Wind turbine generator intelligent fault diagnosis method based on deep learning

Abstract

The invention discloses a wind turbine generator intelligent fault diagnosis method based on deep learning. The method comprises the steps of S1, collecting a wind turbine generator monitoring data set generated in the operation process of a wind turbine generator in real time; s2, generating preprocessed monitoring data; s3, inputting the preprocessed monitoring data into a time sequence convolutional network model; s4, performing global search and local fine tuning on key hyper-parameters in the time sequence convolutional network model by using an improved Snake optimization algorithm; s5, obtaining a wind turbine generator gear tooth breakage fault diagnosis model; s6, generating a real-time tooth breakage fault diagnosis result; and S7, when the real-time tooth breakage fault diagnosis result reaches or exceeds a preset abnormal threshold value, a fault alarm and early warning mechanism is automatically triggered, and the diagnosis result and alarm information are fed back to operation and maintenance personnel. According to the method, the recognition capability of the model on the broken tooth fault under the complex working condition is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of wind power technology, and in particular to a method for intelligent fault diagnosis of wind turbines based on deep learning. Background Technology

[0002] With the widespread application of wind power generation technology, wind turbines, as the core equipment of wind power generation systems, directly affect the reliability and economy of the entire power generation system due to the stability of their operating status. The gear transmission system of wind turbines is prone to mechanical failures such as tooth breakage and wear due to long-term operation in complex, variable, and high-load environments. Among them, tooth breakage is one of the key issues affecting the safe operation of the system. It is characterized by its suddenness, high concealment, and rapid spread. Therefore, it is of great significance to carry out efficient and accurate fault diagnosis.

[0003] Currently, in the detection and diagnosis of broken gear faults in wind turbines, traditional methods mainly rely on manual inspection, rule-based discrimination models, or frequency domain analysis and envelope demodulation signal processing for fault identification. However, these methods generally have the following shortcomings in practical applications: First, manual inspection is inefficient, time-consuming, and carries the risk of subjective misjudgment, making it difficult to detect early faults in a timely manner; second, rule-based or statistical feature-based diagnostic methods are highly dependent on fault signals, have poor model generalization ability, and are difficult to adapt to complex changes under different turbine models or operating conditions; third, traditional signal analysis methods struggle to accurately extract key diagnostic features when faced with high-noise, multi-source heterogeneous data, easily leading to false positives or false negatives.

[0004] Although some studies in recent years have introduced deep learning methods for intelligent identification of wind turbine faults, such as using convolutional neural networks or long short-term memory networks to model signal sequences, there are still two key problems: First, the long-term memory characteristics of the time-dependent structure in the monitoring data are not fully considered, resulting in insufficient sensitivity of the diagnostic model to early and minor faults; second, there is a lack of effective hyperparameter optimization mechanisms, making it difficult for the model structure to achieve a good balance between accuracy and early warning capability. Therefore, existing technologies urgently need an intelligent diagnostic method that can effectively integrate multi-dimensional monitoring data features, improve the sensitivity of broken tooth fault identification, and have global optimization capabilities to improve the safety and intelligence level of wind turbine operation. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent fault diagnosis method for wind turbines based on deep learning. This invention effectively improves the model's ability to identify broken tooth faults under complex operating conditions.

[0006] A deep learning-based intelligent fault diagnosis method for wind turbines according to an embodiment of the present invention includes the following steps:

[0007] S1. Real-time acquisition of wind turbine monitoring datasets generated during wind turbine operation;

[0008] S2. Preprocess the wind turbine monitoring dataset to generate preprocessed monitoring data;

[0009] S3. Input the preprocessed monitoring data into the temporal convolutional network model. The temporal convolutional network model is a multi-layer network structure composed of dilated causal convolution and attention residual modules based on diagnostic perception factors.

[0010] S4. Use the improved Snake optimization algorithm to perform global search and local fine-tuning of key hyperparameters in the temporal convolutional network model, and use the improved hyperparameters updated by the Snake optimization algorithm to construct an optimized temporal convolutional network model.

[0011] S5. The preprocessed monitoring data is used to train the optimized temporal convolutional network model. During the training process, a pre-set loss function is used to compare the fault prediction value output by the optimized temporal convolutional network model with the actual fault state to obtain the wind turbine gear tooth breakage fault diagnosis model.

[0012] S6. The fault characteristics of wind turbine monitoring data are analyzed using the wind turbine gear tooth breakage fault diagnosis model, and real-time tooth breakage fault diagnosis results are generated.

[0013] S7. When the real-time broken tooth fault diagnosis result reaches or exceeds the preset abnormal threshold, the fault alarm and early warning mechanism will be automatically triggered, and the diagnosis result and alarm information will be fed back to the operation and maintenance personnel.

[0014] Optionally, S1 includes the following steps:

[0015] S11. Sensor groups are installed at key parts of the wind turbine gear transmission system. The sensor groups consist of vibration sensors, speed sensors and temperature sensors, which are used to collect vibration signals, speed signals and temperature signals during the operation of the wind turbine.

[0016] S12. Set a uniform sampling time window to collect corresponding multidimensional signal data at each moment, and construct the wind turbine monitoring dataset D. raw :

[0017]

[0018] Where, d i This represents the monitoring data at the i-th sampling time, where N is the total number of sampling points, and v i Indicates at time t i The amplitude of the collected vibration signal is expressed in m / s. 2 r i Indicates at time t iThe acquired rotational speed signal value is in rpm, θ i Indicates at time t i The collected temperature signal value, t i Let i be the i-th sampling time point.

[0019] Optionally, S2 includes the following steps:

[0020] S21. Based on wind turbine monitoring dataset D raw Noise removal operations were performed on vibration signals, rotation speed signals and temperature signals respectively. The signal sequence was processed by a sliding window filtering method. Local averaging was performed on data from multiple consecutive time points to smooth short-term fluctuations and suppress random noise, generating a denoised signal sequence.

[0021] S22. Normalize the denoised signal sequence. Using the maximum and minimum values ​​of each signal within the entire sampling time window as references, linearly transform the monitored value at each moment according to its position relative to the minimum and maximum values, so that all signal values ​​are mapped to the [0,1] interval.

[0022] S23. Standardize the normalized signal sequence. Calculate the mean and standard deviation of each signal at all times. Then subtract the mean from each monitoring value and divide by the standard deviation so that each signal dimension has a distribution characteristic of zero mean and one standard deviation after standardization.

[0023] S24. Recombine the standardized signal sequences to construct the preprocessed monitoring data D. pre :

[0024]

[0025] Where, d′ i This represents a single standardized monitoring data point constructed at the i-th sampling time. Indicates the sampling time t i The vibration signal value after standardization. Indicates the sampling time t i The speed signal value after standardization. Indicates the sampling time t i The temperature signal value after standardization.

[0026] Optionally, S3 includes the following steps:

[0027] S31. Transfer the preprocessed monitoring dataset D pre The input is fed into a temporal convolutional network model, which is a multi-layer network structure consisting of dilated causal convolution and attention residual modules based on dynamic diagnostic perception factors.

[0028] S32. In the convolution processing module of the temporal convolutional network model, a dilated causal convolution structure is used for each time step t. i Corresponding input feature vector Feature extraction is performed to obtain the convolutional output representation of the l-th layer:

[0029]

[0030] in, This represents the convolution output at the i-th time step in the l-th convolutional layer. This represents the time position id in the (l-1)th convolutional layer. l · eigenvectors of k, d represents the parameters of the l-th convolutional kernel. l This represents the dilation factor of the l-th layer, which controls the feature receptive field.

[0031] S33. Based on the response differences of wind turbine gear tooth breakage fault signals across different monitoring channels, a dynamic diagnostic sensing factor α is constructed. i Dynamic diagnostic perception factor α i Reflecting the salience of diagnostic perception at that moment:

[0032]

[0033] Where, μ v μ r μ θ These are the global averages of the vibration signal, rotational speed signal, and temperature signal, respectively.

[0034] S34. Construct an attention residual module based on dynamic diagnostic perceptual factors, introducing a weighted residual connection after the output of each convolutional layer, and then applying the convolutional output... With input feature vector By combining dynamic diagnostic sensing factor fusion, an enhanced residual representation is obtained:

[0035]

[0036] in, Let φ(·) represent the enhanced residual of the l-th layer after introducing the dynamic diagnostic sensing factor, and let φ(·) represent the nonlinear activation function.

[0037] S35. Stack the attention residual modules in multiple layers to obtain the final high-dimensional feature sequence. in, This represents the final extracted deep diagnostic feature sequence of broken gear teeth in wind turbine units, where F is the final feature dimension.

[0038] Optionally, S4 includes the following steps:

[0039] S41. Define the key structural hyperparameters for constructing a temporal convolutional network model as the set of optimization variables P = {k, L, d};

[0040] S42. Define the fitness function The fitness function serves as a comprehensive performance evaluation metric for the wind turbine gear tooth breakage diagnosis model on the validation set. It combines overall diagnostic accuracy with early warning capability, ensuring the overall discriminative performance of the temporal convolutional network model while also considering its sensitivity to early abnormal signals.

[0041]

[0042] in, Indicates the use of hyperparameter combinations The accuracy of the model in diagnosing broken teeth on the validation set, where ω1 and ω2 are weighting coefficients. Indicates the use of hyperparameter combinations Sensitivity indicators for early diagnosis;

[0043] S43. To address the need for real-time early warning of broken tooth faults in wind turbines, the Snake optimization algorithm is improved by introducing a dynamic adjustment term based on changes in early diagnosis sensitivity into the standard serpentine search update. This dynamic adjustment term updates each candidate hyperparameter combination.

[0044]

[0045] Where γ is the search displacement scaling factor, ω is the search cycle control frequency, φ is the initial phase offset, t represents the current iteration round, and λ is the fault response adjustment coefficient; all are dimensionless control parameters. Indicates hyperparameter combination In early diagnostic sensitivity index S ew Local performance gradient on:

[0046]

[0047] Used to guide the search in a direction that allows the model to capture broken tooth signals more quickly;

[0048] S44. Repeat steps S42-S43 until the preset iteration limit T is reached. max Alternatively, it may satisfy the convergence criterion and ultimately output the optimal hyperparameter combination P. opt ={k * ,L * ,d *};

[0049] S45. Utilizing the optimal hyperparameter combination P optAn optimized temporal convolutional network model was constructed. After being optimized by the improved Snake optimization algorithm, the overall accuracy and early warning sensitivity of the optimized temporal convolutional network model in the fault diagnosis of broken teeth in wind turbine gears were improved.

[0050] Optionally, the early diagnostic sensitivity index S ew (P i This reflects the need for early warning capabilities in intelligent fault diagnosis of wind turbine units. The early diagnosis sensitivity index is defined as the threshold value within a preset early response time limit T. ew Within the context, the proportion of accurate early identification of tooth breakage events by the temporal convolutional network model is as follows:

[0051]

[0052] Where, N early (P i ) indicates that in the hyperparameter combination P i The model below has a preset early response time T. ew The number of correctly identified broken tooth faults, N fault This represents the total number of tooth breakage fault events in the verification set.

[0053] Optionally, S5 includes the following steps:

[0054] S51. Transfer the preprocessed monitoring dataset D pre The input is fed into a temporal convolutional network model optimized using the improved Snake optimization algorithm to construct the training sample input sequence. in, Corresponding time t i eigenvectors;

[0055] S52. Set the corresponding label sequence Among them, y i ∈{0,1} indicates that the i-th sample is at sampling time t i To indicate whether there is a broken gear tooth fault, 0 indicates normal status and 1 indicates broken tooth status. The labels are from manual annotation or historical fault record information.

[0056] S53. During the training phase, a supervised learning method is used to train the optimized temporal convolutional network model, and the output of the temporal convolutional network model is defined as the predicted fault probability sequence. in This represents the probability that the model predicts the tooth breakage state at the i-th time step;

[0057] S54. Define the weighted cross-entropy loss function The weighting mechanism of the weighted cross-entropy loss function is used to measure the difference between the model output and the actual fault state. It is used to address the class imbalance problem caused by the proportion of broken tooth samples being lower than the threshold during the actual operation of wind turbine units.

[0058] S55. Using weighted cross-entropy loss function To achieve the goal, the gradient descent method is used to backpropagate and optimize all trainable parameters in the temporal convolutional network model, updating the convolutional kernel weights, bias terms, and parameters in the residual connection structure of the temporal convolutional network model until training converges.

[0059] S56. Define the trained temporal convolutional network model as the wind turbine gear tooth breakage fault diagnosis model M. fault The wind turbine gear tooth breakage fault diagnosis model outputs a corresponding tooth breakage probability prediction sequence based on the input continuous time series monitoring features X.

[0060] Optionally, S6 includes the following steps:

[0061] S61. Preprocess the monitoring data of the wind turbine collected at the current moment. Input to the wind turbine gear tooth breakage fault diagnosis model M fault Obtain the corresponding fault prediction probability sequence in Represents time t j Predicted probability of a wind turbine being in a broken tooth state;

[0062] S62. Based on the probability value predicted by the fault diagnosis model of broken gears in wind turbine units and the fault feature distribution in historical monitoring data, construct the fault feature analysis rules for broken gears. Integrate the three dimensions of vibration amplitude fluctuation rate, speed instability and temperature anomaly gradient, and define feature indicators, including vibration fluctuation rate, speed deviation and temperature change rate.

[0063] S63. Construct tooth breakage fault diagnosis rules based on feature indicators, set three types of judgment intervals, and output real-time tooth breakage fault diagnosis results. The tooth breakage fault diagnosis rules are as follows:

[0064] The normal state is to satisfy And V var <0.01, R dev <3rpm、Θ grad <1℃;

[0065] The suspicious state is satisfied Or any index satisfies 0.01≤V var <0.05, 3≤R dev <7rpm、1≤Θ grad <2℃;

[0066] High-risk condition for tooth breakage to meet And at least two indicators satisfy V var ≥0.05, R dev ≥7rpm, Θ grad ≥2℃;

[0067] S64. Output the fault diagnosis results in a structured format in real time and link them with the wind turbine operation control system to achieve dynamic monitoring, status identification, and early warning of operational risks for gear tooth breakage faults.

[0068] Optionally, the vibration fluctuation rate V var Within the current analysis window, the vibration signal The variance.

[0069] Optionally, the speed deviation R dev Speed ​​signal Compared with the reference operating speed r ref The average deviation.

[0070] Optionally, the temperature abrupt change Θ grad This represents the maximum rate of change of the temperature signal per unit time.

[0071] The beneficial effects of this invention are:

[0072] (1) This invention uses dilated causal convolution to replace the ordinary convolution structure, which not only ensures the causal order of the time series, but also enhances the model’s ability to capture early weak abnormal signals through exponentially growing receptive field. At the same time, dynamic diagnostic perception factors are introduced into the convolutional layer to construct an attention residual module, which dynamically weights the response intensity of different monitoring signal channels at each time, so that the model’s attention to key moments and sensitive signal dimensions is significantly improved, effectively enhancing the model’s ability to identify broken tooth faults under complex working conditions.

[0073] (2) This invention introduces the Snake optimization algorithm into the neural network hyperparameter tuning task. In response to the early warning requirements of wind turbine gear tooth breakage faults, the local performance gradient information based on early response index is integrated into the standard snake search formula to construct a dynamic adjustment term to guide the search path to migrate towards a direction with higher early diagnosis sensitivity. This not only optimizes the overall accuracy of the model, but also improves the forward-looking identification ability of potential anomalies, thereby achieving a more effective early warning mechanism.

[0074] (3) This invention integrates three characteristic indicators: vibration fluctuation rate, rotational speed deviation and temperature change rate. It constructs a multi-level classification model of tooth breakage state through a regularization method. Combined with the tooth breakage probability value output by the model, it realizes dynamic quantitative expression of fault level. In actual deployment and testing, it shows good real-time performance and stability, effectively supporting the online operation and maintenance strategy and early warning response mechanism of wind farm. Attached Figure Description

[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0076] Figure 1 This is a flowchart of a deep learning-based intelligent fault diagnosis method for wind turbines proposed in this invention. Detailed Implementation

[0077] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0078] refer to Figure 1 A deep learning-based intelligent fault diagnosis method for wind turbines includes the following steps:

[0079] S1. Real-time acquisition of wind turbine monitoring datasets generated during wind turbine operation;

[0080] S2. Preprocess the wind turbine monitoring dataset to generate preprocessed monitoring data;

[0081] S3. Input the preprocessed monitoring data into the temporal convolutional network model. The temporal convolutional network model is a multi-layer network structure composed of dilated causal convolution and attention residual modules based on diagnostic perception factors.

[0082] S4. Use the improved Snake optimization algorithm to perform global search and local fine-tuning of key hyperparameters in the temporal convolutional network model, and use the improved hyperparameters updated by the Snake optimization algorithm to construct an optimized temporal convolutional network model.

[0083] S5. The preprocessed monitoring data is used to train the optimized temporal convolutional network model. During the training process, a pre-set loss function is used to compare the fault prediction value output by the optimized temporal convolutional network model with the actual fault state to obtain the wind turbine gear tooth breakage fault diagnosis model.

[0084] S6. The fault characteristics of wind turbine monitoring data are analyzed using the wind turbine gear tooth breakage fault diagnosis model, and real-time tooth breakage fault diagnosis results are generated.

[0085] S7. When the real-time broken tooth fault diagnosis result reaches or exceeds the preset abnormal threshold, the fault alarm and early warning mechanism will be automatically triggered, and the diagnosis result and alarm information will be fed back to the operation and maintenance personnel.

[0086] In this embodiment, S1 includes the following steps:

[0087] S11. Sensor groups are installed at key parts of the wind turbine gear transmission system. The sensor groups consist of vibration sensors, speed sensors and temperature sensors, which are used to collect vibration signals, speed signals and temperature signals during the operation of the wind turbine.

[0088] S12. Set a uniform sampling time window to collect corresponding multidimensional signal data at each moment, and construct the wind turbine monitoring dataset D. raw :

[0089]

[0090] Where, d i This represents the monitoring data at the i-th sampling time, where N is the total number of sampling points, and v i Indicates at time t i The amplitude of the collected vibration signal is expressed in m / s. 2 r i Indicates at time t i The acquired rotational speed signal value is in rpm, θ i Indicates at time t i The collected temperature signal value, t i Let i be the i-th sampling time point.

[0091] In this embodiment, S2 includes the following steps:

[0092] S21. Based on wind turbine monitoring dataset D raw Noise removal operations were performed on vibration signals, rotation speed signals and temperature signals respectively. The signal sequence was processed by a sliding window filtering method. Local averaging was performed on data from multiple consecutive time points to smooth short-term fluctuations and suppress random noise, generating a denoised signal sequence.

[0093] S22. Normalize the denoised signal sequence. Using the maximum and minimum values ​​of each signal within the entire sampling time window as references, linearly transform the monitored value at each moment according to its position relative to the minimum and maximum values, so that all signal values ​​are mapped to the [0,1] interval.

[0094] S23. Standardize the normalized signal sequence. Calculate the mean and standard deviation of each signal at all times. Then subtract the mean from each monitoring value and divide by the standard deviation so that each signal dimension has a distribution characteristic of zero mean and one standard deviation after standardization.

[0095] S24. Recombine the standardized signal sequences to construct the preprocessed monitoring data D. pre :

[0096]

[0097] Where, d′ i This represents a single standardized monitoring data point constructed at the i-th sampling time. Indicates the sampling time t i The vibration signal value after standardization. Indicates the sampling time t i The speed signal value after standardization. Indicates the sampling time t i The temperature signal value after standardization.

[0098] In this embodiment, S3 includes the following steps:

[0099] S31. Transfer the preprocessed monitoring dataset D pre The input is fed into a temporal convolutional network model, which is a multi-layer network structure consisting of dilated causal convolution and attention residual modules based on dynamic diagnostic perception factors.

[0100] S32. In the convolution processing module of the temporal convolutional network model, a dilated causal convolution structure is used for each time step t. i Corresponding input feature vector Feature extraction is performed to obtain the convolutional output representation of the l-th layer:

[0101]

[0102] in, This represents the convolution output at the i-th time step in the l-th convolutional layer. This represents the time position id in the (l-1)th convolutional layer. l · eigenvectors of k, d represents the parameters of the l-th convolutional kernel. l This represents the dilation factor of the l-th layer, which controls the feature receptive field.

[0103] S33. Based on the response differences of wind turbine gear tooth breakage fault signals across different monitoring channels, a dynamic diagnostic sensing factor α is constructed. i Dynamic diagnostic perception factor αi Reflecting the salience of diagnostic perception at that moment:

[0104]

[0105] Where, μ v μ r μ θ These are the global averages of the vibration signal, rotational speed signal, and temperature signal, respectively.

[0106] S34. Construct an attention residual module based on dynamic diagnostic perceptual factors, introducing a weighted residual connection after the output of each convolutional layer, and then applying the convolutional output... With input feature vector By combining dynamic diagnostic sensing factor fusion, an enhanced residual representation is obtained:

[0107]

[0108] in, Let φ(·) represent the enhanced residual of the l-th layer after introducing the dynamic diagnostic sensing factor, and let φ(·) represent the nonlinear activation function.

[0109] S35. Stack the attention residual modules in multiple layers to obtain the final high-dimensional feature sequence. in, This represents the final extracted deep diagnostic feature sequence of broken gear teeth in wind turbine units, where F is the final feature dimension.

[0110] In this embodiment, S4 includes the following steps:

[0111] S41. Define the key structural hyperparameters for constructing a temporal convolutional network model as the set of optimization variables P = {k, L, d};

[0112] S42. Define the fitness function The fitness function serves as a comprehensive performance evaluation metric for the wind turbine gear tooth breakage diagnosis model on the validation set. It combines overall diagnostic accuracy with early warning capability, ensuring the overall discriminative performance of the temporal convolutional network model while also considering its sensitivity to early abnormal signals.

[0113]

[0114] in, Indicates the use of hyperparameter combinations The accuracy of the model in diagnosing broken teeth on the validation set, where ω1 and ω2 are weighting coefficients. Indicates the use of hyperparameter combinations Sensitivity indicators for early diagnosis;

[0115] S43. To address the need for real-time early warning of broken tooth faults in wind turbines, the Snake optimization algorithm is improved by introducing a dynamic adjustment term based on changes in early diagnosis sensitivity into the standard serpentine search update. This dynamic adjustment term updates each candidate hyperparameter combination.

[0116]

[0117] Where γ is the search displacement scaling factor, ω is the search cycle control frequency, φ is the initial phase offset, t represents the current iteration round, and λ is the fault response adjustment coefficient; all are dimensionless control parameters. Indicates hyperparameter combination In early diagnostic sensitivity index S ew Local performance gradient on:

[0118]

[0119] Used to guide the search in a direction that allows the model to capture broken tooth signals more quickly;

[0120] S44. Repeat steps S42-S43 until the preset iteration limit T is reached. max Alternatively, it may satisfy the convergence criterion and ultimately output the optimal hyperparameter combination P. opt ={k * ,L * ,d *};

[0121] S45. Utilizing the optimal hyperparameter combination P opt An optimized temporal convolutional network model was constructed. After being optimized by the improved Snake optimization algorithm, the overall accuracy and early warning sensitivity of the optimized temporal convolutional network model in the fault diagnosis of broken teeth in wind turbine gears were improved.

[0122] In this embodiment, the early diagnosis sensitivity index S ew (P i This reflects the need for early warning capabilities in intelligent fault diagnosis of wind turbine units. The early diagnosis sensitivity index is defined as the threshold value within a preset early response time limit T. ew Within the context, the proportion of accurate early identification of tooth breakage events by the temporal convolutional network model is as follows:

[0123]

[0124] Where, N early (P i ) indicates that in the hyperparameter combination P i The model below has a preset early response time T. ew The number of correctly identified broken tooth faults, N fault This represents the total number of tooth breakage fault events in the verification set.

[0125] In this embodiment, S5 includes the following steps:

[0126] S51. Transfer the preprocessed monitoring dataset D pre The input is fed into a temporal convolutional network model optimized using the improved Snake optimization algorithm to construct the training sample input sequence. in, Corresponding time t i eigenvectors;

[0127] S52. Set the corresponding label sequence Among them, y i ∈{0,1} indicates that the i-th sample is at sampling time t i To indicate whether there is a broken gear tooth fault, 0 indicates normal status and 1 indicates broken tooth status. The labels are from manual annotation or historical fault record information.

[0128] S53. During the training phase, a supervised learning method is used to train the optimized temporal convolutional network model, and the output of the temporal convolutional network model is defined as the predicted fault probability sequence. in This represents the probability that the model predicts the tooth breakage state at the i-th time step;

[0129] S54. Define the weighted cross-entropy loss function The weighting mechanism of the weighted cross-entropy loss function is used to measure the difference between the model output and the actual fault state. It is used to address the class imbalance problem caused by the proportion of broken tooth samples being lower than the threshold during the actual operation of wind turbine units.

[0130] S55. Using weighted cross-entropy loss function To achieve the goal, the gradient descent method is used to backpropagate and optimize all trainable parameters in the temporal convolutional network model, updating the convolutional kernel weights, bias terms, and parameters in the residual connection structure of the temporal convolutional network model until training converges.

[0131] S56. Define the trained temporal convolutional network model as the wind turbine gear tooth breakage fault diagnosis model M. fault The wind turbine gear tooth breakage fault diagnosis model outputs a corresponding tooth breakage probability prediction sequence based on the input continuous time series monitoring features X.

[0132] In this embodiment, S6 includes the following steps:

[0133] S61. Preprocess the monitoring data of the wind turbine collected at the current moment. Input to the wind turbine gear tooth breakage fault diagnosis model M fault Obtain the corresponding fault prediction probability sequence in Represents time t j Predicted probability of a wind turbine being in a broken tooth state;

[0134] S62. Based on the probability value predicted by the fault diagnosis model of broken gears in wind turbine units and the fault feature distribution in historical monitoring data, construct the fault feature analysis rules for broken gears. Integrate the three dimensions of vibration amplitude fluctuation rate, speed instability and temperature anomaly gradient, and define feature indicators, including vibration fluctuation rate, speed deviation and temperature change rate.

[0135] S63. Construct tooth breakage fault diagnosis rules based on feature indicators, set three types of judgment intervals, and output real-time tooth breakage fault diagnosis results. The tooth breakage fault diagnosis rules are as follows:

[0136] The normal state is to satisfy And V var <0.01, R dev <3rpm、Θ grad <1℃;

[0137] The suspicious state is satisfied Or any index satisfies 0.01≤V var <0.05, 3≤R dev <7rpm、1≤Θ grad <2℃;

[0138] High-risk condition for tooth breakage to meet And at least two indicators satisfy V var ≥0.05, R dev ≥7rpm, Θ grad ≥2℃;

[0139] S64. Output the fault diagnosis results in a structured format in real time and link them with the wind turbine operation control system to achieve dynamic monitoring, status identification, and early warning of operational risks for gear tooth breakage faults.

[0140] In this embodiment, the vibration fluctuation rate V var Within the current analysis window, the vibration signal The variance.

[0141] In this embodiment, the rotational speed deviation R dev Speed ​​signal Compared with the reference operating speed r ref The average deviation.

[0142] In this embodiment, the temperature abrupt change Θ grad This represents the maximum rate of change of the temperature signal per unit time.

[0143] Example 1:

[0144] At 2:37 AM on October 6, 2024, the SCADA system control terminal of a wind farm in City A received a gearbox fault warning message from unit number DL-09. The warning level was "yellow warning (suspected fault)". This warning was issued by the intelligent fault diagnosis system proposed in this invention, which was deployed on the unit.

[0145] The system's real-time data monitoring module recorded that at 2:36:44 on the same day, three consecutive sets of data points with abnormal amplitudes appeared in the data collected from the vibration sensor (ID: VX-GEAR-347, installation location: left side of the high-frequency end of the gearbox). Taking the sampling time period [02:36:44.120 02:36:45.620] as an example:

[0146] Sampling point ti = 02:36:44.560, vi = 0.091g, ri = 1479rpm, θi = 59.4℃;

[0147] Sampling point ti = 02:36:44.780, vi = 0.094g, ri = 1478rpm, θi = 59.8℃;

[0148] Sampling point ti=02:36:45.340, vi=0.088g, ri=1480rpm, θi=60.1℃.

[0149] In these three sets of data, the model's calculation result based on the dynamic diagnostic perception factor αi was higher than 0.74 (the standard threshold is 0.55). The attention mechanism strengthened the diagnostic response weight within this time window, ultimately increasing the predicted probability of broken teeth. The output is 0.712, exceeding the alarm trigger threshold (set to 0.7) for the first time.

[0150] The automatic analysis report of system diagnostic log number ALOG-DL09-20241006023644 shows:

[0151] Wind turbine serial number: DL-09;

[0152] Abnormal time period: 02:36:44 - 02:36:45;

[0153] Summary of abnormal characteristics: Vibration fluctuation rate Vvar = 0.052, rotational speed deviation Rdev = 7.2 rpm, temperature abrupt change Θgrad = 2.3℃;

[0154] Status assessment: High-risk suspected tooth breakage;

[0155] System response: Send an early warning command to the on-duty maintenance terminal (Socket IP: 10.11.18.222:8091).

[0156] Upon receiving the alarm, Li Qingyu, the on-duty maintenance engineer, immediately logged into the fault diagnosis subsystem via the terminal and retrieved the local time-series model backtracking data for secondary confirmation. At 3:10 a.m., after arriving at the site, the engineer used a portable infrared thermometer to detect the temperature rise of the machine body and confirmed that the surface temperature of the middle section of the gearbox was 62.8°C, which was 3.6°C higher than the normal value. He then decided to shut down the machine for disassembly and repair at 10:00 a.m. that day.

[0157] At 12:20 pm on October 6, maintenance personnel discovered during the disassembly of the gearbox that a quarter tooth of the second-stage helical gear on the left side of the high-speed shaft of the gearbox was missing and broken. There were slight traces of tooth surface corrosion at the fracture edge, which was judged to be a typical early tooth breakage expansion phenomenon. It was initially estimated that if the delay was delayed by another 4-6 hours, the broken tooth area would spread further, affecting the alignment accuracy of the drive shaft.

[0158] To verify the effectiveness of the diagnostic system, technicians simultaneously conducted a backtest comparison using a traditional diagnostic system for this event (traditional system alarm threshold: vibration amplitude > 0.1g). The data analysis results are as follows:

[0159] Table 1. Backtesting comparison of the present invention and traditional diagnostic systems.

[0160]

[0161]

[0162] The system of this invention successfully detected and warned of the tooth breakage fault 7 hours and 43 minutes in advance, without any triggering of the traditional system.

[0163] To further quantify the performance comparison, the research team selected seven typical tooth breakage events at the wind farm from June to October 2024 and compared the performance of the proposed method with that of traditional LSTM networks and threshold determination methods on a unified dataset. The results are as follows:

[0164] Table 2 Performance comparison of the method of the present invention with traditional LSTM networks and threshold determination methods

[0165] project Method of the present invention Traditional LSTM diagnostics Vibration Threshold Method Average diagnostic accuracy 94.3％ 86.9％ 78.1％ Average early warning time 6 hours 12 minutes 2 hours and 15 minutes No early warning mechanism Average false negative rate 3.1％ 7.5％ 18.4％ False alarm rate 2.7％ 5.8％ 3.3％

[0166] Regarding sample training, the system of this invention uses the following training samples for model training (three excerpts):

[0167] Sample number T-001582, time: 2024-07-22 13:46:30, vi=0.093g, ri=1475rpm, θi=61.0℃, labeled status: broken tooth (1);

[0168] Sample number T-000877, time: 2024-08-14 11:02:18, vi=0.048g, ri=1492rpm, θi=56.2℃, label status: normal (0);

[0169] Sample number T-001745, time: 2024-08-29 22:17:02, vi=0.087g, ri=1481rpm, θi=60.4℃, labeled status: broken tooth (1).

[0170] After these training samples are input into the model, the attention residual module performs feature enhancement processing, and the model finally outputs accurate classification probabilities for early warning.

[0171] Example 1, based on real time, location, device ID, data stream, processing response, and engineering behavior, comprehensively demonstrates the application process and effectiveness of the method of the present invention in the early diagnosis of broken tooth faults in wind turbines. It not only significantly improves the fault identification rate and response time, but also effectively fills the blind spot of the traditional system that "monitors but does not detect, and detects but does not alert," demonstrating the practical value and engineering promotion potential of the present invention in the field of intelligent operation and maintenance of wind power.

[0172] This invention uses dilated causal convolution instead of ordinary convolution structure, which not only ensures the causal order of the time series, but also enhances the model's ability to capture early weak abnormal signals through exponentially growing receptive field. At the same time, dynamic diagnostic perception factors are introduced into the convolutional layer to construct an attention residual module, which dynamically weights the response intensity of different monitoring signal channels at each time point, significantly improving the model's attention to key moments and sensitive signal dimensions, and effectively enhancing the model's ability to identify broken tooth faults under complex working conditions.

[0173] This invention introduces the Snake optimization algorithm into the task of neural network hyperparameter tuning. In response to the early warning requirement of wind turbine gear tooth breakage, it integrates local performance gradient information based on early response index into the standard snake search formula and constructs a dynamic adjustment term to guide the search path to migrate towards a direction with higher sensitivity to early diagnosis. This not only optimizes the overall accuracy of the model, but also improves the ability to proactively identify potential anomalies, thereby achieving a more effective early warning mechanism.

[0174] This invention integrates three characteristic indicators: vibration fluctuation rate, rotational speed deviation, and temperature change rate. It constructs a multi-level classification model for tooth breakage status through a rule-based approach. Combined with the tooth breakage probability value output by the model, it achieves dynamic quantitative expression of the fault level. In actual deployment tests, it has demonstrated good real-time performance and stability, effectively supporting the online operation and maintenance strategy and early warning response mechanism of wind farms.

[0175] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent fault diagnosis method for wind turbine generators, characterized in that, Includes the following steps: S1. Real-time acquisition of wind turbine monitoring datasets generated during wind turbine operation; S2. Preprocess the wind turbine monitoring dataset to generate preprocessed monitoring data; S3. Input the preprocessed monitoring data into the temporal convolutional network model. The temporal convolutional network model is a multi-layer network structure composed of dilated causal convolution and attention residual modules based on diagnostic perception factors. S4. Use the improved Snake optimization algorithm to perform global search and local fine-tuning of key hyperparameters in the temporal convolutional network model, and use the improved hyperparameters updated by the Snake optimization algorithm to construct an optimized temporal convolutional network model. S5. The preprocessed monitoring data is used to train the optimized temporal convolutional network model. During the training process, a pre-set loss function is used to compare the fault prediction value output by the optimized temporal convolutional network model with the actual fault state to obtain the wind turbine gear tooth breakage fault diagnosis model. S6. The fault characteristics of wind turbine monitoring data are analyzed using the wind turbine gear tooth breakage fault diagnosis model, and real-time tooth breakage fault diagnosis results are generated. S7. When the real-time broken tooth fault diagnosis result reaches or exceeds the preset abnormal threshold, the fault alarm and early warning mechanism will be automatically triggered, and the diagnosis result and alarm information will be fed back to the operation and maintenance personnel.

2. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 1, characterized in that, S1 includes the following steps: S11. Sensor groups are installed at key parts of the wind turbine gear transmission system. The sensor groups consist of vibration sensors, speed sensors and temperature sensors, which are used to collect vibration signals, speed signals and temperature signals during the operation of the wind turbine. S12. Set a uniform sampling time window to collect corresponding multidimensional signal data at each moment, and construct the wind turbine monitoring dataset D. raw : Where, d i This represents the monitoring data at the i-th sampling time, where N is the total number of sampling points, and v i Indicates at time t i The amplitude of the collected vibration signal, r i Indicates at time t i The acquired rotational speed signal value, θ i Indicates at time t i The collected temperature signal value, t i Let i be the i-th sampling time point.

3. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 2, characterized in that, S2 includes the following steps: S21. Based on wind turbine monitoring dataset D raw Noise removal operations were performed on vibration signals, rotation speed signals and temperature signals respectively. The signal sequence was processed by a sliding window filtering method. Local averaging was performed on data from multiple consecutive time points to smooth short-term fluctuations and suppress random noise, generating a denoised signal sequence. S22. Normalize the denoised signal sequence. Using the maximum and minimum values ​​of each signal within the entire sampling time window as references, linearly transform the monitored value at each moment according to its position relative to the minimum and maximum values, so that all signal values ​​are mapped to the [0,1] interval. S23. Standardize the normalized signal sequence. Calculate the mean and standard deviation of each signal at all times. Then subtract the mean from each monitoring value and divide by the standard deviation so that each signal dimension has a distribution characteristic of zero mean and one standard deviation after standardization. S24. Recombine the standardized signal sequences to construct the preprocessed monitoring data D. pre : Where, d′ i This represents a single standardized monitoring data point constructed at the i-th sampling time. Indicates the sampling time t i The vibration signal value after standardization. Indicates the sampling time t i The speed signal value after standardization. Indicates the sampling time t i The temperature signal value after standardization.

4. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 3, characterized in that, S3 includes the following steps: S31. Transfer the preprocessed monitoring dataset D pre The input is fed into a temporal convolutional network model, which is a multi-layer network structure consisting of dilated causal convolution and attention residual modules based on dynamic diagnostic perception factors. S32. In the convolution processing module of the temporal convolutional network model, a dilated causal convolution structure is used for each time step t. i Corresponding input feature vector Feature extraction is performed to obtain the convolutional output representation of the l-th layer: in, This represents the convolution output at the i-th time step in the l-th convolutional layer. This represents the time position id in the (l-1)th convolutional layer. l · eigenvectors of k, Let d represent the weight coefficient of the l-th convolutional kernel at the k-th position. l represents the dilation factor of the l-th layer, and K represents the kernel size of each dilated convolution layer; S33. Based on the response differences of wind turbine gear tooth breakage fault signals across different monitoring channels, a dynamic diagnostic sensing factor α is constructed. i Dynamic diagnostic perception factor α i Reflecting the salience of diagnostic perception at that moment: Where, μ v μ r μ θ These are the global averages of the vibration signal, rotational speed signal, and temperature signal, respectively. S34. Construct an attention residual module based on dynamic diagnostic perceptual factors, introducing a weighted residual connection after the output of each convolutional layer, and then applying the convolutional output... With input feature vector By combining dynamic diagnostic sensing factor fusion, an enhanced residual representation is obtained: in, Let φ(·) represent the enhanced residual of the l-th layer after introducing the dynamic diagnostic sensing factor, and let φ(·) represent the nonlinear activation function. S35. Stack the attention residual modules in multiple layers to obtain the final high-dimensional feature sequence. in, This represents the final extracted deep diagnostic feature sequence of broken gear teeth in wind turbine units, where F is the final feature dimension and L is the maximum depth of the attention residual module.

5. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 4, characterized in that, S4 includes the following steps: S41. Define the key structural hyperparameters for constructing a temporal convolutional network model as the set of optimization variables P = {k, L, d}; S42. Define the fitness function The fitness function serves as a comprehensive performance evaluation metric for the wind turbine gear tooth breakage diagnosis model on the validation set. It combines overall diagnostic accuracy with early warning capability, ensuring the overall discriminative performance of the temporal convolutional network model while also considering its sensitivity to early abnormal signals. in, Indicates the use of hyperparameter combinations The accuracy of the model in diagnosing broken teeth on the validation set, where ω1 and ω2 are weighting coefficients. Indicates the use of hyperparameter combinations Sensitivity indicators for early diagnosis; S43. To address the need for real-time early warning of broken tooth faults in wind turbines, the Snake optimization algorithm is improved by introducing a dynamic adjustment term based on changes in early diagnosis sensitivity into the standard serpentine search update. This dynamic adjustment term updates each candidate hyperparameter combination. Where γ is the search displacement scaling factor, ω is the search cycle control frequency, φ is the initial phase offset, t represents the current iteration round, and λ is the fault response adjustment coefficient; all are dimensionless control parameters. Indicates hyperparameter combination In early diagnostic sensitivity index S ew Local performance gradient on: Used to guide the search in a direction that allows the model to capture broken tooth signals more quickly; S44. Repeat steps S42-S43 until the preset iteration limit T is reached. max Alternatively, it may satisfy the convergence criterion and ultimately output the optimal hyperparameter combination P. opt ={k * ,L * ,d * }; S45. Utilizing the optimal hyperparameter combination P opt An optimized temporal convolutional network model was constructed. After being optimized by the improved Snake optimization algorithm, the overall accuracy and early warning sensitivity of the optimized temporal convolutional network model in the fault diagnosis of broken teeth in wind turbine gears were improved.

6. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 5, characterized in that, The early diagnostic sensitivity index S ew (P i This reflects the need for early warning capabilities in intelligent fault diagnosis of wind turbine units. The early diagnosis sensitivity index is defined as the threshold value within a preset early response time limit T. ew Within the context, the proportion of accurate early identification of tooth breakage events by the temporal convolutional network model is as follows: Where, N early (P i ) indicates that in the hyperparameter combination P i The model below has a preset early response time T. ew The number of correctly identified broken tooth faults, N fault This represents the total number of tooth breakage fault events in the verification set.

7. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 1, characterized in that, S5 includes the following steps: S51. Transfer the preprocessed monitoring dataset D pre The input is fed into a temporal convolutional network model optimized using the improved Snake optimization algorithm to construct the training sample input sequence. in, Corresponding time t i eigenvectors; S52. Set the corresponding label sequence Among them, y i ∈{0,1} indicates that the i-th sample is at sampling time t i To indicate whether there is a broken gear tooth fault, 0 indicates normal status and 1 indicates broken tooth status. The labels are from manual annotation or historical fault record information. S53. During the training phase, a supervised learning method is used to train the optimized temporal convolutional network model, and the output of the temporal convolutional network model is defined as the predicted fault probability sequence. in This represents the probability that the model predicts the tooth breakage state at the i-th time step; S54. Define the weighted cross-entropy loss function The weighting mechanism of the weighted cross-entropy loss function is used to measure the difference between the model output and the actual fault state. It is used to address the class imbalance problem caused by the proportion of broken tooth samples being lower than the threshold during the actual operation of wind turbine units. S55. Using weighted cross-entropy loss function To achieve the goal, the gradient descent method is used to backpropagate and optimize all trainable parameters in the temporal convolutional network model, updating the convolutional kernel weights, bias terms, and parameters in the residual connection structure of the temporal convolutional network model until training converges. S56. Define the trained temporal convolutional network model as the wind turbine gear tooth breakage fault diagnosis model M. fault The wind turbine gear tooth breakage fault diagnosis model outputs a corresponding tooth breakage probability prediction sequence based on the input continuous time series monitoring features X.

8. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 7, characterized in that, S6 includes the following steps: S61. Preprocess the monitoring data of the wind turbine collected at the current moment. Input to the wind turbine gear tooth breakage fault diagnosis model M fault Obtain the corresponding fault prediction probability sequence in Represents time t j Predicted probability of a wind turbine being in a broken tooth state; S62. Based on the probability values ​​predicted by the wind turbine gear tooth breakage fault diagnosis model and the fault characteristic distribution in historical monitoring data, construct tooth breakage fault characteristic analysis rules. Integrate three dimensions—vibration amplitude fluctuation rate, speed instability, and temperature anomaly gradient—definition characteristic indicators, including vibration fluctuation rate V. var Speed ​​deviation R dev and temperature change Θ grad ; S63. Construct tooth breakage fault diagnosis rules based on feature indicators, set three types of judgment intervals, and output real-time tooth breakage fault diagnosis results. The tooth breakage fault diagnosis rules are as follows: The normal state is to satisfy And V var <0.01, R dev <3rpm、Θ grad <1℃; The suspicious state is satisfied Or any index satisfies 0.01≤V var <0.05, 3≤R dev <7rpm、1≤Θ grad <2℃; High-risk condition for tooth breakage to meet And at least two indicators satisfy V var ≥0.05, R dev ≥7rpm, Θ grad ≥2℃; S64. Output the fault diagnosis results in a structured form in real time and link them with the wind turbine operation control system to realize dynamic monitoring, status identification and early warning of operation risk of gear tooth breakage fault.

9. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 8, characterized in that, The vibration fluctuation rate V var Within the current analysis window, the vibration signal The variance.

10. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 8, characterized in that, The speed deviation R dev Speed ​​signal Compared with the reference operating speed r ref The average deviation.

11. The intelligent fault diagnosis method for wind turbines based on deep learning according to claim 8, characterized in that, The temperature abrupt change Θ grad This represents the maximum rate of change of the temperature signal per unit time.

Images