating conditions of the wind turbine, and classify and standardize the sample dataset based on the structural characteristics of the wind turbine to generate standardized operating data corresponding to each structural category; When the wind turbine is running in real time, if the wind turbine's built-in normal judgment logic triggers an initial abnormal alarm event, the wind turbine structure category corresponding to the initial abnormal alarm event is identified, and unsupervised deep learning technology is used to perform dimensionality reduction processing on the standardized operating data of the corresponding wind turbine structure category to obtain dimensionality-reduced low-dimensional feature data. The low-dimensional feature data is re-verified using a pre-defined anomaly detection logic. If the re-verification determines that the wind turbine has a genuine anomaly, an initial anomaly alarm event is reported; otherwise, the initial anomaly alarm event is determined to be a false alarm and is automatically filtered out. Specifically, this includes: Define anomaly detection logic, wherein the anomaly detection logic adopts a rate of change fluctuation detection method or a normal weighted neural network detection method; Calculate the instantaneous information entropy of the low-dimensional feature data at the current moment, and determine the anomaly type corresponding to the initial anomaly alarm event based on the numerical characteristics of the instantaneous information entropy; Based on the error cost parameters of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, the expected error loss of the two judgment methods is calculated respectively. If the expected misjudgment loss of the rate of change fluctuation judgment method is greater than the expected misjudgment loss of the normal weighted neural network judgment method, then the normal weighted neural network judgment method will be used as the current anomaly judgment method, and the low-dimensional feature data will be reconfirmed through the normal weighted neural network judgment method. If the expected misjudgment loss of the normal weighted neural network judgment method is greater than the expected misjudgment loss of the rate of change fluctuation judgment method, then the rate of change fluctuation judgment method will be used as the current anomaly judgment method, and the low-dimensional feature data will be reconfirmed through the rate of change fluctuation judgment method. If the expected misjudgment loss of the normal weighted neural network judgment method is equal to the expected misjudgment loss of the rate of change fluctuation judgment method, then any judgment method can be selected as the current anomaly judgment method. If the secondary confirmation results indicate that the fan has a real abnormality, the initial abnormality alarm event information will be reported; otherwise, the initial abnormality alarm event will be determined to be a false alarm and automatically filtered out.

2. The method for judging abnormal wind turbine states based on unsupervised learning according to claim 1, characterized in that, The method of using unsupervised deep learning technology to perform dimensionality reduction on standardized operating data of corresponding wind turbine structure categories, resulting in dimensionality-reduced low-dimensional feature data, includes: Standardized operating data for the corresponding wind turbine structure category is extracted, and the operating parameters at each data acquisition time are combined into a high-dimensional operating feature vector. The high-dimensional operating feature vectors at all times are used to form a high-dimensional feature vector set for unsupervised deep learning training. The high-dimensional feature vector set is input into a linear autoencoder network to train and obtain the dimensionality reduction transformation matrix and inverse transformation matrix corresponding to the current dimension. The dimensionality reduction transformation matrix and inverse transformation matrix are combined to generate a low-dimensional feature vector set and verify the effectiveness of dimensionality reduction. The validated low-dimensional feature vector set is used as the high-dimensional input feature vector set for the new round of training. The single-level training and validation process is repeated to obtain the intermediate dimensionality reduction transformation matrix corresponding to each dimension level. The iteration process is terminated when the reconstruction error can no longer meet the preset engineering threshold. The intermediate dimensionality reduction transformation matrices corresponding to each dimension level are multiplied sequentially according to the training order to obtain the final dimensionality reduction transformation matrix of the corresponding wind turbine structure category, and at the same time, the low-dimensional feature data of the corresponding wind turbine structure category are obtained.

3. The method for judging abnormal wind turbine states based on unsupervised learning according to claim 2, characterized in that, The process of inputting a high-dimensional feature vector set into a linear autoencoder network, training it to obtain the dimensionality reduction transformation matrix and inverse transformation matrix corresponding to the current dimension, and combining the dimensionality reduction transformation matrix and inverse transformation matrix to generate a low-dimensional feature vector set and verify the effectiveness of dimensionality reduction includes: Define a linear autoencoder network that includes an encoder and a decoder, and input the high-dimensional feature vector set of the current dimension into the linear autoencoder network to perform an unsupervised training process; In each round of training, the dimensionality reduction transformation matrix and the inverse transformation matrix corresponding to the current dimension are obtained; the encoder compresses the input vector set of the current dimension using the dimensionality reduction transformation matrix and outputs the dimensionality-reduced low-dimensional feature vector set; the decoder reconstructs the dimensionality-reduced low-dimensional feature vector set back to the original dimension using the inverse transformation matrix and calculates the reconstruction error. The reconstruction error is calculated to verify the effectiveness of dimensionality reduction. If the reconstruction error is less than the preset engineering threshold, the dimensionality reduction is considered successful, the current low-dimensional feature vector set is retained, and the current low-dimensional feature vector set is used as the input for the next round of training. If the reconstruction error is greater than or equal to the preset engineering threshold, the dimensionality reduction is considered unsuccessful, and the dimensionality reduction iteration process is terminated.

4. The method for judging abnormal wind turbine states based on unsupervised learning according to claim 1, characterized in that, The misjudgment cost parameters for the rate of change fluctuation-based judgment method and the normal weighted neural network judgment method under the current anomaly type are calculated separately, including the expected misjudgment loss for each method: Obtain the historical false alarm rate statistics of the current anomaly type in the same type of wind turbine structure from the target database, and use it as the prior probability; The performance differences between the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type are analyzed, and a set of misjudgment cost parameters for the two judgment methods are set according to the performance differences. Using prior probability and scenario-specific misjudgment cost parameter sets as input, the minimum risk decision algorithm is used to calculate the expected false alarm loss for choosing either the rate of change fluctuation judgment method or the normal weighted neural network judgment method under the current anomaly type.

5. The method for judging abnormal wind turbine states based on unsupervised learning according to claim 4, characterized in that, The analysis compares the response performance differences between the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type. Based on the response performance differences, false negative and false positive cost parameters are set for the two judgment methods, and the integrated set of scenario-specific false positive cost parameters includes: The false negative rate, false positive rate and response delay of the change rate fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type are statistically analyzed to form a set of performance indicators for different scenarios. Based on the actual loss data caused by the current anomaly type in the historical operation and maintenance records, a unit loss benchmark is defined and a unit loss benchmark table is formed. The unit loss mechanism includes unit loss due to equipment damage, unit loss due to downtime operation and maintenance, unit loss due to safety risks, and unit loss due to invalid inspections. The time series features of low-dimensional feature data at the current moment are obtained. Based on the performance index set of different scenarios and the unit loss benchmark table, and combined with the fault evolution results reflected by the time series features, the correlation between fault detection timing and loss degree is dynamically quantified. Based on the correlation obtained by dynamic quantization, each performance indicator in the scenario-specific performance indicator set is multiplied by the corresponding unit loss, and the calculation results are weighted and corrected according to the fault evolution results to obtain the missed detection cost parameters and false detection cost parameters of the two judgment methods under the current anomaly type. The missed detection cost parameters and false detection cost parameters are integrated to obtain the scenario-specific misjudgment cost parameter set of the two judgment methods.

6. The method for determining abnormal wind turbine states based on unsupervised learning according to claim 5, characterized in that, The calculation of the expected false alarm loss for selecting either the rate of change fluctuation judgment method or the normal weighted neural network judgment method under the current anomaly type, using prior probability and scenario-specific misjudgment cost parameter sets as input and employing a minimum risk decision algorithm, includes: The prior probabilities are mapped to the initial class weights of the decision tree nodes, and the misjudgment cost parameter set for each scenario is used as the misjudgment cost coefficient of the decision tree loss function; an initial decision tree structure is constructed that includes two decision branches: judgment of the rate of change fluctuation and judgment of the normal weighted neural network. The initial decision tree structure is used as the optimization object, with the objective function being to minimize the expected misjudgment loss. An adaptive optimization algorithm is used to perform global pruning optimization on the initial decision tree structure. The pruning process starts from the leaf nodes of the initial decision tree and traverses each leaf node layer by layer from bottom to top. At each layer, the subtree branches under the current leaf node are attempted to be pruned. After each pruning attempt, the total loss function value of the decision tree is recalculated. If the total loss function value decreases, the pruning operation is retained and the next leaf node is traversed until the root node is reached. The final decision tree structure is used as the optimized hybrid decision tree model. The performance index set of the two judgment methods under the current anomaly type is used as the input feature of the hybrid decision tree model. Through node splitting and classification reasoning of the hybrid decision tree model, the comprehensive misjudgment probability of choosing the rate of change fluctuation judgment method and the comprehensive misjudgment probability of choosing the normal weighted neural network judgment method are output respectively. By combining the combined false positive probability of the two judgment methods with the prior probability of the current anomaly type, the expected loss of underreporting and the expected loss of false positive for the two judgment methods are calculated and summed to obtain the total expected false positive loss of the rate of change fluctuation judgment method and the total expected false positive loss of the normal weighted neural network judgment method.

7. The method for determining abnormal wind turbine states based on unsupervised learning according to claim 6, characterized in that, The method of using node splitting and classification reasoning in a hybrid decision tree model to output the combined misjudgment probability of selecting the rate of change fluctuation judgment method and the combined misjudgment probability of selecting the normal weighted neural network judgment method includes: Obtain the false negative rate and false positive rate of the rate of change fluctuation judgment method and the normal weighted neural network judgment method under the current anomaly type, combine the false negative rate and false positive rate into an input feature vector, and input it into the root node of the hybrid decision tree model; The hybrid decision tree model performs layer-by-layer classification reasoning on the input feature vector according to the splitting rules of each node. At each non-leaf node, it selects a branch path based on the comparison between the false negative rate or false positive rate and the corresponding splitting threshold until it reaches the leaf node. Extract the pre-stored comprehensive misjudgment probability value from the leaf nodes, and output the first comprehensive misjudgment probability corresponding to the selection of the rate of change fluctuation judgment method and the second comprehensive misjudgment probability corresponding to the selection of the normal weighted neural network judgment method.

8. The method for judging abnormal wind turbine states based on unsupervised learning according to claim 1, characterized in that, The secondary confirmation of low-dimensional feature data through the method of judging the rate of change fluctuation includes: Based on the low-dimensional feature data of the corresponding wind turbine structure category, low-dimensional feature vectors are extracted for the current time, the previous time, and the second time before that. Based on the low-dimensional feature vector at the current moment and the low-dimensional feature vector at the previous moment, the relative change of the current moment relative to the previous moment is calculated in each dimension, and all relative changes are summed to obtain the first rate of change; using the same calculation method, based on the low-dimensional feature vector at the previous moment and the low-dimensional feature vector at the second preceding moment, the second rate of change is calculated. Calculate the absolute value of the difference between the first rate of change and the second rate of change, and use it as the rate of change fluctuation value; compare the rate of change fluctuation value with a preset fluctuation threshold. If the rate of change fluctuation is less than the preset fluctuation threshold, the fan status is determined to be normal and the initial abnormal alarm event is a false alarm; if the rate of change fluctuation is greater than or equal to the preset fluctuation threshold, the fan is determined to have a real abnormality.

9. The method for judging abnormal wind turbine states based on unsupervised learning according to claim 8, characterized in that, The secondary confirmation of low-dimensional feature data using a normal weighted neural network includes: Define a standard normal function for calculating time decay weights; Based on the preset time window duration and the data collection time interval for a single set, determine the historical time window range corresponding to the current moment, as well as the historical time sequence within that window that is used for calculation; Based on the low-dimensional feature data of the corresponding wind turbine structure category, extract the low-dimensional feature data of the current moment and the low-dimensional feature data corresponding to each historical moment in the historical time sequence; The time decay weight of each historical moment relative to the current moment is calculated using the standard normal function, and the time decay weight of all historical moments is normalized to obtain the normalized weighted weight sequence. Using the normalized weighted sequence, the low-dimensional feature data corresponding to each historical moment in the historical time sequence are weighted and summed to obtain the time-series weighted feature vector of the current moment. The time-weighted feature vector is input into a pre-trained neural network model, and the neural network model outputs the anomaly detection result at the current time. If the anomaly detection result indicates that the fan is in normal condition, the initial anomaly alarm event is determined to be a false alarm; if the anomaly detection result indicates that the fan is in abnormal condition, the fan is determined to have a real anomaly.

Images