Fan running state anomaly detection method

Abstract

The invention discloses a fan operation state anomaly detection method. The method comprises the following steps: S1, generating a fan operation voiceprint signal data set; s2, preprocessing the fan operation voiceprint signal data set; s3, constructing a feature data set; s4, performing preliminary classification on the fan operation voiceprint signals by using a hierarchical clustering algorithm based on the feature data set; s5, performing secondary subdivision in each fan operation state cluster by further using a hierarchical clustering algorithm; s6, establishing a fan normal operation voiceprint signal feature model based on the fan voiceprint signal feature data of the normal state sub-cluster, and establishing a fan potential abnormal state voiceprint signal feature model based on the voiceprint signal feature data of the potential abnormal state sub-cluster; s7, judging the running state of the fan by the system; and S8, when the running state of the fan is detected to be abnormal, the system sends out an alarm signal. The method has remarkable beneficial effects in the aspects of fan operation state monitoring and anomaly detection.

Description

Technical Field

[0001] The present invention relates to the technical field of fans, and in particular to a method for detecting abnormal operation status of a fan. Background Art

[0002] In the prior art, the monitoring of the operating status of fans and the detection of abnormalities mainly rely on physical sensors, such as vibration sensors and temperature sensors. Vibration sensors and temperature sensors determine whether the fan is in normal operation by monitoring the mechanical vibration or temperature changes of the fan. Although vibration sensor and temperature sensor technologies have been widely used, there are still some limitations. First, vibration sensors and temperature sensors are easily affected by external environmental factors, such as wind speed, temperature changes and electromagnetic interference, which may lead to false alarms or missed alarms. In addition, physical sensors require regular maintenance and calibration, which increases the operating and maintenance costs. The problems of sensor failure and decreased accuracy are particularly prominent during the long-term operation of the fan.

[0003] In addition, traditional anomaly detection methods are often based on simple statistical analysis or threshold judgment, which makes it difficult to capture complex mechanical failures or tiny abnormal signals in wind turbine operation. They have limitations when dealing with complex mechanical systems and cannot accurately identify tiny abnormal signals in the early fault detection of key components such as wind turbine blades and bearings. As the scale and complexity of wind turbines continue to increase, traditional physical sensor monitoring methods and anomaly detection methods based on threshold judgment have been unable to meet the needs of modern wind turbine operation monitoring.

[0004] The fan operation status monitoring and anomaly detection method based on voiceprint signals provides a non-destructive and non-contact solution. By collecting and analyzing the voiceprint signals generated during the operation of the fan, the reliance on physical sensors can be effectively avoided. However, the application of existing voiceprint signal monitoring technology in the field of fans is still in its early stages. Simple signal processing methods are usually used and cannot fully utilize the rich information in the voiceprint signals. In terms of complex changes in the fan operation status and small anomaly detection, the existing technology has the problems of insufficient detection accuracy and limited feature extraction capabilities. In addition, the existing fan voiceprint monitoring technology lacks a complete model establishment and anomaly recognition capabilities, making it difficult to accurately classify and analyze different fan operation statuses.

[0005] In summary, the existing technology has the following main problems in monitoring the operating status of wind turbines: first, traditional physical sensors are highly dependent and easily interfered by environmental factors, which increases maintenance costs and failure risks; second, the anomaly detection method based on simple threshold judgment is difficult to capture tiny fault signals, affecting the accuracy of fault detection; finally, the existing voiceprint signal monitoring technology has deficiencies in feature extraction and state classification, and it is difficult to cope with complex wind turbine operating conditions and potential abnormal situations. Summary of the invention

[0006] An object of the present invention is to provide a method for detecting abnormal operation status of a fan. The present invention has significant beneficial effects in monitoring the operation status of the fan and detecting abnormalities.

[0007] A method for detecting abnormal operation status of a fan according to an embodiment of the present invention comprises the following steps:

[0008] S1. Multiple audio acquisition devices are arranged at multiple key locations of the fan to collect the fan blade rotation signal, bearing noise signal and fan mechanical vibration signal generated during the operation of the fan in real time, and generate a fan operation soundprint signal data set;

[0009] S2. Preprocess the fan operation voiceprint signal dataset to remove environmental noise, filter background sound signals irrelevant to the fan operation, and normalize the voiceprint signal to obtain a preprocessed fan operation voiceprint signal dataset;

[0010] S3, extracting features from the preprocessed fan operation soundprint signal data set to construct a feature data set;

[0011] S4. Based on the feature data set, a hierarchical clustering algorithm is used to preliminarily classify the fan operation soundprint signal, and the fan operation soundprint signal data under different operating conditions are divided into multiple fan operation state clusters, each state cluster corresponds to a different fan operation mode;

[0012] S5. Further use a hierarchical clustering algorithm to perform secondary subdivision in each wind turbine operation status cluster, identify and cluster the slight feature changes in the voiceprint signal, and further subdivide the voiceprint signal feature data in each operation status cluster into a normal state sub-cluster and a potential abnormal state sub-cluster;

[0013] S6. Establishing a normal operation soundprint signal feature model of the fan based on the soundprint signal feature data of the normal state sub-cluster, and establishing a potential abnormal state soundprint signal feature model of the fan based on the soundprint signal feature data of the potential abnormal state sub-cluster;

[0014] S7. During the actual operation of the fan, the voiceprint signal of the fan's current operation is collected in real time, and the extracted voiceprint signal features are compared with the pre-established normal operation voiceprint signal feature model of the fan. When a significant difference is detected between the voiceprint signal features and the normal operation voiceprint signal feature model of the fan, the system determines that the fan has entered an abnormal operation state;

[0015] S8. When the operating status of the fan is detected as abnormal, the system sends out an alarm signal and analyzes the specific fault type of the fan based on the characteristic model of the soundprint signal of the fan's potential abnormal state, including fan blade failure, bearing damage or loose mechanical parts.

[0016] Optionally, the S1 specifically includes:

[0017] S11. Arrange multiple audio collection devices near the blades, bearings and other key mechanical parts of the fan for real-time collection of soundprint signals generated during the operation of the fan;

[0018] S12. The voiceprint signal includes the following three types of signals:

[0019] The blade rotation signals are collected at the blade root, middle and tip of the blade:

[0020] X l (t) = {x l1 (t),x l2 (t),…,x ln (t)};

[0021] Among them, X l (t) is the blade rotation signal set, x li (t) is the blade rotation signal collected by the collection device located at the i-th position of the blade at time t;

[0022] Bearing noise signals are collected from the main bearing and support bearing:

[0023] X b (t) = {x b1 (t),x b2 (t),…,x bm (t)};

[0024] Among them, X b (t) is the bearing noise signal set, x bi (t) is the noise signal collected by the collection device located at the i-th position of the bearing at time t;

[0025] Mechanical vibration signals are collected in the gearbox and nacelle:

[0026] X v (t) = {x v1 (t),x v2 (t),…,x vk (t)};

[0027] Among them, X v (t) is the mechanical vibration signal set, x vi (t) is the mechanical vibration signal collected by the collection device located at the i-th position of the gearbox or nacelle at time t;

[0028] S13, by real-time storage of the collected blade rotation signal, bearing noise signal and mechanical vibration signal, a fan operation soundprint signal data set X(t) is generated:

[0029] X(t) = {X l (t),X b (t),X v (t)}.

[0030] Optionally, the S2 specifically includes:

[0031] S21, performing noise filtering on the fan operation voiceprint signal dataset X(t) to remove the environmental noise N(t) in the collected signal, where the environmental noise includes the natural noise around the fan and the equipment interference signal, and obtaining the processed fan operation voiceprint signal dataset X'(t);

[0032] S22, filtering out background sound signals B(t) irrelevant to the operation of the fan, and removing the background signals from the denoised voiceprint signals through a signal separation algorithm, to obtain a fan operation voiceprint signal dataset X"(t) after background signal filtering;

[0033] S23, normalizing the filtered fan operation soundprint signal dataset X”(t), and normalizing the amplitude values ​​of the soundprint signals collected by different collection devices to obtain a normalized fan operation soundprint signal dataset X n (t);

[0034] S24, the normalized fan operation soundprint signal dataset X is processed by signal enhancement technology n (t) Improve key acoustic features, enhance specific frequency and time domain feature signals, and obtain the preprocessed fan operation soundprint signal dataset X p (t).

[0035] Optionally, the S3 specifically includes:

[0036] S31, short-time Fourier transform is used to pre-process the fan operation soundprint signal dataset X p (t) Perform time-frequency analysis to obtain the time-frequency feature matrix S(f,t);

[0037] S32. Extract the key features of energy spectral density P(f) and Mel-frequency cepstrum coefficient M(n) from the time-frequency feature matrix S(f,t) and construct the feature vector c i :

[0038] c i =[P(f i ),M(n i ),…],i=1,2,…,N;

[0039] Where N is the number of eigenvectors, P(f i ) is the frequency f i The energy spectral density ati ) is the nth i Mel-frequency cepstral coefficients, c i is the characteristic vector of the fan operation soundprint signal at the i-th moment;

[0040] S33, all the eigenvectors c i Composition feature data set C = {c1, c2, ..., c N}, standardize the feature data set C to eliminate the influence of different feature dimensions and scales, and obtain the final fan operation soundprint signal feature data set C f .

[0041] Optionally, the S4 specifically includes:

[0042] S41, based on the final fan operation soundprint signal feature data set C f Calculate the similarity matrix D between feature vectors = [d ij ],d ij represents the feature vector c i With c j The Mahalanobis distance between them is used to eliminate the influence of different feature dimensions and correlation on the distance calculation, and is used to measure the differences between the feature vectors of the soundprint signals of different fans:

[0043]

[0044] Among them, c i and c j is the characteristic vector of the voiceprint signal collected by the fan at different time points, Σ is the characteristic data set C f The covariance matrix of reflects the correlation between acoustic features, (·) T Represents a vector transpose operation;

[0045] S42. Use the similarity matrix D to perform hierarchical clustering analysis, and use the agglomerative hierarchical clustering algorithm to build a clustering tree T of the fan operation soundprint signal. Starting from each feature vector, the closest clusters are merged step by step, and the feature data set C is clustered by setting the clustering threshold θ. f Divided into m wind turbine operation status clusters G = {G1, G2, ..., G m}, the clustering threshold θ is determined according to the distribution characteristics of the distance values ​​in the similarity matrix D:

[0046] θ=μ D +λ·σ D ;

[0047] Among them, μ D is the average value of all distance values ​​in the similarity matrix D, σ D is the standard deviation of the distance value, and λ is an adjustment parameter used to control the granularity of clustering;

[0048] By adjusting the value of λ, a balance is achieved between the number of clusters and the similarity within the clusters, so that the clustering results are consistent with the actual distribution of the fan operation soundprint signal;

[0049] S43, for each fan operation status cluster G i Calculate its feature center As the representative voiceprint feature vector of the running state cluster:

[0050]

[0051] Among them, |G i | is the state cluster G i The number of eigenvectors in , is the state cluster G i The feature center of each state cluster G i Corresponding to a fan operation mode, including a normal operation mode, a slight abnormality mode or a serious abnormality mode;

[0052] S44. Feature centers based on different state clusters For each state cluster G i The corresponding fan operation mode is given a specific meaning, and each operation status cluster is defined as the corresponding fan operation mode:

[0053] Mode definition = {G1→normal mode, G2→slightly abnormal mode, …, G m →Severe abnormal mode}.

[0054] Optionally, the S5 specifically includes:

[0055] S51, for each wind turbine operating status cluster G i Calculate the weighted cosine distance similarity matrix between feature vectors within a cluster in Represents the feature vector and The weighted cosine distance between them is used to measure the small differences between the feature vectors of the voiceprint signals within the cluster:

[0056]

[0057] in, and is the state cluster G i The characteristic vector of the voiceprint signal in reflects the running voiceprint characteristics of the fan at different time points, w i For state cluster G iThe feature weight vector is weighted according to the importance of each acoustic feature for anomaly detection. ⊙ represents the Hadamard product of the corresponding elements, ∥·∥ represents the bi-norm of the vector, and · represents the vector dot product operation.

[0058] S52, using the similarity matrix D i For state cluster G i The characteristic data in the voiceprint are analyzed by secondary hierarchical clustering, and the density peak clustering algorithm is used to identify the density peak points of the data in the cluster to capture the small characteristic changes in the voiceprint signal;

[0059] Calculate each eigenvector The local density ρ k :

[0060]

[0061] Among them, |G i | is the state cluster G i The number of eigenvectors in , d c is the cutoff distance, which determines the influence range of local density calculation. Usually, the similarity matrix D i The mean of the median distance values, ρ k Reflects the eigenvector The density of surrounding data, used to identify areas with high density;

[0062] Calculate each eigenvector The minimum distance δ k :

[0063]

[0064] Among them, δ k Represents the feature vector The minimum distance between all eigenvectors with higher density than it, when When the local density is the highest, let δ k is the similarity matrix D i The maximum distance value in k To measure the feature vector Whether it is at the peak density position;

[0065] S53, according to the local density ρ k and the minimum distance δ k The distribution characteristics of the threshold value ρ i and δ i , the state cluster G i Subdivided into normal state sub-cluster G iN and potential abnormal state sub-cluster G iA , when the condition ρ is satisfied k >ρ i And δk >δ i When , the feature vector Marked as the density peak point, belonging to the potential abnormal state subcluster G iA , indicating that there are abnormal voiceprint features, and the remaining feature vectors belong to the normal state subcluster G iN ;

[0066] S54, complete the state cluster G i The secondary subdivision of iN and potential abnormal state sub-cluster G iA , further optimize the ability to identify minor anomalies in voiceprint signal feature data;

[0067] S55, calculate the normal state sub-cluster G respectively iN and potential abnormal state sub-cluster G iA The feature center of the normal state subcluster

[0068]

[0069] Potential abnormal state sub-cluster feature center

[0070]

[0071] Among them, |G iN | and |G iA | are the number of eigenvectors in the normal state sub-cluster and the potential abnormal state sub-cluster, respectively. and is the characteristic center of the corresponding sub-cluster, representing the typical soundprint characteristics of the fan in normal operation and potential abnormal state.

[0072] Optionally, the S6 specifically includes:

[0073] S61, based on normal state sub-cluster G iN The fan soundprint signal feature data is used to establish the normal operation soundprint signal feature model M of the fan N , calculate the normal state subcluster G iN The mean of the eigenvectors in and the covariance matrix Σ iN , and the mean vector and the covariance matrix Σ iN Construct the characteristic model M of the soundprint signal of the normal operation of the fan N = It is used to describe the statistical characteristics of the fan soundprint signal under normal operating conditions;

[0074] S62, based on potential abnormal state sub-cluster G iAThe characteristic data of the fan soundprint signal is used to establish the characteristic model M of the fan potential abnormal state soundprint signal A , calculate the potential abnormal state sub-cluster G iA The mean of the eigenvectors in and the covariance matrix Σ iA , the mean vector and the covariance matrix Σ iA Construct the characteristic model M of the soundprint signal of the potential abnormal state of the fan A =(c iA ,Σ iA ), which is used to describe the statistical characteristics of the fan soundprint signal under potential abnormal conditions.

[0075] Optionally, the S7 specifically includes:

[0076] S71. During the actual operation of the fan, the current voiceprint signal of the fan is collected in real time, and the collected voiceprint signal is preprocessed to obtain a preprocessed voiceprint signal data set of the current operation;

[0077] S72, extracting feature vectors from the currently running pre-processed voiceprint signal data set to construct a feature data set of the current fan operation voiceprint signal, which is used to compare the fan normal operation voiceprint signal feature model M in real time N and potential abnormal state voiceprint signal feature model M A ;

[0078] S73, calculating the characteristic vector of the soundprint signal of the currently running fan and the characteristic model M of the soundprint signal of the normal operation of the fan N The Mahalanobis distance d of the feature mean N (t), determine whether the current voiceprint signal is in a normal operating state;

[0079] S74, when the Mahalanobis distance d N (t) exceeds the preset threshold θ N When the system determines that the current fan operation status deviates from the normal range, it enters an abnormal state:

[0080] d N (t)>θ N ;

[0081] Among them, θ N The abnormal detection threshold is set according to the distribution of voiceprint feature data in normal operating state, and is used to determine whether the fan has entered an abnormal state;

[0082] S75. If the soundprint signal of the fan currently in operation is judged to be in an abnormal state, further calculate the current feature vector and the feature model M of the soundprint signal of the fan potential abnormal state A The Mahalanobis distance d of the feature mean A(t), determine the specific type of anomaly;

[0083] S76, according to the Mahalanobis distance d A (t) size and specific abnormal state model M A Based on the distribution characteristics of the wind turbine, the system classifies and determines the type of anomaly, including wind turbine blade failure, bearing damage, or loose mechanical parts.

[0084] The beneficial effects of the present invention are:

[0085] (1) The present invention introduces a voiceprint signal feature extraction method that combines short-time Fourier transform and Mel-frequency cepstrum coefficients, which effectively improves the accuracy of feature extraction. The time-frequency features are obtained by short-time Fourier transform and combined with energy spectrum density and MFCC to extract key features, so that even the slight changes in voiceprint signals during the operation of the fan can be accurately captured. Compared with the traditional physical sensor method, the present invention achieves high-sensitivity detection of the operating status of the fan without contacting the device. By standardizing and enhancing the extracted voiceprint signal, the problem of inconsistent signal amplitudes in different environments is overcome, thereby ensuring the uniformity and accuracy of the signal and greatly improving the robustness of the system.

[0086] (2) The present invention adopts an innovative combination of hierarchical clustering algorithm and density peak clustering algorithm to perform fine-grained classification and anomaly detection on the fan operation status. The fan operation voiceprint signal is divided into multiple operation status clusters through hierarchical clustering, and each status cluster is further subdivided using the density peak clustering algorithm to accurately identify small changes in the voiceprint signal. The fan operation status is further subdivided into normal state subclusters and potential abnormal state subclusters by calculating the local density and the minimum distance to identify the density peak point, thereby improving the accuracy of anomaly detection. In the early fault detection of fan blades and bearing components, potential anomalies can be identified in advance to reduce false alarms and missed alarms.

[0087] (3) The present invention realizes comparative analysis of the real-time operating status of the fan by establishing a characteristic model of the normal operating soundprint signal of the fan and a characteristic model of the potential abnormal state soundprint signal. When the system collects the soundprint signal of the current fan in real time, it can compare the Mahalanobis distance with the pre-established normal operating model to accurately determine whether the fan has entered an abnormal state. By calculating the characteristic mean and covariance matrix under the abnormal state, the present invention further refines the identification of abnormal types, and can effectively distinguish specific problems such as fan blade failure, bearing damage or loose mechanical parts, thereby improving the efficiency and accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0089] Figure 1 A flow chart of a method for detecting abnormal operation status of a fan proposed by the present invention;

[0090] Figure 2 This is a schematic diagram of cluster classification of fan operating status based on voiceprint signal feature extraction and hierarchical clustering algorithm in a fan operating status abnormality detection method proposed by the present invention. DETAILED DESCRIPTION

[0091] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0092] refer to Figure 1-2 A method for detecting abnormal operation status of a fan comprises the following steps:

[0093] S1. Multiple audio acquisition devices are arranged at multiple key locations of the fan to collect the fan blade rotation signal, bearing noise signal and fan mechanical vibration signal generated during the operation of the fan in real time, and generate a fan operation soundprint signal data set;

[0094] S2. Preprocess the fan operation voiceprint signal dataset to remove environmental noise, filter background sound signals irrelevant to the fan operation, and normalize the voiceprint signal to obtain a preprocessed fan operation voiceprint signal dataset;

[0095] S3, extracting features from the preprocessed fan operation soundprint signal data set to construct a feature data set;

[0096] S4. Based on the feature data set, a hierarchical clustering algorithm is used to preliminarily classify the fan operation soundprint signal, and the fan operation soundprint signal data under different operating conditions are divided into multiple fan operation state clusters, each state cluster corresponds to a different fan operation mode;

[0097] S5. Further use a hierarchical clustering algorithm to perform secondary subdivision in each wind turbine operation status cluster, identify and cluster the slight feature changes in the voiceprint signal, and further subdivide the voiceprint signal feature data in each operation status cluster into a normal state sub-cluster and a potential abnormal state sub-cluster;

[0098] S6. Establishing a normal operation soundprint signal feature model of the fan based on the soundprint signal feature data of the normal state sub-cluster, and establishing a potential abnormal state soundprint signal feature model of the fan based on the soundprint signal feature data of the potential abnormal state sub-cluster;

[0099] S7. During the actual operation of the fan, the voiceprint signal of the fan's current operation is collected in real time, and the extracted voiceprint signal features are compared with the pre-established normal operation voiceprint signal feature model of the fan. When a significant difference is detected between the voiceprint signal features and the normal operation voiceprint signal feature model of the fan, the system determines that the fan has entered an abnormal operation state;

[0100] S8. When the operating status of the fan is detected as abnormal, the system sends out an alarm signal and analyzes the specific fault type of the fan based on the characteristic model of the soundprint signal of the fan's potential abnormal state, including fan blade failure, bearing damage or loose mechanical parts.

[0101] In this implementation, S1 specifically includes:

[0102] S11. Arrange multiple audio collection devices near the blades, bearings and other key mechanical parts of the fan for real-time collection of soundprint signals generated during the operation of the fan;

[0103] S12. Voiceprint signals include the following three types of signals:

[0104] The blade rotation signals are collected at the blade root, middle and tip of the blade:

[0105] X l (t) = {x l1 (t),x l2 (t),…,x ln (t)};

[0106] Among them, X l (t) is the blade rotation signal set, x li (t) is the blade rotation signal collected by the collection device located at the i-th position of the blade at time t;

[0107] Bearing noise signals are collected from the main bearing and support bearing:

[0108] X b (t) = {x b1 (t),x b2 (t),…,x bm (t)};

[0109] Among them, X b (t) is the bearing noise signal set, x bi (t) is the noise signal collected by the collection device located at the i-th position of the bearing at time t;

[0110] Mechanical vibration signals are collected in the gearbox and nacelle:

[0111] X v (t) = {x v1 (t),xv2 (t),…,x vk (t)};

[0112] Among them, X v (t) is the mechanical vibration signal set, x vi (t) is the mechanical vibration signal collected by the collection device located at the i-th position of the gearbox or nacelle at time t;

[0113] S13, by real-time storage of the collected blade rotation signal, bearing noise signal and mechanical vibration signal, a fan operation soundprint signal data set X(t) is generated:

[0114] X(t) = {X l (t),X b (t),X v (t)}.

[0115] In this implementation, S2 specifically includes:

[0116] S21, performing noise filtering on the fan operation voiceprint signal dataset X(t) to remove the environmental noise N(t) in the collected signal, where the environmental noise includes the natural noise around the fan and the equipment interference signal, and obtaining the processed fan operation voiceprint signal dataset X'(t);

[0117] S22, filtering out background sound signals B(t) irrelevant to the operation of the fan, and removing the background signals from the denoised voiceprint signals through a signal separation algorithm, to obtain a fan operation voiceprint signal dataset X"(t) after background signal filtering;

[0118] S23, normalizing the filtered fan operation soundprint signal dataset X”(t), and normalizing the amplitude values ​​of the soundprint signals collected by different collection devices to obtain a normalized fan operation soundprint signal dataset X n (t);

[0119] S24, the normalized fan operation soundprint signal dataset X is processed by signal enhancement technology n (t) Improve key acoustic features, enhance specific frequency and time domain feature signals, and obtain the preprocessed fan operation soundprint signal dataset X p (t).

[0120] In this implementation, S3 specifically includes:

[0121] S31, short-time Fourier transform is used to pre-process the fan operation soundprint signal dataset X p (t) Perform time-frequency analysis to obtain the time-frequency feature matrix S(f,t);

[0122] S32. Extract the key features of energy spectral density P(f) and Mel frequency cepstrum coefficient M(n) from the time-frequency feature matrix S(f,t) and construct the feature vector c i :

[0123] c i =[P(f i ),M(n i ),…],i=1,2,…,N;

[0124] Where N is the number of eigenvectors, P(f i ) is the frequency f i The energy spectral density at i ) is the nth i Mel-frequency cepstral coefficients, c i is the characteristic vector of the fan operation soundprint signal at the i-th moment;

[0125] S33, all the eigenvectors c i Composition feature data set C = {c1, c2, ..., c N}, standardize the feature data set C to eliminate the influence of different feature dimensions and scales, and obtain the final fan operation soundprint signal feature data set C f .

[0126] In this implementation, S4 specifically includes:

[0127] S41, based on the final fan operation soundprint signal feature data set C f Calculate the similarity matrix D between feature vectors = [d ij ],d ij represents the feature vector c i With c j The Mahalanobis distance between them is used to eliminate the influence of different feature dimensions and correlation on the distance calculation, and is used to measure the differences between the feature vectors of the soundprint signals of different fans:

[0128]

[0129] Among them, c i and c j is the characteristic vector of the voiceprint signal collected by the fan at different time points, Σ is the characteristic data set C f The covariance matrix of reflects the correlation between acoustic features, (·) T Represents a vector transpose operation;

[0130] S42. Use the similarity matrix D to perform hierarchical clustering analysis, and use the agglomerative hierarchical clustering algorithm to build a clustering tree T of the fan operation soundprint signal. Starting from each feature vector, the closest clusters are merged step by step, and the feature data set C is clustered by setting the clustering threshold θ. f Divided into m wind turbine operation status clusters G = {G1, G2, ..., G m}, the clustering threshold θ is determined according to the distribution characteristics of the distance values ​​in the similarity matrix D:

[0131] θ=μ D +λ·σ D ;

[0132] Among them, μ D is the average value of all distance values ​​in the similarity matrix D, σ D is the standard deviation of the distance value, and λ is an adjustment parameter used to control the granularity of clustering;

[0133] By adjusting the value of λ, a balance is achieved between the number of clusters and the similarity within the clusters, so that the clustering results are consistent with the actual distribution of the fan operation soundprint signal;

[0134] S43, for each fan operation status cluster G i Calculate its feature center c i , as the representative voiceprint feature vector of the running state cluster:

[0135]

[0136] Among them, |G i | is the state cluster G i The number of eigenvectors in , is the state cluster G i The feature center of each state cluster G i Corresponding to a fan operation mode, including a normal operation mode, a slight abnormality mode or a serious abnormality mode;

[0137] S44, Feature Center c Based on Different State Clusters i For each state cluster G i The corresponding fan operation mode is given a specific meaning, and each operation status cluster is defined as the corresponding fan operation mode:

[0138] Mode definition = {G1→normal mode, G2→slightly abnormal mode, …, G m →Severe abnormal mode}.

[0139] In this implementation, S5 specifically includes:

[0140] S51, for each wind turbine operating status cluster G i Calculate the weighted cosine distance similarity matrix between feature vectors within a cluster in Represents the feature vector and The weighted cosine distance between them is used to measure the small differences between the feature vectors of the voiceprint signals within the cluster:

[0141]

[0142] in, and is the state cluster G i The characteristic vector of the voiceprint signal in reflects the running voiceprint characteristics of the fan at different time points, w i For state cluster G i The feature weight vector is weighted according to the importance of each acoustic feature for anomaly detection. ⊙ represents the Hadamard product of the corresponding elements, ∥·∥ represents the bi-norm of the vector, and · represents the vector dot product operation.

[0143] S52, using the similarity matrix D i For state cluster G i The characteristic data in the voiceprint are analyzed by secondary hierarchical clustering, and the density peak clustering algorithm is used to identify the density peak points of the data in the cluster to capture the small characteristic changes in the voiceprint signal;

[0144] Calculate each eigenvector The local density ρ k :

[0145]

[0146] Among them, |G i | is the state cluster G i The number of eigenvectors in , d c is the cutoff distance, which determines the influence range of local density calculation. Usually, the similarity matrix D i The mean of the median distance values, ρ k Reflects the eigenvector The density of surrounding data, used to identify areas with high density;

[0147] Calculate each eigenvector The minimum distance δ k :

[0148]

[0149] Among them, δ k Represents the feature vector The minimum distance between and all eigenvectors with higher density than it, when When the local density is the highest, let δ k is the similarity matrix D iThe maximum distance value in k To measure the feature vector Whether it is at the peak density position;

[0150] S53, according to the local density ρ k and the minimum distance δ k The distribution characteristics of the threshold value ρ i and δ i , the state cluster G i Subdivided into normal state sub-cluster G iN and potential abnormal state sub-cluster G iA , when the condition ρ is satisfied k >ρ i And δ k >δ i When , the feature vector Marked as the density peak point, belonging to the potential abnormal state subcluster G iA , indicating that there are abnormal voiceprint features, and the remaining feature vectors belong to the normal state subcluster G iN ;

[0151] S54, complete the state cluster G i The secondary subdivision of iN and potential abnormal state sub-cluster G iA , further optimize the ability to identify minor anomalies in voiceprint signal feature data;

[0152] S55, calculate the normal state sub-cluster G respectively iN and potential abnormal state sub-cluster G iA The feature center of the normal state subcluster

[0153]

[0154] Potential abnormal state sub-cluster feature center

[0155]

[0156] Among them, |G iN | and |G iA | are the number of eigenvectors in the normal state sub-cluster and the potential abnormal state sub-cluster, respectively. and is the characteristic center of the corresponding sub-cluster, representing the typical soundprint characteristics of the fan in normal operation and potential abnormal state.

[0157] In this implementation, S6 specifically includes:

[0158] S61, based on normal state sub-cluster G iNThe fan soundprint signal feature data is used to establish the normal operation soundprint signal feature model M of the fan N , calculate the normal state subcluster G iN The mean of the eigenvectors in and the covariance matrix Σ iN , and the mean vector and the covariance matrix Σ iN Construct the characteristic model M of the soundprint signal of the normal operation of the fan N = It is used to describe the statistical characteristics of the fan soundprint signal under normal operating conditions;

[0159] S62, based on potential abnormal state sub-cluster G iA The characteristic data of the fan soundprint signal is used to establish the characteristic model M of the fan potential abnormal state soundprint signal A , calculate the potential abnormal state sub-cluster G iA The mean of the eigenvectors in and the covariance matrix Σ iA , the mean vector and the covariance matrix Σ iA Constructing the characteristic model of the soundprint signal of the potential abnormal state of the fan Used to describe the statistical characteristics of fan soundprint signal characteristics under potential abnormal conditions.

[0160] In this implementation, S7 specifically includes:

[0161] S71. During the actual operation of the fan, the current voiceprint signal of the fan is collected in real time, and the collected voiceprint signal is preprocessed to obtain a preprocessed voiceprint signal data set of the current operation;

[0162] S72, extracting feature vectors from the currently running pre-processed voiceprint signal data set to construct a feature data set of the current fan operation voiceprint signal, which is used to compare the fan normal operation voiceprint signal feature model M in real time N and potential abnormal state voiceprint signal feature model M A ;

[0163] S73, calculating the characteristic vector of the soundprint signal of the currently running fan and the characteristic model M of the soundprint signal of the normal operation of the fan N The Mahalanobis distance d of the feature mean N (t), determine whether the current voiceprint signal is in a normal operating state;

[0164] S74, when the Mahalanobis distance d N (t) exceeds the preset threshold θ N When the system determines that the current fan operation status deviates from the normal range, it enters an abnormal state:

[0165] d N(t)>θ N ;

[0166] Among them, θ N The abnormal detection threshold is set according to the distribution of voiceprint feature data in normal operating state, and is used to determine whether the fan has entered an abnormal state;

[0167] S75. If the soundprint signal of the fan currently in operation is judged to be in an abnormal state, further calculate the current feature vector and the feature model M of the soundprint signal of the fan potential abnormal state A The Mahalanobis distance d of the feature mean A (t), determine the specific type of anomaly;

[0168] S76, according to the Mahalanobis distance d A (t) size and specific abnormal state model M A Based on the distribution characteristics of the wind turbine, the system classifies and determines the type of anomaly, including wind turbine blade failure, bearing damage, or loose mechanical parts.

[0169] Embodiment 1:

[0170] Example 1 is applied in a wind farm, which is equipped with 30 large wind turbines. The wind turbines run continuously 24 hours a day and are subjected to complex natural environmental conditions, including strong winds, humidity, temperature fluctuations and sea salt erosion. Due to long-term operation, the blades, bearings and gearbox key components of the wind turbines are prone to wear, looseness and failure. The stable operation of the wind turbines directly affects the output efficiency and operating costs of electricity. In order to improve the fault detection accuracy of the wind turbines and reduce maintenance costs, the wind farm introduced the abnormal operation status detection method of the wind turbines of the present invention, which aims to monitor the operation status of the wind turbines in real time through non-contact and non-destructive voiceprint signal monitoring technology.

[0171] After the soundprint signal collection equipment is installed in the power plant, high-sensitivity audio collectors are placed in the key parts of the blades, bearings and gearboxes of each wind turbine. The collectors can obtain the soundprint signals during the operation of the wind turbine in real time, and collect the mechanical noise generated by the rotation of the wind turbine blades, the noise generated by the operation of the bearings, and the vibration sound of the gearbox 24 hours a day.

[0172] The collected original voiceprint signal data is first subjected to noise filtering and preprocessing to filter out environmental noise (wind noise, bird calls) and background sound signals unrelated to the operation of the fan. Then, the preprocessed voiceprint signal is subjected to short-time Fourier transform and Mel-frequency cepstrum coefficient extraction. Through the feature extraction method, the system can construct a voiceprint feature model that reflects the normal operating status of the fan.

[0173] The whole process of training sample simulation:

[0174] From June 1, 2023 to June 30, 2023, the system collects voiceprint signal data of the fan in normal operation every day. The collection time is from 2 am to 4 am every day. The environmental conditions are relatively stable and the fan is in normal operation. After processing, 1,000 voiceprint signal sample data sets in normal operation are obtained. These data are used to construct the voiceprint signal feature model M of the fan in normal operation N , the eigenmean vector and covariance matrix are:

[0175] Feature mean 2.54,3.67,1.25,0.89,4.01];

[0176] Covariance matrix

[0177] From July 1 to July 10, 2023, the system collected voiceprint signals of wind turbines in simulated abnormal states. The fault simulations included: tiny cracks in wind turbine blades, slight wear of bearings, and loose gearboxes. Each abnormal state lasted for 1 hour. The system collected a total of 300 voiceprint signal sample data of abnormal states. Through these data, a potential abnormal state voiceprint signal feature model M was constructed. A :

[0178] Feature mean 5.20,1.89,2.35,4.70,3.10];

[0179] Covariance matrix

[0180] On July 15, 2023, the system collected a set of real-time voiceprint signals during the actual operation of the fan, and extracted the feature vector after preprocessing. In the comparative analysis process, the Mahalanobis distance between the current voiceprint signal characteristics and the normal operation model was first calculated, and it was found that the Mahalanobis distance d N (t) = 6.12, which obviously exceeds the preset threshold value θ N =4.5, the system determines that the current fan status is abnormal.

[0181] Then, the system calculates the Mahalanobis distance between the current voiceprint signal feature and the potential abnormal state model, and obtains the distance value d A (t) = 2.80, which is close to the standard range of abnormal state characteristics. The system further determines that the current fan may have tiny cracks in the blades or bearing wear, and sends out an alarm signal in time.

[0182] To verify the effectiveness of the present invention, from July 1 to July 31, 2023, the power plant compared the performance of the method of the present invention and the traditional physical sensor monitoring method in wind turbine fault detection. The specific data are shown in Table 1 below:

[0183] Table 1 Comparison data of fan fault detection methods

[0184]

[0185] From the data in Table 1 above, it can be seen that the traditional vibration sensor method has high missed alarm rate and high false alarm rate when detecting early faults of wind turbines, and the fault response time is long, usually taking 10 hours to find the problem. The detection method based on the voiceprint signal feature extraction and hierarchical clustering algorithm of the present invention has a detection accuracy of up to 96%, a significantly reduced false alarm rate of 2%, a missed alarm rate of only 5%, and a shortened fault detection response time of 2 hours, which can timely detect potential abnormalities of wind turbine blades, bearings or gearboxes.

[0186] Through actual application in wind farms and verification of simulation processes, the abnormal operation status detection method of the wind turbine of the present invention effectively solves the problems of susceptibility to external interference, low detection accuracy and long fault response time in traditional sensor methods. Experimental data show that the algorithm improvement of the present invention significantly improves the accuracy, sensitivity and real-time performance of wind turbine fault detection, reduces the risks of false alarms and missed alarms, and provides strong technical support for fault prevention and maintenance of wind turbines.

[0187] The present invention introduces a voiceprint signal feature extraction method that combines short-time Fourier transform and Mel-frequency cepstrum coefficients, which effectively improves the accuracy of feature extraction. The time-frequency features are obtained by short-time Fourier transform and combined with energy spectral density and MFCC to extract key features, so that even tiny voiceprint signal changes during fan operation can be accurately captured. Compared with traditional physical sensor methods, the present invention achieves high-sensitivity detection of the fan operating status without contacting the device. By standardizing and enhancing the extracted voiceprint signal, the problem of inconsistent signal amplitudes in different environments is overcome, thereby ensuring the uniformity and accuracy of the signal and greatly improving the robustness of the system.

[0188] The present invention adopts an innovative combination of hierarchical clustering algorithm and density peak clustering algorithm to perform fine-grained classification and anomaly detection on the fan operating status. The fan operating voiceprint signal is divided into multiple operating status clusters through hierarchical clustering, and each status cluster is further subdivided for secondary segmentation using the density peak clustering algorithm to accurately identify slight changes in the voiceprint signal. The fan operating status is further subdivided into normal status subclusters and potential abnormal status subclusters by calculating the local density and minimum distance to identify the density peak point, thereby improving the accuracy of anomaly detection. In the early fault detection of fan blades and bearing components, potential anomalies can be identified in advance to reduce false alarms and missed alarms.

[0189] The present invention realizes comparative analysis of the real-time operating status of the fan by establishing a characteristic model of the normal operating soundprint signal of the fan and a characteristic model of the potential abnormal state soundprint signal. When the system collects the soundprint signal of the current fan in real time, it can compare the Mahalanobis distance with the pre-established normal operating model to accurately determine whether the fan has entered an abnormal state. By calculating the characteristic mean and covariance matrix under the abnormal state, the present invention further refines the identification of abnormal types, and can effectively distinguish specific problems such as fan blade failure, bearing damage or loose mechanical parts, thereby improving the efficiency and accuracy of fault diagnosis.

[0190] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting abnormal operation status of a fan, characterized in that: The steps include: S1. Multiple audio acquisition devices are arranged at multiple key locations of the fan to collect the fan blade rotation signal, bearing noise signal and fan mechanical vibration signal generated during the operation of the fan in real time, and generate a fan operation soundprint signal data set; S2. Preprocess the fan operation voiceprint signal dataset to remove environmental noise, filter background sound signals irrelevant to the fan operation, and normalize the voiceprint signal to obtain a preprocessed fan operation voiceprint signal dataset; S3, extracting features from the preprocessed fan operation soundprint signal data set to construct a feature data set; S4. Based on the feature data set, a hierarchical clustering algorithm is used to preliminarily classify the fan operation soundprint signal, and the fan operation soundprint signal data under different operating conditions are divided into multiple fan operation state clusters, each state cluster corresponds to a different fan operation mode; S5. Further use a hierarchical clustering algorithm to perform secondary subdivision in each wind turbine operation status cluster, identify and cluster the slight feature changes in the voiceprint signal, and further subdivide the voiceprint signal feature data in each operation status cluster into a normal state sub-cluster and a potential abnormal state sub-cluster; S6. Establishing a normal operation soundprint signal feature model of the fan based on the soundprint signal feature data of the normal state sub-cluster, and establishing a potential abnormal state soundprint signal feature model of the fan based on the soundprint signal feature data of the potential abnormal state sub-cluster; S7. During the actual operation of the fan, the voiceprint signal of the fan's current operation is collected in real time, and the extracted voiceprint signal features are compared with the pre-established normal operation voiceprint signal feature model of the fan. When a significant difference is detected between the voiceprint signal features and the normal operation voiceprint signal feature model of the fan, the system determines that the fan has entered an abnormal operation state; S8. When the operating status of the fan is detected as abnormal, the system sends out an alarm signal and analyzes the specific fault type of the fan based on the characteristic model of the soundprint signal of the fan's potential abnormal state, including fan blade failure, bearing damage or loose mechanical parts.

2. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S1 specifically includes: S11. Arrange multiple audio collection devices near the blades, bearings and other key mechanical parts of the fan for real-time collection of soundprint signals generated during the operation of the fan; S12. The voiceprint signal includes the following three types of signals: The blade rotation signals are collected at the blade root, middle and tip of the blade: X l (t)={x l1 (t),x l2 (t),…,x ln (t)}; Among them, X l (t) is the blade rotation signal set, x li (t) is the blade rotation signal collected by the collection device located at the i-th position of the blade at time t; Bearing noise signals are collected from the main bearing and support bearing: X b (t)={x b1 (t),x b2 (t),…,x bm (t)}; Among them, X b (t) is the bearing noise signal set, x bi (t) is the noise signal collected by the collection device located at the i-th position of the bearing at time t; Mechanical vibration signals are collected in the gearbox and nacelle: X v (t)={x v1 (t),x v2 (t),…,x vk (t)}; Among them, X v (t) is the mechanical vibration signal set, x vi (t) is the mechanical vibration signal collected by the collection device located at the i-th position of the gearbox or nacelle at time t; S13, by real-time storage of the collected blade rotation signal, bearing noise signal and mechanical vibration signal, a fan operation soundprint signal data set X(t) is generated: X(t)={X l (t),X b (t),X v (t)}。 3. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S2 specifically includes: S21, performing noise filtering on the fan operation voiceprint signal dataset X(t) to remove the environmental noise N(t) in the collected signal, where the environmental noise includes the natural noise around the fan and the equipment interference signal, and obtaining the processed fan operation voiceprint signal dataset X'(t); S22, filtering out background sound signals B(t) irrelevant to the operation of the fan, and removing the background signals from the denoised voiceprint signals through a signal separation algorithm, to obtain a fan operation voiceprint signal dataset X"(t) after background signal filtering; S23, normalizing the filtered fan operation soundprint signal dataset X”(t), and normalizing the amplitude values ​​of the soundprint signals collected by different collection devices to obtain a normalized fan operation soundprint signal dataset X n (t); S24, the normalized fan operation soundprint signal dataset X is processed by signal enhancement technology n (t) Improve key acoustic features, enhance specific frequency and time domain feature signals, and obtain the preprocessed fan operation soundprint signal dataset X p (t).

4. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S3 specifically includes: S31, short-time Fourier transform is used to pre-process the fan operation soundprint signal dataset X p (t) Perform time-frequency analysis to obtain the time-frequency feature matrix S(f,t); S32. Extract the key features of energy spectral density P(f) and Mel-frequency cepstrum coefficient M(n) from the time-frequency feature matrix S(f,t) and construct the feature vector c i : c i =[P(f i ),M(n i ),…],i=1,2,…,N; Where N is the number of eigenvectors, P(f i ) is the frequency f i The energy spectral density at i ) is the nth i Mel-frequency cepstral coefficients, c i is the characteristic vector of the fan operation soundprint signal at the i-th moment; S33, all the eigenvectors c i Composition feature data set C = {c1, c2, ..., c N }, standardize the feature data set C to eliminate the influence of different feature dimensions and scales, and obtain the final fan operation soundprint signal feature data set C f .

5. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S4 specifically includes: S41, based on the final fan operation soundprint signal feature data set C f Calculate the similarity matrix D between feature vectors = [d ij ],d ij represents the feature vector c i With c j The Mahalanobis distance between them is used to eliminate the influence of different feature dimensions and correlation on the distance calculation, and is used to measure the differences between the feature vectors of the soundprint signals of different fans: Among them, c i and c j is the characteristic vector of the voiceprint signal collected by the fan at different time points, Σ is the characteristic data set C f The covariance matrix of reflects the correlation between acoustic features, (·) T Represents a vector transpose operation; S42. Use the similarity matrix D to perform hierarchical clustering analysis, and use the agglomerative hierarchical clustering algorithm to build a clustering tree T of the fan operation soundprint signal. Starting from each feature vector, the closest clusters are merged step by step, and the feature data set C is clustered by setting the clustering threshold θ. f Divided into m wind turbine operation status clusters G = {G1, G2, ..., G m }, the clustering threshold θ is determined according to the distribution characteristics of the distance values ​​in the similarity matrix D: θ=μ D +l·s D ; Among them, μ D is the average value of all distance values ​​in the similarity matrix D, σ D is the standard deviation of the distance value, and λ is an adjustment parameter used to control the granularity of clustering; By adjusting the value of λ, a balance is achieved between the number of clusters and the similarity within the clusters, so that the clustering results are consistent with the actual distribution of the fan operation soundprint signal; S43, for each fan operation status cluster G i Calculate its feature center As the representative voiceprint feature vector of the running state cluster: Among them, |G i | is the state cluster G i The number of eigenvectors in , is the state cluster G i The feature center of each state cluster G i Corresponding to a fan operation mode, including a normal operation mode, a slight abnormality mode or a serious abnormality mode; S44. Feature centers based on different state clusters For each state cluster G i The corresponding fan operation mode is given a specific meaning, and each operation status cluster is defined as the corresponding fan operation mode: Mode definition = {G1→normal mode, G2→slightly abnormal mode, …, G m →Severe abnormal mode}.

6. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S5 specifically includes: S51, for each wind turbine operating status cluster G i Calculate the weighted cosine distance similarity matrix between feature vectors within a cluster in Represents the feature vector and The weighted cosine distance between them is used to measure the small differences between the feature vectors of the voiceprint signals within the cluster: in, and is the state cluster G i The characteristic vector of the voiceprint signal in reflects the running voiceprint characteristics of the fan at different time points, w i For state cluster G i The feature weight vector is weighted according to the importance of each acoustic feature for anomaly detection. ⊙ represents the Hadamard product of the corresponding elements, ∥·∥ represents the bi-norm of the vector, and · represents the vector dot product operation. S52, using the similarity matrix D i For state cluster G i The characteristic data in the voiceprint are analyzed by secondary hierarchical clustering, and the density peak clustering algorithm is used to identify the density peak points of the data in the cluster to capture the tiny characteristic changes in the voiceprint signal; Calculate each eigenvector The local density ρ k : Among them, |G i | is the state cluster G i The number of eigenvectors in , d c is the cutoff distance, which determines the influence range of local density calculation. Usually, the similarity matrix D i The mean of the median distance values, ρ k Reflects the eigenvector The density of surrounding data, used to identify areas with high density; Calculate each eigenvector The minimum distance δ k : Among them, δ k Represents the feature vector The minimum distance between and all eigenvectors with higher density than it, when When the local density is the highest, let δ k is the similarity matrix D i The maximum distance value in k To measure the feature vector Whether it is at the peak density position; S53, according to the local density ρ k and the minimum distance δ k The distribution characteristics of the threshold value ρ i and δ i , the state cluster G i Subdivided into normal state sub-cluster G iN and potential abnormal state sub-cluster G iA , when the condition ρ is satisfied k >ρ i And δ k >δ i When , the feature vector Marked as the density peak point, belonging to the potential abnormal state subcluster G iA , indicating that there are abnormal voiceprint features, and the remaining feature vectors belong to the normal state subcluster G iN ; S54, complete the state cluster G i The secondary subdivision of iN and potential abnormal state sub-cluster G iA , further optimize the ability to identify minor anomalies in voiceprint signal feature data; S55, calculate the normal state sub-cluster G respectively iN and potential abnormal state sub-cluster G iA The feature center of the normal state subcluster Potential abnormal state sub-cluster feature center Among them, |G iN | and |G iA | are the number of eigenvectors in the normal state sub-cluster and the potential abnormal state sub-cluster, respectively. and is the characteristic center of the corresponding sub-cluster, representing the typical soundprint characteristics of the fan in normal operation and potential abnormal state.

7. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S6 specifically includes: S61, based on normal state sub-cluster G iN The fan soundprint signal feature data is used to establish the normal operation soundprint signal feature model M of the fan N , calculate the normal state subcluster G iN The mean of the eigenvectors in and the covariance matrix Σ iN , and the mean vector and the covariance matrix Σ iN Constructing the characteristic model of the soundprint signal of the normal operation of the fan It is used to describe the statistical characteristics of the fan soundprint signal under normal operating conditions; S62, based on potential abnormal state sub-cluster G iA The characteristic data of the fan soundprint signal is used to establish the characteristic model M of the fan potential abnormal state soundprint signal A , calculate the potential abnormal state sub-cluster G iA The mean of the eigenvectors in and the covariance matrix Σ iA , the mean vector and the covariance matrix Σ iA Constructing the characteristic model of the soundprint signal of the potential abnormal state of the fan Used to describe the statistical characteristics of fan soundprint signal characteristics under potential abnormal conditions.

8. A method for detecting abnormal operation status of a fan according to claim 1, characterized in that: The S7 specifically includes: S71. During the actual operation of the fan, the current voiceprint signal of the fan is collected in real time, and the collected voiceprint signal is preprocessed to obtain a preprocessed voiceprint signal data set of the current operation; S72, extracting feature vectors from the currently running pre-processed voiceprint signal data set to construct a feature data set of the current fan operation voiceprint signal, which is used to compare the fan normal operation voiceprint signal feature model M in real time N and potential abnormal state voiceprint signal feature model M A ; S73, calculating the characteristic vector of the soundprint signal of the currently running fan and the characteristic model M of the soundprint signal of the normal operation of the fan N The Mahalanobis distance d of the feature mean N (t), determine whether the current voiceprint signal is in a normal operating state; S74, when the Mahalanobis distance d N (t) exceeds the preset threshold θ N When the system determines that the current fan operation status deviates from the normal range, it enters an abnormal state: d N (t)>θ N ; Among them, θ N The abnormal detection threshold is set according to the distribution of voiceprint feature data in normal operating state, and is used to determine whether the fan has entered an abnormal state; S75. If the soundprint signal of the fan currently in operation is judged to be in an abnormal state, further calculate the current feature vector and the feature model M of the soundprint signal of the fan potential abnormal state A The Mahalanobis distance d of the feature mean A (t), determine the specific type of anomaly; S76, according to the Mahalanobis distance d A (t) size and specific abnormal state model M A Based on the distribution characteristics of the wind turbine, the system classifies and determines the type of anomaly, including wind turbine blade failure, bearing damage, or loose mechanical parts.

Images