A wind turbine noise monitoring method, system and electronic device

Abstract

The present application relates to the technical field of noise monitoring, and discloses a wind turbine noise monitoring method, system and electronic device, the method comprising: acquiring audio streams of a wind turbine at multiple preset noise monitoring points; performing spatial sound pressure judgment on the audio streams to screen out pending audio segments in the audio streams; screening out abnormal audio segments from the pending audio segments based on differences between the pending audio segments and reference audio; and inputting the abnormal audio segments into an abnormal voiceprint library and a sound large model for parallel identification to obtain a noise identification result of the wind turbine. Through the cooperative identification of spatial sound pressure judgment, abnormal preliminary screening, a voiceprint library and a large model, the present application realizes efficient and accurate wind turbine noise monitoring and improves the accuracy and reliability of noise monitoring under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of noise monitoring technology, specifically to a method, system, and electronic equipment for monitoring the noise of wind turbine generators. Background Technology

[0002] Excessive noise from wind turbines is often an acoustic manifestation of early problems such as gearbox wear, blade damage, or yaw system malfunctions. Failure to identify and address these issues promptly can lead not only to complaints from nearby residents and compliance risks, but also to performance degradation and even major mechanical failures. Therefore, effective noise monitoring of wind turbines is a problem that needs to be addressed.

[0003] Currently, while machine learning-based automatic noise monitoring solutions have achieved initial applications in the wind power sector, they still face technical bottlenecks when applied in complex industrial sound environments. Firstly, environmental noise sources (such as transport vehicles, agricultural machinery, and human activity) are typically located in the near-field region of the monitoring point, where their signal strength is higher than that of wind turbine noise in the far field, causing the target signal to be submerged and resulting in a low signal-to-noise ratio. Secondly, the types of noise generated by natural environments and human activities are diverse, with significant differences in characteristics. For example, sudden bird calls, continuous wind and rain sounds, and rhythmic mechanical sounds have completely different time-frequency distributions, placing extremely high demands on the generalization and robustness of noise detection models. Thirdly, high-quality abnormal wind turbine audio data for training is scarce. Existing datasets are mostly collected under ideal conditions with clean backgrounds, failing to fully reproduce the complex working conditions of various noises mixed together on-site. This leads to a decrease in model recognition accuracy and an increase in false alarm rate after actual deployment, severely restricting its engineering reliability and large-scale promotion. Summary of the Invention

[0004] This invention provides a method, system, and electronic device for monitoring noise in wind turbine generators, in order to solve one of the technical problems existing in the prior art.

[0005] In a first aspect, the present invention provides a method for monitoring noise in a wind turbine generator set, the method comprising: Acquire audio streams from wind turbine generators at multiple preset noise monitoring points; Spatial sound pressure level is determined in the audio stream to filter out undetermined audio segments. Based on the differences between the pending audio segments and the baseline audio, abnormal audio segments are filtered out from the pending audio segments; Abnormal audio segments are input into the abnormal voiceprint database and the large sound model for parallel recognition to obtain the noise recognition results of the wind turbine.

[0006] This invention acquires audio streams from wind turbines at multiple preset noise monitoring points and performs spatial sound pressure assessment. Without relying on signal strength, it pre-identifies the sound source properties at a physical level, reducing invalid data upload rates and network bandwidth consumption, thus alleviating computational burden in subsequent identification stages. Furthermore, it filters out anomalous audio segments deviating from the baseline audio from the pending audio segments, effectively reducing the amount of data processed in subsequent stages and improving identification efficiency. Finally, the anomalous audio segments are input into an anomalous speaker database and a large-scale sound model for dual-channel parallel identification. This allows the two channels to complement each other in terms of knowledge sources and identification mechanisms, effectively overcoming the problem of insufficient reliability of a single model under conditions of scarce anomalous data. This achieves efficient and accurate wind turbine noise monitoring, improving the accuracy and reliability of noise monitoring under complex operating conditions.

[0007] In one optional implementation, after acquiring the audio streams of the wind turbine at multiple preset noise monitoring points, the method further includes: Preprocess the audio stream; The preprocessed audio stream is divided into multiple audio segments.

[0008] This invention improves data quality through preprocessing and then divides the audio stream into audio segments, laying a structured data foundation for subsequent processing.

[0009] In one optional implementation, spatial sound pressure level is determined in the audio stream to filter out undetermined audio segments, including: For each audio segment in the audio stream, extract the equivalent continuous sound pressure level at each preset noise monitoring point; Based on the preset distance and equivalent continuous sound pressure level of each preset noise monitoring point, a sound pressure-distance curve is constructed; Calculate the fitting error between the sound pressure-distance curve and the theoretical radiation attenuation model; Audio segments with fitting errors less than a first preset threshold are identified as undetermined audio segments.

[0010] This invention extracts the equivalent continuous sound pressure level at each point to construct a measured sound pressure-distance curve, calculates the fitting error between the curve and the theoretical radiation attenuation model, and compares it with a first preset threshold to achieve automatic screening. This allows for pre-judgment of the sound source properties at the physical level, effectively reducing the invalid data upload rate and the computational load of subsequent identification processes.

[0011] In one optional implementation, abnormal audio segments are filtered from the undetermined audio segments based on the differences between the undetermined audio segments and the reference audio, including: Extract the first spectrogram of the audio segment to be determined; The first spectrogram is input into the autoencoder to obtain the second spectrogram, which is obtained by pre-training the autoencoder based on the reference audio. Calculate the reconstruction error between the first and second spectrograms; Audio segments whose reconstruction error exceeds the second preset threshold are identified as abnormal audio segments.

[0012] This invention extracts the spectrogram of an audio segment to be determined and inputs it into an autoencoder trained based on a benchmark audio to obtain a reconstructed spectrogram. By calculating the reconstruction error between the original spectrogram and the reconstructed spectrogram, the degree to which the audio segment deviates from the normal pattern is quantified. Segments with reconstruction errors greater than a second preset threshold are then identified as abnormal and pushed to subsequent dual-channel recognition, reducing the computational load of subsequent steps and improving recognition efficiency.

[0013] In one optional implementation, the abnormal audio segments are input into an abnormal speaker database and a large sound model for parallel recognition to obtain the noise recognition results of the wind turbine, including: Abnormal audio segments are identified using an abnormal voiceprint database to obtain the first identification result; A large sound model is used to identify abnormal audio segments, resulting in a second identification result; Based on the first and second identification results, a collaborative decision is made to obtain the noise identification result of the wind turbine.

[0014] This invention employs a dual-channel parallel recognition approach using a voiceprint database and a large model. These two approaches complement each other in terms of knowledge sources and recognition mechanisms. Based on the recognition results of both approaches, collaborative decision-making is performed to obtain the final noise recognition result. This effectively overcomes the problem of insufficient reliability of single-model recognition, reduces the risk of false alarms and false negatives, and improves the accuracy of recognition.

[0015] In one optional implementation, an abnormal audio segment is identified using an abnormal voiceprint database to obtain a first identification result, including: An encoder is used to convert abnormal audio segments into abnormal feature vectors; Calculate the similarity between the abnormal feature vector and the feature vector of each sample in the abnormal voiceprint database; The noise classification label of the sample feature vector corresponding to the highest similarity is used as the noise recognition result, and the highest similarity is used as the confidence level to form the first recognition result.

[0016] This invention converts abnormal audio segments into abnormal feature vectors using an encoder, calculates similarity with samples in a voiceprint database, and performs identification based on the similarity results, thus achieving fast and accurate noise identification based on local prior knowledge.

[0017] In one optional implementation, a large sound model is used to identify abnormal audio segments to obtain a second identification result, including: The sound big data model is used to identify abnormal audio segments and obtain the probability distribution of each candidate category; The candidate classification corresponding to the highest probability in the probability distribution is taken as the noise recognition result, and the highest probability is taken as the confidence level to form the second recognition result.

[0018] This invention achieves noise recognition based on prior knowledge of the large sound model by using a large sound model for recognition. This complements the voiceprint database channel and improves the accuracy of noise recognition.

[0019] In one optional implementation, both the first identification result and the second identification result include noise identification results and confidence levels; Based on the first and second identification results, a collaborative decision is made to obtain the noise identification results of the wind turbine, including: When the noise identification results of the first identification result and the second identification result are consistent, the noise identification result shall be taken as the noise identification result of the wind turbine. When the noise identification results of the first identification result and the second identification result are inconsistent, and the confidence level of the first identification result or the second identification result is greater than the third preset threshold, the noise identification result corresponding to the higher confidence level shall be taken as the noise identification result of the wind turbine. When the noise identification results of the first identification result and the second identification result are inconsistent, and the confidence levels of both the first identification result and the second identification result are less than the third preset threshold, the abnormal audio segment is marked as a conflict sample.

[0020] This invention utilizes the recognition results from a voiceprint database and a large model for collaborative decision-making, effectively suppressing the risk of misjudgment from a single channel and improving recognition efficiency and reliability.

[0021] Secondly, the present invention provides a wind turbine noise monitoring system, the system comprising: The perception layer is used to acquire the audio stream of the wind turbine at multiple preset noise monitoring points; The edge layer is used to determine the spatial sound pressure level of the audio stream and filter out the audio segments to be determined in the audio stream. The algorithm layer is used to filter out abnormal audio segments from the undetermined audio segments based on the difference between the undetermined audio segments and the reference audio. The abnormal audio segments are then input into the abnormal voiceprint library and the sound big model for parallel recognition to obtain the noise recognition results of the wind turbine.

[0022] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the wind turbine noise monitoring method of the first aspect or any corresponding embodiment described above.

[0023] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the wind turbine noise monitoring method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of a wind turbine noise monitoring system according to an embodiment of the present invention; Figure 2 This is a flowchart of a wind turbine noise monitoring method according to an embodiment of the present invention; Figure 3 This is a flowchart of another wind turbine noise monitoring method according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0028] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0029] Figure 1 This is a schematic diagram of a wind turbine noise monitoring system according to an embodiment of the present invention, as shown below. Figure 1 As shown, the system includes: a perception layer, used to acquire audio streams of wind turbines at multiple preset noise monitoring points; an edge layer, used to perform spatial sound pressure judgment on the audio streams and filter out undetermined audio segments in the audio streams; and an algorithm layer, used to filter out abnormal audio segments from the undetermined audio segments based on the differences between the undetermined audio segments and the reference audio, and input the abnormal audio segments into the abnormal voiceprint library and the sound big model for parallel recognition to obtain the noise recognition results of the wind turbines.

[0030] Specifically, the perception layer constitutes the data acquisition front end of the entire system, consisting of intelligent monitoring terminals strictly deployed in the wind farm according to acoustic standards. This ensures that the acquired audio streams are highly representative, comparable, and reliable, laying a solid data foundation for upper-level intelligent analysis. The edge layer is the computing hub deployed at the monitoring site, responsible for real-time processing of the audio streams reported by the perception layer, performing spatial sound pressure judgment, and achieving data cleaning and condensation. Only when anomalies are found are the corresponding audio segments uploaded to the upper layer, optimizing network bandwidth utilization and reducing cloud computing pressure. The algorithm layer is the core of the system's intelligent processing. It receives the filtered audio segments reported by the edge layer, first initially screening out abnormal audio segments, then performing a collaborative recognition process on them, and outputting noise recognition results after comprehensive decision-making.

[0031] Optionally, the system also includes an interaction layer, which serves as the interface for presenting the system's value and facilitating user interaction. This layer transforms the structured noise recognition results output by the algorithm layer into diverse business services and an intuitive user interface. It includes human-computer interaction functions such as alarms, data display, and requests for manual review, facilitating operation.

[0032] Through a hierarchical architecture that integrates cloud, edge, and device, the system progresses step by step through three levels: physical screening, initial model screening, and collaborative identification. This approach fully leverages the real-time advantages of edge computing and the cognitive capabilities of large cloud models while reducing the risks of false alarms and missed alarms, thus achieving efficient and reliable noise monitoring of wind turbine units.

[0033] According to an embodiment of the present invention, a method for monitoring noise of a wind turbine is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] This embodiment provides a method for monitoring wind turbine noise, which can be used in the aforementioned wind turbine noise monitoring system. Figure 2 This is a flowchart of a wind turbine noise monitoring method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain the audio stream of the wind turbine at multiple preset noise monitoring points.

[0035] Specifically, intelligent monitoring terminals are deployed at key locations such as the near field, mid field, far field, and background-sensitive area of ​​the wind turbine as preset noise monitoring points to form a multi-point acoustic monitoring array covering different spatial distances around the wind turbine. Each intelligent monitoring terminal consists of a high-performance omnidirectional MEMS (Micro-Electro-Mechanical Systems) microphone, a noise monitoring unit compliant with Class 1 sound level meters, a controller unit with edge computing and communication capabilities, and an all-weather protective enclosure. It is installed on a wind-resistant metal bracket and supports intelligent power supply from solar energy and mains power to ensure long-term stable operation in harsh environments.

[0036] The perception layer synchronously and continuously acquires the raw audio stream through these preset noise monitoring points and transmits it to the intelligent devices in the edge layer in real time, laying a high-quality data foundation for upper-layer intelligent analysis.

[0037] Step S202: Perform spatial sound pressure judgment on the audio stream and filter out the undetermined audio segments in the audio stream.

[0038] Specifically, the edge layer performs spatial sound pressure assessment on the audio stream based on the geometric diffusion attenuation and atmospheric absorption laws of sound waves in free / semi-free fields. The core principle of this assessment is that the sound pressure level radiated by a point sound source decreases logarithmically with increasing propagation distance. By assessing spatial sound pressure, it can be determined whether the current audio event is a suspicious sound event emitted by the wind turbine itself. If so, the corresponding audio segment is retained and pushed to the algorithm layer, thereby pre-judging the nature of the sound source from a physical perspective without relying on signal strength. This significantly reduces the rate of invalid data uploads and network bandwidth consumption, alleviating the computational burden on subsequent identification stages.

[0039] Step S203: Based on the difference between the pending audio segments and the reference audio, abnormal audio segments are selected from the pending audio segments.

[0040] Specifically, the algorithm layer receives the spatially filtered audio segments to be determined, and based on the difference between them and the baseline audio, it initially filters out abnormal audio segments from the audio segments to achieve rapid filtering of normal baseline audio and accurate retention of suspicious noise segments, effectively reducing the amount of data processed in subsequent stages.

[0041] Step S204: Input the abnormal audio segments into the abnormal voiceprint library and the large sound model respectively for parallel recognition to obtain the noise recognition results of the wind turbine.

[0042] Specifically, the algorithm layer inputs the initially screened abnormal audio segments into two channels—an abnormal speaker database and a large sound model—for parallel recognition. The recognition results from both channels are then combined to obtain the noise recognition result for the wind turbine. By constructing a dual-channel parallel recognition architecture, the two channels complement each other in terms of knowledge sources and recognition mechanisms, effectively overcoming the problem of insufficient reliability of a single model under conditions of scarce abnormal data, and achieving efficient and accurate noise monitoring.

[0043] The wind turbine noise monitoring method provided in this embodiment acquires audio streams from multiple preset noise monitoring points of the wind turbine and performs spatial sound pressure assessment. Without relying on signal strength, it pre-determines the nature of the sound source at a physical level, reducing the rate of invalid data uploads and network bandwidth consumption, thus alleviating the computational burden on subsequent identification stages. Furthermore, it filters out abnormal audio segments that deviate from the baseline audio from the pending audio segments, effectively reducing the amount of data processed in subsequent stages and improving identification efficiency. Finally, the abnormal audio segments are input into an abnormal speaker database and a large sound model for dual-channel parallel identification. This allows the two channels to complement each other in terms of knowledge sources and identification mechanisms, effectively overcoming the problem of insufficient reliability of a single model under conditions of scarce abnormal data. This achieves efficient and accurate wind turbine noise monitoring, improving the accuracy and reliability of noise monitoring under complex operating conditions.

[0044] This embodiment provides a method for monitoring wind turbine noise, which can be used in the aforementioned wind turbine noise monitoring system. The method specifically includes the following steps: Step S301: Acquire the audio streams of the wind turbine at multiple preset noise monitoring points. For details, please refer to [link to relevant documentation]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0045] Step S302: Preprocess the audio stream.

[0046] Specifically, the perception layer performs noise reduction and frame segmentation on the acquired raw audio stream to improve data quality. Then, each frame of data is assigned a precise timestamp and location label to ensure the temporal synchronization and spatial traceability of the data at each location.

[0047] Step S303: Divide the preprocessed audio stream into multiple audio segments.

[0048] Specifically, the preprocessed audio stream is divided into multiple audio segments according to a preset time window, with each audio segment corresponding to a continuous duration of sound signal. It should be noted that since the audio stream is obtained through synchronous acquisition from multiple points, each audio segment is essentially a collection of audio data from each preset noise monitoring point within the time window of that audio segment, that is, sound signals from different locations within the same time period.

[0049] Taking a preset time window of 1 second and 4 preset noise monitoring points as an example: After the audio stream is divided into audio segments according to the preset time window, in the audio segment of T=0~1 second, point 1, point 2, point 3 and point 4 each generate a 1 second audio segment.

[0050] Step S304: Perform spatial sound pressure judgment on the audio stream and filter out the undetermined audio segments in the audio stream.

[0051] Specifically, step S304 includes: Step S3041: For each audio segment in the audio stream, extract the equivalent continuous sound pressure level of each preset noise monitoring point.

[0052] Specifically, for each audio segment, the equivalent continuous sound pressure level (Leq) of each preset noise monitoring point within the time interval of that audio segment is extracted. The equivalent continuous sound pressure level is a statistical quantity obtained by averaging the instantaneous sound pressure over a period of time, which can stably reflect the average sound energy level of that point within that time window.

[0053] Step S3042: Based on the preset distance and equivalent continuous sound pressure level of each preset noise monitoring point, construct a sound pressure-distance curve.

[0054] Specifically, the preset distance for each preset noise monitoring point represents the distance between the point and the wind turbine, which is a fixed value measured and calibrated in advance during the deployment of the points. Based on the preset distance of each point and the calculated equivalent continuous sound pressure level, a sound pressure-distance curve corresponding to the current audio segment is constructed with the preset distance as the abscissa and the equivalent continuous sound pressure level as the ordinate. This curve reflects the actual distribution of sound pressure energy with spatial distance within the same time window.

[0055] Step S3043: Calculate the fitting error between the sound pressure-distance curve and the theoretical radiation attenuation model.

[0056] Specifically, the fitting error between the sound pressure-distance curve of the current audio segment and the theoretical radiation attenuation model is calculated using the following formula:

[0057] In the formula, i and j represent the indices of the preset noise monitoring points; n represents the number of preset noise monitoring points; This represents the equivalent continuous sound pressure level at preset noise monitoring points i and j; This represents the theoretical radiation attenuation model; This indicates the preset distance between preset noise monitoring points i and j and the wind turbine; This represents the atmospheric absorption coefficient.

[0058] The above formula determines whether a sound source originates from a wind turbine by comparing the consistency between the "actual attenuation pattern of sound propagation" and the "attenuation pattern predicted by physical laws." As a fixed-point sound source, the sound pressure attenuation with distance from a wind turbine is predictable. However, other types of sound sources have variable locations and propagation paths, resulting in attenuation patterns that differ significantly from those of wind turbine sources. This fitting error correction does not require prior knowledge of the absolute intensity of the sound source; it only needs to examine whether the trend of sound pressure level differences between points matches the theoretical model to determine the nature of the sound source at a physical level, thus exhibiting stronger engineering robustness and field applicability.

[0059] Step S3044: Audio segments with fitting errors less than a first preset threshold are identified as undetermined audio segments.

[0060] Specifically, if the fitting error is less than the first preset threshold, it indicates that the multi-point sound pressure distribution of the current audio segment conforms to the spatial attenuation characteristics of the airborne sound source, and may be the sound emitted by the wind turbine itself. Therefore, the current audio segment is identified as a pending audio segment for further identification. If the fitting error is not less than the first preset threshold, it indicates that the measured sound pressure attenuation trend deviates from the theoretical model, and there is an abnormal sound pressure gradient (such as a sudden increase in far-field sound pressure, failure to reach the threshold in the near field, or no spatial correlation between sound pressure points). This is determined to be an independent environmental interference source (such as distant road vehicles, nearby agricultural machinery, etc.) and is directly filtered out, not entering the subsequent identification process. By pre-judging the nature of the sound source at the physical level, a large number of non-airborne interference sources are effectively eliminated, reducing the computational load and false alarm risk in the subsequent identification process.

[0061] Step S305: Based on the difference between the pending audio segments and the reference audio, filter out the abnormal audio segments from the pending audio segments.

[0062] Specifically, step S305 includes: Step S3051: Extract the first spectrogram of the audio segment to be determined.

[0063] Specifically, the log-Mel spectrum of the audio segment to be determined, i.e. the first spectrogram, is extracted as a standard feature input representing the time-frequency characteristics of the audio.

[0064] Step S3052: Input the first spectrogram into the autoencoder to obtain the second spectrogram, which is obtained by pre-training the autoencoder based on the reference audio.

[0065] Specifically, the reference audio includes ambient audio and audio from wind turbines during normal operation. The autoencoder is a lightweight sound classification model employing an encoder-decoder symmetric structure. During training, the log-Mel spectrum of the reference audio is used as both input and output targets. By minimizing the mean square error between input and output, the encoder is forced to extract the most essential low-dimensional features of normal acoustic patterns, while the decoder learns the ability to recover the original spectrum from these low-dimensional features. This results in the autoencoder ultimately learning efficient compression representation and reconstruction capabilities for both normal and background sound features. Inputting the first spectrogram into the autoencoder yields the corresponding reconstructed spectrogram, i.e., the second spectrogram.

[0066] Step S3053: Calculate the reconstruction error between the first spectrogram and the second spectrogram.

[0067] Specifically, the mean square error between the first and second spectrograms is calculated as the reconstruction error.

[0068] Step S3054: The audio segments to be determined with reconstruction errors greater than the second preset threshold are identified as abnormal audio segments.

[0069] Specifically, the second preset threshold is set based on the statistical distribution of reconstruction error of the autoencoder on the validation set. If the reconstruction error of the audio segment to be identified is greater than the second preset threshold, it indicates that the acoustic features of the audio segment significantly deviate from the pattern distribution of the normal and background categories, and it is identified as an abnormal audio segment and pushed to the subsequent dual-channel parallel recognition process; if the reconstruction error is not greater than the second preset threshold, it indicates that the audio segment to be identified conforms to the acoustic patterns of the normal and background categories in the training data, and it is judged as normal audio and directly filtered out. By performing initial screening based on reconstruction error, the processing load of subsequent stages is effectively reduced, and the recognition efficiency is improved.

[0070] Step S306: Input the abnormal audio segments into the abnormal voiceprint library and the large sound model respectively for parallel recognition to obtain the noise recognition result of the wind turbine.

[0071] Specifically, step S306 includes: Step S3061: The abnormal audio segment is identified using the abnormal voiceprint database to obtain the first identification result.

[0072] In some optional implementations, step S3061 above includes: Step a1: Use an encoder to convert the abnormal audio segments into abnormal feature vectors.

[0073] Specifically, the abnormal voiceprint database is a pre-built local feature library. Each sample in the database is extracted into a high-dimensional voiceprint embedding vector using a deep neural network encoder, and then associated and stored with its corresponding noise classification label, covering various wind turbine faults and common local environmental noises. This deep neural network encoder is trained using a metric learning method to ensure that feature vectors of similar samples are close to each other in the vector space, while feature vectors of dissimilar samples are far apart. To facilitate subsequent similarity calculation and matching, the same deep neural network encoder used when building the voiceprint database is employed to convert abnormal audio segments into abnormal feature vectors with the same format as the sample feature vectors in the database.

[0074] Step a2: Calculate the similarity between the abnormal feature vector and the feature vector of each sample in the abnormal voiceprint database.

[0075] Specifically, the cosine similarity between the abnormal feature vector and the feature vector of each sample in the abnormal voiceprint database is calculated. The closer the value is to 1, the more similar the two vectors are at the acoustic feature level, thereby quantifying the similarity between the abnormal feature vector and various known anomalies.

[0076] Step a3: The noise classification label of the sample feature vector corresponding to the highest similarity is taken as the noise recognition result, and the highest similarity is taken as the confidence level to form the first recognition result.

[0077] Specifically, the highest similarity value is selected from all calculated similarities, and the noise classification label of the sample feature vector corresponding to this highest similarity value is used as the noise recognition result. This noise classification label is a pre-labeled noise category name from the samples in the database, such as gear wear or blade damage. Simultaneously, the highest similarity value is used as the confidence level of this recognition result, together forming the first recognition result. Recognition using a local speaker database achieves fast and accurate noise recognition based on local prior knowledge.

[0078] Step S3062: Use a large sound model to identify abnormal audio segments and obtain a second identification result.

[0079] Specifically, step S3062 includes: Step b1: Use a large sound model to identify abnormal audio segments and obtain the probability distribution of each candidate category.

[0080] Specifically, large-scale sound models (such as wav2vec and BEATs) are large-scale neural network models pre-trained on massive, highly labeled general-purpose audio datasets covering hundreds of thousands of hours. Their core capability lies in learning the general acoustic representations of massive amounts of natural and artificial sounds, enabling them to extract deep features containing advanced semantic information from raw audio or spectrograms, and possessing strong recognition capabilities for sound categories, attributes, and scenes. When an abnormal audio segment is input into this large-scale sound model, leveraging its zero-shot / few-shot generalization capabilities and massive prior knowledge of environmental acoustics, the model outputs the probability distribution of the abnormal audio segment belonging to each candidate category after forward inference. For example, the model might output: wind and rain sound: 0.85, birdsong sound: 0.08, vehicle engine sound: 0.04, windmill gear wear: 0.02, others: 0.01.

[0081] Step b2: The candidate classification corresponding to the highest probability in the probability distribution is taken as the noise recognition result, and the highest probability is taken as the confidence level to form the second recognition result.

[0082] Specifically, the candidate category corresponding to the highest probability in the probability distribution is taken as the noise recognition result, and this highest probability is used as the confidence level to form a second recognition result. By using a large sound model for recognition, noise recognition based on the prior knowledge of the large model is achieved, complementing the voiceprint database channel.

[0083] It should be noted that the above steps S3061 and S3062 are executed in parallel, that is, the abnormal audio segments are simultaneously identified by both the voiceprint database and the large model.

[0084] Step S3063: Based on the first identification result and the second identification result, a collaborative decision is made to obtain the noise identification result of the wind turbine. Both the first identification result and the second identification result include the noise identification result and the confidence level.

[0085] Specifically, step S3063 includes: Step c1: When the noise identification results of the first identification result and the second identification result are consistent, the noise identification result is taken as the noise identification result of the wind turbine.

[0086] Specifically, the noise identification results in the first identification result and the second identification result are compared. If they are consistent, it means that the noise classification of the abnormal audio segment by the voiceprint library and the large model is mutually verified. The noise identification result in either identification result is taken as the noise identification result of the wind turbine.

[0087] Step c2: When the noise identification results of the first identification result and the second identification result are inconsistent, and the confidence level of the first identification result or the second identification result is greater than the third preset threshold, the noise identification result corresponding to the higher confidence level shall be taken as the noise identification result of the wind turbine.

[0088] Specifically, if the noise identification results in the first and second identification results are inconsistent, it indicates a discrepancy between the voiceprint database and the large model in classifying the abnormal audio segment as noise. In this case, the confidence level is further assessed. If at least one of the confidence levels is greater than a third preset threshold, the noise identification result with the higher confidence level is adopted as the noise identification result for the wind turbine. The third preset threshold needs to be set relatively high to avoid misjudgment.

[0089] Step c3: When the noise recognition results of the first recognition result and the second recognition result are inconsistent, and the confidence levels of both the first recognition result and the second recognition result are less than the third preset threshold, the abnormal audio segment is marked as a conflict sample.

[0090] Specifically, if the noise identification results in the first and second identification results are inconsistent, and the confidence levels of both are less than the third preset threshold, it indicates that neither the voiceprint database nor the large model can provide sufficiently reliable judgment criteria, and the system cannot automatically make a credible decision regarding this abnormal audio segment. In this case, the abnormal audio segment is marked as a low-confidence conflict sample and retained as a difficult case.

[0091] Optionally, the interaction layer can execute different response strategies based on the identification results of step S306. If the noise identification result belongs to fan noise, a structured alarm event containing information such as the noise identification result and confidence level is generated; if the noise identification result belongs to known environmental noise, it is recorded and archived without alarm; if it is a low-confidence conflict sample, it is stored in the verification queue and a notification is triggered for manual verification. After manual verification, it can be added to the abnormal voiceprint database for incremental updates.

[0092] In some alternative implementations, Figure 3 This is a flowchart of another wind turbine noise monitoring method according to an embodiment of the present invention, such as... Figure 3 As shown, the audio stream of the wind turbine is acquired and preprocessed. Spatial sound pressure level is assessed to filter out undetermined audio segments. An autoencoder trained on a benchmark audio source is used to filter out abnormal audio segments from the undetermined segments. These abnormal audio segments are simultaneously input into an abnormal speaker database and a large-scale sound model for parallel recognition. Based on the recognition results from both, a collaborative decision is made to obtain the noise recognition result of the wind turbine. If the noise recognition result is wind turbine noise, an alarm is triggered; if it is known environmental noise, no alarm is triggered; and if it is a conflicting sample, manual verification is performed.

[0093] The wind turbine noise monitoring method provided in this embodiment acquires audio streams from multiple preset noise monitoring points of the wind turbine and performs spatial sound pressure assessment. Without relying on signal strength, it pre-determines the nature of the sound source at a physical level, reducing the rate of invalid data uploads and network bandwidth consumption, thus alleviating the computational burden on subsequent identification stages. Furthermore, it filters out abnormal audio segments that deviate from the baseline audio from the pending audio segments, effectively reducing the amount of data processed in subsequent stages and improving identification efficiency. Finally, the abnormal audio segments are input into an abnormal speaker database and a large sound model for dual-channel parallel identification. This allows the two channels to complement each other in terms of knowledge sources and identification mechanisms, effectively overcoming the problem of insufficient reliability of a single model under conditions of scarce abnormal data. This achieves efficient and accurate wind turbine noise monitoring, improving the accuracy and reliability of noise monitoring under complex operating conditions.

[0094] The present invention has the following beneficial effects: (1) Constructing a multi-recognition architecture: Constructing a multi-collaborative recognition architecture that includes sound pressure judgment, lightweight model inference, voiceprint comparison, and large sound model analysis. At the physical level, interference sound sources are filtered out, and at the algorithm level, normal sound, fan malfunctions, and environmental noise are effectively distinguished, greatly reducing false alarms and false negatives; (2) Optimize human-machine collaboration efficiency: The decision-making mechanism based on confidence management can automatically handle the vast majority of routine events and submit only a very small number of high uncertainty samples for manual verification, thereby greatly reducing the reliance on expert experience and the frequency of manual intervention, and achieving overall optimization of operation and maintenance costs and response efficiency; (3) Introducing prior knowledge of large models: Innovatively utilizing large sound models pre-trained in massive amounts of general audio data as recognition components to enhance the system's ability to identify and eliminate diverse unknown environmental noise; (4) Construct an edge-cloud collaborative computing architecture: Lightweight computing tasks such as sound pressure comprehensive judgment and normal sound separation can be completed in the edge computing unit to achieve real-time response and data filtering, while voiceprint database comparison, large model inference and collaborative decision-making that require a lot of computing can be carried out in the cloud server, balancing real-time performance, accuracy, bandwidth consumption and computing cost, making large-scale deployment possible.

[0095] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0096] The following is a detailed reference. Figure 4This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from memory 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device. The processor 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0097] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 408 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0098] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 409, or installed from a memory 408, or installed from a ROM 402. When the computer program is executed by the processor 401, it performs the functions defined in the wind turbine noise monitoring method of the embodiments of the present invention.

[0099] Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0100] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the wind turbine noise monitoring method shown in the above embodiments is implemented.

[0101] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for monitoring noise in wind turbine generators, characterized in that, The method includes: Acquire audio streams from wind turbine generators at multiple preset noise monitoring points; Spatial sound pressure level is determined in the audio stream to filter out undetermined audio segments in the audio stream; Based on the difference between the undetermined audio segment and the reference audio, abnormal audio segments are filtered out from the undetermined audio segments; The abnormal audio segments are input into the abnormal voiceprint library and the sound model respectively for parallel recognition to obtain the noise recognition result of the wind turbine.

2. The method according to claim 1, characterized in that, After acquiring the audio streams of the wind turbine at multiple preset noise monitoring points, the method further includes: The audio stream is preprocessed; The preprocessed audio stream is divided into multiple audio segments.

3. The method according to claim 2, characterized in that, The step of performing spatial sound pressure level determination on the audio stream and filtering out undetermined audio segments from the audio stream includes: For each audio segment in the audio stream, extract the equivalent continuous sound pressure level at each preset noise monitoring point; Based on the preset distance and equivalent continuous sound pressure level of each preset noise monitoring point, a sound pressure-distance curve is constructed; Calculate the fitting error between the sound pressure-distance curve and the theoretical radiation attenuation model; Audio segments with fitting errors less than a first preset threshold are identified as the audio segments to be determined.

4. The method according to claim 1, characterized in that, The step of filtering out abnormal audio segments from the undetermined audio segments based on the differences between the undetermined audio segments and the reference audio includes: Extract the first spectrogram of the audio segment to be determined; The first spectrogram is input into the autoencoder to obtain the second spectrogram, the autoencoder being based on... The reference audio is obtained through pre-training; Calculate the reconstruction error between the first spectrogram and the second spectrogram; Audio segments whose reconstruction error is greater than a second preset threshold are identified as abnormal audio segments.

5. The method according to claim 1, characterized in that, The step of inputting the abnormal audio segments into an abnormal voiceprint database and a large sound model for parallel recognition to obtain the noise recognition result of the wind turbine includes: The abnormal audio segment is identified using the aforementioned abnormal voiceprint database to obtain a first identification result; The abnormal audio segment is identified using the aforementioned sound model to obtain a second identification result; Based on the first identification result and the second identification result, a collaborative decision is made to obtain the noise identification result of the wind turbine.

6. The method according to claim 5, characterized in that, The step of using the abnormal voiceprint database to identify the abnormal audio segment and obtaining a first identification result includes: An encoder is used to convert the abnormal audio segments into abnormal feature vectors; Calculate the similarity between the abnormal feature vector and the feature vector of each sample in the abnormal voiceprint database; The noise classification label of the sample feature vector corresponding to the highest similarity is used as the noise recognition result, and the highest similarity is used as the confidence level to form the first recognition result.

7. The method according to claim 5, characterized in that, The step of using the aforementioned large sound model to identify the abnormal audio segments and obtaining a second identification result includes: The abnormal audio segments are identified using the aforementioned sound model to obtain the probability distribution of each candidate category; The candidate classification corresponding to the highest probability in the probability distribution is taken as the noise recognition result, and the highest probability is taken as the confidence level to form the second recognition result.

8. The method according to claim 5, characterized in that, Both the first and second identification results include noise identification results and confidence levels; The step of making a collaborative decision based on the first identification result and the second identification result to obtain the noise identification result of the wind turbine includes: When the noise identification result of the first identification result is consistent with the noise identification result of the second identification result, the noise identification result shall be used as the noise identification result of the wind turbine. When the noise identification results of the first identification result and the second identification result are inconsistent, and the confidence level of the first identification result or the second identification result is greater than the third preset threshold, the noise identification result corresponding to the higher confidence level shall be taken as the noise identification result of the wind turbine. When the noise identification results of the first identification result and the second identification result are inconsistent, and the confidence levels of both the first identification result and the second identification result are less than the third preset threshold, the abnormal audio segment is marked as a conflict sample.

9. A wind turbine noise monitoring system, characterized in that, The system includes: The perception layer is used to acquire the audio stream of the wind turbine at multiple preset noise monitoring points; An edge layer is used to perform spatial sound pressure level determination on the audio stream and filter out undetermined audio segments in the audio stream. The algorithm layer is used to filter out abnormal audio segments from the undetermined audio segments based on the difference between the undetermined audio segments and the reference audio, and input the abnormal audio segments into the abnormal voiceprint library and the sound big model respectively for parallel recognition to obtain the noise recognition result of the wind turbine.

10. An electronic device, characterized in that, include: A memory and a processor are interconnected, the memory stores computer instructions, and the processor executes the wind turbine noise monitoring method according to any one of claims 1 to 8 by executing the computer instructions.

Images