Hydroelectric generating set voiceprint anomaly analysis method and system based on self-supervised image algorithm

By installing acoustic sensors in key parts of the hydroelectric unit and using self-supervised image algorithms, the online monitoring problem without shutdown and manual arrangement of measurement points is solved, and the accurate identification and analysis of soundprint faults of the hydroelectric unit is achieved, improving the identification accuracy and Adaptability reduces operating costs.

CN119091180APending Publication Date: 2024-12-06STATE GRID XINYUAN GRP CO LTD +3
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410981441.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing technology is difficult to realize online monitoring without shutdown and manual arrangement of measurement points, and it is impossible to accurately identify the soundprint failure of the water-power unit. The traditional method has low recognition rate and poor generalization ability when facing nonlinear inputs, and does not consider the equipment type and location. and environmental factors.

Method used

Using a method based on self-supervised image algorithm, by installing acoustic sensors in key parts of the hydroelectric unit, collecting sound pattern samples and converting them into spectral diagrams, establishing a visual model for training and reasoning, and fine-tuning self-supervised learning using mask strategies. Voiceprint recognition and abnormal analysis were performed, and distance calculation was calculated using a prototype network.

Benefits of technology

It realizes online monitoring without shutdown and manual arrangement of measurement points, improves the accuracy and adaptability of identification of soundprint failures of hydropower units, reduces operating costs, and is suitable for units of different forms and operating times.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091180A_ABST
    Figure CN119091180A_ABST
Patent Text Reader

Abstract

The invention discloses a hydroelectric generating set voiceprint anomaly analysis method and system based on a self-supervised image algorithm, and relates to the field of hydroelectric generating set voiceprint monitoring and artificial intelligence, and the method comprises the steps: converting a voiceprint sample collected by a hydroelectric generating set into a spectrogram, and carrying out the self-supervised learning through a pre-training model, backbone network self-supervised learning is carried out on the voiceprint spectrogram, after representation of the backbone network, a prototype network is used for calculating the distance between each sample and all types of prototypes in the embedded space, and small sample abnormity early warning analysis is carried out on the voiceprint according to the distance; on the basis of online monitoring data, shutdown and manual measurement point arrangement are not needed, so that the economic benefit of the power plant is greatly increased, meanwhile, multi-source data of the unit can be comprehensively utilized for modeling, hidden features of the unit are automatically mined, manual judgment by experts is not needed, and the method is suitable for different forms of units with different operation time, and has high practicability. And the method has adaptivity and objectivity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of hydropower unit voiceprint monitoring and artificial intelligence, and in particular to a hydropower unit voiceprint anomaly analysis method and system based on a self-supervisory image algorithm. Background Art

[0002] With the continuous construction of UHV projects, hydropower units are increasingly shifting from traditional power generation functions to peak and frequency regulation. The frequency and depth of regulation are much higher than before. During operation, they are more affected by erosion, wear, fatigue and other factors. The performance of each component will gradually deteriorate with the increase of service time, leading to the occurrence of serious accidents. Therefore, it is urgent to grasp the operating status of the unit in real time and accurately, so as to make accurate and scientific decision-making basis for the operation of the unit.

[0003] At present, most hydropower stations still implement a joint maintenance mode of "planned maintenance" and "post-maintenance". This maintenance mode can ensure the normal operation of the unit in a short period of time, but it consumes too much manpower and material resources. Based on this problem, many hydropower stations have tried to change from a planned maintenance mode to a condition-based maintenance mode. Formulating maintenance plans based on the operating status of the unit can not only reduce operating costs and maintenance costs, but also extend the operating time of the unit. After a long period of operation, the performance of the hydropower unit will gradually decrease. As the operating time gets longer and longer, the more serious the performance degradation of the equipment will have a significant impact on the stability of the unit. Timely grasping the degree of equipment performance degradation can make a more reasonable and effective maintenance plan for the equipment. However, when traditional methods such as vibration and water pressure are used for fault monitoring, it is often necessary to deploy sensors, collectors, and connect them, which can only be limited to specific parts, resulting in many fault characteristics that cannot be captured. However, because of its non-contact deployment, sound can be installed at any part. At the same time, many faults are also accompanied by signs of sound. At the same time, equipment voiceprint recognition, that is, measuring the fault characteristics of the unit through the acoustic algorithm of the equipment, is one of the current hot research directions. Traditional methods based on feature rules have low recognition rates and poor generalization capabilities when facing nonlinear inputs. The collected signals of hydropower units often contain a lot of noise and show great nonlinearity. The noise caused by different equipment represents different operating states. In addition, traditional recognition methods do not take into account the actual application of hydropower stations, including factors such as equipment type, equipment location, and surrounding environment, so they cannot meet the actual engineering needs of hydropower unit voiceprint fault recognition. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to achieve online monitoring without stopping the machine and manually arranging measuring points, and identify faults through voiceprint feature images.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a method for analyzing anomaly of a hydropower unit voiceprint based on a self-supervised image algorithm, comprising:

[0008] Acoustic sensors are installed in the stator, waterwheel chamber, +X, +Y, -X, -Y directions and tailwater gate cone of the hydropower unit, and identification and sampling are carried out when the guide vane opening exceeds 70% and the load is less than 80% of the rated load;

[0009] Acquire voiceprint samples from the sampling results, convert them into spectrograms through the spectrogram algorithm, and use them as training samples; collect historical sound data of each part as historical fault samples;

[0010] Establishing a visual big model for voiceprint anomaly analysis, segmenting the spectrogram into a non-overlapping slice sequence and flattening it, passing it into the visual big model, and optimizing it using the pre-trained knowledge of the visual big model;

[0011] The visual big model is trained and deployed for inference. In the training stage, the mask strategy is used to fine-tune the visual big model for self-supervised learning and to perform water and electricity voiceprint recognition training. In the inference deployment stage, the training samples and historical fault samples are respectively passed into the trained visual big model and features are extracted. The metric distance calculation is then performed on each feature. The metric distance calculation is implemented using a prototype network.

[0012] Based on the metric distance, the soundprint anomaly analysis of the hydropower unit is carried out.

[0013] As a preferred solution for the hydropower unit voiceprint anomaly analysis method based on the self-supervised image algorithm, the following is a preferred solution:

[0014] The conversion into a spectrogram by a spectrogram algorithm includes:

[0015] The waveforms of the +X, +Y, -X, -Y directions of the waterwheel chamber and the tailwater gate cone pipe were collected. Each waveform lasted for 10 seconds. Feature extraction was performed through a 128-dimensional logarithmic Mel filter group to obtain a 128×1024 voiceprint spectrogram of these five parts.

[0016] As a preferred solution for the hydropower unit soundprint anomaly analysis method based on the self-supervised image algorithm, the following is a preferred solution:

[0017] The method of fine-tuning the large visual model through mask strategy for self-supervised learning and conducting water and electricity voiceprint recognition training includes:

[0018] Process the slices: randomly mask the segmented slice sequence, add identification bits to the masked slices, bring them into the linear projection mapping, and output 100 after mapping. Position encode each slice and add them.

[0019] As a preferred solution for the hydropower unit soundprint anomaly analysis method based on the self-supervised image algorithm, the following is a preferred solution:

[0020] The method of fine-tuning the large visual model through mask strategy for self-supervised learning and conducting water and electricity voiceprint recognition training also includes:

[0021] The processed slice results are brought into the visual model for training. Between the input and output of the visual model, the reconstruction mean square error is used as the accuracy loss, which is expressed as:

[0022]

[0023] Among them, L g Indicates the loss of precision, x i is the i-th slice, r i Corresponding to the slice x i is the result reconstructed by the large visual model, and N is the total number of slices.

[0024] As a preferred solution for the hydropower unit soundprint anomaly analysis method based on the self-supervised image algorithm, the following is a preferred solution:

[0025] After the training samples and the historical fault samples are respectively input into the trained visual large model and features are extracted, the metric distance calculation of the respective features includes:

[0026] Based on historical fault samples, we obtain samples of five categories: knock, scratch, friction, cavitation and normal, and form a data set (x1, y1), (x2, y2),…, (x n ,y n ), where x is the vector representation of the sample and y is the classification label;

[0027] Collect each type of historical fault samples x as the support set S; bring x into the visual large model for characterization, and obtain the feature Class Prototype (c), which is used as the prototype identifier and expressed as:

[0028]

[0029] As a preferred solution for the hydropower unit soundprint anomaly analysis method based on the self-supervised image algorithm, the following is a preferred solution:

[0030] After the training samples and the historical fault samples are respectively transferred into the trained visual large model and features are extracted, the metric distance calculation of the respective features also includes:

[0031] Collect the current sample y as the query set Q, expressed as:

[0032]

[0033] Substitute y into the large visual model for representation and obtain the feature f(φ).

[0034] As a preferred solution for the hydropower unit voiceprint anomaly analysis method based on the self-supervised image algorithm, the following is a preferred solution:

[0035] After the training samples and the historical fault samples are respectively input into the trained visual large model and features are extracted, the metric distance calculation of the respective features also includes:

[0036] Based on the feature f(φ), the distance between y and each type ClasPrototype(c) in the support set is calculated to obtain the probability p of the sample belonging to each fault φ , expressed as:

[0037]

[0038] Where k represents the probability that y belongs to the fault sample x.

[0039] In a second aspect, an embodiment of the present invention provides a hydropower unit voiceprint anomaly analysis system based on a self-supervised image algorithm, characterized in that it includes:

[0040] The sampling module is used to install acoustic sensors in the stator of the hydropower unit, the +X, +Y, -X, -Y directions of the waterwheel chamber and the tailwater gate cone, and to identify and sample when the guide vane opening exceeds 70% and the load is less than 80% of the rated load;

[0041] The sample generation module is used to obtain voiceprint samples from the sampling results, convert them into spectrograms through the spectrogram algorithm, and use them as training samples; collect historical sound data of various parts as historical fault samples;

[0042] A model building module is used to build a visual big model for voiceprint anomaly analysis, segment the spectrogram into a non-overlapping slice sequence and flatten it, pass it into the visual big model, and optimize it using the pre-trained knowledge of the visual big model;

[0043] The training and deployment module is used to train and reason about the large visual model. In the training phase, the large visual model is fine-tuned for self-supervised learning through the mask strategy to perform water and electricity voiceprint recognition training. In the reasoning and deployment phase, the training samples and historical fault samples are respectively passed into the trained large visual model and features are extracted. The metric distance calculation is performed on each feature. The metric distance calculation is implemented using the prototype network.

[0044] The analysis module is used to perform anomaly analysis of the soundprint of the hydropower unit based on the metric distance.

[0045] In a third aspect, an embodiment of the present invention provides a computing device, including:

[0046] Memory and processor;

[0047] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for analyzing the soundprint anomaly of a hydropower unit based on a self-supervised image algorithm as described in any embodiment of the present invention.

[0048] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, can implement the method for analyzing the anomaly of a hydropower unit voiceprint based on a self-supervised image algorithm.

[0049] The beneficial effects of the present invention are as follows: the present invention is based on online monitoring data, and does not require shutdown or manual arrangement of measuring points, thereby greatly increasing the economic benefits of the power plant. At the same time, it can comprehensively utilize multi-source data of the unit for modeling, automatically mine its implicit features, and does not require manual judgment by experts. It is adaptable to units of different forms and different operating times, and has adaptability and objectivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0051] Figure 1 It is an overall flow chart of the method for analyzing abnormality of voiceprint of hydropower units based on self-supervised image algorithm according to the first embodiment of the present invention;

[0052] Figure 2 It is a spectrogram calculation flow chart of the method for analyzing the abnormality of the voiceprint of a hydropower unit based on a self-supervised image algorithm according to the first embodiment of the present invention;

[0053] Figure 3 It is a schematic diagram of a visual large model of a method for analyzing abnormal soundprints of hydropower units based on a self-supervised image algorithm according to the first embodiment of the present invention;

[0054] Figure 4It is a prototype network recognition principle diagram of the hydropower unit voiceprint anomaly analysis method based on the self-supervised image algorithm described in the first embodiment of the present invention.

[0055] Figure 5 It is the original waveform diagram of the simulated knocking in the simulation example of the method for analyzing the abnormality of the soundprint of a hydropower unit based on the self-supervised image algorithm described in the second embodiment of the present invention;

[0056] Figure 6 It is an original waveform diagram of simulated scratching in a simulation example of a method for analyzing abnormal soundprint of a hydropower unit based on a self-supervised image algorithm according to a second embodiment of the present invention;

[0057] Figure 7 It is an original waveform diagram simulating continuous point collisions in a simulation example of a method for analyzing anomaly soundprints of a hydropower unit based on a self-supervised image algorithm according to a second embodiment of the present invention;

[0058] Figure 8 It is the original waveform diagram of intermittent point collision in the simulation example of the method for analyzing the abnormality of soundprint of a hydropower unit based on a self-supervised image algorithm according to the second embodiment of the present invention;

[0059] Fig. 9 It is a spectrogram of simulated knocking in a simulation example of a method for analyzing abnormality of a hydropower unit voiceprint based on a self-supervised image algorithm according to a second embodiment of the present invention;

[0060] Fig.10 It is a spectrogram of simulated scratching in a simulation example of a method for analyzing abnormal soundprint of a hydropower unit based on a self-supervised image algorithm according to a second embodiment of the present invention;

[0061] Fig.11 It is a spectrogram simulating continuous point collisions in a simulation example of a method for analyzing anomaly soundprints of a hydropower unit based on a self-supervised image algorithm according to a second embodiment of the present invention;

[0062] Fig.12 It is a spectrogram of intermittent point collisions in a simulation example of a method for analyzing anomaly soundprints of a hydropower unit based on a self-supervised image algorithm as described in the second embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0064] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0065] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0066] Example 1

[0067] Reference Figure 1-Figure 4 , which is the first embodiment of the present invention, and provides a method for analyzing abnormality of voiceprint of a hydropower unit based on a self-supervised image algorithm, comprising:

[0068] S1: Install acoustic sensors in the +X, +Y, -X, -Y directions of the stator and waterwheel chamber of the hydropower unit and the tailwater gate cone, and identify and sample when the guide vane opening exceeds 70% and the load is less than 80% of the rated load;

[0069] It should be noted that the acoustic sensors should be installed at the key parts of the hydropower unit, the stator, different directions of the waterwheel room (+X, +Y, -X, -Y) and the tailwater gate cone. These locations are usually where abnormal sounds may be generated during the operation of the unit, and the choice of installation location is crucial to capture abnormal sounds; the guide vane opening exceeds 70% and the load is less than 80% of the rated load. Under these conditions, the unit may be more prone to abnormal sounds, such as sound changes caused by component wear, looseness or fluid instability.

[0070] S2: Obtain voiceprint samples from the sampling results, convert them into spectrograms through the spectrogram algorithm, and use them as training samples; collect historical sound data of each part as historical fault samples;

[0071] In the embodiment of the present application, converting to a spectrogram by a spectrogram algorithm includes:

[0072] The waveforms of the +X, +Y, -X, -Y directions of the waterwheel chamber and the tailwater gate cone pipe were collected. Each waveform lasted for 10 seconds. Feature extraction was performed through a 128-dimensional logarithmic Mel filter group to obtain a 128×1024 voiceprint spectrogram of these five parts.

[0073] It should be noted that the two frequency bands of 20Hz~500Hz and 2000Hz~20kHz can be selected for the collection of voiceprint samples. These frequency bands cover a wide range from low frequency to high frequency, and can capture various sound characteristics that may be generated during the operation of the unit, including mechanical vibration, fluid movement, etc. Determine the appropriate sampling rate and bit depth to ensure that the collected sound data has high quality and detailed information, which is convenient for subsequent spectrogram algorithm processing and analysis.

[0074] S3: Establish a visual big model for voiceprint anomaly analysis, segment the spectrogram into a non-overlapping slice sequence and flatten it, pass it into the visual big model, and optimize it using the pre-trained knowledge of the visual big model;

[0075] Specifically, the spectrogram is divided into a non-overlapping 16×16 slice sequence, one sample contains 512 sequence slices, and flattened to 768 dimensions; the visual large model structure is a Transformer encoder-decoder structure, and before entering the network, sinusoidal position encoding is used, and the input embedding layer uses two one-dimensional convolutional encodings and GELU;

[0076] There are 6 encoders in the model, and the encoder structure is a self-attention mechanism + MLP mode

[0077] There are 6 decoders in the model, and the decoder is fired in the mode of cross attention + self-attention + MLP. Its training loss function is the image structure similarity index SSIM.

[0078] S4: Train and reason about the big visual model. In the training phase, the mask strategy is used to fine-tune the big visual model for self-supervised learning and to train water and electricity voiceprint recognition. In the reasoning and deployment phase, the training samples and historical fault samples are respectively passed into the trained big visual model and features are extracted. The metric distance calculation is then performed on each feature. The metric distance calculation is implemented using a prototype network.

[0079] In the embodiment of the present application, the mask strategy is used to fine-tune the visual large model for self-supervised learning, and the water and electricity voiceprint recognition training includes:

[0080] Process the slices: randomly mask the segmented slice sequence, add identification bits to the masked slices, bring them into the linear projection mapping, and output 100 after mapping. Position encode each slice and add them.

[0081] The processed slice results are brought into the visual model for training. Between the input and output of the visual model, the reconstruction mean square error is used as the accuracy loss, which is expressed as:

[0082]

[0083] Among them, L g Indicates the loss of precision, x i is the i-th slice, r i Corresponding to the slice x i is the result reconstructed by the large visual model, and N is the total number of slices.

[0084] After the training samples and historical fault samples are respectively passed into the trained visual model and features are extracted, the metric distance calculation of each feature includes:

[0085] Based on historical fault samples, we obtain samples of five categories: knock, scratch, friction, cavitation and normal, and form a data set (x1, y1), (x2, y2),…, (x n ,y n ), where x is the vector representation of the sample and y is the classification label;

[0086] Collect each type of historical fault samples x as the support set S; bring x into the visual large model for characterization, and obtain the feature Class Prototype (c), which is used as the prototype identifier and expressed as:

[0087]

[0088] Collect the current sample y as the query set Q, expressed as:

[0089]

[0090] Substitute y into the large visual model for representation and obtain the feature f(φ).

[0091] Based on the feature f(φ), the distance between y and each type ClasPrototype(c) in the support set is calculated to obtain the probability p of the sample belonging to each fault φ , expressed as:

[0092]

[0093] Where k represents the probability that y belongs to the fault sample x.

[0094] S5: Based on the metric distance, the soundprint anomaly analysis of the hydropower unit is performed.

[0095] It should be noted that determining a suitable distance measurement method, such as Euclidean distance, cosine distance or a more complex kernel-based distance measurement, can more accurately measure the similarity or difference between samples, which helps to effectively distinguish normal sounds from various abnormal sounds.

[0096] In another possible implementation, a prototype identifier for each type of fault is established in the sound feature space to ensure that these prototypes can represent the typical sound features of each type of fault. In this way, by calculating the distance between the query sample and each type of prototype, its abnormality and possible fault type can be effectively evaluated; the probability that the query sample belongs to each type of fault is calculated based on the distance metric. An appropriate threshold is set to determine whether the sample belongs to a certain type of fault, or a probability distribution is used to evaluate the abnormality of the sample. This method can improve the accuracy and reliability of anomaly detection;

[0097] In another possible implementation, a real-time monitoring system is established to continuously receive and process new sound data. By analyzing the abnormal probability calculated in real time, an alarm or suggestion is promptly issued to the operator or monitoring system. This real-time feedback mechanism helps to quickly respond to and handle potential fault problems and reduce the risk of unit damage; the voiceprint abnormality analysis model is regularly optimized and updated, and the accuracy and adaptability of the model are continuously improved through actual case data collected by the feedback mechanism. This continuous improvement can make the sound abnormality analysis system more intelligent and effective.

[0098] The above is a schematic scheme of a method for analyzing the voiceprint anomaly of a hydropower unit based on a self-supervised image algorithm in this embodiment. It should be noted that the technical scheme of the hydropower unit voiceprint anomaly analysis system based on a self-supervised image algorithm and the technical scheme of the method for analyzing the voiceprint anomaly of a hydropower unit based on a self-supervised image algorithm belong to the same concept. For details not described in detail in the technical scheme of the hydropower unit voiceprint anomaly analysis system based on a self-supervised image algorithm in this embodiment, please refer to the description of the technical scheme of the method for analyzing the voiceprint anomaly of a hydropower unit based on a self-supervised image algorithm.

[0099] The hydropower unit voiceprint anomaly analysis system based on the self-supervised image algorithm in this embodiment includes:

[0100] The sampling module is used to install acoustic sensors in the stator of the hydropower unit, the +X, +Y, -X, -Y directions of the waterwheel chamber and the tailwater gate cone, and to identify and sample when the guide vane opening exceeds 70% and the load is less than 80% of the rated load;

[0101] The sample generation module is used to obtain voiceprint samples from the sampling results, convert them into spectrograms through the spectrogram algorithm, and use them as training samples; collect historical sound data of various parts as historical fault samples;

[0102] A model building module is used to build a visual big model for voiceprint anomaly analysis, segment the spectrogram into a non-overlapping slice sequence and flatten it, pass it into the visual big model, and optimize it using the pre-trained knowledge of the visual big model;

[0103] The training and deployment module is used to train and reason about the large visual model. In the training phase, the large visual model is fine-tuned for self-supervised learning through the mask strategy to perform water and electricity voiceprint recognition training. In the reasoning and deployment phase, the training samples and historical fault samples are respectively passed into the trained large visual model and features are extracted. The metric distance calculation is performed on each feature. The metric distance calculation is implemented using the prototype network.

[0104] The analysis module is used to perform anomaly analysis of the soundprint of the hydropower unit based on the metric distance.

[0105] This embodiment further provides a computing device, which is applicable to the case of a method for analyzing abnormality of a hydropower unit voiceprint based on a self-supervised image algorithm, and includes:

[0106] A memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the hydropower unit voiceprint anomaly analysis method based on the self-supervised image algorithm as proposed in the above embodiment.

[0107] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for analyzing the abnormality of the sound print of a hydropower unit based on a self-supervised image algorithm as proposed in the above embodiment is implemented.

[0108] The storage medium proposed in this embodiment and the method for analyzing the anomaly of the soundprint of a hydropower unit based on a self-supervised image algorithm proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0109] Example 2

[0110] Refer to Table 1 and Figure 5-Figure 12 , which is an embodiment of the present invention, provides a method for analyzing the abnormality of the soundprint of a hydropower unit based on a self-supervised image algorithm. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0111] The field monitoring data of a power station from May 1, 2013 to May 21, 2013 is used as a research case, including the X / Y swing of the water guide and the X / Y / Z vibration data of the top cover. The degradation degree of the data monitored every half an hour is calculated by the method proposed in this paper, and the degradation trend sequence of 1008 sample data is obtained. The characteristics of the data set are as follows Figure 5-Figure 12 As shown, the number of samples is 20;

[0112] 1) Data processing and cleaning: First, collect the monitoring data of the hydropower unit equipment. Data collection and cleaning: First, install acoustic sensors on the stator, +X, +Y, -X, -Y directions of the waterwheel room and the tailwater gate cone of the hydropower unit, and sample at a sampling frequency of 20Hz to 20Khz. The steady-state operating conditions of more than 70% guide vane opening and 80% to the rated load in the operating condition information are identified.

[0113] 2) By acquiring voiceprint samples in two frequency bands, 20Hz to 500Hz and 2000Hz to 20Khz, combined with the spectrogram algorithm, the spectrogram is a kind of speech graph that uses two-dimensional coordinates to express the three-dimensional spatial information characteristics of the sound signal, with the vertical axis y being the frequency and the horizontal axis x being the time. The point (x, y) corresponding to the two-dimensional coordinate plane represents the sound intensity.

[0114] The voiceprint samples of the stator of the hydropower unit, the +X, +Y, -X, -Y directions of the waterwheel room and the tailwater gate cone are collected and converted into spectrograms as training samples.

[0115] 3) Construct voiceprint spectrogram samples and pass them into the visual big model, and use the self-supervised comparative learning strategy to fine-tune the visual big model. The voiceprint spectrogram training is regarded as the image mask restoration problem of the spectrogram.

[0116] 4) Combined with historical fault samples, obtain samples of five categories: knock, scratch, friction, cavitation and normal. After learning the encoding representation of each sample, sum and average all sample encodings under each category, and use the result as the prototype representation of the classification.

[0117] 5) Actual operation stage

[0118] In the actual operation stage, after the voiceprint spectrogram is passed into the visual large model and features are extracted, the features are substituted into the prototype sample for small sample classification, and early warning analysis is performed by measuring the distance between the small sample and various samples. The voiceprint feature image of the hydropower unit is obtained through the spectrogram, and the feature generalization ability of small sample classification is introduced to identify fault sounds. The test results are shown in Table 1.

[0119] Table 1 Test results

[0120] Identification method CNN Transform Siamese Network This method result 85.62±2.3 93.25±1.4 96.14±0.7 96.71±0.6

[0121] Compared with traditional deep learning methods, it can make full use of the transfer learning ability of deep learning and has better recognition efficiency for small sample voiceprints.

[0122] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for analyzing abnormality of soundprint of hydropower units based on self-supervised image algorithm, characterized in that: include: Acoustic sensors are installed in the stator, waterwheel chamber, +X, +Y, -X, -Y directions and tailwater gate cone of the hydropower unit, and identification and sampling are carried out when the guide vane opening exceeds 70% and the load is less than 80% of the rated load; Obtain voiceprint samples from the sampling results, convert them into spectrograms through the spectrogram algorithm, and use them as training samples; Collect historical sound data of each part as historical fault samples; Establishing a visual big model for voiceprint anomaly analysis, segmenting the spectrogram into a non-overlapping slice sequence and flattening it, passing it into the visual big model, and optimizing it using the pre-trained knowledge of the visual big model; Train and reason about the large visual model. During the training phase, the large visual model is fine-tuned for self-supervised learning through masking strategies to perform water and electricity voiceprint recognition training. In the inference deployment phase, the training samples and historical fault samples are respectively passed into the trained visual model and features are extracted. Then, the metric distance calculation is performed on each feature. The metric distance calculation is implemented using the prototype network. Based on the metric distance, the soundprint anomaly analysis of the hydropower unit is carried out.

2. The method for training a hydropower unit voiceprint model based on a self-supervised image algorithm according to claim 1, characterized in that: The conversion into a spectrogram by a spectrogram algorithm includes: The waveforms of the +X, +Y, -X, -Y directions of the waterwheel chamber and the tailwater gate cone pipe were collected. Each waveform lasted for 10 seconds. Feature extraction was performed through a 128-dimensional logarithmic Mel filter group to obtain a 128×1024 voiceprint spectrogram of these five parts.

3. The method for analyzing abnormality of soundprint of a hydropower unit based on a self-supervised image algorithm according to claim 2, characterized in that: The method of fine-tuning the large visual model through mask strategy for self-supervised learning and conducting water and electricity voiceprint recognition training includes: Process the slices: randomly mask the segmented slice sequence, add identification bits to the masked slices, bring them into the linear projection mapping, and output 100 after mapping. Position encode each slice and add them.

4. The method for analyzing the abnormality of the soundprint of a hydropower unit based on a self-supervised image algorithm according to claim 3 is characterized in that: The method of fine-tuning the large visual model through mask strategy for self-supervised learning and conducting water and electricity voiceprint recognition training also includes: The processed slice results are brought into the visual model for training. Between the input and output of the visual model, the reconstruction mean square error is used as the accuracy loss, which is expressed as: Among them, L g Indicates the loss of precision, x i is the i-th slice, r i Corresponding to the slice x i is the result reconstructed by the large visual model, and N is the total number of slices.

5. The method for analyzing the abnormality of the soundprint of a hydropower unit based on a self-supervised image algorithm according to claim 4 is characterized in that: After the training samples and the historical fault samples are respectively input into the trained visual large model and features are extracted, the metric distance calculation of the respective features includes: Based on historical fault samples, we obtain samples of five categories: knock, scratch, friction, cavitation and normal, and form a data set (x1, y1), (x2, y2),…, (x n ,y n ), where x is the vector representation of the sample and y is the classification label; Collect each type of historical fault samples x as the support set S; bring x into the visual big model for characterization, and obtain the feature ClassPrototype (c), which is used as the prototype identifier and expressed as:

6. The method for analyzing the abnormality of the soundprint of a hydropower unit based on a self-supervised image algorithm according to claim 5 is characterized in that: After the training samples and the historical fault samples are respectively transferred into the trained visual large model and features are extracted, the metric distance calculation of the respective features also includes: Collect the current sample y as the query set Q, expressed as: Substitute y into the large visual model for representation and obtain the feature f(φ).

7. The method for analyzing abnormality of soundprint of a hydropower unit based on a self-supervised image algorithm according to claim 6, characterized in that: After the training samples and the historical fault samples are respectively transferred into the trained visual large model and features are extracted, the metric distance calculation of the respective features also includes: Based on the feature f(φ), the distance between y and each type ClasPrototype(c) in the support set is calculated to obtain the probability p of the sample belonging to each fault φ , expressed as: Where k represents the probability that y belongs to the fault sample x.

8. A system using the method for analyzing abnormality of voiceprint of a hydropower unit based on a self-supervised image algorithm as claimed in any one of claims 1 to 7, characterized in that: include: The sampling module is used to install acoustic sensors in the stator of the hydropower unit, the +X, +Y, -X, -Y directions of the waterwheel chamber and the tailwater gate cone, and to identify and sample when the guide vane opening exceeds 70% and the load is less than 80% of the rated load; The sample generation module is used to obtain voiceprint samples from the sampling results and convert them into spectrograms through the spectrogram algorithm as training samples; Collect historical sound data of each part as historical fault samples; A model building module is used to build a visual big model for voiceprint anomaly analysis, segment the spectrogram into a non-overlapping slice sequence and flatten it, pass it into the visual big model, and optimize it using the pre-trained knowledge of the visual big model; The training and deployment module is used to train and reason about the large visual model. During the training phase, the mask strategy is used to fine-tune the large visual model for self-supervised learning and to perform water and electricity voiceprint recognition training. In the inference deployment phase, the training samples and historical fault samples are respectively passed into the trained visual model and features are extracted. Then, the metric distance calculation is performed on each feature. The metric distance calculation is implemented using the prototype network. The analysis module is used to perform anomaly analysis of the soundprint of the hydropower unit based on the metric distance.

9. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for analyzing the abnormality of the soundprint of a hydropower unit based on a self-supervised image algorithm as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for analyzing the abnormality of a soundprint of a hydropower unit based on a self-supervised image algorithm as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Equipment fault voiceprint monitoring system based on cloud edge collaboration

    CN114495983A

  • Transformer voiceprint signal fault unsupervised detection method based on combination of feature clustering and generative learning

    CN117437933A

  • Water-turbine generator set abnormal sound detection method based on self-supervision and weak supervision

    CN117524250A

  • Pumped storage power station time series data cleaning method based on comparative learning and storage medium

    CN117609698A