A wind turbine fault diagnosis method and system based on voiceprint recognition

Abstract

The application discloses a wind turbine fault diagnosis method and system based on voiceprint recognition. The method collects sound signals from multiple locations and converts them into electronic data format. After filtering to remove background noise, the purified audio information is segmented and processed. The frequency distribution characteristics of each segment of audio are calculated to form a sound pattern that can represent different time points. Then, a dynamic voiceprint identifier is constructed, which not only reflects the audio characteristics of the normal operation of the wind turbine, but also has the ability to adjust over time to adapt to new operating conditions. Finally, by comparing with the pre-stored abnormal voiceprint identifier library, the possible fault source is quickly and accurately identified, and the corresponding warning information and maintenance suggestions are generated. This method improves the accuracy of fault diagnosis, reduces the false positive rate, and realizes early fault warning, thereby helping to improve the operation efficiency and safety of the wind turbine.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of voiceprint recognition, and particularly relates to a wind turbine fault diagnosis method and system based on voiceprint recognition. BACKGROUND

[0002] As an important part of renewable energy power generation systems, the monitoring and maintenance of wind turbines is crucial. Traditionally, the fault diagnosis of wind turbines relies on periodic manual inspection and vibration analysis-based methods. Manual inspection is not only time-consuming and labor-intensive, but also difficult to capture early potential faults. Although the vibration analysis-based method can provide some fault warnings, it is sensitive to installation location, sensor type and environmental factors, and has limited effectiveness for detecting certain types of faults.

[0003] In recent years, with the development of voiceprint recognition technology, researchers have begun to explore the possibility of using sound signals for mechanical fault diagnosis. Some existing solutions attempt to collect noise generated during equipment operation by installing sound sensors at key locations, and then apply spectral analysis or machine learning algorithms to identify abnormal patterns. However, these methods usually focus on post-event analysis, i.e., identifying after the fault has occurred, and cannot achieve true preventive maintenance. In addition, existing voiceprint recognition technologies often lack effective adaptability to complex and variable working environments, resulting in high false alarm rates.

[0004] To address the above problems, a significant technical challenge is how to achieve early fault warning while ensuring a low false alarm rate. Traditional voiceprint recognition methods, when faced with large mechanical equipment such as wind turbines, due to the dynamic changes in the working environment (such as wind speed, temperature, etc.), and the large differences in normal noise generated under different operating conditions, are prone to misjudging normal operating changes as faults, or failing to identify early fault characteristics in a timely manner. SUMMARY

[0005] The present application aims to provide a wind turbine fault diagnosis method and system based on voiceprint recognition, which improves the accuracy of fault diagnosis, reduces the false alarm rate, and achieves early fault warning, thereby helping to improve the operational efficiency and safety of wind turbines, to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application proposes a wind turbine fault diagnosis method based on voiceprint recognition, comprising the following steps:

[0007] During the operation of the wind turbine, sound signals are collected from multiple locations, and the sound signals are converted into electronic data format. According to the obtained data, background noise is removed through filtering processing to purify the audio information;

[0008] The purified audio information is segmented to ensure that each segment contains sufficient information to reflect the working state of the unit;

[0009] The frequency distribution characteristics of each segment of audio are calculated to form sound patterns representing different time points. A dynamic voiceprint identifier is constructed using the obtained sound patterns, which can reflect the audio characteristics of the wind turbine during normal operation;

[0010] The established voiceprint identifier is compared with a pre-stored abnormal voiceprint identifier library to find out if there is a matching item. If a matching abnormal voiceprint identifier is found, a warning message is generated and the possible source of the fault is indicated;

[0011] Maintenance recommendations are provided according to the source of the fault, and the voiceprint identifier library is updated.

[0012] Preferably, the sound signals are collected from multiple positions and converted into electronic data format, including:

[0013] N different positions in the wind turbine are selected as listening points, where N is a positive integer, and an acoustic sensor is deployed at each listening point. Each acoustic sensor captures an analog signal of the sound amplitude varying with time;

[0014] The analog signal is discretized by sampling rate to obtain a series of discrete time-amplitude data pairs;

[0015] Based on the discrete data pairs, the time-domain signal is converted to the frequency domain by applying the Fourier transform formula to obtain the spectral intensity at the corresponding frequency, thereby forming a frequency domain representation in the electronic data format;

[0016] The generated frequency domain data is grouped according to the pre-set frequency interval, and the mean and standard deviation of all spectral intensities in each group are calculated to construct a multi-dimensional vector that can reflect the audio characteristics of each listening point.

[0017] Preferably, the background noise is removed by filtering to purify the audio information, including:

[0018] The multi-dimensional vector is received, and an adaptive threshold is applied to determine the upper limit of the background noise level. By comparing the mean value of each frequency band with the adaptive threshold, the data segment exceeding the threshold is identified and marked as a potential target sound feature;

[0019] For the marked data segment, a weighting matrix is constructed, and the weighting matrix is applied to the original multi-dimensional vector to generate a new vector to highlight the sound features after filtering.

[0020] Preferably, the purified audio information is segmented, including:

[0021] receiving a new vector, the vector contains the filtered and enhanced audio features;

[0022] defining a time window and an overlap ratio, to determine the length of each segment and the overlap between adjacent segments, and calculating the start time of each segment;

[0023] for each defined segment, calculating the cumulative energy of the audio features in that segment, and identifying segments with significant energy changes;

[0024] marking the segments that meet the conditions as valid segments, and recording the corresponding start time and end time, and for consecutive valid segments, merging them to form a longer time period, to ensure that each segment contains enough information to reflect the working state of the unit.

[0025] Preferably, the calculation of the frequency distribution characteristics of each segment of audio forms a sound pattern that represents different time points, including:

[0026] receiving the marked valid segments and their start time and end time, and extracting the corresponding audio data for each valid segment;

[0027] for the extracted audio data segment, applying a fast Fourier transform to convert the time domain signal to a frequency domain representation;

[0028] based on the obtained frequency domain representation, calculating the power spectral density at each frequency to quantify the energy distribution of each frequency component, then determining a frequency range and dividing multiple frequency bands at fixed intervals within the range, and calculating the average power spectral density in each frequency band;

[0029] forming a vector of the calculated average power spectral density of each frequency band, and this vector is used as a feature descriptor that represents the sound pattern at that time point.

[0030] Preferably, the construction of a dynamic voiceprint identifier includes:

[0031] receiving the sound pattern feature descriptor vector, applying a time weighting factor to each sound pattern feature descriptor to generate a weighted feature descriptor, then calculating the cumulative value of all valid segment weighted feature descriptors to form a global representation of the audio features of the wind turbine unit over a period of time;

[0032] based on the cumulative value, constructing a dynamic voiceprint identifier, and adding for each new sound pattern feature descriptor.

[0033] Preferably, the comparison of the established voiceprint identifier with the pre-stored abnormal voiceprint identifier library includes:

[0034] receiving the constructed dynamic voiceprint identifier and loading a pre-stored abnormal voiceprint identifier library containing a plurality of voiceprint identifiers under abnormal conditions;

[0035] For each abnormal voiceprint identifier, calculating a similarity score between it and the dynamic voiceprint identifier, using the Euclidean distance formula to quantify the difference between the two;

[0036] Determining a preset threshold, comparing all calculated similarity scores with the threshold, and if any similarity score is less than or equal to the threshold, marking the abnormal voiceprint identifier as a potential match and recording the corresponding index;

[0037] For all abnormal voiceprint identifiers marked as potential matches, finding detailed information in the abnormal voiceprint identifier library according to their indexes, including but not limited to fault type, possible causes and recommended inspection measures, forming a report for subsequent analysis.

[0038] Preferably, the generated warning information and the possible source of the fault include:

[0039] Receiving the abnormal voiceprint identifier marked as a potential match and its corresponding index, and obtaining the associated detailed information, including the fault type and recommended inspection measures;

[0040] For each marked abnormal voiceprint identifier, constructing a warning information template containing a fixed part and a variable part, where the variable part is filled according to the detailed information to form specific warning content, ensuring that each warning information clearly indicates the possible source of the fault;

[0041] Calculating an urgency score, combining the generated specific warning content with the calculated urgency score to create a final warning notification, which is sent to maintenance personnel or monitoring systems through pre-set communication channels.

[0042] Preferably, the maintenance recommendations are provided according to the source of the fault, and the voiceprint identifier library is updated, including:

[0043] Receiving the final warning notification, developing a targeted maintenance action plan based on the warning content and urgency score, ensuring that each recommendation is specific and feasible based on fault type, historical maintenance records and current operating conditions;

[0044] After executing the maintenance action, collecting maintenance result feedback, evaluating the maintenance effect, integrating the obtained maintenance result feedback and score change into the voiceprint identifier library, updating the corresponding abnormal voiceprint identifier and its associated information, and for normal voiceprint identifiers generated under new or improved operating conditions.

[0045] In another aspect, the present application provides a wind turbine fault diagnosis system based on voiceprint recognition, comprising:

[0046] The sound signal collection and preliminary processing module is used for collecting sound signals from multiple positions during the operation of the wind turbine generator and converting the sound signals into electronic data format, removing background noise through filtering processing based on the obtained data, and purifying audio information;

[0047] The audio information segmentation processing module is used for segmenting the purified audio information to ensure that each segment contains sufficient information to reflect the working state of the unit;

[0048] The frequency distribution characteristic analysis and voiceprint identification construction module is used for calculating the frequency distribution characteristics of each segment of audio, forming a sound pattern that can represent different time points, and constructing a dynamic voiceprint identification using the obtained sound pattern, which can reflect the audio characteristics of the normal operation of the wind turbine generator;

[0049] The abnormality detection and warning generation module is used for comparing the established voiceprint identification with the pre-stored abnormal voiceprint identification library to find out whether there is a matching item, and if a matching abnormal voiceprint identification is found, generating warning information and indicating the possible fault source;

[0050] The maintenance suggestion and voiceprint library updating module is used for providing maintenance suggestions according to the fault source and updating the voiceprint identification library.

[0051] The technical effects and advantages of the present application: the wind turbine generator fault diagnosis method and system based on voiceprint recognition proposed in the present application have the following advantages compared with the prior art:

[0052] The present application collects sound signals from multiple positions and converts them into electronic data format, removes background noise through filtering processing, and then segments the purified audio information, calculates the frequency distribution characteristics of each segment of audio, and forms a sound pattern that can represent different time points. Subsequently, a dynamic voiceprint identification is constructed, which not only reflects the audio characteristics of the normal operation of the wind turbine generator, but also has the ability to adjust over time to adapt to new operating conditions. Finally, by comparing with the pre-stored abnormal voiceprint identification library, the possible fault source is quickly and accurately identified, and the corresponding warning information and maintenance suggestions are generated. This method improves the accuracy of fault diagnosis, reduces the false alarm rate, realizes early fault warning, and thus helps to improve the operation efficiency and safety of the wind turbine generator. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The flowchart of the wind turbine generator fault diagnosis method based on voiceprint recognition of the present application;

[0054] Figure 2 The block diagram of the wind turbine generator fault diagnosis system based on voiceprint recognition of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. The specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the scope of protection of the present application.

[0056] The present application provides a wind turbine fault diagnosis method based on voiceprint recognition, aiming to solve the problems of insufficient early fault warning and high false alarm rate in the prior art. By collecting sound signals from multiple positions and converting them into electronic data format, background noise is removed through filtering processing, and then the purified audio information is segmented for processing, the frequency distribution characteristics of each segment of audio are calculated, and the sound patterns representing different time points are formed. Subsequently, a dynamic voiceprint identifier is constructed, which not only reflects the audio characteristics of the normal operation of the wind turbine, but also has the ability to adjust over time to adapt to new operating conditions. Finally, by comparing with the pre-stored abnormal voiceprint identifier library, the possible fault sources are quickly and accurately identified, and the corresponding warning information and maintenance suggestions are generated.

[0057] This method improves the accuracy of fault diagnosis, reduces the false alarm rate, and realizes early fault warning, thereby helping to improve the operation efficiency and safety of the wind turbine.

[0058] As shown in Figure 1 The wind turbine fault diagnosis method based on voiceprint recognition in the present embodiment includes the following steps:

[0059] During the operation of the wind turbine, sound signals are collected from multiple positions, and the sound signals are converted into electronic data format. According to the obtained data, background noise is removed through filtering processing to purify the audio information; specifically including:

[0060] Select N different positions in the wind turbine as listening points, where N is a positive integer, and deploy acoustic sensors at each listening point. Each acoustic sensor will capture the analog signal of the sound amplitude A changing with time t; by deploying acoustic sensors at multiple different positions in the wind turbine, the sound signals during the operation of the unit can be comprehensively captured, ensuring that the noise sources of key components are covered. This not only improves the comprehensiveness of data collection, but also enhances the monitoring capability of potential fault points.

[0061] The discretization process is performed by a sampling rate Fs, obtaining a series of discrete time-amplitude data pairs {(T1, A1), (T2, A2),..., (Tn, An)}, where Ti represents the i-th sampling time, and Ai represents the sound amplitude at that time; the discretization process is performed by a fixed sampling rate Fs, converting the continuous sound amplitude A varying with time t into a series of discrete time-amplitude data pairs. This process ensures the digitization of the sound signal, providing a basis for subsequent data analysis and processing.

[0062] Based on the discrete data pairs, the Fourier transform formula F(Ω) = ∫[A(T)*e^(-jΩT)]dT is applied to convert the time-domain signal to the frequency domain to obtain the spectral intensity F(Ω) at the corresponding frequency Ω, thereby forming a frequency-domain representation in electronic data format; this step enables the audio information to be converted from time-domain characteristics to frequency-domain characteristics, facilitating the identification of specific frequency components and their intensity distribution, thereby better understanding the essential characteristics of the sound.

[0063] The generated frequency-domain data is grouped according to the pre-set frequency interval, and the mean value Mg and standard deviation Sg of all spectral intensities in each group are calculated to construct a multi-dimensional vector V = [Mg1, Sg1, Mg2, Sg2,..., Mgn, Sgn] that reflects the audio characteristics of each listening point. This operation helps to extract the core information reflecting the audio characteristics of each listening point, reduces redundant data, and improves the efficiency of subsequent analysis.

[0064] The multi-dimensional vector V = [Mg1, Sg1, Mg2, Sg2,..., Mgn, Sgn] is received, where each element represents the mean value and standard deviation of the listening point in a specific frequency interval; an adaptive threshold Thresh = α*max(Mg1, Mg2,..., Mgn) is applied, where α is a positive coefficient less than 1, used to determine the upper limit of the background noise level, and by comparing the mean value Mgi of each frequency band with Thresh, the data segment exceeding this threshold is identified and marked as a potential target sound feature; by comparing the mean value Mgi of each frequency band with Thresh, background noise and target sound features can be effectively distinguished, reducing the false positive rate.

[0065] For the marked data segment, a weighted matrix W is constructed, where the weight wj is calculated by the formula wj = exp(-β*(Sgj-avg(Sg))^2), and avg(Sg) is the average of all standard deviations, and β is a regulation parameter; the application of this matrix can enhance the target sound feature while weakening the influence of background noise, improving the clarity and recognizability of the sound feature.

[0066] Apply the weighting matrix W to the original multi-dimensional vector V to generate a new vector V' = V * W, highlighting the filtered sound features. This process effectively enhances the purified sound features, providing higher quality data support for subsequent fault diagnosis.

[0067] The implementation is as follows:

[0068] Suppose a wind farm is being monitored in a mountainous area, and multiple wind turbines are installed on the site. To ensure the safe and stable operation of these units, it is decided to use the above method for real-time fault diagnosis, as follows:

[0069] Multi-location acoustic sensor deployment:

[0070] Deploy acoustic sensors at selected N different locations (such as gearboxes, generators, blades, etc. critical parts). Each sensor is responsible for capturing the sound signal at its location and converting it into an analog signal.

[0071] Discretization process:

[0072] Discretize the captured analog signal of sound amplitude A as a function of time t with a fixed sampling rate Fs (e.g. 44.1 kHz) to obtain a series of discrete time-amplitude data pairs {(T1, A1), (T2, A2),..., (Tn, An)}.

[0073] Fourier transform:

[0074] Apply the Fourier transform formula F(Ω) = ∫[A(T) * e^(-jΩT)] dT to convert the discretized time-domain signal to a frequency-domain representation, obtaining the spectral intensity F(Ω) at corresponding frequency Ω.

[0075] Frequency band grouping and statistical calculation:

[0076] According to the predefined frequency interval (such as 0-100Hz, 100-500Hz, etc.), group the frequency domain data and calculate the mean value Mg and standard deviation Sg of all spectral intensities in each group, forming a multi-dimensional vector V = [Mg1, Sg1, Mg2, Sg2,..., Mgn, Sgn].

[0077] Adaptive threshold setting:

[0078] Set the adaptive threshold Thresh = α * max(Mg1, Mg2,..., Mgn), where α is set to 0.8. Compare the mean value Mgi of each frequency band with Thresh, and mark the data segment that exceeds this threshold as a potential target sound feature.

[0079] Weighting matrix construction:

[0080] For the marked data segments, the weight wj is calculated according to the formula wj = exp(-β*(Sgj-avg(Sg))^2), where β is set to 0.5, and a weighted matrix W is constructed.

[0081] A new vector is generated:

[0082] The weighted matrix W is applied to the original multi-dimensional vector V to generate a new vector V' = V*W, which highlights the filtered sound features and prepares for the next step of fault diagnosis.

[0083] Through this embodiment, valuable feature information can be effectively extracted from the operating sound of the wind turbine, thereby realizing early fault warning and ensuring the safe and stable operation of the wind turbine.

[0084] The purified audio information is segmented to ensure that each segment contains sufficient information to reflect the working state of the unit; specifically including:

[0085] A new vector V' = V*W is received, which contains the filtered and enhanced audio features; this step ensures that the subsequent segmentation processing is based on purified and enhanced audio data, improving the quality and reliability of the signal.

[0086] A time window Δτ and an overlap ratio ρ, 0 < ρ < 1, are defined to determine the time length of each segment and the overlapping part between adjacent segments, and the starting time point τi of each segment is calculated as τi = i*Δτ*(1-ρ), where i is an integer index starting from 0, representing the i-th segment; this method not only ensures the integrity of each segment of audio information, but also enhances the ability to capture continuous trends by appropriate overlap, avoiding the loss of important information due to too short segmentation.

[0087] For each defined segment, the cumulative energy Ei = Σ[A(T)^2] of the audio features in that time period is calculated, where A(T) is the sound amplitude at the sampling time within the segment. By comparing Ei with the set energy threshold Eth, segments with significant energy changes can be effectively identified, which often reflect important changes in the working state of the unit or the occurrence of potential faults.

[0088] Segments that meet the conditions are marked as valid segments, and the corresponding starting time τi and ending time τf = τi + Δτ are recorded. For consecutive valid segments, they are combined to form a longer time period, ensuring that each segment contains sufficient information to reflect the working state of the unit. This step ensures that each segment contains sufficient information to reflect the working state of the unit, while reducing redundant data and improving analysis efficiency.

[0089] The specific implementation is as follows:

[0090] Suppose a wind turbine located in a wind farm is being monitored, equipped with the acoustic sensor system mentioned in the above method for real-time monitoring of operating conditions. Now the implementation of the segmentation process is handled in segments to ensure that each segment of audio information fully reflects the working state of the unit.

[0091] Receive filtered and enhanced audio features:

[0092] Receive the new vector V' = V * W, which contains filtered and enhanced audio features. These features have been optimized through previous steps such as multi-position acoustic sensor deployment, discretization processing, Fourier transform, frequency band grouping statistical calculation, adaptive threshold setting, and weight matrix construction.

[0093] Define time window and overlap ratio:

[0094] Define a time window Δτ as 5 seconds and set the overlap ratio ρ as 0.5. This means that each segment has a length of 5 seconds, and there is a 2.5 second overlap between adjacent segments. According to the formula τi = i * Δτ * (1-ρ), calculate the starting time point of each segment, for example, the first segment starts at 0 seconds, the second segment starts at 2.5 seconds, and so on.

[0095] Calculate cumulative energy and identify significant changes:

[0096] For each defined segment, calculate the cumulative energy E i = Σ[A(T)^2] of the audio features within that time period, where A(T) is the sound amplitude at the sampling time within the segment. Set the energy threshold Eth as a predetermined empirical value or through historical data analysis. By comparing E i with Eth, identify segments with significant energy changes. These segments may indicate changes in the working state of the unit or the occurrence of potential faults.

[0097] Label and merge valid segments:

[0098] Label segments that meet the conditions as valid segments and record the corresponding starting time τi and ending time τf = τi + Δτ. If multiple consecutive segments are labeled as valid, merge them into a longer time period. For example, if the first, second, and third segments are valid, merge them into a long time period from 0 seconds to 7.5 seconds. This ensures that each segment contains enough information to reflect the working state of the unit, while also simplifying the subsequent analysis process.

[0099] Through this embodiment, the purified audio information can be effectively segmented and processed to ensure that each segment of audio information contains enough information to reflect the working state of the wind turbine. This method not only improves the accuracy of fault diagnosis, but also provides reliable data support for preventive maintenance.

[0100] The frequency distribution characteristics of each segment of audio are calculated to form sound patterns representing different time points, and a dynamic voiceprint identifier is constructed using the obtained sound patterns, which can reflect the audio characteristics of the normal operation of the wind turbine; specifically comprising:

[0101] The marked valid segments and their start time τi and end time τf are received, and the corresponding audio data is extracted for each valid segment; this step ensures that the processed data segments are filtered and have significant energy changes or potential fault characteristics, thereby improving the relevance and accuracy of subsequent analysis.

[0102] For the extracted audio data segment, a fast Fourier transform (FFT) is applied to calculate the spectral intensity of the segment at different frequencies Ω through the formula F(Ω) = Σ[A(T)*e^(-jΩT)], where A(T) is the sound amplitude at time T, and Ω represents the angular frequency. This process converts the time-domain signal into a frequency-domain representation, allowing audio information to be analyzed from a frequency perspective, making it easier to identify specific frequency components and their intensity distribution.

[0103] Based on the obtained frequency-domain representation, the power spectral density PSD(Ω) = |F(Ω)|^2 at each frequency Ω is calculated to quantify the energy distribution of each frequency component. Then, a frequency range Φ is determined, and multiple frequency bands are divided within the range at a fixed interval δΦ. The average power spectral density APSD(b) = (1 / |b|)Σ[PSD(Ω)] is calculated for each frequency band, where b represents a frequency band, and |b| is the number of frequency points in the frequency band. This step helps to extract core information reflecting the energy distribution of each frequency band, reducing redundant data and improving subsequent analysis efficiency.

[0104] The average power spectral density APSD(b) of each frequency band is combined into a vector S = [APSD(b1), APSD(b2),..., APSD(bm)], where m is the total number of frequency bands. This vector S serves as a characteristic descriptor representing the sound pattern at that time point. This vector S reflects the frequency-domain characteristics of the audio data, providing a basis for constructing a dynamic voiceprint identifier.

[0105] The sound pattern characteristic descriptor vector S = [APSD(b1), APSD(b2),..., APSD(bm)] is received, where each element represents the average power spectral density of a different frequency band.

[0106] S = [APSD(b1), APSD(b2),..., APSD(bm)], where each element represents the average power spectral density of a different frequency band. For the received sound pattern characteristic descriptor, a time weighting factor Wτ = exp(-γ*Δτ) is calculated, where Δτ is the time interval and γ is a positive adjustment parameter. This factor reflects the change in the importance of sound characteristics over time, ensuring that recent sound information has a greater impact on the voiceprint identifier.

[0107] Apply a time weighting factor to each sound pattern feature descriptor to generate a weighted feature descriptor Sw=S*Wτ, then compute the cumulative value C=∑[Sw] of all valid segment weighted feature descriptors to form an overall representation of the audio features of the wind turbine over a period of time; this operation enhances the time dimension of the feature descriptors and improves the reliability of long-term monitoring.

[0108] Based on the cumulative value C, construct a dynamic voiceprint identifier D, for each new sound pattern feature descriptor added, update D as D'=(1-λ)*D+λ*C, where λ is an update coefficient between 0 and 1, ensuring that the voiceprint identifier D evolves smoothly with the addition of new data while retaining the memory of historical data. This method ensures that the voiceprint identifier D evolves smoothly with the addition of new data while retaining the memory of historical data, adapting to changing operating conditions.

[0109] The specific implementation is as follows:

[0110] Suppose a wind turbine located in a wind farm is being monitored, which has completed the steps of multi-position acoustic sensor deployment, discretization processing, filter enhancement, segmentation processing, etc. Now the process of frequency distribution feature calculation and dynamic voiceprint identifier construction will be implemented.

[0111] Receive valid segment audio data:

[0112] Receive the labeled valid segment and its start time τi and end time τf, for example, the first valid segment from 0 seconds to 5 seconds. Extract the corresponding audio data for each valid segment, which has been processed in the early stage and contains important audio features.

[0113] Fast Fourier Transform (FFT):

[0114] For the extracted audio data segment, apply Fast Fourier Transform (FFT) to calculate the spectral intensity of the segment audio at different frequencies Ω by the formula F(Ω)=∑[A(T)*e^(-jΩT)]. For example, for the first valid segment, calculate its spectral intensity at different frequencies to obtain the corresponding frequency domain representation.

[0115] Calculate the power spectral density:

[0116] Based on the obtained frequency domain representation, calculate the power spectral density PSD(Ω)=|F(Ω)|^2 at each frequency Ω. Set the frequency range Φ as 0-2000Hz, and divide it into multiple frequency bands with a fixed interval δΦ=10Hz. Calculate the average power spectral density APSD(b) in each frequency band. For example, calculate the average power spectral density in the first frequency band (0-10Hz).

[0117] Forming sound pattern feature descriptors:

[0118] The calculated average power spectral density APSD(b) for each frequency band is formed into a vector S = [APSD(b1), APSD(b2), ..., APSD(bm)]. For the above example, the formed vector S will contain the average power spectral density of multiple frequency bands, such as S = [APSD(0-10Hz), APSD(10-20Hz), ..., APSD(1990-2000Hz)].

[0119] Calculate the time weighting factor:

[0120] For the received sound pattern feature descriptor vector S, a time weighting factor Wτ = exp(-γ * Δτ) is calculated, where Δτ is the time interval (for example, once every 5 minutes) and γ is set to 0.01. This step takes time into account and ensures that the latest audio features occupy a more important position in subsequent analysis.

[0121] Generate weighted feature descriptors:

[0122] Apply the temporal weighting factor to each sound pattern feature descriptor to generate a weighted feature descriptor Sw = S * Wτ. Next, calculate the cumulative value C = Σ[Sw] of all valid segment weighted feature descriptors. For example, if there are multiple valid segments, calculate their weighted cumulative value to form an overall representation.

[0123] Build dynamic voiceprint identification:

[0124] Based on the accumulated value C, a dynamic voiceprint identifier D is constructed. Each time a new voice pattern feature descriptor is added, D is updated to D' = (1-λ)D + λC, where λ is set to 0.8. This method ensures that the voiceprint identifier D can evolve smoothly with the addition of new data, while retaining a memory of historical data and adapting to changing operating conditions.

[0125] This embodiment effectively calculates the frequency distribution characteristics of each audio segment, forming sound patterns that represent different time points. These patterns can then be used to construct a dynamic voiceprint. This approach not only improves the accuracy of fault diagnosis but also provides reliable data support for preventive maintenance, ensuring the safe and stable operation of wind turbines.

[0126] Compare the established voiceprint identifier with the pre-stored abnormal voiceprint identifier library to find out whether there is a match. If a matching abnormal voiceprint identifier is found, a warning message is generated and the possible source of the fault is indicated; specifically, the following are included:

[0127] Receive the constructed dynamic voiceprint identifier D and load the pre-stored abnormal voiceprint identifier library L, where L contains a plurality of voiceprint identifiers Li under abnormal conditions, i = 1, 2,..., N, N being the number of abnormal voiceprint identifiers in the library; this step ensures that the system can accurately compare the current operating state with known abnormal patterns, providing a scientific basis for subsequent fault diagnosis.

[0128] For each abnormal voiceprint identifier Li, calculate the similarity score Si between it and the dynamic voiceprint identifier D, using the Euclidean distance formula S i = sqrt(Σ[(Di-Li)^2]) to quantify the difference between the two; Di and Li represent the values of the corresponding positions of the dynamic voiceprint identifier and the i-th abnormal voiceprint identifier in the library, respectively; this method effectively filters out irrelevant abnormal patterns, improving the accuracy and efficiency of fault recognition.

[0129] Determine a preset threshold Thresh, and compare all calculated similarity scores Si with Thresh; if any S i is less than or equal to Thresh, mark the abnormal voiceprint identifier L i as a potential match and record the corresponding index i;

[0130] For all abnormal voiceprint identifiers L i marked as potential matches, find detailed information in the abnormal voiceprint identifier library L according to their indexes i, including but not limited to fault type, possible causes, and recommended inspection measures, to form a report for subsequent analysis; this step provides detailed fault background information, which helps maintenance personnel quickly locate the problem and take appropriate measures.

[0131] Receive the abnormal voiceprint identifier Li marked as a potential match and its corresponding index i, and obtain the associated detailed information, including fault type and recommended inspection measures;

[0132] For each marked abnormal voiceprint identifier Li, construct a warning information template W, which contains a fixed part and a variable part; the variable part is filled according to the detailed information of L i, forming specific warning content C = W(L i), ensuring that each warning message clearly indicates the possible source of the fault; this process ensures that each warning message clearly indicates the possible source of the fault, facilitating the rapid response of the maintenance team.

[0133] Calculate an urgency score U = Σ[wj*Fj], where wj is the weight coefficient for different fault types, and Fj is the frequency factor of the corresponding fault type extracted from the abnormal voiceprint identifier library L; this score is used to determine the priority of the warning information and affects the urgency of subsequent processing, helping to allocate resources reasonably.

[0134] The generated specific warning content C is combined with the calculated urgency score U to create a final warning notification N = [C, U], which is sent to maintenance personnel or monitoring systems through pre-set communication channels. This notification is sent to maintenance personnel or monitoring systems through pre-set communication channels, ensuring timely communication of fault information and facilitating rapid response and preventive maintenance.

[0135] The implementation is as follows:

[0136] Suppose a wind turbine located in a wind farm is being monitored, and the turbine has completed the initial steps of sound signal collection, filtering and enhancement, segmentation processing, frequency distribution characteristic calculation, and dynamic voiceprint identification construction. Now the process of voiceprint identification comparison and warning information generation will be implemented.

[0137] Load the voiceprint identification library and calculate the similarity:

[0138] Receive the constructed dynamic voiceprint identification D, and load the pre-stored abnormal voiceprint identification library L, which contains multiple abnormal voiceprint identifications Li (i = 1, 2,..., N) under abnormal conditions. For each abnormal voiceprint identification Li, use the Euclidean distance formula Si = sqrt(Σ[(Di - Li)^2]) to calculate the similarity score between it and the dynamic voiceprint identification D. For example, calculate the similarity score S1 between the first abnormal voiceprint identification L1 and the dynamic voiceprint identification D.

[0139] Set a threshold for matching determination:

[0140] Determine a pre-set threshold Thresh, for example 0.8. Compare all calculated similarity scores Si with Thresh. If any Si is less than or equal to Thresh, mark the abnormal voiceprint identification Li as a potential match and record the corresponding index i. For example, if S1 ≤ 0.8, mark L1 as a potential match and record index 1.

[0141] Find detailed information to form a report:

[0142] For all abnormal voiceprint identifications Li marked as potential matches, find detailed information in the abnormal voiceprint identification library L according to their index i, including but not limited to fault type, possible cause, and recommended inspection measures, to form a report for subsequent analysis. For example, for L1 marked as a potential match, find its detailed information, such as noise increase caused by insufficient lubrication of the gearbox, and recommend checking the lubricating oil level and quality.

[0143] Construct a warning information template:

[0144] For each labeled abnormal acoustic signature L i, construct a warning message template W. This template contains a fixed part (e.g., "Warning: Possible malfunction detected") and a variable part (filled according to the detailed information of L i), forming a specific warning content C = W(L i). For example, for L1, the generated warning content might be: "Warning: Possible malfunction detected: Gearbox lubrication insufficient, please check lubrication level and quality."

[0145] Calculate the urgency score:

[0146] Calculate an urgency score U = Σ[w j * F j], where w j is the weight coefficient for different fault types, and F j is the frequency factor of the corresponding fault type extracted from the abnormal acoustic signature library L. For example, for the gearbox lubrication insufficient fault type, the weight coefficient w j is set to 0.7, and the frequency factor F j is 0.5, then the calculated urgency score is U = 0.7 * 0.5 = 0.35.

[0147] Create the final warning notification:

[0148] Combine the generated specific warning content C with the calculated urgency score U to create the final warning notification N = [C, U]. For example, the final warning notification N = ["Warning: Possible malfunction detected: Gearbox lubrication insufficient, please check lubrication level and quality.", 0.35]. This notification is sent to maintenance personnel or monitoring systems through pre-set communication channels (such as email, SMS or monitoring system interface), ensuring timely communication of fault information and promoting rapid response and preventive maintenance.

[0149] Through this embodiment, the established acoustic signature can be effectively compared with the pre-stored abnormal acoustic signature library to find out if there is a matching item, and when a matching item is found, a warning message is generated to indicate the possible source of failure. This method not only improves the accuracy of fault diagnosis, but also provides reliable data support for preventive maintenance, ensuring the safe and stable operation of wind turbines.

[0150] Provide maintenance recommendations according to the source of failure, and update the acoustic signature library; specifically including:

[0151] Receive the final warning notification N = [C, U], where C is the specific warning content and U is the urgency score; this step ensures that the system can obtain the latest fault warning information in time and provide the basis for subsequent maintenance recommendations and acoustic signature library updates.

[0152] Based on the warning content C and the urgency score U, a set of targeted maintenance action plans P is developed, which is based on the fault type, historical maintenance records, and current operating conditions, ensuring that each suggestion is specific and feasible. For each fault source i, a set of maintenance suggestions Si is generated, and these suggestions are integrated into the maintenance action plan P = {S1, S2,..., Sn}. This process improves the targeting and efficiency of maintenance work, reducing unnecessary maintenance costs.

[0153] After performing the maintenance actions, maintenance result feedback R is collected, and the effectiveness of the maintenance is evaluated by calculating the change in the urgency score before and after maintenance using the formula ΔU = Uafter - Ubefore, where Ubefore is the urgency score before maintenance and Uafter is the score after maintenance. This step quantifies the effectiveness of the maintenance, helping to verify the effectiveness of the maintenance measures and providing data support for future maintenance decisions.

[0154] The obtained maintenance result feedback R and score change ΔU are integrated into the voiceprint identification library L, updating the corresponding abnormal voiceprint identification Li and its associated information. At the same time, for normal voiceprint identification Dnew generated under new or improved operating conditions, the voiceprint identification library is expanded according to the formula L' = L ∪ {Dnew}. This method ensures the dynamic updating of the voiceprint identification library, enhancing the adaptability and accuracy of the system, and helping to continuously optimize the fault diagnosis capability.

[0155] The specific implementation is as follows:

[0156] Suppose a wind turbine located in a wind farm is being monitored, and the turbine has completed the initial steps of sound signal collection, filtering and enhancement, segmentation processing, frequency distribution characteristic calculation, dynamic voiceprint identification construction, and abnormal matching. Now the process of providing maintenance suggestions and updating the voiceprint identification library according to the fault source will be implemented.

[0157] Receive the final warning notification:

[0158] Receive the final warning notification N = [C, U], for example N = ["Warning: Possible fault detected: Gearbox lubrication insufficient, please check the lubricating oil level and quality.", 0.35]. This notification contains specific warning content C (gearbox lubrication insufficient) and urgency score U (0.35), ensuring that the current fault situation can be quickly understood.

[0159] Develop targeted maintenance action plans:

[0160] Based on the warning content C (gearbox lubrication insufficient) and the urgency score U (0.35), combined with historical maintenance records and current operating conditions, a set of targeted maintenance action plans P is developed. For the fault source of gearbox lubrication insufficient, a set of maintenance suggestions S1 is generated, such as:

[0161] Check if the lubricating oil level is within the normal range.

[0162] Sample and analyze the quality of the lubricating oil to confirm the presence of impurities or signs of aging.

[0163] If insufficient lubricating oil or quality issues are found, replenish or replace the lubricating oil in a timely manner.

[0164] Integrate these recommendations into the maintenance action plan P = {S1} to ensure that each recommendation is specific and feasible, facilitating the maintenance team's rapid response.

[0165] Evaluate the maintenance effect:

[0166] After performing the above maintenance actions, collect the maintenance result feedback R, such as the lubricating oil level returning to normal and the lubricating oil quality being good. Evaluate the maintenance effect by calculating the change in urgency score before and after maintenance using the formula ΔU = Uafter - Ubefore. Assuming the urgency score before maintenance is 0.35 and the score after maintenance is 0.1, then ΔU = 0.1 - 0.35 = -0.25. This change indicates that the maintenance measures effectively reduce the risk of failure.

[0167] Update the voiceprint identification library:

[0168] Integrate the obtained maintenance result feedback R (lubricating oil level returns to normal, lubricating oil quality is good) and score change ΔU = -0.25 into the voiceprint identification library L, updating the corresponding abnormal voiceprint identification Li and its associated information. For example, update the abnormal voiceprint identification L1 about insufficient lubrication of the gearbox, including the maintenance result and the score change.

[0169] At the same time, for the normal voiceprint identification Dnew generated under new or improved operating conditions, extend the voiceprint identification library according to the formula L' = L ∪ {Dnew}. For example, if the unit runs more stably after maintenance, the normal voiceprint identification at that time can be added to the voiceprint identification library to reflect the audio features under the new operating conditions.

[0170] Through this embodiment, maintenance recommendations can be effectively provided based on the fault source, and the voiceprint identification library is updated synchronously. This method not only improves the pertinence and efficiency of maintenance work, but also enhances the adaptability and accuracy of the system, ensuring the safe and stable operation of the wind turbine. In addition, the continuously updated voiceprint identification library helps to continuously optimize the fault diagnosis capability, achieving the goal of preventive maintenance.

[0171] On the other hand, the present application proposes a wind turbine fault diagnosis system based on voiceprint recognition, such as Figure 2The system comprises a sound signal collection and preliminary processing module, an audio information segmentation processing module, a frequency distribution characteristic analysis and voiceprint identification construction module, an abnormality detection and warning generation module, and a maintenance suggestion and voiceprint library updating module.

[0172] The sound signal collection and preliminary processing module is used for collecting sound signals from multiple positions during the operation of the wind turbine generator and converting the sound signals into electronic data format, removing background noise through filtering processing according to the obtained data, and performing purification of audio information.

[0173] The audio information segmentation processing module is used for segmenting the purified audio information to ensure that each segment contains sufficient information to reflect the working state of the unit.

[0174] The frequency distribution characteristic analysis and voiceprint identification construction module is used for calculating the frequency distribution characteristics of each segment of audio, forming a sound pattern that can represent different time points, and constructing a dynamic voiceprint identification using the obtained sound pattern, which can reflect the audio characteristics of the normal operation of the wind turbine generator.

[0175] The abnormality detection and warning generation module is used for comparing the established voiceprint identification with a pre-stored abnormal voiceprint identification library to find out whether there is a matching item, and if a matching abnormal voiceprint identification is found, generating warning information and indicating the possible source of failure.

[0176] The maintenance suggestion and voiceprint library updating module is used for providing maintenance suggestions according to the source of failure and updating the voiceprint identification library.

[0177] In addition, the sound signal collection and preliminary processing module, the audio information segmentation processing module, the frequency distribution characteristic analysis and voiceprint identification construction module, the abnormality detection and warning generation module, and the maintenance suggestion and voiceprint library updating module are also used to implement other steps of the wind turbine generator fault diagnosis method based on voiceprint recognition when they are executed, which will not be described here.

[0178] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the above-mentioned embodiments of the present application have been described in detail, for those skilled in the art, it still can be modified to the technical solutions recorded in the above-mentioned embodiments, or equivalent replacement of some technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A wind turbine fault diagnosis method based on voiceprint recognition, characterized in that: The following steps are involved: During the operation of the wind turbine, sound signals are collected from multiple locations and converted into electronic data format. Based on the obtained data, background noise is removed through filtering to purify the audio information. The purified audio information is segmented to ensure that each segment contains sufficient information to reflect the working status of the unit; Calculate the frequency distribution characteristics of each audio segment to form sound patterns that can represent different time points. Use the obtained sound patterns to construct a dynamic voiceprint identifier that can reflect the audio characteristics of the wind turbine when it is operating normally. Compare the established voiceprint identification with the pre-stored abnormal voiceprint identification library to find out whether there is a match. If a matching abnormal voiceprint identification is found, a warning message is generated and the corresponding fault source is indicated; Provide maintenance suggestions based on the fault source and update the voiceprint identification library; The step of calculating the frequency distribution characteristics of each audio segment to form a sound pattern that can represent different time points includes: Receive the marked valid segments and their start and end times, and extract the corresponding audio data for each valid segment; For the extracted audio data segments, fast Fourier transform is applied to convert the time domain signal into frequency domain representation; Based on the obtained frequency domain representation, the power spectral density at each frequency is calculated to quantify the energy distribution of each frequency component. Then, a frequency range is determined and multiple frequency bands are divided at fixed intervals within the range. The average power spectral density in each frequency band is calculated. The calculated average power spectrum density of each frequency band is formed into a vector, which is used as a feature descriptor that can represent the sound pattern at that time point; The construction of a dynamic voiceprint identification includes: receiving a sound pattern feature descriptor vector, applying a temporal weighting factor to each sound pattern feature descriptor to generate a weighted feature descriptor, and then calculating a cumulative value of all valid segmented weighted feature descriptors to form an overall representation that can reflect the audio characteristics of the wind turbine over a period of time; Based on the accumulated values, a dynamic voiceprint identifier is constructed, for each new voice pattern feature descriptor added; The step of comparing the established voiceprint identification with a pre-stored abnormal voiceprint identification library includes: Receive the constructed dynamic voiceprint identification and load the pre-stored abnormal voiceprint identification library, which contains multiple voiceprint identifications under abnormal circumstances; For each abnormal voiceprint identifier, calculate the similarity score between it and the dynamic voiceprint identifier, and use the Euclidean distance formula to quantify the difference between the two; Determine a preset threshold and compare all calculated similarity scores with the threshold. If any similarity score is less than or equal to the threshold, mark the abnormal voiceprint as a potential match and record the corresponding index; For all abnormal voiceprint identifications marked as potential matches, detailed information is searched in the abnormal voiceprint identification library based on their indexes, including the fault type, corresponding cause, and recommended inspection measures, and a report is generated for subsequent analysis; Generating a warning message and indicating the corresponding fault source includes: Receive abnormal voiceprint identifications marked as potential matches and their corresponding indexes, and obtain detailed information associated with them, including fault types and recommended inspection measures; For each marked abnormal voiceprint identifier, a warning message template is constructed. The template contains a fixed part and a variable part. The variable part is filled in according to the detailed information to form a specific warning content, ensuring that each warning message clearly points out the corresponding fault source; Calculate an urgency score, combine the generated specific warning content with the calculated urgency score, and create a final warning notification, which is sent to maintenance personnel or a monitoring system through a pre-set communication channel; Providing maintenance suggestions based on the fault source and updating the voiceprint identification library at the same time include: Receive the final warning notification and develop a targeted maintenance action plan based on the warning content and urgency score. This plan is based on the fault type, historical maintenance records, and current operating conditions to ensure that each suggestion is specific and feasible; After performing the maintenance action, collect maintenance result feedback, evaluate the maintenance effect, integrate the maintenance result feedback and score changes into the voiceprint identification library, update the corresponding abnormal voiceprint identification and its associated information, and at the same time, for the normal voiceprint identification generated under new or improved operating conditions.

2. A wind turbine fault diagnosis method based on voiceprint recognition according to claim 1, characterized in that: The method of collecting sound signals from multiple locations and converting the sound signals into electronic data format includes: Select N different locations within the wind turbine as monitoring points, where N is a positive integer, and deploy acoustic sensors at each monitoring point. Each acoustic sensor captures an analog signal representing the time-varying sound amplitude. Through discretization processing of the sampling rate, a series of discrete time-amplitude data pairs are obtained; Based on the discrete data pairs, the Fourier transform formula is applied to convert the time domain signal into the frequency domain to obtain the spectrum intensity at the corresponding frequency, thereby forming a frequency domain representation in the electronic data format; The generated frequency domain data is grouped according to preset frequency intervals, and the mean and standard deviation of all spectrum intensities in each group are calculated to construct a multidimensional vector that can reflect the audio characteristics of each monitoring point.

3. A wind turbine fault diagnosis method based on voiceprint recognition according to claim 2, characterized in that: The filtering process to remove background noise and purify audio information includes: Receive a multidimensional vector from, apply an adaptive threshold to determine the upper limit of the background noise level, and identify and mark the data segments exceeding the threshold as potential target sound features by comparing the mean of each frequency band with the adaptive threshold; For the marked data segments, a weighting matrix is ​​constructed and applied to the original multidimensional vector to generate a new vector to highlight the sound features after filtering processing.

4. A wind turbine fault diagnosis method based on voiceprint recognition according to claim 3, characterized in that: The segmentation processing of the purified audio information includes: Receive a new vector containing the audio features after filtering and enhancement; Define a time window and overlap ratio to determine the time length of each segment and the overlap between adjacent segments, and calculate the starting time point of each segment; For each defined segment, calculate the cumulative energy of the audio features within that period and identify segments with significant energy changes; Mark the segments that meet the conditions as valid segments and record the corresponding start time and end time. For consecutive valid segments, merge them to form a longer time period to ensure that each segment contains sufficient information to reflect the working status of the unit.

5. A wind turbine fault diagnosis system based on voiceprint recognition for implementing the method according to any one of claims 1 to 4, characterized in that: include: The sound signal acquisition and preliminary processing module is used to collect sound signals from multiple locations during the operation of the wind turbine and convert the sound signals into electronic data format. Based on the obtained data, background noise is removed through filtering to purify the audio information; The audio information segmentation processing module is used to segment the purified audio information to ensure that each segment contains sufficient information to reflect the working status of the unit; The frequency distribution characteristics analysis and voiceprint identification construction module is used to calculate the frequency distribution characteristics of each audio segment, forming sound patterns that can represent different time points. The obtained sound patterns are used to construct a dynamic voiceprint identification, which can reflect the audio characteristics of the wind turbine when it is operating normally. The anomaly detection and warning generation module is used to compare the established voiceprint identification with the pre-stored abnormal voiceprint identification library to find out whether there is a match. If a matching abnormal voiceprint identification is found, a warning message is generated and the corresponding fault source is indicated; The maintenance suggestion and voiceprint library update module is used to provide maintenance suggestions based on the fault source and update the voiceprint identification library at the same time.

Images