A wind turbine gearbox fault diagnosis method based on time-shift cosine similarity entropy

Abstract

The application discloses a wind power gear box fault diagnosis method based on time shift cosine similarity entropy, belongs to the technical field of wind turbine gear box fault diagnosis, and comprises the following steps: collecting vibration signals of a wind turbine gear box under different health states; key multi-scale features capable of reflecting different health states are mined from the vibration signals by using time shift cosine similarity entropy; the multi-scale features are proportionally divided into a training set and a test set; an intelligent fault diagnosis model based on an extreme learning machine is constructed, and the training set is input to perform model training; then, the test set is input into the trained extreme learning machine, and gear box fault diagnosis is completed. The application can accurately identify the specific fault type of the wind power gear box, provides a strong basis for discovering the position of the gear box fault and performing maintenance, and ensures reliable and stable operation of the wind turbine.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine gearbox fault diagnosis technology, and more specifically, relates to a wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy. Background Technology

[0002] In recent years, wind power technology has developed rapidly in my country, playing a vital role in optimizing the energy structure. For investors, the primary goal of wind turbines is to maximize economic benefits. Therefore, wind turbines need to meet requirements for higher yield power and less downtime. In other words, the research objective is to improve productivity and reduce operating costs. However, various faults in wind turbines have become a significant obstacle to achieving this goal. Due to their long operating hours, complex structure, and harsh working conditions, wind turbine gearbox fault diagnosis has always been a major challenge, especially in reducing fault frequency and downtime. More specifically, as a core component of wind power systems, gearboxes not only account for a large proportion of the total cost of wind turbines but are also consistently considered the component with the highest failure rate. Therefore, wind turbine gearbox fault diagnosis is an indispensable part of the further development of wind turbines. How to provide mature technologies to significantly reduce downtime and avoid catastrophic accidents is an urgent problem to be solved.

[0003] When wind turbine components malfunction, their vibration signals exhibit strong non-stationary characteristics due to the influence of surrounding noise. Conventional time-domain and frequency-domain analysis methods struggle to extract valuable fault features. Signal processing-based fault diagnosis methods, such as wavelet analysis and empirical mode decomposition, are overly complex and difficult for wind farm personnel to master. Furthermore, the analysis results are easily affected by noise frequencies, often reducing the performance of the diagnostic model on the target task. Summary of the Invention

[0004] To address the shortcomings of existing diagnostic techniques, such as insufficient utilization of signals and difficulty in effectively characterizing the dynamic evolution of faults, a wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy is proposed. By combining the rapid learning characteristics of extreme learning machines, a high-efficiency and high-precision diagnosis can be achieved, providing strong guidance for ensuring the reliable, safe, and economical operation of wind farms.

[0005] To achieve the above objectives, this invention provides a wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy, comprising:

[0006] (1) Collect vibration signals of the wind turbine gearbox under different health conditions;

[0007] (2) Use time-shifted cosine similarity entropy to extract multi-scale fault features from vibration signals and form a fault feature space for diagnosis;

[0008] (3) Divide the feature space into a training set and a validation set according to the proportions;

[0009] (4) Construct an intelligent fault diagnosis model based on Extreme Learning Machine (ELM) and input the training set for model training;

[0010] (5) Input the test set into the trained intelligent fault diagnosis model to complete the fault diagnosis of the gearbox.

[0011] In some optional implementations, the vibration signals collected in step 1 are the gearboxes of the operating wind turbines and historical data recorded by the wind farm, wherein each set of vibration signal data includes the corresponding gearbox health status and a corresponding health status label is set.

[0012] In some alternative implementations, step (2) includes:

[0013] Each set of vibration signal data is divided into several subsequences according to the time shift scale, and the phase space is reconstructed for each subsequence to obtain the embedded sample sequence in the phase space corresponding to each set of vibration signal data.

[0014] Calculate the angular distance between all pairs of embedded sample sequences in the phase space. Given a decision threshold, when the angular distance is less than or equal to the decision threshold, count the number of similar patterns under the time-shifting scale, and calculate the probability of local similar patterns and global similar patterns.

[0015] The cosine similarity entropy at the time-shifted scale is calculated from the global similarity pattern, and then the time-shifted cosine similarity entropy at the maximum time scale is calculated.

[0016] In some alternative implementations, by Each set of vibration signal data is divided into several subsequences according to the time shift scale. The vibration signal data is represented by x = {x(i), i = 1, 2, ..., N}, where β and α are positive integers, β = 1, 2, ..., α, β is the starting point of the sample sequence, and α is the time shift scale, representing the time interval. This indicates the floor function.

[0017] In some alternative implementations, by For each subsequence, phase space reconstruction is performed, with temporal embedding dimension m and time delay d.

[0018] In some alternative implementations, by Calculate the angular distance between all pairs of embedded sample sequences in the phase space.

[0019] In some alternative implementations, by Calculate the time-shifted cosine similarity entropy at the maximum time scale, αmax CSE(α) is the maximum value on the time scale, and CSE(α) is the cosine similarity entropy at scale α. CSE(α) = -(G α log2G α +(1-G α log2(1-G α )), G α This is a globally similar pattern.

[0020] In some alternative implementations, in step 3, the feature space is divided into two parts, wherein 80% of the features are randomly selected from the feature space as the training set and the remaining 20% ​​is used as the test set to verify the fault diagnosis performance of the model.

[0021] In some alternative implementations, step (4) includes:

[0022] An ELM fault diagnosis model is constructed, and the number of input nodes, hidden layer nodes, output nodes, and activation function of the ELM are determined. The number of input layer nodes is the same as the dimension of the input feature vector, the number of hidden layer nodes is set to twice the number of input layer nodes, and the number of output layer nodes is the number of categories of the health status of the wind turbine gearbox. The radial basis function is selected as the activation function of the ELM fault diagnosis model.

[0023] The ELM fault diagnosis model is trained using the training feature set.

[0024] In some alternative implementations, training the ELM fault diagnosis model using a training feature set includes:

[0025] Input training features, set the number of hidden layer nodes, randomly generate input weights and hidden layer threshold values, and then calculate the hidden layer output matrix corresponding to the training features;

[0026] Solve for the Moore-Penrose inverse of the hidden layer output matrix; the input weights and hidden layer thresholds are automatically adjusted during training.

[0027] Then, the output weight matrix of the output node and the hidden layer node is calculated.

[0028] In some alternative implementations, step (5) includes:

[0029] Input test features to obtain the corresponding hidden layer output matrix, and obtain the output of the ELM model from the hidden layer output matrix;

[0030] The ELM output is compared with the actual label category of the gearbox to identify the specific fault type, obtain the diagnostic accuracy of each type of fault, verify the ELM fault identification model, and complete the diagnosis of multiple types of faults.

[0031] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0032] This invention can be used for fault diagnosis of wind turbine gearboxes. To effectively characterize the dynamic evolution of faults, a time-shifted cosine similarity entropy is innovatively proposed for extracting key fault features. This entropy is then organically integrated with an Extreme Learning Machine (ELM), resulting in a wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy and ELM. In this method, time-shifted cosine similarity entropy can effectively extract fault features with high expressive power and separability. The parameters in the ELM are randomly generated and do not require adjustment during training; by setting the number of hidden layer neurons, multi-class fault state recognition can be directly achieved; the ELM recognition model has fast iteration speed and high accuracy. Therefore, the wind turbine gearbox fault diagnosis method integrating time-shifted cosine similarity entropy and ELM can achieve rapid and accurate fault diagnosis. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the implementation of a wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy, provided by an embodiment of the present invention.

[0034] Figure 2 This is a flowchart of a time-shifted cosine similarity entropy implementation provided by an embodiment of the present invention;

[0035] Figure 3 This is a structural diagram of an extreme learning machine provided in an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0037] The purpose of this invention is to provide a novel diagnostic method for the health status of wind turbine gearboxes, which offers higher assessment efficiency and accuracy, and addresses the problem that traditional methods struggle to effectively characterize the dynamic evolution of faults.

[0038] like Figure 1 As shown, the wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy according to an embodiment of the present invention includes two stages: the first stage is to extract key fault features using time-shifted cosine similarity entropy; the second stage is to perform fault diagnosis using an extreme learning machine based on the extracted features. The method includes the following steps:

[0039] Step 1: First, collect vibration signals from the wind turbine gearbox under different health conditions;

[0040] Specifically, this includes some common health status types: normal, broken tooth, cracked, scratched, mixed fault, bearing failure, etc. These data are collected from operating transformers and power company test data, used as input for time-shifted cosine similarity entropy, and the corresponding health status labels are set as {C1, C2, C3, C4, C5, C6}.

[0041] Step 2: Use time-shifted cosine similarity entropy to extract features from the collected multi-type vibration signals to obtain a feature space that can highly represent the original data, which is then used as the input to the ELM classifier;

[0042] The parameter settings for the time-shifted cosine similarity entropy in step 2 are shown in Table 1.

[0043] Table 1. Time-shifted cosine similarity entropy parameter settings

[0044]

[0045] like Figure 2 As shown, feature extraction is performed on the input data using time-shifted cosine similarity entropy, specifically as follows:

[0046] Assuming a vibration signal sample sequence x = {x(i), i = 1, 2, ..., N}, and letting β and α be positive integers, where β = 1, 2, ..., α, then α subsequences can be obtained, as shown by the formula:

[0047]

[0048] Where β is the starting point of the sample sequence; α is the time shift scale, representing the time interval; This indicates the floor function.

[0049] Given a temporal embedding dimension m and a time delay d, reconstructing the phase space for each subsequence yields the corresponding embedded sample sequence y in the phase space, as shown in the formula:

[0050]

[0051] The angular distance between all paired embedded sample sequences in the phase space is calculated using the following formula:

[0052]

[0053] Given a decision threshold r CSE When the criterion AngDis(i,j)≤r is satisfied CSE At that time, count the number P of similar patterns at scale α. α (k), and calculate the probability B of the occurrence of local similar patterns.α (k), its formula is:

[0054]

[0055] The formula for calculating global similarity patterns is:

[0056]

[0057] The formula for calculating the cosine similarity entropy at scale α is:

[0058] CSE(α)=-(G α log2G α +(1-G α log2(1-G α ))

[0059] Given a time scale α max The formula for calculating the time-shifted cosine similarity entropy is:

[0060]

[0061] Where x(i) is the i-th value in the time series x, and N is the length of the time series.

[0062] Step 3: Divide the feature space into training and validation sets proportionally;

[0063] Step 3 divides the dataset into two parts: 80% is used as the training set to train the ELM diagnostic model, and 20% is used as the test set to verify the diagnostic model's ability to classify and identify health status.

[0064] Step 4: Construct an intelligent fault diagnosis model based on Extreme Learning Machine (ELM), and input the training set to train the diagnostic model;

[0065] like Figure 3 As shown, the construction method of the ELM network in step 4 is as follows: First, construct a 3-layer ELM network; determine the number of input nodes, hidden layer nodes, and output nodes of the ELM fault diagnosis model; the number of input layer nodes must correspond to the dimension of the input feature vector; the number of output layer nodes is the number of fault types in step 1; the number of hidden layer nodes is twice the dimension of the feature vector; the activation function is the radial basis function, and the formula is:

[0066]

[0067] Where x is the input feature; w i For input weights; b i This is the threshold value for the hidden layer.

[0068] The pre-built ELM fault diagnosis model is then trained using the input training set to obtain the wind turbine gearbox health status assessment model.

[0069] The specific training process of the fault diagnosis model is as follows:

[0070] Input k training features, set the number of hidden layer nodes to n; randomly generate input weights w. i and the value of the hidden layer threshold b i Then, the hidden layer output matrix corresponding to the training features is calculated:

[0071]

[0072] Solve for the Moore-Penrose inverse H of the hidden layer output matrix H. + Input weight w i and hidden layer threshold b i Automatically adjusts during training;

[0073] Then, the output weight matrix A = H of the output node and the hidden layer node is calculated. + T, where T is the output label vector.

[0074] For simplicity, the accuracy rate can be calculated using the following formula:

[0075]

[0076] Wherein, TP refers to the number of positive classes predicted as positive, TN refers to the number of positive classes predicted as negative, FP refers to the number of negative classes predicted as positive, and FN refers to the number of negative classes predicted as negative.

[0077] Step 5: Substitute the samples from the test set into the trained ELM classifier to verify their categories, compare the ELM recognition results with the actual operation of the wind turbine, calculate the diagnostic accuracy, and complete the verification of the ELM fault recognition model.

[0078] Step 5 specifically includes:

[0079] Input q test features to obtain the corresponding hidden layer output matrix:

[0080]

[0081] The output of the ELM model is calculated as follows:

[0082]

[0083] The ELM output is compared with the actual label category of the gearbox to identify the specific fault type, obtain the diagnostic accuracy of each type of fault, verify the ELM fault identification model, and complete the diagnosis of multiple types of faults.

[0084] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0085] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wind turbine gearbox fault diagnosis method based on time-shifted cosine similarity entropy, characterized in that, include: (1) Collect vibration signals of the wind turbine gearbox under different health conditions; (2) Use time-shifted cosine similarity entropy to extract multi-scale fault features from vibration signals and form a fault feature space for diagnosis; (3) Divide the feature space into a training set and a validation set according to the proportions; (4) Construct an intelligent fault diagnosis model based on Extreme Learning Machine (ELM) and input the training set for model training; (5) Input the test set into the trained intelligent fault diagnosis model to complete the fault diagnosis of the gearbox; The vibration signals collected in step 1 are the gearboxes of the wind turbines in operation and the historical data recorded by the wind farm. Each set of vibration signal data includes the corresponding gearbox health status and sets a corresponding health status label. Step (2) includes: Each set of vibration signal data is divided into several subsequences according to the time shift scale, and the phase space is reconstructed for each subsequence to obtain the embedded sample sequence in the phase space corresponding to each set of vibration signal data. Calculate the angular distance between all pairs of embedded sample sequences in the phase space. Given a decision threshold, when the angular distance is less than or equal to the decision threshold, count the number of similar patterns under the time-shifting scale, and calculate the probability of local similar patterns and global similar patterns. The cosine similarity entropy at the time-shifting scale is calculated from the global similarity pattern, and then the time-shifting cosine similarity entropy at the maximum time scale is calculated. By Each group of vibration signal data is divided into several subsequences according to the time shift scale, wherein the vibration signal data is x = { x ( i ), i = 1,2,…, N}, β and α is a positive integer, β = 1,2,…, α , β is the starting point of the sample sequence, α is the time shift scale, indicating the interval of time, indicates the floor operation. Depend on Phase space reconstruction is performed on each subsequence, with temporal embedding dimension of . m The time delay is d ; Depend on Calculate the angular distance between all pairs of embedded sample sequences in the phase space; Depend on Calculate the time-shifted cosine similarity entropy at the maximum time scale. α max The maximum value on the time scale. for α Cosine similarity entropy at the scale, , This is a globally similar pattern.

2. The method according to claim 1, characterized in that, Step (4) includes: An ELM fault diagnosis model is constructed, and the number of input nodes, hidden layer nodes, output nodes, and activation function of the ELM are determined. The number of input layer nodes is the same as the dimension of the input feature vector, the number of hidden layer nodes is set to twice the number of input layer nodes, and the number of output layer nodes is the number of categories of the health status of the wind turbine gearbox. The radial basis function is selected as the activation function of the ELM fault diagnosis model. The ELM fault diagnosis model is trained using the training feature set.

3. The method according to claim 2, characterized in that, The process of training the ELM fault diagnosis model using a training feature set includes: Input training features, set the number of hidden layer nodes, randomly generate input weights and hidden layer threshold values, and then calculate the hidden layer output matrix corresponding to the training features; Solve for the Moore-Penrose inverse of the hidden layer output matrix; the input weights and hidden layer thresholds are automatically adjusted during training. Then, the output weight matrices of the output nodes and hidden layer nodes are calculated.

4. The method according to claim 3, characterized in that, Step (5) includes: Input test features to obtain the corresponding hidden layer output matrix, and obtain the output of the ELM model from the hidden layer output matrix; The ELM output is compared with the actual label category of the gearbox to identify the specific fault type, obtain the diagnostic accuracy of each type of fault, verify the ELM fault identification model, and complete the diagnosis of multiple types of faults.

Images