An AI-based method for detecting anomalies in rotating components of wind turbines

Abstract

This invention discloses an artificial intelligence-based method for detecting anomalies in rotating wind turbine components. The method comprises the following steps: S1, constructing a set of sound signals from the rotating wind turbine components; S2, generating a set of feature vectors from the sound signals of the rotating wind turbine components; S3, training a deep neural network model for detecting abnormal states of the rotating wind turbine components; S4, using a hierarchical reinforcement learning algorithm to analyze and classify the feature vector set of the rotating wind turbine components. A higher-level strategy determines whether to activate a lower-level strategy based on the wind turbine's operating environment, load conditions, and the overall trend of the sound signal set. The lower-level strategy further determines whether an abnormal state exists based on the feature vector set of the rotating wind turbine components; S5, adjusting the denoising and filtering parameters of the sound signal set in real time; and S6, automatically triggering an alarm mechanism when the lower-level strategy identifies an abnormal state through analysis of the feature vector set. This invention not only reduces the cost of wind turbine anomaly detection but also improves detection accuracy and intelligence.

Description

Technical Field

[0001] This invention relates to the field of wind turbine technology, and in particular to an artificial intelligence-based method for detecting anomalies in rotating components of wind turbines. Background Technology

[0002] In existing technologies, the detection of abnormalities in the rotating components of wind turbines mainly relies on vibration sensors and other mechanical sensing devices. The operating status of the wind turbine is determined by monitoring the vibration signals of the turbine blades, rotor, and bearings. However, the installation and maintenance costs of traditional vibration sensors are high. In large-scale wind farm applications, their hardware configuration is complex and the overall maintenance cost is high. At the same time, due to the limited detection sensitivity of vibration sensors, traditional detection methods can often only detect abnormalities when they are already relatively obvious, making it difficult to provide early warning of faults. This causes potential small problems to gradually evolve into larger equipment failures when not resolved in time, increasing the maintenance difficulty and downtime risk of wind turbines.

[0003] In addition, existing anomaly detection methods make little use of the acoustic signals of the rotating parts of the fan. They usually only target mechanical vibration signals and ignore the noise characteristics generated by the fan during operation. In fact, the rotating parts of the fan generate specific acoustic signals during operation. These acoustic signals are closely related to the operating state of the parts and can reflect early signs of anomalies. Therefore, the existing technology fails to make full use of the acoustic signals of the fan operation for anomaly detection, which limits its detection accuracy and sensitivity.

[0004] Furthermore, the existing technology has a low level of intelligence and mainly relies on manually set thresholds and simple signal analysis methods, lacking adaptive capabilities. With the increasing complexity of wind turbine equipment and the diversification of operating environments, it is difficult to cope with the changes in wind turbines under different loads, wind speeds and environmental noise conditions. This often leads to false alarms and missed alarms, affecting the overall operating efficiency of wind farms.

[0005] In summary, the existing technologies have the following main drawbacks: First, the installation and maintenance costs of traditional vibration sensors are high, making them difficult to adapt to large-scale wind farm applications; second, existing methods are unable to capture early abnormal signals of rotating components of wind turbines, making timely warnings impossible; and finally, existing technologies lack effective utilization of sound signals generated during wind turbine operation and have a low level of intelligence, making them unable to adapt to complex changes in the operating environment. Summary of the Invention

[0006] One objective of this invention is to propose an artificial intelligence-based method for detecting anomalies in rotating components of wind turbines. This invention not only reduces the cost of wind turbine anomaly detection but also improves detection accuracy and intelligence.

[0007] An artificial intelligence-based method for detecting anomalies in rotating components of a wind turbine, according to an embodiment of the present invention, includes the following steps:

[0008] S1. Multiple acoustic sensor nodes are arranged in several key parts of the wind turbine to collect wind turbine blade signals, rotor vibration signals and bearing noise signals generated by the rotating parts of the wind turbine in real time during operation, and to construct a sound signal set of the rotating parts of the wind turbine.

[0009] S2. Preprocess the collected sound signal set of the rotating parts of the wind turbine to remove environmental noise and irrelevant signals, and perform spectrum analysis and time-frequency analysis on the sound signal set of the rotating parts of the wind turbine to obtain the frequency domain characteristics and time domain characteristics of the operation of the rotating parts of the wind turbine, and generate a feature vector set of the sound signal of the rotating parts of the wind turbine.

[0010] S3. Based on the historical operating data of the wind turbine and the feature vector set of the sound signals of the rotating parts of the wind turbine, a hierarchical reinforcement learning algorithm is used to train a deep neural network model. The training process includes a high-level strategy that uses the overall analysis of the sound signal set of the rotating parts of the wind turbine to determine whether further analysis of the feature vector set is needed, and a low-level strategy that performs fine-grained classification and analysis of the feature vector set under the instruction of the high-level strategy, thereby training a deep neural network model for detecting abnormal states of the rotating parts of the wind turbine.

[0011] S4. Input the currently generated feature vector set into the trained deep neural network model, and use the hierarchical reinforcement learning algorithm to analyze and classify the feature vector set of the rotating parts of the wind turbine. The high-level strategy determines whether to activate the low-level strategy based on the wind turbine's operating environment, load conditions, and the overall trend of the sound signal set. The low-level strategy further determines whether there is an abnormal state based on the feature vector set of the rotating parts of the wind turbine.

[0012] S5. Dynamically adjust the sensitivity and operating frequency of the acoustic sensor nodes according to the operating environment parameters of the fan, optimize the acquisition quality of the sound signal set of the fan rotating parts, and adjust the noise reduction and filtering parameters of the sound signal set in real time.

[0013] S6. When the low-level strategy identifies an abnormal state through the analysis of the feature vector set, the system automatically triggers the alarm mechanism, generates a report containing the abnormal type, abnormal location and possible cause, and sends a warning signal to the wind farm's operation and maintenance team.

[0014] Optionally, S1 specifically includes the following steps:

[0015] S11. Arrange acoustic sensor nodes {S1, S2, ..., S...} at multiple key locations of the wind turbine. n}, where n is the number of acoustic sensor nodes, and S is the number of acoustic sensor nodes. iIt is used to collect real-time signals of fan blades, rotor vibration, and bearing noise generated by the rotating components of the fan during operation;

[0016] S12. The acoustic sensor node collects the sound signal of the rotating parts of the fan at fixed time intervals Δt to obtain a time-series signal:

[0017] {x1(t),x2(t),…,x n (t)};

[0018] Where, x i (t) represents the time t when the acoustic sensor node S... i The collected sound signals;

[0019] S13. Construct the sound signal set X(t) of the rotating components of the wind turbine:

[0020] X(t) = [x1(t), x2(t), ..., x n (t)];

[0021] Wherein, X(t) is the set of sound signals of the rotating parts of the wind turbine at time t, which includes multi-channel sound signals collected by n acoustic sensor nodes;

[0022] S14, For each acoustic sensor node S i The acquired signal x i (t extracts the corresponding blade signal, rotor vibration signal, and bearing noise signal, which are denoted as follows)

[0023] S15. Combine the different types of sound signals collected by each acoustic sensor node to construct a multi-channel sound signal set X for the rotating components of the wind turbine. c (t):

[0024]

[0025] Among them, X c (t) is the combined signal set of the rotating parts of the wind turbine at time t, which includes different types of sound signals collected by all acoustic sensor nodes;

[0026] S16. Generate a complete dataset D of acoustic signals from the rotating components of the wind turbine:

[0027] D = {X} c (t1),X c (t2),..,X c (t m )}.

[0028] Optionally, S2 specifically includes the following steps:

[0029] S21. Preprocess the sound signal dataset D of the rotating parts of the fan, and apply a bandpass filter to filter each group of sound signals. The filtering range is [f min ,f max ], where f min f is the lowest effective frequency. max The highest effective frequency is used to remove irrelevant environmental noise, resulting in the filtered acoustic signal dataset D of the fan rotating components. f ;

[0030] S22. Data set D of the filtered sound signal from the rotating components of the fan. f Normalization is performed to ensure that the amplitude of each sound signal is within the same numerical range, resulting in the normalized sound signal dataset D of the fan rotating component. n ;

[0031] S23. The normalized sound signal dataset D of the rotating components of the wind turbine. n Spectrum analysis was performed, and the time-domain signal was converted into a time-frequency domain signal using short-time Fourier transform to obtain the frequency domain characteristics of the rotating components of the wind turbine.

[0032] S24, Data set of sound signals from the rotating parts of the fan (D) n Perform time-domain analysis to extract instantaneous energy E t Temporal characteristics of signal envelope and amplitude variation;

[0033] S25. The frequency domain feature vector V of the rotating parts of the fan. f and time-domain feature vector V t Combined, they constitute the sound signal feature vector set V of the rotating components of the wind turbine. c .

[0034] Optionally, S3 specifically includes the following steps:

[0035] S31. Utilize the historical operation dataset H of the wind turbine and the feature vector set of acoustic signals from the rotating components of the wind turbine V. c Construct a dataset D for training deep neural network models. train :

[0036]

[0037] in, Let be the sound signal feature vector of the i-th rotating component of the wind turbine, containing the sound signal feature vector of the rotating component of the wind turbine in the frequency domain. and time-domain wind turbine rotating component sound signal feature vector y (i) This corresponds to the wind turbine operating status label, with a value of normal (y). (i)=0) or abnormal state (y (i) =1); N is the number of training samples;

[0038] S32. Establish a deep neural network model M for a hierarchical reinforcement learning algorithm, including a high-level policy network π. h and low-level policy network π l :

[0039] High-level policy network π h Feature vector of the input acoustic signal of the rotating part of the wind turbine. A comprehensive assessment was conducted on the high-level policy network π. h Receive the sound signal feature vector of the rotating part of the wind turbine As input, calculate the high-level policy action. The probability value is used to determine whether further analysis is needed:

[0040]

[0041] Among them, W h b is the weight matrix of the high-level policy network, representing the degree of influence of each feature on high-level decision-making; h σ is the bias vector; σ(·) is the activation function; This indicates the probability that higher-level strategies will determine further analysis.

[0042] when When δ is a preset threshold, the higher-level policy determines the activation of the lower-level policy network π. l The sound signal feature vector of the rotating parts of the wind turbine Perform fine-grained classification and analysis;

[0043] Low-level policy network π l When high-level strategy decisions require further analysis, the acoustic signal feature vectors of the rotating components of the wind turbine are examined. Fine-grained classification and abnormal state identification are performed, and the low-level policy network computes low-level policy actions. Predict the specific abnormal conditions of the rotating components of the wind turbine:

[0044]

[0045] Among them, W l b is the weight matrix of the low-level policy network, representing the influence of each feature on different abnormal states; l It is the bias vector; Let K be the probability distribution vector of the K possible abnormal states of the rotating components of the wind turbine;

[0046] S33. Define the overall loss function L, and optimize the parameters of the deep neural network model by combining the losses of high-level and low-level policies:

[0047]

[0048] The first term is the cross-entropy loss of the high-level policy, y (i) The label indicates whether further analysis is needed; λ is the trade-off coefficient; L l For the loss of the low-level strategy, when Time calculation:

[0049]

[0050] Where N' is satisfied The number of samples; Let be the true label of the i-th sample in the k-th type of abnormal state; The probability of the k-th type of abnormal state predicted by the low-level strategy;

[0051] S34. By minimizing the loss function L, gradient descent is used to update the parameters {W} of the high-level policy network. h ,b h} and low-level policy network parameters {W l ,b l}, and a deep neural network model M for detecting abnormal states of rotating components of the wind turbine is obtained through training.

[0052] Optionally, S4 specifically includes the following steps:

[0053] S41. The feature vector set of the sound signal generated by the current rotating component of the wind turbine. The input is fed into the trained deep neural network model M;

[0054] S42, High-Level Policy Network π h Receive the current feature vector And based on the current environmental parameters E of the wind turbine operation and the overall changing trend ΔV of the sound signal set of the rotating parts of the wind turbine. c Calculate high-level strategy actions The probability value is used to determine whether to activate the lower-level strategy for further analysis;

[0055] S43, When the high-level strategy is output When δ is a preset threshold, the low-level policy network π is activated. l Current feature vector of the rotating component of the wind turbine Fine-grained analysis and classification are performed to further determine whether there are any abnormal conditions in the rotating parts of the fan;

[0056] S44, Low-level policy network π l Receive the current feature vector And output the low-level strategy action. Predict the specific or abnormal conditions of the rotating components of the wind turbine;

[0057] S45. Based on the state probability distribution output by the low-level policy network Determine if the rotating parts of the fan are in an abnormal state, and the probability of a certain state. satisfy Where ∈ is the threshold for identifying abnormal states, the rotating components of the wind turbine are considered to be in an abnormal state.

[0058] Optionally, S6 specifically includes the following steps:

[0059] S61, when the low-level policy network π l Based on the feature vector of the current rotating components of the wind turbine When an abnormal state is detected, the system automatically triggers an alarm mechanism and generates an alarm report containing abnormal information. The alarm report includes the following:

[0060] Exception type T error : Probability distribution output by the low-level policy network The abnormal state type T identified in the middle error The highest probability corresponding to the abnormal state Where ∈ is the threshold for anomaly detection;

[0061] Abnormal location P error Determine the abnormal location P based on the characteristics of the abnormal signal source in the sound signal of the rotating parts of the fan. error ;

[0062] Anomaly occurrence time t error Record the time point t when the anomaly is identified. error To assist the operations and maintenance team in tracking the timeline of abnormal events;

[0063] Possible causes of the abnormality R error Based on eigenvector analysis, possible causes of anomalies, R, are inferred. error By combining historical data and low-level strategies, the identification of abnormal features can provide possible inferences about the causes of failures.

[0064] S62, Report the alarm containing the abnormal information to R alarm The report is sent to the wind farm operation and maintenance team and transmitted to the operation and maintenance management system via the communication network, and relevant personnel are notified in the form of a warning signal.

[0065] S63. After a warning signal is generated, the system records and stores information about the abnormal event, including the abnormality type T. error Abnormal location P error Time of occurrence t error and possible reasons R errorThe system automatically updates the historical operating dataset H of the wind turbines as reference data for future wind turbine maintenance and anomaly analysis;

[0066] S64. After the maintenance team receives a warning signal and takes relevant maintenance measures, the system continuously monitors the operating status of the rotating parts of the wind turbine. If the abnormal status is repaired, the system will generate a status recovery report and record the status after maintenance. If the abnormality is not repaired, the system will continue to send reminder signals.

[0067] Optionally, the identification of possible causes of the anomaly includes when the low-level policy network π l When an abnormal state is detected in the rotating parts of the fan, the sound signal feature vector of the rotating parts in the current abnormal state is extracted:

[0068]

[0069] in, This represents the frequency domain feature vector of the abnormal state. This represents the temporal feature vector of the abnormal state;

[0070] Extract a set of historical anomaly feature vectors from the historical operation dataset H of wind turbines:

[0071]

[0072] in, Let M be the sound signal feature vector of the rotating component of the wind turbine for the h-th historical anomaly, and M be the number of historical anomaly records.

[0073] Calculate the current anomaly feature vector With each historical anomaly feature vector Similarity S between h Determine the historical anomaly feature vector with the highest similarity. in,

[0074] Extract from historical anomaly records Corresponding abnormal reasons Consider this as a possible cause of the current anomaly (R). error The inferred possible causes of the anomaly R error Log to an anomaly report R alarm middle.

[0075] The beneficial effects of this invention are:

[0076] (1) This invention optimizes the deep neural network model by using a hierarchical reinforcement learning algorithm, and achieves efficient multi-level decision-making when detecting abnormalities in the rotating parts of the wind turbine. The high-level strategy decides in real time whether to activate the low-level strategy for fine-grained analysis based on the overall operating status of the wind turbine and environmental conditions. Under the guidance of the high-level strategy, the low-level strategy accurately judges the abnormal state through fine-grained analysis of the feature vector. The multi-level decision-making mechanism not only improves the sensitivity of detection and can capture the early abnormal signals of the wind turbine in a timely manner, avoiding the missed and false alarms caused by the reliance on a single threshold in traditional methods, but also significantly improves the accuracy and efficiency of wind turbine abnormality detection.

[0077] (2) This invention innovatively introduces voiceprint recognition technology and combines deep learning models to perform refined analysis of the sound signals of the rotating parts of the wind turbine. By extracting frequency and time domain features from the collected sound signals to form a high-dimensional feature vector set, it can not only analyze minute acoustic changes but also effectively filter environmental noise interference. The combination of voiceprint recognition and multi-dimensional feature extraction significantly enhances the sensitivity of anomaly detection, enabling the system to provide early warning of potential problems in the early stages of wind turbine failure by identifying subtle changes in the voiceprint signal, thus avoiding the limitation of traditional vibration detection methods that cannot detect anomalies in the early stages.

[0078] (3) Based on feature vector analysis and historical abnormal dataset, this invention intelligently infers the possible causes of the current abnormality of the rotating parts of the wind turbine through similarity calculation and generates an abnormality report. By comparing the similarity between the current abnormality features and historical data, it quickly provides the inference of the cause of the fault, reducing the time and effort required for traditional manual diagnosis and improving the efficiency of fault location and handling. The report details the abnormality type, location, time and cause, which makes it easy for the operation and maintenance team to quickly take targeted maintenance measures, thereby significantly improving the overall operation efficiency of the wind farm and reducing the downtime and maintenance cost of the wind turbine. Attached Figure Description

[0079] 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:

[0080] Figure 1 This is a flowchart of an artificial intelligence-based method for detecting anomalies in rotating components of a wind turbine, as proposed in this invention.

[0081] Figure 2 This is a schematic diagram of the training process of a deep neural network model based on a hierarchical reinforcement learning algorithm in an artificial intelligence-based method for detecting anomalies in rotating components of a wind turbine proposed in this invention. Detailed Implementation

[0082] 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.

[0083] refer to Figure 1-2 An artificial intelligence-based method for detecting anomalies in rotating components of a wind turbine includes the following steps:

[0084] S1. Multiple acoustic sensor nodes are arranged in several key parts of the wind turbine to collect wind turbine blade signals, rotor vibration signals and bearing noise signals generated by the rotating parts of the wind turbine in real time during operation, and to construct a sound signal set of the rotating parts of the wind turbine.

[0085] S2. Preprocess the collected sound signal set of the rotating parts of the wind turbine to remove environmental noise and irrelevant signals, and perform spectrum analysis and time-frequency analysis on the sound signal set of the rotating parts of the wind turbine to obtain the frequency domain characteristics and time domain characteristics of the operation of the rotating parts of the wind turbine, and generate a feature vector set of the sound signal of the rotating parts of the wind turbine.

[0086] S3. Based on the historical operating data of the wind turbine and the feature vector set of the sound signals of the rotating parts of the wind turbine, a hierarchical reinforcement learning algorithm is used to train a deep neural network model. The training process includes a high-level strategy that uses the overall analysis of the sound signal set of the rotating parts of the wind turbine to determine whether further analysis of the feature vector set is needed, and a low-level strategy that performs fine-grained classification and analysis of the feature vector set under the instruction of the high-level strategy, thereby training a deep neural network model for detecting abnormal states of the rotating parts of the wind turbine.

[0087] S4. Input the currently generated feature vector set into the trained deep neural network model, and use the hierarchical reinforcement learning algorithm to analyze and classify the feature vector set of the rotating parts of the wind turbine. The high-level strategy determines whether to activate the low-level strategy based on the wind turbine's operating environment, load conditions, and the overall trend of the sound signal set. The low-level strategy further determines whether there is an abnormal state based on the feature vector set of the rotating parts of the wind turbine.

[0088] S5. Dynamically adjust the sensitivity and operating frequency of the acoustic sensor nodes according to the operating environment parameters of the fan, optimize the acquisition quality of the sound signal set of the fan rotating parts, and adjust the noise reduction and filtering parameters of the sound signal set in real time.

[0089] S6. When the low-level strategy identifies an abnormal state through the analysis of the feature vector set, the system automatically triggers the alarm mechanism, generates a report containing the abnormal type, abnormal location and possible cause, and sends a warning signal to the wind farm's operation and maintenance team.

[0090] In this embodiment, S1 specifically includes the following steps:

[0091] S11. Arrange acoustic sensor nodes {S1, S2, ..., S...} at multiple key locations of the wind turbine. n}, where n is the number of acoustic sensor nodes, and S is the number of acoustic sensor nodes. i It is used to collect real-time signals of fan blades, rotor vibration, and bearing noise generated by the rotating components of the fan during operation;

[0092] S12. The acoustic sensor node collects the sound signal of the rotating parts of the fan at fixed time intervals Δt to obtain a time-series signal:

[0093] {x1(t),x2(t),…,x n (t)};

[0094] Where, x i (t) represents the time t when the acoustic sensor node S... i The collected sound signals;

[0095] S13. Construct the sound signal set X(t) of the rotating components of the wind turbine:

[0096] X(t) = [x1(t), x2(t), ..., x n (t)];

[0097] Wherein, X(t) is the set of sound signals of the rotating parts of the wind turbine at time t, which includes multi-channel sound signals collected by n acoustic sensor nodes;

[0098] S14, For each acoustic sensor node S i The acquired signal x i (t extracts the corresponding blade signal, rotor vibration signal, and bearing noise signal, which are denoted as follows)

[0099] S15. Combine the different types of sound signals collected by each acoustic sensor node to construct a multi-channel sound signal set X for the rotating components of the wind turbine. c (t):

[0100]

[0101] Among them, X c (t) is the combined signal set of the rotating parts of the wind turbine at time t, which includes different types of sound signals collected by all acoustic sensor nodes;

[0102] S16. Generate a complete dataset D of acoustic signals from the rotating components of the wind turbine:

[0103] D = {X} c (t1),X c (t2),…,X c(t m )}.

[0104] In this embodiment, S2 specifically includes the following steps:

[0105] S21. Preprocess the sound signal dataset D of the rotating parts of the fan, and apply a bandpass filter to filter each group of sound signals. The filtering range is [f min ,f max ], where f min f is the lowest effective frequency. max The highest effective frequency is used to remove irrelevant environmental noise, resulting in the filtered acoustic signal dataset D of the fan rotating components. f ;

[0106] S22. Data set D of the filtered sound signal from the rotating components of the fan. f Normalization is performed to ensure that the amplitude of each sound signal is within the same numerical range, resulting in the normalized sound signal dataset D of the fan rotating component. n ;

[0107] S23. The normalized sound signal dataset D of the rotating components of the wind turbine. n Spectrum analysis was performed, and the time-domain signal was converted into a time-frequency domain signal using short-time Fourier transform to obtain the frequency domain characteristics of the rotating components of the wind turbine.

[0108] S24, Data set of sound signals from the rotating parts of the fan (D) n Perform time-domain analysis to extract instantaneous energy E t Temporal characteristics of signal envelope and amplitude variation;

[0109] S25. The frequency domain feature vector V of the rotating parts of the fan. f and time-domain feature vector V t Combined, they constitute the sound signal feature vector set V of the rotating components of the wind turbine. c .

[0110] In this embodiment, S3 specifically includes the following steps:

[0111] S31. Utilize the historical operation dataset H of the wind turbine and the feature vector set of acoustic signals from the rotating components of the wind turbine V. c Construct a dataset D for training deep neural network models. train :

[0112]

[0113] in, Let be the sound signal feature vector of the i-th rotating component of the wind turbine, containing the sound signal feature vector of the rotating component of the wind turbine in the frequency domain. and time-domain wind turbine rotating component sound signal feature vector y (i) This corresponds to the wind turbine operating status label, with a value of normal (y). (i) =0) or abnormal state (y (i) =1); N is the number of training samples;

[0114] S32. Establish a deep neural network model M for a hierarchical reinforcement learning algorithm, including a high-level policy network π. h and low-level policy network π l :

[0115] High-level policy network π h Feature vector of the input acoustic signal of the rotating part of the wind turbine. A comprehensive assessment was conducted on the high-level policy network π. h Receive the sound signal feature vector of the rotating part of the wind turbine As input, calculate the high-level policy action. The probability value is used to determine whether further analysis is needed:

[0116]

[0117] Among them, W h b is the weight matrix of the high-level policy network, representing the degree of influence of each feature on high-level decision-making; h σ is the bias vector; σ(·) is the activation function; This indicates the probability that higher-level strategies will determine further analysis.

[0118] when When δ is a preset threshold, the higher-level policy determines the activation of the lower-level policy network π. l The sound signal feature vector of the rotating parts of the wind turbine Perform fine-grained classification and analysis;

[0119] Low-level policy network π l When high-level strategy decisions require further analysis, the acoustic signal feature vectors of the rotating components of the wind turbine are examined. Fine-grained classification and abnormal state identification are performed, and the low-level policy network computes low-level policy actions. Predict the specific abnormal conditions of the rotating components of the wind turbine:

[0120]

[0121] Among them, W l b is the weight matrix of the low-level policy network, representing the influence of each feature on different abnormal states; l It is the bias vector; Let K be the probability distribution vector of the K possible abnormal states of the rotating components of the wind turbine;

[0122] S33. Define the overall loss function L, and optimize the parameters of the deep neural network model by combining the losses of high-level and low-level policies:

[0123]

[0124] The first term is the cross-entropy loss of the high-level policy, y (i) The label indicates whether further analysis is needed; λ is the trade-off coefficient; L l For the loss of the low-level strategy, when Time calculation:

[0125]

[0126] Where N' is satisfied The number of samples; Let be the true label of the i-th sample in the k-th type of abnormal state; The probability of the k-th type of abnormal state predicted by the low-level strategy;

[0127] S34. By minimizing the loss function L, gradient descent is used to update the parameters {W} of the high-level policy network. h ,b h} and low-level policy network parameters {W l ,b l}, and a deep neural network model M for detecting abnormal states of rotating components of the wind turbine is obtained through training.

[0128] In this embodiment, S4 specifically includes the following steps:

[0129] S41. The feature vector set of the sound signal generated by the current rotating component of the wind turbine. The input is fed into the trained deep neural network model M;

[0130] S42, High-Level Policy Network π h Receive the current feature vector And based on the current environmental parameters E of the wind turbine operation and the overall changing trend ΔV of the sound signal set of the rotating parts of the wind turbine. c Calculate high-level strategy actions The probability value is used to determine whether to activate the lower-level strategy for further analysis;

[0131] S43, When the high-level strategy is output When δ is a preset threshold, the low-level policy network π is activated. l Current feature vector of the rotating component of the wind turbine Fine-grained analysis and classification are performed to further determine whether there are any abnormal conditions in the rotating parts of the fan;

[0132] S44, Low-level policy network π l Receive the current feature vector And output the low-level strategy action. Predict the specific or abnormal conditions of the rotating components of the wind turbine;

[0133] S45. Based on the state probability distribution output by the low-level policy network Determine if the rotating parts of the fan are in an abnormal state, and the probability of a certain state. satisfy Where ∈ is the threshold for identifying abnormal states, the rotating components of the wind turbine are considered to be in an abnormal state.

[0134] In this embodiment, S6 specifically includes the following steps:

[0135] S61, when the low-level policy network π l Based on the feature vector of the current rotating components of the wind turbine When an abnormal state is detected, the system automatically triggers an alarm mechanism and generates an alarm report containing abnormal information. The alarm report includes the following:

[0136] Exception type T error : Probability distribution output by the low-level policy network The abnormal state type T identified in the middle error The highest probability corresponding to the abnormal state Where ∈ is the threshold for anomaly detection;

[0137] Abnormal location P error Determine the abnormal location P based on the characteristics of the abnormal signal source in the sound signal of the rotating parts of the fan. error ;

[0138] Anomaly occurrence time t error Record the time point t when the anomaly is identified. error To assist the operations and maintenance team in tracking the timeline of abnormal events;

[0139] Possible causes of the abnormality R error Based on eigenvector analysis, possible causes of anomalies, R, are inferred. error By combining historical data and low-level strategies, the identification of abnormal features can provide possible inferences about the causes of failures.

[0140] S62, Report the alarm containing the abnormal information to R alarm The report is sent to the wind farm operation and maintenance team and transmitted to the operation and maintenance management system via the communication network, and relevant personnel are notified in the form of a warning signal.

[0141] S63. After a warning signal is generated, the system records and stores information about the abnormal event, including the abnormality type T.error Abnormal location P error Time of occurrence t error and possible reasons R error The system automatically updates the historical operating dataset H of the wind turbines as reference data for future wind turbine maintenance and anomaly analysis;

[0142] S64. After the maintenance team receives a warning signal and takes relevant maintenance measures, the system continuously monitors the operating status of the rotating parts of the wind turbine. If the abnormal status is repaired, the system will generate a status recovery report and record the status after maintenance. If the abnormality is not repaired, the system will continue to send reminder signals.

[0143] In this embodiment, the identification of possible causes of anomalies includes when the lower-level policy network π l When an abnormal state is detected in the rotating parts of the fan, the sound signal feature vector of the rotating parts in the current abnormal state is extracted:

[0144]

[0145] in, This represents the frequency domain feature vector of the abnormal state. This represents the temporal feature vector of the abnormal state;

[0146] Extract a set of historical anomaly feature vectors from the historical operation dataset H of wind turbines:

[0147]

[0148] in, Let M be the sound signal feature vector of the rotating component of the wind turbine for the h-th historical anomaly, and M be the number of historical anomaly records.

[0149] Calculate the current anomaly feature vector With each historical anomaly feature vector Similarity S between h Determine the historical anomaly feature vector with the highest similarity. in,

[0150] Extract from historical anomaly records Corresponding abnormal reasons Consider this as a possible cause of the current anomaly (R). error The inferred possible causes of the anomaly R error Log to an anomaly report R alarm middle.

[0151] Example 1:

[0152] To verify the feasibility and effectiveness of this invention in detecting anomalies in rotating components of wind turbines, researchers used two real wind turbine operation datasets: WindFarm-Dataset-A and WindFarm-Dataset-B. These two datasets recorded the operating status of wind turbines in two different wind farms. WindFarm-Dataset-A contained wind turbine operation data from January to December 2021, including wind speed, load, vibration signals, and sound signals, totaling data from 50 wind turbines. WindFarm-Dataset-B contained operating data from July 2020 to July 2021 from 100 wind turbines, primarily recording sound signals and operating environment information of the rotating components of the wind turbines. The researchers divided the two datasets into training and testing sets in an 8:2 ratio.

[0153] In terms of sound signal feature extraction, the researchers used a pre-trained ResNet50 model to extract features from the acoustic signature signal during wind turbine operation, generating an embedding dimension of 2048. To adapt to the deep learning model of this invention, the researchers introduced a 2D convolutional layer to convert the acoustic signature feature dimension from 2048 to 256, and combined the results of frequency domain feature extraction to construct the final feature vector set.

[0154] In terms of text data, it includes the wind turbine's operation logs and environmental parameter records (wind speed, temperature, and load information). The researchers used a pre-trained BERT model to extract features from the operation logs, and the extracted text embedding dimension was 768.

[0155] The deep learning model of this invention is based on a hierarchical reinforcement learning algorithm. The high-level policy network of the model is responsible for analyzing the overall operating status of the wind turbine and determining whether further analysis of the feature vector is needed. The low-level policy network performs fine-grained analysis of the feature vector after the high-level policy is activated, and finally outputs the anomaly detection result. The experimental model was trained using the Adam optimizer for a total of 100 epochs with an initial learning rate of 0.001 and a batch size of 128.

[0156] To verify the detection effect of this invention, the researchers first selected a wind turbine that had been running for many years from the WindFarm-Dataset-A dataset as the experimental object. On the training set, the method of this invention combined the features of sound signals and vibration signals to perform multimodal feature extraction and classification training. In the test set, the wind turbine operated under different wind speeds and load conditions. After two months of continuous monitoring, the detection model of this invention successfully detected an abnormal bearing wear event during the operation of the wind turbine rotor. The specific process is as follows:

[0157] In October 2021, when the wind turbine was running under high load, the system collected abnormal rotor vibration and sound signals. The vibration amplitude and sound frequency peaks fluctuated, and the instantaneous energy value increased significantly. The high-level strategy of hierarchical reinforcement learning determined that the anomaly required further analysis. The low-level strategy was activated to perform fine-grained classification of the feature vector set. By comparing with historical anomaly datasets, the model identified signal features that were highly similar to bearing wear. Finally, an anomaly report was generated, and it was inferred that the bearing may be excessively worn and needs to be lubricated or replaced immediately.

[0158] The alarm report generated by the system was pushed to the wind farm operation and maintenance team within 2 hours after the anomaly was detected. The team carried out maintenance in a timely manner, which ultimately prevented the wind turbine from being shut down for a long time. In contrast, the traditional vibration sensor only issued an alarm signal when a fault was about to occur. The operation and maintenance team failed to react in time, which caused another wind turbine monitored by the traditional method to be shut down for 8 hours.

[0159] To demonstrate the superiority of the method of this invention, the researchers compared the detection performance of the method of this invention with that of traditional vibration sensor methods on two datasets, WindFarm-Dataset-A and WindFarm-Dataset-B. The researchers used accuracy, precision, recall, and weighted F1 score as evaluation metrics.

[0160] Table 1 compares the results of the proposed method with the traditional method on WindFarm-Dataset-A.

[0161] Monitoring methods accuracy accuracy Recall rate Weighted F1 score Traditional vibration sensor methods 84.5％ 81.2％ 76.8％ 78.9％ The method of the present invention 93.2％ 91.5％ 90.8％ 91.1％

[0162] Table 2 shows the comparison results between the method of this invention and the traditional method on WindFarm-Dataset-B.

[0163] Monitoring methods accuracy accuracy Recall rate Weighted F1 score Traditional vibration sensor methods 82.1％ 79.7％ 74.3％ 76.9％ The method of the present invention 94.6％ 92.8％ 91.3％ 92.0％

[0164] As can be seen from Tables 1 and 2, the method of the present invention outperforms the traditional vibration sensor method on both datasets, especially in terms of recall and weighted F1 score. This indicates that the present invention can not only detect abnormalities in the rotating parts of the fan more accurately, but also reduce false alarms and missed alarms, thus greatly improving the reliability and effectiveness of the system.

[0165] The above experiments demonstrate that the method of this invention significantly outperforms traditional vibration sensor methods in detecting anomalies during actual wind turbine operation. This invention can not only effectively capture minute anomalies during wind turbine operation through multimodal feature extraction, but also achieve more intelligent multi-level decision-making through hierarchical reinforcement learning algorithms, thereby improving the sensitivity and accuracy of detection. In data experiments from multiple real wind farms, the method of this invention has shown good detection performance, effectively reducing wind turbine downtime and lowering maintenance costs.

[0166] Taking WindFarm-Dataset-A as an example, the wind farm's operation and maintenance team successfully identified and handled four abnormal events involving the rotating components of the wind turbine within two months by applying the method of this invention, avoiding at least 24 hours of wind turbine downtime. Compared with traditional methods, this improved maintenance efficiency and reduced operation and maintenance costs by about 15%. The effectiveness of this invention has been fully demonstrated in practical applications, providing strong technical support for the safe and efficient operation of wind farms.

[0167] This invention optimizes a deep neural network model using a hierarchical reinforcement learning algorithm, achieving efficient multi-level decision-making when detecting anomalies in rotating components of a wind turbine. The high-level strategy determines in real time whether to activate the low-level strategy for fine-grained analysis based on the overall operating status of the wind turbine and environmental conditions. Under the guidance of the high-level strategy, the low-level strategy accurately judges the abnormal state through fine-grained analysis of feature vectors. The multi-level decision-making mechanism not only improves the sensitivity of detection and can capture early abnormal signals of the wind turbine in a timely manner, avoiding the missed and false alarms caused by the reliance on a single threshold in traditional methods, but also significantly improves the accuracy and efficiency of wind turbine anomaly detection.

[0168] This invention innovatively introduces voiceprint recognition technology, combining it with a deep learning model to perform refined analysis of the sound signals of the rotating components of a wind turbine. By extracting frequency and time domain features from the collected sound signals to form a high-dimensional feature vector set, it can not only analyze minute acoustic changes but also effectively filter environmental noise interference. The combination of voiceprint recognition and multi-dimensional feature extraction significantly enhances the sensitivity of anomaly detection, enabling the system to provide early warnings of potential problems in the early stages of wind turbine failure by identifying subtle changes in the voiceprint signal, thus avoiding the limitations of traditional vibration detection methods that cannot detect anomalies in the early stages.

[0169] This invention, based on feature vector analysis combined with historical anomaly datasets, intelligently infers the possible causes of current wind turbine rotating component anomalies through similarity calculations and generates anomaly reports. By comparing the similarity between current anomaly features and historical data, it quickly provides fault cause inferences, reducing the time and effort required for traditional manual diagnosis and improving the efficiency of fault location and handling. The report details the anomaly type, location, time, and cause, facilitating the operation and maintenance team to quickly take targeted maintenance measures, thereby significantly improving the overall operational efficiency of the wind farm and reducing wind turbine downtime and maintenance costs.

[0170] 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 method for detecting anomalies in rotating components of a wind turbine based on artificial intelligence, characterized in that, Comprising the following steps: S1, arranging a plurality of acoustic sensor nodes at a plurality of key positions of the fan, for collecting fan blade signals, rotor vibration signals and bearing noise signals generated by the fan rotating components during operation in real time, and constructing a fan rotating component sound signal set; S2, preprocessing the collected fan rotating component sound signal set to remove environmental noise and irrelevant signals, and performing frequency spectrum analysis and time-frequency analysis on the fan rotating component sound signal set to obtain frequency domain features and time domain features of the fan rotating component operation, and generating a fan rotating component sound signal feature vector set; S3, training a deep neural network model based on fan historical operation data and fan rotating component sound signal feature vector set using a hierarchical reinforcement learning algorithm, the training process including a high-level strategy determining whether further analysis of the feature vector set is needed through overall analysis of the fan rotating component sound signal set, and a low-level strategy performing fine-grained classification and analysis of the feature vector set under the instruction of the high-level strategy, and training a deep neural network model for detecting abnormal states of the fan rotating component; S3 specifically comprises the following steps: S31, utilize the fan historical operation data set H and the fan rotating component sound signal feature vector set , construct a data set for training a deep neural network model : ； wherein, is the i-th fan rotating component sound signal feature vector, including a frequency domain fan rotating component sound signal feature vector and a time domain fan rotating component sound signal feature vector ; is the corresponding fan operating state label, taking a normal state or an abnormal state ; N is the number of training samples; S32, a deep neural network model M of the hierarchical reinforcement learning algorithm is established, including a high-level policy network and a low-level policy network : High-level strategy network Feature vector of the input acoustic signal of the rotating part of the wind turbine. Conduct an overall assessment, high-level strategy network Receive the sound signal feature vector of the rotating part of the wind turbine As input, calculate the high-level policy action. The probability value is used to determine whether further analysis is needed: ； wherein, is a weight matrix of the high-level policy network, representing the influence degree of each feature on the high-level decision; is a bias vector; is an activation function; represents the probability of the high-level policy decision to further analyze. When wherein is a preset threshold, and a high-level policy determines to activate a low-level policy network , the fan rotating part sound signal feature vector is subjected to fine-grained classification and analysis; Low-level policy network When the high-level policy decision needs further analysis, the fan rotating part sound signal feature vector is classified in fine granularity and the abnormal state is identified, the low-level policy network calculates the low-level policy action , and the specific abnormal state of the fan rotating part is predicted. ； wherein, is a weight matrix of the low-level policy network, representing the influence of each feature on different abnormal states; is a bias vector; is a probability distribution vector of K possible abnormal states of the rotating part of the fan. S33, defining an overall loss function L, and optimizing the deep neural network model parameters by combining the loss of the high-level strategy and the low-level strategy: ； where the first term is the cross-entropy loss of the high-level policy, is a label for whether further analysis is actually needed or not; is a trade-off coefficient; is the loss of the low-level policy, when is calculated as: ； wherein N' is the number of samples satisfying is the true label of the kth abnormal state of the ith sample; is the probability of the kth abnormal state predicted by the low-level strategy.​ S34, update the high-level policy network parameters by minimizing the loss function L using gradient descent and the low-level policy network parameters , to obtain a deep neural network model M for detecting abnormal states of rotating components of a fan. S4, inputting the currently generated feature vector set into the trained deep neural network model, and analyzing and classifying the fan rotating component feature vector set using a hierarchical reinforcement learning algorithm, the high-level strategy determining whether to activate the low-level strategy according to the operating environment, load conditions and overall trend of the sound signal set, and the low-level strategy further determining whether there is an abnormal state according to the fan rotating component feature vector set; S5, dynamically adjusting the sensitivity and working frequency of the acoustic sensor nodes according to the operating environment parameters of the fan, optimizing the collection quality of the fan rotating component sound signal set, and adjusting the denoising and filtering parameters of the sound signal set in real time; S6, when the low-level strategy identifies an abnormal state through analysis of the feature vector set, the system automatically triggers an alarm mechanism, generates a report containing the abnormal type, abnormal position and possible causes, and sends a warning signal to the operation and maintenance team of the wind farm. 2.The fan rotating component abnormality detection method based on artificial intelligence according to claim 1, characterized in that, S1 specifically comprises the following steps: S11, arranging acoustic sensor nodes at multiple key positions of the fan wherein n is the number of acoustic sensor nodes, the acoustic sensor nodes are used to collect fan blade signals, rotor vibration signals and bearing noise signals generated by rotating parts of the fan during operation in real time. S12, the acoustic sensor node acquires the sound signal of the rotating part of the fan at a fixed time interval The sound signal of the rotating part of the fan is collected to obtain a time series signal: ； wherein, represents a sound signal collected at time t by an acoustic sensor node ​ S13, constructing a sound signal set of the rotating component of the fan : ； wherein, a set of sound signals of the rotating part of the fan at time t, comprising multi-channel sound signals collected by n acoustic sensor nodes; S14, for each acoustic sensor node collected signal x i (t) extract the corresponding blade signal, rotor vibration signal and bearing noise signal, respectively denoted as ; S15, combine different types of sound signals collected by each acoustic sensor node to construct a multi-channel sound signal set of the rotating part of the fan : ； wherein, is a combined signal set of the fan rotating components at time t, containing different types of sound signals collected by all acoustic sensor nodes; S16, generating a complete fan rotating component sound signal data set D: 。 3.The fan rotating component abnormality detection method based on artificial intelligence according to claim 1, characterized in that, S2 specifically comprises the following steps: S21, pre-process the fan rotating part sound signal data set D, apply a band-pass filter to each group of sound signals, the filtering range is , wherein, is the lowest effective frequency, is the highest effective frequency, used to remove irrelevant environmental noise, to obtain the filtered fan rotating part sound signal data set ; S22, the fan rotating component sound signal dataset after filtering processing The normalization processing is performed to make the amplitude of each sound signal in the same numerical range, and the normalized fan rotating component sound signal dataset is obtained ; S23, normalizing the fan rotating component sound signal dataset Spectrums are analyzed, time domain signals are converted into time-frequency domain signals by using short-time Fourier transform, and frequency domain features of the fan rotating component are obtained. S24, wind turbine rotating component sound signal data set Temporal analysis is performed to extract the instantaneous energy , signal envelope and amplitude variation time domain features; S25, the frequency domain feature vector of the fan rotating component and the time domain feature vector are combined to form a feature vector set of the fan rotating component sound signal . 4.The fan rotating component abnormality detection method based on artificial intelligence according to claim 1, characterized in that, S4 specifically comprises the following steps: S41, generating a fan rotating component sound signal feature vector set of the current fan rotating component into the trained deep neural network model M; S42, high-level policy network receiving the current feature vector and according to the environmental parameters E of the current fan operation and the overall trend of the fan rotating component sound signal set calculating the probability value of the high-level policy action to determine whether to activate the low-level policy for further analysis S43, When the high-level strategy is output At that time, among them Activate the low-level policy network based on the preset threshold. Current feature vector of the rotating component of the wind turbine Fine-grained analysis and classification are performed to further determine whether there are any abnormal conditions in the rotating parts of the fan; S44, low-level policy network receiving the current feature vector and outputting a low-level policy action , predicting a specific state or an abnormal state of the rotating component of the fan S45, state probability distribution output by the low layer policy network , determine whether the rotating component of the fan is in an abnormal state, if the probability of a state satisfies , wherein is a recognition threshold value of the abnormal state, and it is determined that the rotating component of the fan is in an abnormal state. 5.The fan rotating component abnormality detection method based on artificial intelligence according to claim 1, characterized in that, S6 specifically comprises the following steps: S61、when the low-layer policy network According to the feature vector of the current fan rotating part When the abnormal state is identified, the system automatically triggers an alarm mechanism, generates an alarm report containing abnormal information, and the alarm report includes the following contents: Anomaly type : probability distribution output by the low-level policy network Anomaly state type identified in Highest probability corresponding to the anomaly state wherein Threshold for anomaly identification; Abnormal position : determining an abnormal position according to the characteristics of the abnormal signal source in the sound signal of the fan rotating part ; Time of abnormal occurrence : Time point when the abnormality is identified , to assist the operation and maintenance team to track the timeline of the abnormal event; Possible causes of anomaly : Inferred possible causes of anomaly according to feature vector analysis , combined with historical data and low-level strategy to identify anomaly features to provide possible fault cause inference; S62, reporting an alarm containing abnormal information is sent to the wind farm operation and maintenance team, reported to the operation and maintenance management system through the communication network and notified to the relevant personnel in the form of a warning signal; S63, after the generation of the warning signal, the system records and stores information of the abnormal event, including the abnormal type , the abnormal location , the occurrence time and the possible cause As the reference data for future wind turbine maintenance and abnormal analysis, the system automatically updates the historical operation data set H of the wind turbine; S64, after the operation and maintenance team receives the warning signal and takes relevant maintenance measures, the system continuously monitors the operating state of the fan rotating component, if the abnormal state is repaired, the system will generate a state recovery report and record the state after maintenance, if the abnormality is not repaired, the system will continue to send a reminder signal. 6.The fan rotating component abnormality detection method based on artificial intelligence according to claim 1, characterized in that, The identification of the possible causes of the anomaly includes when the low-level policy network When the fan rotating component is identified as being in an abnormal state, a fan rotating component sound signal feature vector of the current abnormal state is extracted: ； wherein, a frequency domain feature vector of the abnormal state, a time domain feature vector of the abnormal state; Extracting a set of historical abnormal feature vectors from the fan historical operation data set H: ； wherein, is the fan rotating component sound signal feature vector for the hth historical anomaly, and M is the number of historical anomaly records. computing a current anomaly feature vector a similarity between each historical anomaly feature vector determining a historical anomaly feature vector with the highest similarity wherein ;​ extracting from historical anomaly records corresponding anomaly causes as a possible cause of the current anomaly the inferred possible anomaly causes into the anomaly report .

Images