Rotating machine fault early warning diagnosis method based on voice large model fine tuning

By applying the voice big model fine-tuning technology in rotary mechanical fault warning, using pre-trained voice big model and lightweight autoencoder, the problems of poor generalization ability and high training cost in the existing technology are solved, and the more Efficient fault warning and feature extraction.

CN118918923APending Publication Date: 2024-11-08BEIJING UNIV OF CHEM TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411137477.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-08

Smart Images

  • Figure CN118918923A_ABST
    Figure CN118918923A_ABST
Patent Text Reader

Abstract

The invention provides a rotating machine fault early warning and diagnosis method based on voice large model fine tuning. The method comprises the following steps: preprocessing a data set; loading a pre-trained voice large model encoder part, and freezing parameters; adding a logic averaging layer and a lightweight auto-encoder; calculating the reconstruction error of the lightweight auto-encoder as a fault early warning index; and calculating a threshold value of fault early warning to obtain a result of fault early warning. According to the method, the advanced technology in the field of voice processing is innovatively applied to mechanical fault diagnosis, richer features are extracted by utilizing the generalization ability of the pre-trained voice large model, and only the lightweight auto-encoder needs to be finely adjusted, so that the training time and cost are remarkably reduced. Meanwhile, an early warning threshold value is set by adopting gamma distribution and a quartile method, so that the diagnosis accuracy is improved. According to the invention, cross-domain knowledge migration is realized, and an efficient and accurate new method is provided for rotating machinery fault early warning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of rotating machinery fault warning, and in particular to a rotating machinery fault warning diagnosis method based on fine-tuning of a large speech model. Background Art

[0002] Rotating machinery, as an indispensable key equipment in modern industrial production, is widely used in wind power generation, petrochemical industry, metallurgy, energy and other fields. They realize energy conversion, transmission or material processing through rotational motion, which is of great significance to improving production efficiency and ensuring production safety. However, with the long-term operation of rotating machinery, its internal parts will be affected by various factors such as wear and corrosion, resulting in equipment performance degradation and frequent failures. These failures will not only affect the normal operation of the equipment, but may also cause significant losses to production and even threaten the safety of personnel. Therefore, it is particularly important to provide fault warning for rotating machinery.

[0003] Rotating machinery fault early warning technology collects equipment operation data in real time through various sensors, such as vibration, temperature, pressure, etc., and then uses data analysis, pattern recognition and other methods to analyze and process the collected data to determine whether the equipment has a fault and its type and degree. This technology can issue a warning signal in advance before the equipment fails, so that operators can take timely measures to repair and maintain the equipment to avoid further damage to the equipment or the expansion of the fault. In previous methods, if only a specific data set is relied on to directly train the early warning model, the model may overfit to this specific data set, resulting in limited early warning effect when facing new data or equipment under different operating conditions; at the same time, because the features extracted by the model are not sufficient, its generalization ability will also be poor. When the model parameters are large, all parameters need to be trained every time facing a different data set, which increases the time and cost of training. Summary of the invention

[0004] The present invention aims to solve the problems existing in the existing rotating machinery fault warning technology, and proposes a rotating machinery fault warning diagnosis method based on fine-tuning of a large speech model. This method innovatively applies advanced technology in the field of speech processing to mechanical fault diagnosis, and realizes cross-domain knowledge transfer. The main technical solutions of the present invention are as follows:

[0005] A method for early warning and diagnosis of rotating machinery faults based on fine-tuning of a large speech model is provided, comprising:

[0006] Step S1, preprocessing the data set of the rotating machinery;

[0007] Step S2, loading the encoder part of the pre-trained large speech model into the fault warning system, and freezing all parameters of the encoder part;

[0008] Step S3, adding a logical averaging layer and a lightweight autoencoder after the encoder part;

[0009] Step S4, using the lightweight autoencoder to calculate a reconstruction error, and using the reconstruction error as a fault warning indicator;

[0010] Step S5, calculating a threshold of fault warning based on the reconstruction error, and obtaining a result of the fault warning according to the threshold.

[0011] Furthermore, the data set is preprocessed in step S1, specifically including:

[0012] Perform a short-time Fourier transform on the data set;

[0013] When the data is a sound signal, the logarithmic Mel spectrum is obtained for the data to meet the input requirements of the large speech model.

[0014] Furthermore, in step S2, the pre-trained speech large model encoder part is loaded and the parameters are frozen, specifically including:

[0015] Use speech signals to train a large speech model, take the encoder part and freeze the parameters as a feature extractor.

[0016] Furthermore, in step S3, a logical averaging layer is added, which specifically includes:

[0017] Add a logical calculation part to the averaging layer to determine the average of the output values ​​of the speech model encoder along the specified numerical axis.

[0018] Furthermore, the calculation of the lightweight autoencoder reconstruction error in step S4 specifically includes:

[0019] The lightweight autoencoder is trained using the mean eigenvalues ​​of normal working condition samples;

[0020] The mean square error between the mean eigenvalue of the sample and the output value of the lightweight autoencoder is calculated as the reconstruction error.

[0021] Furthermore, the calculation of the threshold value of the fault warning in step S5 specifically includes:

[0022] The reconstruction error of the sample is modeled using the gamma distribution;

[0023] Quartiles are used to calculate the fault warning threshold.

[0024] Furthermore, the calculation formula of the fault warning threshold is:

[0025] thr=Q3+k*IQR

[0026] Among them, Q3 is the third quartile, IQR is the interquartile range, and k is an adjustable constant.

[0027] Optionally, the rotating machinery fault early warning diagnosis method based on fine-tuning of a large speech model further includes:

[0028] Input new sample data into the trained model and calculate its reconstruction error;

[0029] The calculated reconstruction error is compared with the fault warning threshold to determine whether there is a fault.

[0030] The beneficial effects of the present invention are: loading the pre-trained speech large model encoder part, freezing the parameters; adding a logical averaging layer; calculating the threshold of the fault warning, and obtaining the result of the fault warning. This application uses a universal pre-trained speech large model, which only needs to be adaptively adjusted for specific tasks in the fine-tuning stage, training lightweight autoencoders, reducing data processing time, and reducing the complexity and cost of developing and maintaining different models; in addition, the model is expected to extract richer and deeper features through pre-training of massive data sets, and has stronger generalization capabilities, thereby reducing production risks and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the method for early warning and diagnosis of rotating machinery faults based on fine-tuning of a large speech model in the present invention;

[0032] Figure 2 It is a schematic diagram of the model in the rotating machinery fault early warning diagnosis method based on fine-tuning of the speech large model in the present invention;

[0033] Figure 3 A schematic diagram of the Mel spectrum of normal and abnormal working conditions obtained by data preprocessing in the present invention;

[0034] Figure 4 This is a schematic diagram of the structure of the pre-trained speech model in the present invention;

[0035] Figure 5 A schematic diagram of the structure of a lightweight autoencoder for training a fan dataset in the present invention;

[0036] Figure 6 Loss graph for training lightweight autoencoder for fan dataset in this invention;

[0037] Figure 7 Reconstructing error value histograms and probability density graphs for some normal operating conditions of the fan data set in the present invention;

[0038] Figure 8 Schematic diagram of the confusion matrix of the fan data set test set in the present invention;

[0039] Fig. 9 Schematic diagram of the ROC curve of the fan data set test set in the present invention;

[0040] Fig.10 It is a schematic diagram of the reconstruction error scatter points of the fan data set in the present invention;

[0041] Fig.11 It is a schematic diagram of the time-frequency diagram of the gas turbine data set in the present invention;

[0042] Fig.12 A schematic diagram of the structure of a lightweight autoencoder for training a gas turbine dataset in the present invention;

[0043] Fig.13 Loss map for training lightweight autoencoder for gas turbine in the present invention;

[0044] Fig.14 The error value histogram and probability density diagram for data reconstruction of the gas turbine before March 1, 2023 in the present invention;

[0045] Fig.15 It is a schematic diagram of the reconstruction error scatter points of the gas turbine data set in the present invention. DETAILED DESCRIPTION

[0046] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0047] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Instead, they are only examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims.

[0048] like Figure 1 As shown, this embodiment provides a rotating machinery fault early warning diagnosis method based on fine-tuning of a large speech model, including:

[0049] Step S1, preprocessing the data set of the rotating machinery;

[0050] Step S2, loading the encoder part of the pre-trained large speech model into the fault warning system, and freezing all parameters of the encoder part;

[0051] Step S3, adding a logical averaging layer and a lightweight autoencoder after the encoder part;

[0052] Step S4, using the lightweight autoencoder to calculate a reconstruction error, and using the reconstruction error as a fault warning indicator;

[0053] Step S5, calculating a threshold of fault warning based on the reconstruction error, and obtaining a result of the fault warning according to the threshold.

[0054] Optionally, the rotating machinery fault state is determined by the following steps:

[0055] For the target rotating machinery dataset, according to the input shape requirements of the speech model, a short-time Fourier transform is performed on each sample to obtain a time-frequency spectrum, and a Mel transform is performed on the obtained time-frequency spectrum to obtain the Mel spectrum of the target dataset. The following formula is the Mel spectrum frequency conversion formula:

[0056]

[0057] Among them, f is the original frequency of the sample, and m represents the converted Mel frequency.

[0058] The mel spectrum of each sample is input into the encoder of the pre-trained speech model to obtain the feature matrix of each sample;

[0059] The feature matrix of each sample is input into the logical mean layer, and the mean feature value of each sample is calculated along the axis of the Mel spectrum time direction;

[0060] The lightweight autoencoder is trained using the mean eigenvalues ​​of the normal working condition samples. After the training, the mean eigenvalues ​​of the normal working condition samples are input into the lightweight autoencoder to obtain the output eigenvalues. The reconstruction error between the mean eigenvalues ​​of the normal samples and the output eigenvalues ​​is calculated to obtain the reconstruction error value of the normal working condition samples. The reconstruction error value calculation process is as follows:

[0061]

[0062] Among them, A(X) is the reconstruction error value of the sample, f(x i ) is the output value of the lightweight autoencoder, x i The mean eigenvalue of the sample.

[0063] The gamma distribution is often used to model the autoencoder reconstruction error because the reconstruction error under normal conditions is usually skewed and has a heavy tail, which is the natural shape of the gamma distribution. The probability density function of the gamma distribution is as follows:

[0064]

[0065] Where x is the value of the random variable, α and β are the shape and scale parameters of the gamma distribution, respectively, x>0, β>0, α>0. Γ(α) represents the gamma function.

[0066] The fault warning result is obtained by fitting the gamma distribution curve to the reconstruction error of the sample and using the quartile to calculate the fault warning threshold.

[0067] If a value greater than or equal to a fault warning threshold exists in the reconstruction error, the state of the target rotating machine is faulty.

[0068] All samples of the target rotating machinery dataset are input into the trained lightweight autoencoder, the reconstruction error is calculated, and the sample working condition is obtained according to the fault warning threshold to complete the warning. The specific model structure is as follows: Figure 2 shown.

[0069] In this embodiment, the rotating machinery data sets used include data of the fan part at a signal-to-noise ratio of 6 dB and a gas turbine data set. The two data sets respectively verify the feasibility of the present invention.

[0070] The fan dataset sound is recorded by an eight-channel microphone array with a sampling rate of 16kHz. The number of samples in the dataset is 5550, including 4075 normal samples and 1475 abnormal samples. The sample length is 10s.

[0071] The mel spectrum of the sound signal is calculated according to the sound signal, including:

[0072] The sound signal is resampled to 15 kHz and then subjected to short-time Fourier transform with a 30 ms window and a 3.3 ms step size;

[0073] Mel transform is performed on the time-frequency graph, and an 80-channel Mel filter group is used to obtain the Mel spectrum of the target data set. At this time, the shape of the Mel spectrum matrix is ​​[80, 3000], which meets the input requirements of the pre-trained speech model; the shape of the entire data set is [5550, 80, 3000]. At this time, the first dimension of the data set is the sample number, the second dimension is the Mel spectrum, and the third dimension is time. Figure 3 As shown in the figure, normal represents the Mel-spectrogram under normal working conditions, and abnormal represents the Mel-spectrogram under abnormal working conditions.

[0074] In this embodiment, the structure of the speech model is as follows: Figure 4As shown in the figure, the model is divided into two parts, the encoder part and the decoder part. The encoder part consists of two one-dimensional convolutional layers Conv1D, a sinusoidal positional encoding Sinusoidal Positional Encoding, a dimension conversion layer permute, and four encoder blocks Encoder Block, which include a linear layer MLP and a self-attention layer self-attention; the reason for adding the dimension conversion layer is to adapt to the input requirements of the subsequent Encoder block. The model can make full use of the local features extracted by the convolution layer and the global dependencies captured by the Encoder blocks, thereby improving the performance and accuracy of the model; the decoder part consists of the positional encoding Learned Positional Encoding learned from the encoder part and four decoding blocks Decoder Block, which include MLP, self-attention, and cross-attention. Sinusoidal positional encoding is to introduce position information, and the form of positional encoding is usually a combination of a sine function and a cosine function. The formula is as follows:

[0075]

[0076] Among them, PE (pos,i) Represents the value of the posth position and the i-th dimension in the position encoding matrix; d model Represents the dimension of the model embedding vector; i represents the value of the i-th dimension in the position encoding matrix. This position encoding method can introduce position information, allowing the Transformer model to process sequence data.

[0077] Furthermore, the speech model is trained using multiple translations from "any language" to English, short speech clips and speech data labeled in multiple languages. This data is resampled to 16 kHz and the 80-channel log Mel spectrogram representation is calculated using a 25 ms window and 10 ms stride. The input audio is split into 30 second segments and optimized using AdamW with a learning rate of 10. -4 , the batch size is 32.

[0078] In the Transformer model, the encoder is responsible for converting the input sequence into a context-aware representation, while the decoder is responsible for generating the output sequence. Therefore, the pre-trained speech model encoder part is used for feature extraction to improve the effect of subsequent anomaly detection and freeze its parameters. The model parameter file size of the speech model encoding part is 33M.

[0079] The mel spectrum of each sample is input into the encoder of the pre-trained speech model to obtain the feature matrix of each sample. At this time, the shape of the feature matrix is ​​[1500, 384], and the shape of the entire data set is [5550, 1500, 384].

[0080] In this embodiment, since the third dimension of the target rotating machinery data set is the time dimension, and the target data set undergoes dimensional transformation when passing through the large speech model, according to logical judgment, when the feature matrix of the data set is input into the logical mean layer, the mean eigenvalue of each sample is calculated along the numerical axis of the Mel spectrum time direction. At this time, the feature matrix of the sample becomes a one-dimensional matrix, and the shape of the entire data set is [5550,384].

[0081] In this embodiment, the lightweight autoencoder has a shape as follows: Figure 5 As shown in the figure, the autoencoder contains an input layer inputLayer with a dimension of 384, an output layer output with a dimension of 384, 5 batch normalization layers batch_normalization, and 5 fully connected layers, among which the fully connected layer contains a bottleneck layer Bottleneck with a dimension of 32, and the remaining fully connected layers Dense with a dimension of 512. The number of iterations epochs is 400, the batch size batch_size is 512, and the initial learning rate learning_rate is 10 -3 The learning rate shows a gradient decrease trend. Every 100 epochs, the learning rate becomes 0.1 times the current one. The optimizer is Adam, and the loss is the mean square error. The batch normalization layer is used to normalize the hidden layer in the batch dimension during each training. The parameters are constrained in magnitude and scale to alleviate overfitting. The model parameter file size is 11M.

[0082] Among them, the one-dimensional eigenvalues ​​of some normal working conditions are used as training sets and verification sets, and the one-dimensional eigenvalues ​​of the remaining normal working conditions and all abnormal working conditions are used as test sets. The loss curve of the training process is as follows: Figure 6 As shown, training_loss is the training set loss and val_loss is the validation set loss. It can be seen that the model converges quickly.

[0083] In this embodiment, according to the lightweight autoencoder, the reconstruction error of the above-mentioned part of normal working conditions is calculated, and the histogram and probability density diagram of the reconstruction error are fitted, such as Figure 7 shown.

[0084] Among them, quartiles are used to calculate the threshold of fault warning. Quartiles refer to the values ​​at three split points in statistics, which are arranged from small to large and divided into four equal parts. That is, the values ​​at 25%, 50% and 75% after a set of data is sorted. They are the first quartile Q1, the second quartile Q2 and the third quartile Q3 respectively. The gap between Q3 and Q1 is also called the interquartile range IQR, as shown below:

[0085] IQR=Q3-Q1

[0086] Based on this, the threshold is calculated as follows:

[0087] thr=Q3+k*IQR

[0088] Among them, k is a constant used to adjust the strictness of the threshold. After multiple experiments, the k value was selected as 1.1, and the threshold was 0.006, that is, Figure 7 The dotted line position in .

[0089] In this embodiment, the test set is used to verify the model. Figure 8 The confusion matrix in the figure, the left figure is the confusion diagram of the number of samples, and the right figure is the confusion diagram of the classification accuracy. "Normal" is the normal working condition label of the test set, and "abnormal" is the abnormal working condition label of the test set; Fig. 9 This is the ROC curve of the test set. The AUC value of the test set is calculated to be 0.988; Figure 8 and Fig. 9 It can be seen that the threshold setting is reasonable and the fault warning is normal. The reconstruction error scatter plot is as follows Fig.10 As shown, A_train, A_val, A_test_normal, and A_test_abnormal are the reconstruction error values ​​of the training set, validation set, normal working condition of the test set, and abnormal working condition of the test set, respectively. It can be seen that the model fits well.

[0090] In this embodiment, the gas turbine data set is a vibration signal, with a total of 18523 sets of data, a sampling rate of 64kHz, each sample contains 16384 points, the sample collection time intervals are not equal, and the data set does not determine the specific time point when the fault occurs.

[0091] The time spectrum of the vibration signal is calculated according to the vibration signal, including:

[0092] For the vibration signal, since each sample is short, 0.256s, a window length of 5998 points and a stride of 210 points are used to perform a short-time Fourier transform on the sample. At this time, the shape of the time-frequency matrix is ​​[3000,80]. After transposing the sample, the shape of the entire data set is [18523,80,3000], which meets the input requirements of the pre-trained speech model. At this time, the first dimension of the data set is the sample number, the second dimension is time, and the third dimension is amplitude. For example, the time-frequency diagram of normal and suspected abnormal conditions of a gas turbine Fig.11 As shown, the left figure represents the time-frequency diagram under normal working conditions, and the right figure represents the time-frequency diagram under suspected abnormal working conditions. The horizontal axis of the time-frequency diagram is time, and the vertical axis is frequency.

[0093] The time-frequency spectrum of each sample is input into the encoder of the pre-trained speech model to obtain the feature matrix of each sample. At this time, the shape of the feature matrix is ​​[1500, 384], and the shape of the entire data set is [18523, 1500, 384].

[0094] In this embodiment, since the second dimension of the target rotating machinery data set is the time dimension, and the target data set undergoes dimensional transformation when passing through the large speech model, according to logical judgment, when the feature matrix of the data set is input into the logical mean layer, the mean eigenvalue of each sample is calculated along the numerical axis in the time direction of the time-frequency spectrum. At this time, the feature matrix of the sample becomes a one-dimensional matrix, and the shape of the entire data set is [18523,1500].

[0095] In this embodiment, the lightweight autoencoder has a shape as follows: Fig.12 As shown in the figure, the autoencoder contains an input layer inputLayer with a dimension of 1500, an output layer output with a dimension of 1500, 5 batch normalization layers batch_normalization, and 5 fully connected layers, among which the fully connected layer contains a bottleneck layer Bottleneck with a dimension of 32, and the remaining fully connected layers Dense with a dimension of 256. The number of iterations epochs is 400, the batch size batch_size is 512, and the initial learning rate learning_rate is 10 -3 The learning rate shows a gradient decrease trend. Every 100 epochs, the learning rate becomes 0.1 times of the current state. The optimizer is Adam and the loss is the mean square error. The batch normalization layer is used to normalize the hidden layer in the batch dimension during each training. The parameters are constrained in magnitude and scale to alleviate overfitting. The model parameter file size is 11M.

[0096] The data before March 1, 2023 is used as the training set, and the rest is used as the test set. The loss curve of the training process is as follows: Fig.13 As shown, training_loss is the training set loss and val_loss is the validation set loss. It can be seen that the model converges quickly.

[0097] In this embodiment, according to the lightweight autoencoder, the reconstruction error of the above-mentioned part of normal working conditions is calculated, and the histogram and probability density diagram of the reconstruction error are fitted, such as Fig.14 Based on the above threshold selection method, after multiple experiments, k=2.5 was selected, and the threshold calculation result was 2.4721*10 -4 , the scatter plot of the reconstruction error value of the gas turbine data set is as follows Fig.15 As shown, it means that the gas turbine may have a minor failure between June 1, 2023 and July 1, 2023, and may have a failure after July 1, 2023.

[0098] In this embodiment, samples of the two data sets are input into the speech model from different directions to extract features, and the fault warning diagnosis method has clear diagnosis results.

[0099] In summary, the present invention has the following advantages and innovations:

[0100] 1. Cross-domain knowledge transfer: By utilizing a pre-trained large speech model, advanced technologies in the field of speech processing are transferred to the field of mechanical fault diagnosis, improving the ability and efficiency of feature extraction.

[0101] 2. Improve generalization ability: The pre-trained large speech model has strong generalization ability, which helps to extract more general and robust features and is suitable for different types of rotating machinery.

[0102] 3. Reduce data dependence: Through transfer learning, the dependence on large amounts of labeled data is reduced, solving the problem of scarce fault samples in industrial environments.

[0103] 4. Improve real-time performance: By freezing the large model parameters and only training lightweight autoencoders, the computational complexity is greatly reduced and the real-time performance of fault warning is improved.

[0104] 5. Enhanced interpretability: By analyzing the reconstruction error, the decision-making process of the model can be intuitively explained, which improves the credibility of the method.

[0105] 6. Strong adaptability: By dynamically adjusting the structure of the lightweight autoencoder, this method can adapt to different types of rotating machinery data.

[0106] 7. Improve accuracy: Combining the reconstruction errors of multiple time windows and using the sliding average method, the stability and accuracy of fault warning are improved.

[0107] The present invention provides a novel, efficient and accurate method for early warning of rotating machinery faults, which is expected to be widely used in the industrial field and provide strong support for equipment management and preventive maintenance.

[0108] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for early warning and diagnosis of rotating machinery faults based on fine-tuning of a large speech model, characterized in that: include: Step S1, preprocessing the data set of the rotating machinery; Step S2, loading the encoder part of the pre-trained large speech model into the fault warning system, and freezing all parameters of the encoder part; Step S3, adding a logical averaging layer and a lightweight autoencoder after the encoder part; Step S4, using the lightweight autoencoder to calculate a reconstruction error, and using the reconstruction error as a fault warning indicator; Step S5, calculating a threshold of fault warning based on the reconstruction error, and obtaining a result of the fault warning according to the threshold.

2. The method according to claim 1, characterized in that: The data set is preprocessed in step S1, specifically including: Perform a short-time Fourier transform on the data set; When the data is a sound signal, the logarithmic Mel spectrum is obtained for the data to meet the input requirements of the large speech model.

3. The method according to claim 1, characterized in that: The step S2 loads the pre-trained speech model encoder part and freezes the parameters, specifically including: Use speech signals to train a large speech model, take the encoder part and freeze the parameters as a feature extractor.

4. The method according to claim 1, characterized in that: The step S3 of adding a logical averaging layer specifically includes: Add a logical calculation part to the averaging layer to determine the average of the output values ​​of the speech model encoder along the specified numerical axis.

5. The method according to claim 1, characterized in that: The calculation of the lightweight autoencoder reconstruction error in step S4 specifically includes: The lightweight autoencoder is trained using the mean eigenvalues ​​of normal working condition samples; The mean square error between the mean eigenvalue of the sample and the output value of the lightweight autoencoder is calculated as the reconstruction error.

6. The method according to claim 1, characterized in that: The calculation of the fault warning threshold in step S5 specifically includes: The reconstruction error of the sample is modeled using the gamma distribution; Quartiles are used to calculate the fault warning threshold.

7. The method according to claim 6, characterized in that: The calculation formula of the fault warning threshold is: thr=Q3+k*IQR Among them, Q3 is the third quartile, IQR is the interquartile range, and k is an adjustable constant.

8. The method according to any one of claims 1 to 7, characterized in that: Also includes: Input new sample data into the trained model and calculate its reconstruction error; The calculated reconstruction error is compared with the fault warning threshold to determine whether there is a fault.

Citation Information

Cited By

  • Power equipment fault detection method, device, equipment and medium

    CN122050436A