Water pump fault diagnosis method based on two-stage generative adversarial network
Through two-stage generative adversarial network, the submersible pump soundprint data is enhanced and noise-reduced, which solves the problem of difficulty in extracting the soundprint feature of the submersible pump in complex noise environments, and achieves efficient and accurate fault diagnosis.
Patent Information
- Application Number
- CN202510005498.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-03
- Publication Date
- 2025-04-25
AI Technical Summary
In complex noise environments, existing noise reduction technologies are difficult to accurately extract the soundprint characteristics of submersible pumps, resulting in inaccuracy and inefficiency of pump fault diagnosis.
The method based on a two-stage generation adversarial network is adopted, and the voiceprint data of the submersible pump is enhanced through the first-stage generation adversarial network, and the noise-free voiceprint data is reduced through the second-stage generation adversarial network, thereby extracting the characteristic voiceprints of the submersible pump failure.
It significantly improves the accuracy and efficiency of submersible pump fault diagnosis, can effectively remove noise in complex noise environments, extract clear voiceprint features, reduce noise interference, and improve the reliability of fault detection.
Smart Images
Figure CN119889349A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of water pump soundprint noise reduction, and in particular to a water pump fault diagnosis method based on a two-stage generative adversarial network. Background Art
[0002] The submersible pump in the tunnel pump room plays a vital role in the tunnel drainage system. Its failure may lead to a decrease in drainage capacity, and in severe cases, it may cause tunnel flooding, endangering safe operation. Since the submersible pump is in an underwater working environment, its working environment is complex, mixed with various noises such as the sound of mechanical equipment running, water flow, and vehicles passing by. It is a strong noise environment, which brings challenges to the voiceprint recognition of the submersible pump. Therefore, it is necessary to develop an effective voiceprint noise reduction technology to improve the accuracy of water pump voiceprint recognition and detect water pump failures in time.
[0003] Existing noise reduction technologies often cannot accurately extract target sound features in complex noise environments, especially when there are many types of noise and complex spectra, traditional filtering or signal processing methods have limited effects. Therefore, how to effectively remove strong noise and enhance the characteristics of water pump voiceprint data, thereby improving the accuracy of voiceprint recognition, is a technical problem that technicians in this field need to solve urgently. Summary of the invention
[0004] The present invention provides a water pump fault diagnosis method based on a two-stage generative adversarial network to solve the above technical problems.
[0005] In order to solve the above technical problems, the present invention provides a water pump fault diagnosis method based on a two-stage generative adversarial network, comprising the following steps:
[0006] Step 1: Collecting the operating sound data of the submersible pump in the tunnel pump room, wherein the operating sound data includes normal operating data and fault operating data in a noise-free environment and a noise-containing environment;
[0007] Step 2: preprocessing the collected operation sound data to obtain frequency domain features;
[0008] Step 3: construct a one-stage generative adversarial network, and use the frequency domain features to train the one-stage generative adversarial network, and output enhanced submersible pump fault time-frequency diagram and normal time-frequency diagram;
[0009] Step 4: construct a two-stage generative adversarial network, and use the frequency domain features and the output of the first-stage generative adversarial network to train the two-stage generative adversarial network, and output the submersible pump fault time-frequency diagram and normal time-frequency diagram after noise reduction;
[0010] Step 5: Detect the submersible pump fault based on the noise-reduced submersible pump fault time-frequency diagram and the normal time-frequency diagram output in step 4.
[0011] Preferably, the pre-processing step comprises:
[0012] Step 2.1: normalizing the running sound data;
[0013] Step 2.2: Frame the continuous sound signal according to the preset time window;
[0014] Step 2.3: Perform short-time Fourier transform on the sound signal of each frame to generate a time-frequency graph and extract the time-frequency features of the sound;
[0015] Step 2.4: Calculate the amplitude of the converted spectrum data to generate a two-dimensional time-frequency diagram.
[0016] Preferably, the preset time window is 20 ms.
[0017] Preferably, step 3 comprises:
[0018] Step 3.1: construct a first generator, take the two-dimensional time-frequency graph as input, perform data enhancement on the noisy and noise-free fault operation data, and generate an enhanced time-frequency graph;
[0019] Step 3.2: construct a first discriminator, take the two-dimensional time-frequency map and the generated enhanced time-frequency map as input, and distinguish the source of the time-frequency map;
[0020] Step 3.3: construct a first adversarial loss function, and optimize the adversarial process between the first generator and the first discriminator based on the first adversarial loss function;
[0021] Step 3.4: define a reconstruction loss, the first generator uses the reconstruction loss to ensure that the generated time-frequency graph maintains the original characteristics;
[0022] Step 3.5: Construct the first optimization objective;
[0023] Step 3.6: Based on the first optimization objective, an enhanced submersible pump fault time-frequency diagram and a normal time-frequency diagram are obtained and output.
[0024] Preferably, the reconstruction loss is:
[0025]
[0026] Among them, [||x-G1(x)||1] is the L1 norm, which represents the difference between the input sample x and the generated sample G1(x); Represents the expected value of the difference between the input sample x and the generated sample G1(x).
[0027] Preferably, the first optimization objective is:
[0028] The optimization objective of the first generator is to minimize the adversarial loss and the reconstruction loss;
[0029] The optimization objective of the first discriminator is to maximize the adversarial loss.
[0030] Preferably, step 4 includes:
[0031] Step 4.1: construct a second generator, taking the enhanced time-frequency graph generated by the first-stage generative adversarial network and the time-frequency graph of the running sound data in the noisy environment as input, and generating a denoised time-frequency graph;
[0032] Step 4.2: construct a second discriminator, take the noise-free time-frequency graph in the two-dimensional time-frequency graph and the noise-reduced time-frequency graph generated by the second generator as input, and distinguish the source of the time-frequency graph;
[0033] Step 4.3: construct a second adversarial loss function, and optimize the adversarial process between the second generator and the second discriminator based on the second adversarial loss function;
[0034] Step 4.4: defining a noise reduction loss, wherein the second generator removes noise by using the noise reduction loss;
[0035] Step 3.5: Construct the second optimization objective;
[0036] Step 3.6: Based on the second optimization objective, obtain the noise-reduced submersible pump fault time-frequency diagram and normal time-frequency diagram and output them.
[0037] Preferably, the second optimization objective is:
[0038] The optimization objective of the second generator is to minimize the adversarial loss and the denoising loss;
[0039] The optimization objective of the second discriminator is to maximize the adversarial loss.
[0040] Preferably, the noise reduction loss is:
[0041]
[0042] Among them, [||x-G2(x)||1] is the L1 norm, which represents the difference between the input sample x and the generated sample G2(x); Represents the expected value of the difference between the input sample x and the generated sample G2(x).
[0043] Preferably, step 5 includes: utilizing the noise-reduced submersible pump fault time-frequency diagram and normal time-frequency diagram output by the second-stage generative adversarial network, performing fault classification through machine learning or voiceprint recognition technology, and obtaining the submersible pump fault detection result.
[0044] Compared with the prior art, the pump fault diagnosis method based on the two-stage generative adversarial network provided by the present invention has the following advantages:
[0045] 1. The present invention first collects the normal operation and fault operation data of the submersible pump in noise-free and noisy environments, then uses the first-stage generative adversarial network for data enhancement, and uses the second-stage generative adversarial network for noise reduction. It can effectively remove the noise of mechanical operation, water flow, vehicle driving and other noises in a complex and noisy environment such as a tunnel pump room, extract the core soundprint features of the submersible pump, and significantly improve the noise reduction capability;
[0046] 2. After removing the noise, the present invention makes the soundprint characteristics of the water pump clearer, avoids the interference of strong noise on the recognition result, and thus improves the accuracy of soundprint recognition in the fault state of the submersible pump. In the fault diagnosis of the submersible pump in the tunnel pump room, especially at the critical moment of the drainage system, the operation abnormality of the submersible pump can be detected in time, which greatly improves the efficiency and accuracy of the fault diagnosis of the submersible pump and ensures the safe and reliable operation of the tunnel pump room system;
[0047] 3. The present invention uses a one-stage generative adversarial network to perform data enhancement and generate more data with submersible pump fault characteristics, which not only increases the diversity of fault samples, but also improves the generalization ability of the voiceprint recognition model and can adapt to more diverse noises and working environments;
[0048] 4. The present invention is non-contact. Even if the submersible pump is in an underwater operating environment, it can extract the sound signal of the submersible pump and realize efficient fault diagnosis while real-time monitoring, which is helpful for the automatic management of the tunnel pump room and the remote monitoring of the submersible pump;
[0049] 5. The present invention enhances the adaptability of the system to different noise environments through a two-stage generative adversarial network, especially in complex environments where multiple noise sources overlap, and shows higher robustness. It can still maintain a high noise reduction effect when the noise types are complex and the spectrum differences are large, which is significantly better than traditional filtering or frequency domain processing methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 This is a flow chart of a water pump fault diagnosis method based on a two-stage generative adversarial network in a specific embodiment of the present invention. DETAILED DESCRIPTION
[0051] In order to describe the technical solution of the above invention in more detail, specific embodiments are listed below to demonstrate the technical effects; it should be emphasized that these embodiments are used to illustrate the present invention but not to limit the scope of the present invention.
[0052] The pump fault diagnosis method based on the two-stage generative adversarial network provided by the present invention is as follows: Figure 1 As shown, the following steps are included:
[0053] Step 1: Collect the operating sound data of the submersible pump in the tunnel pump room, which includes normal operation data and fault operation data in a noise-free environment and a noise environment. The operating sound data of the submersible pump in the tunnel pump room includes the mechanical vibration sound of the submersible pump, the operation sound, the water flow sound, the noise generated by the passing of vehicles, etc.
[0054] Step 2: Preprocess the collected running sound data and convert the time domain signal into frequency domain features to facilitate subsequent training of the adversarial network.
[0055] Step 3: construct a one-stage generative adversarial network, and use the frequency domain features to train the one-stage generative adversarial network, and output enhanced submersible pump fault time-frequency diagram and normal time-frequency diagram;
[0056] Step 4: construct a two-stage generative adversarial network, and use the frequency domain features and the output of the first-stage generative adversarial network to train the two-stage generative adversarial network, and output the submersible pump fault time-frequency diagram and normal time-frequency diagram after noise reduction;
[0057] Step 5: Detect the submersible pump fault based on the noise-reduced submersible pump fault time-frequency diagram and the normal time-frequency diagram output in step 4.
[0058] The present invention first collects the normal operation and fault operation data of the submersible pump in noise-free and noisy environments, then uses a one-stage generative adversarial network to perform data enhancement, and uses a two-stage generative adversarial network to perform noise reduction, which greatly improves the efficiency and accuracy of submersible pump fault diagnosis and ensures the safe and reliable operation of the tunnel pump room system.
[0059] In some embodiments, the pre-processing step comprises:
[0060] Step 2.1: The running sound data is normalized to reduce the amplitude differences caused by different volumes and different equipment collection.
[0061] Step 2.2: Frame the continuous sound signal according to a preset time window to capture the local features of the sound. In some embodiments, the preset time window may be 20 ms.
[0062] Step 2.3: Perform short-time Fourier transform (STFT) on the sound signal of each frame to generate a time-frequency diagram and extract the time-frequency features of the sound. The specific formula is as follows:
[0063]
[0064] Among them, X(t, f) is a time-frequency diagram that displays the signal information at time t and frequency f, x[n] is the discrete time series of the original signal, w[nt] is the window function that determines the weighting of the local time region, and e[n] is the window function that determines the weighting of the local time region. -j2πfn It is the complex exponential term of the Fourier transform, which represents the frequency component of the signal.
[0065] Step 2.4: Calculate the amplitude of the converted spectrum data to generate a two-dimensional time-frequency graph as the input of the subsequent generative adversarial network.
[0066] In some embodiments, step 3 is used for data enhancement, specifically including:
[0067] Step 3.1: Construct the first generator (Generator 1, G1), and take the two-dimensional time-frequency graph as input, which includes normal and faulty soundprint time-frequency graphs of submersible pumps with and without noise; then perform data enhancement on the faulty operation data with and without noise to generate an enhanced time-frequency graph, which includes faulty time-frequency graph samples and normal time-frequency graph samples.
[0068] Step 3.2: Construct the first discriminator (Discriminator 1, D1), take the two-dimensional time-frequency graph (ie, real data) and the generated enhanced time-frequency graph (ie, generated data) as input, distinguish the source of the time-frequency graph, that is, distinguish whether the input time-frequency graph comes from real data or generated data.
[0069] Step 3.3: Construct a first adversarial loss function (GAN Loss), and optimize the adversarial process between the first generator and the first discriminator based on the first adversarial loss function. As shown in the following formula:
[0070]
[0071] Among them, x is the real time-frequency graph sample, p data (x) represents the probability distribution of the real time-frequency graph samples, represents the expected value of the first discriminator for distinguishing the real data as the “true” class. G1(z) is the time-frequency diagram generated by the first generator based on the noise z. represents the expected value of the first discriminator classifying the generated data as the “fake” class.
[0072] Step 3.4: define a reconstruction loss, wherein the first generator ensures that the generated time-frequency graph maintains the original fault or normal characteristics through the reconstruction loss;
[0073] In this embodiment, the reconstruction loss is:
[0074]
[0075] Among them, [||x-G1(x)||1] is the L1 norm, that is, the sum of absolute errors, which represents the difference between the input sample x and the generated sample G1(x); Represents the expected value of the difference between the input sample x and the generated sample G1(x).
[0076] Step 3.5: Construct the first optimization objective.
[0077] In this embodiment, the first optimization goal is:
[0078] The optimization goal of the first generator is to minimize the adversarial loss L GAN1 and the reconstruction loss L rec1 ;
[0079] The optimization goal of the first discriminator is to maximize the adversarial loss L GAN1 .
[0080] Step 3.6: Based on the first optimization objective, an enhanced submersible pump fault time-frequency diagram and a normal time-frequency diagram are obtained and output for further data processing.
[0081] In some embodiments, step 4 is used to extract the fault characteristics of the submersible pump from the noisy voiceprint frequency graph and remove the environmental noise, which specifically includes:
[0082] Step 4.1: Construct a second generator (Generator 2, G2), take the enhanced time-frequency graph generated by the first-stage generative adversarial network and the time-frequency graph of the operating sound data in a noisy environment as input, generate a noise-reduced time-frequency graph, and ensure that the fault characteristics of the submersible pump are retained.
[0083] Step 4.2: Construct a second discriminator (Discriminator 2, D2), take the noise-free time-frequency graph in the two-dimensional time-frequency graph (i.e., the real noise-free time-frequency graph) and the denoised time-frequency graph generated by the second generator (i.e., the generated noise-free time-frequency graph) as input, distinguish the source of the time-frequency graph, that is, distinguish whether the input time-frequency graph comes from the real noise-free data.
[0084] Step 4.3: Construct a second adversarial loss function (GAN Loss), and optimize the adversarial process between the second generator and the second discriminator based on the second adversarial loss function. As follows:
[0085]
[0086] Among them, x is the real time-frequency graph sample, p data (x) represents the probability distribution of the real time-frequency graph samples, represents the expected value of the second discriminator for distinguishing the real data as the “true” class. G2(z) is the time-frequency diagram generated by the second generator based on the noise z. represents the expected value of the second discriminator classifying the generated data as the “fake” class.
[0087] Step 4.4: Define noise suppression loss, and the second generator removes noise through the noise suppression loss.
[0088] In this embodiment, the noise reduction loss is defined as:
[0089]
[0090] Among them, [||x-G2(x)||1] is the L1 norm, that is, the sum of absolute errors, which represents the difference between the input sample x and the generated sample G2(x); Represents the expected value of the difference between the input sample x and the generated sample G2(x).
[0091] This loss measures the difference between the generated time-frequency graph and the true noise-free time-frequency graph, ensuring that the noise is effectively removed.
[0092] Step 3.5: Construct the second optimization objective.
[0093] In this embodiment, the second optimization objective is:
[0094] The optimization goal of the second generator is to minimize the adversarial loss L GAN2 and noise reduction loss L denoise ;
[0095] The optimization goal of the second discriminator is to maximize the adversarial loss L GAN2 .
[0096] Step 3.6: Based on the second optimization objective, the fault time-frequency diagram and the normal time-frequency diagram of the submersible pump after noise reduction are obtained and output for subsequent fault diagnosis.
[0097] In some embodiments, step 5 includes: using the two-stage generated adversarial network output to generate the noise-reduced submersible pump fault time-frequency diagram and the normal time-frequency diagram, obtaining the noise-reduced submersible pump fault voiceprint data and the normal voiceprint data, performing fault classification through machine learning or voiceprint recognition technology, and obtaining the submersible pump fault detection result, thereby timely detecting the submersible pump fault and ensuring the safe operation of the tunnel pump room.
[0098] In summary, the water pump fault diagnosis method based on the two-stage generative adversarial network provided by the present invention includes the following steps: Step 1: Collecting the operating sound data of the submersible pump in the tunnel pump room, the operating sound data including normal operating data and fault operating data in a noise-free environment and a noisy environment; Step 2: Preprocessing the collected operating sound data to convert the time domain signal into a frequency domain feature; Step 3: Constructing a one-stage generative adversarial network, and training the one-stage generative adversarial network with the frequency domain features, and outputting enhanced submersible pump fault time-frequency diagrams and normal time-frequency diagrams; Step 4: Constructing a two-stage generative adversarial network, and training the two-stage generative adversarial network with the frequency domain features and the output of the one-stage generative adversarial network, and outputting the submersible pump fault time-frequency diagram and normal time-frequency diagram after noise reduction; Step 5: Detecting the submersible pump fault based on the submersible pump fault time-frequency diagram and normal time-frequency diagram after noise reduction output in step 4. The present invention first collects the operating soundprint data of the submersible pump in the tunnel pump room, including the soundprint under normal conditions and the soundprint when a fault occurs, and then enhances the fault soundprint through a one-stage generative adversarial network. On this basis, a two-stage generative adversarial network is used to perform noise reduction processing on the soundprint with mechanical noise, water flow noise and vehicle noise, and the characteristic soundprint of the submersible pump fault is successfully extracted, thereby accurately identifying the fault state of the submersible pump.
[0099] Obviously, those skilled in the art can make various changes and modifications to the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.
Claims
1. A water pump fault diagnosis method based on a two-stage generative adversarial network, characterized in that: The steps include: Step 1: Collecting the operating sound data of the submersible pump in the tunnel pump room, wherein the operating sound data includes normal operating data and fault operating data in a noise-free environment and a noise-containing environment; Step 2: preprocessing the collected operation sound data to obtain frequency domain features; Step 3: construct a one-stage generative adversarial network, and use the frequency domain features to train the one-stage generative adversarial network, and output enhanced submersible pump fault time-frequency diagram and normal time-frequency diagram; Step 4: construct a two-stage generative adversarial network, and use the frequency domain features and the output of the first-stage generative adversarial network to train the two-stage generative adversarial network, and output the submersible pump fault time-frequency diagram and normal time-frequency diagram after noise reduction; Step 5: Detect the submersible pump fault based on the noise-reduced submersible pump fault time-frequency diagram and the normal time-frequency diagram output in step 4.
2. The water pump fault diagnosis method based on a two-stage generative adversarial network according to claim 1 is characterized in that: The pre-processing steps include: Step 2.1: normalizing the running sound data; Step 2.2: Frame the continuous sound signal according to the preset time window; Step 2.3: Perform short-time Fourier transform on the sound signal of each frame to generate a time-frequency graph and extract the time-frequency features of the sound; Step 2.4: Calculate the amplitude of the converted spectrum data to generate a two-dimensional time-frequency diagram.
3. The water pump fault diagnosis method based on a two-stage generative adversarial network as claimed in claim 2 is characterized in that: The preset time window is 20ms.
4. The water pump fault diagnosis method based on a two-stage generative adversarial network as claimed in claim 2 is characterized in that: Step 3 includes: Step 3.1: construct a first generator, take the two-dimensional time-frequency graph as input, perform data enhancement on the noisy and noise-free fault operation data, and generate an enhanced time-frequency graph; Step 3.2: construct a first discriminator, take the two-dimensional time-frequency map and the generated enhanced time-frequency map as input, and distinguish the source of the time-frequency map; Step 3.3: construct a first adversarial loss function, and optimize the adversarial process between the first generator and the first discriminator based on the first adversarial loss function; Step 3.4: define a reconstruction loss, the first generator uses the reconstruction loss to ensure that the generated time-frequency graph retains the original characteristics; Step 3.5: Construct the first optimization objective; Step 3.6: Based on the first optimization objective, an enhanced submersible pump fault time-frequency diagram and a normal time-frequency diagram are obtained and output.
5. The water pump fault diagnosis method based on a two-stage generative adversarial network as claimed in claim 4 is characterized in that: The reconstruction loss is: Among them, [||x-G1(x)||1] is the L1 norm, which represents the difference between the input sample x and the generated sample G1(x); Represents the expected value of the difference between the input sample x and the generated sample G1(x).
6. The water pump fault diagnosis method based on a two-stage generative adversarial network as claimed in claim 4 is characterized in that: The first optimization goal is: The optimization objective of the first generator is to minimize the adversarial loss and the reconstruction loss; The optimization objective of the first discriminator is to maximize the adversarial loss.
7. The water pump fault diagnosis method based on a two-stage generative adversarial network as claimed in claim 4 is characterized in that: Step 4 includes: Step 4.1: construct a second generator, taking the enhanced time-frequency graph generated by the first-stage generative adversarial network and the time-frequency graph of the running sound data in the noisy environment as input, and generating a denoised time-frequency graph; Step 4.2: construct a second discriminator, take the noise-free time-frequency graph in the two-dimensional time-frequency graph and the noise-reduced time-frequency graph generated by the second generator as input, and distinguish the source of the time-frequency graph; Step 4.3: construct a second adversarial loss function, and optimize the adversarial process between the second generator and the second discriminator based on the second adversarial loss function; Step 4.4: defining a noise reduction loss, wherein the second generator removes noise by using the noise reduction loss; Step 3.5: Construct the second optimization objective; Step 3.6: Based on the second optimization objective, obtain the noise-reduced submersible pump fault time-frequency diagram and normal time-frequency diagram and output them.
8. The water pump fault diagnosis method based on a two-stage generative adversarial network according to claim 7 is characterized in that: The second optimization objective is: The optimization objective of the second generator is to minimize the adversarial loss and the denoising loss; The optimization objective of the second discriminator is to maximize the adversarial loss.
9. The water pump fault diagnosis method based on a two-stage generative adversarial network according to claim 7 is characterized in that: The noise reduction loss is: Among them, [||x-G2(x)||1] is the L1 norm, which represents the difference between the input sample x and the generated sample G2(x); Represents the expected value of the difference between the input sample x and the generated sample G2(x).
10. The water pump fault diagnosis method based on a two-stage generative adversarial network according to claim 7, characterized in that: Step 5 includes: using the noise-reduced submersible pump fault time-frequency diagram and normal time-frequency diagram output by the second-stage generative adversarial network, classifying faults through machine learning or voiceprint recognition technology, and obtaining submersible pump fault detection results.