Rotating equipment fault open set diagnosis method, system and device based on voiceprint

Abstract

The invention belongs to the technical field of rotating equipment fault diagnosis, and particularly relates to a voiceprint-based rotating equipment fault open set diagnosis method, system and device. The method comprises the following steps: acquiring a voiceprint signal of a rotating device; preprocessing the voiceprint signal, extracting a Mel-frequency cepstrum coefficient and a sound spectrum centroid feature based on the preprocessed voiceprint signal, and fusing the Mel-frequency cepstrum coefficient and the sound spectrum centroid feature to obtain a fused feature; mapping the fusion feature to a potential space through an encoder, decoupling the fusion feature into a classification variable and a noise variable, and reconstructing the fusion feature through a decoder to obtain a reconstructed feature; calculating a prototype center of each fault category based on the classification variables; training a variational auto-encoder model based on the total loss function; and measuring the distance between the to-be-detected sample and the center of the known category prototype through the Euclidean distance and the dot product similarity, and judging whether the to-be-detected sample is an unknown fault category in combination with a preset threshold value. According to the invention, fault diagnosis of the rotating equipment is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating equipment fault diagnosis, and in particular relates to a method, system and device for open-set diagnosis of rotating equipment faults based on voiceprint. Background Art

[0002] Rotating equipment, due to its exceptional efficiency, excellent dynamic response characteristics, and high power density, has been widely used in a number of key areas, including modern industry, transportation, and aerospace. However, with its gradual promotion in various complex application scenarios, the stability of rotating equipment's operating status has also attracted increasing attention. Especially under harsh operating conditions, potential failures of rotating equipment may not only cause equipment downtime and production interruptions, but may even trigger serious safety accidents, with far-reaching negative impacts on overall production and use. Therefore, ensuring the reliable operation of rotating equipment and performing timely and accurate fault detection and diagnosis are particularly critical to ensure production efficiency and operational safety, while effectively reducing maintenance costs and the risk of unplanned downtime.

[0003] Although existing fault diagnosis methods for rotating equipment have shown remarkable performance under ideal conditions, they still face the following challenges in practical applications.

[0004] (1) High cost of traditional methods: Traditional fault detection methods usually rely on the participation of professional technicians, such as regular maintenance and troubleshooting. This type of manual operation is not only time-consuming, but also requires a high level of technical skills, increasing the cost of personnel training and maintenance. In addition, the diagnostic process is easily affected by human factors, which may lead to untimely problem discovery or misjudgment, thereby increasing indirect costs. At the same time, some traditional detection methods require direct contact with the equipment, such as current analysis and vibration analysis. These invasive methods not only increase the cost of installation and maintenance, but may also cause damage to the equipment itself.

[0005] (2) Inadequate early fault detection capabilities: Early faults in rotating equipment often manifest as minor wear and fatigue in the mechanical structure, or minor anomalies in the electrical system. These changes are often difficult to detect during normal operation of the rotating equipment. In industrial environments, early equipment failures are often a sign of gradual equipment deterioration. If faults cannot be accurately identified at an early stage, they will lead to further deterioration, resulting in more serious downtime or equipment damage.

[0006] (3) Lack of ability to identify unknown faults: Traditional fault identification methods usually assume that all possible fault types have been labeled during the training process. Therefore, they often perform poorly when faced with completely new types of faults. The lack of ability to effectively handle unknown fault categories makes these methods less adaptable and generalizable in complex industrial environments. Summary of the Invention

[0007] The present invention aims to address the problems of the prior art and proposes a voiceprint-based open-set fault diagnosis method, system, and device for rotating equipment. By combining latent space unwrapping technology with a prototype representation method, an effective fault diagnosis mechanism for rotating equipment voiceprint signals is constructed. On the one hand, latent space unwrapping enables accurate differentiation of different fault characteristics, improving the variational autoencoder model's ability to identify unknown faults. On the other hand, a classification strategy based on prototype representation improves the classification accuracy and diagnostic robustness of known fault categories, ultimately achieving accurate, real-time, and efficient detection of rotating equipment fault states.

[0008] In order to achieve the above object, the technical solution provided by the present invention is: In the first aspect, the present invention provides a method for open-set diagnosis of rotating equipment faults based on voiceprints, the method comprising: Acquire a voiceprint signal of the rotating equipment, wherein the voiceprint signal includes a fault signal and a non-fault signal, wherein the fault signal includes multiple known fault categories; Preprocess the voiceprint signal, extract the Mel-frequency cepstral coefficient and the sound spectrum centroid features based on the preprocessed voiceprint signal, and fuse them to obtain the fusion feature; Construct a variational autoencoder model, map the fused features to the latent space through the encoder in the variational autoencoder model, decouple them into classification variables and noise variables, and reconstruct the fused features through the decoder in the variational autoencoder model to obtain reconstructed features; Calculate the prototype center for each known fault category based on the categorical variables; An unknown category detection loss function is constructed with reconstruction error and KL divergence as optimization targets, and a known category classification loss function is constructed based on the prototype center. The parameters of the variational autoencoder model are optimized based on the unknown category detection loss function and the known category classification loss function to obtain a trained variational autoencoder model. The sample to be tested is input into the trained variational autoencoder model, the reconstruction error value is calculated, and the reconstruction error value is compared with the preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the sample to be tested is judged to be an unknown fault category. Otherwise, the distance between the sample to be tested and the prototype center is measured to judge that the sample to be tested belongs to the known fault category corresponding to the nearest prototype center.

[0009] Furthermore, the pre-processing of the voiceprint signal includes: The pre-emphasis processing of the voiceprint signal is expressed as follows: ,in, Represents the index of the global sampling point, express The emphasized signal at the sampling point, express The original signal at the sampling point, express The original signal at the sampling point, Indicates the pre-emphasis coefficient; Divide the pre-emphasized voiceprint signal into short time frames; Apply a Hamming window function to each frame for windowing.

[0010] Furthermore, the variational autoencoder model is based on the ResNet18 structure. The encoder includes multiple residual convolution units, each of which includes a convolution layer, a batch normalization layer and a Leaky ReLU activation function. The structure of the decoder mirrors that of the encoder.

[0011] Furthermore, the categorical variables and noise variables are fused through reparameterization techniques to generate latent variables, which are then input into the decoder and expressed as follows:

[0012]

[0013]

[0014] in, represents a categorical variable, represents the noise variable, represents the latent variable, Indicates the total number of frames. Indicates the The fusion features of the frame, Indicates that the weight parameter is The encoder pair fusion feature The latent space representation after nonlinear mapping, is the weight adjustment value, represents noise sampled from a standard normal distribution.

[0015] Furthermore, the prototype center of each known fault category is calculated based on the classification variable, which is expressed as follows:

[0016] in, Indicates known fault categories The center of the prototype in the latent space, Indicates known fault categories The number of samples, , represents a categorical variable.

[0017] Furthermore, the unknown category detection loss function is constructed with reconstruction error and KL divergence as optimization targets, which is expressed as follows:

[0018]

[0019]

[0020] in, represents the reconstruction error, represents the KL divergence, represents the unknown category detection loss function, Indicates the total number of frames. Indicates the The fusion features of the frame, Indicates the Reconstructed features of the frame, represents the total number of dimensions of the latent space, Represents the latent space The mean of the dimensions, Represents the latent space The standard deviation of each dimension.

[0021] Furthermore, the construction of a known category classification loss function based on the prototype center includes:

[0022]

[0023]

[0024]

[0025] in, represents the categorical variable of the current sample, Indicates known fault categories The center of the prototype in the latent space, represents the L1 norm, Represents the classification variable of the current sample and the known fault category The Euclidean distance between the centers of the prototypes in the latent space, Represents the classification variable of the current sample and the known fault category The dot product distance between prototype centers in latent space, Represents the classification variable of the current sample and the known fault category The integrated distance between prototype centers in the latent space, represents the total number of dimensions of the latent space, represents the known category classification loss function, Indicates the total number of samples in the current training batch, , Indicates the fault category to which the current sample belongs, Represents the comprehensive distance between the classification variable of the current sample and the prototype center of the fault category to which the current sample belongs in the latent space, represents the activation function, represents the hyperparameter, Indicates the preset distance threshold.

[0026] Furthermore, the optimization of the parameters of the variational autoencoder model based on the unknown category detection loss function and the known category classification loss function includes: The total loss function is obtained based on the unknown category detection loss function and the known category classification loss function. The formula is expressed as:

[0027] in, represents the total loss function, represents the unknown category detection loss function, represents the known category classification loss function, represents a hyperparameter; Optimize the parameters of the variational autoencoder model based on the total loss function.

[0028] In the second aspect, the present invention provides a rotating equipment fault open set diagnosis system based on voiceprint, comprising: a feature extraction module, a decoupling and reconstruction module, a prototype center calculation module, an optimization training module and a detection module, wherein: The feature extraction module is used to obtain the voiceprint signal of the rotating device; preprocess the voiceprint signal, extract the Mel frequency cepstral coefficient and the sound spectrum centroid feature based on the preprocessed voiceprint signal, and fuse them to obtain a fusion feature; The decoupling and reconstruction module is used to construct a variational autoencoder model, map the fused features to the latent space through the encoder in the variational autoencoder model, decouple them into classification variables and noise variables, and reconstruct the fused features through the decoder in the variational autoencoder model to obtain reconstructed features; The prototype center calculation module is used to calculate the prototype center of each known fault category based on the classification variables; The optimization training module is used to construct an unknown category detection loss function with reconstruction error and KL divergence as optimization targets, construct a known category classification loss function based on the prototype center, and optimize the parameters of the variational autoencoder model based on the unknown category detection loss function and the known category classification loss function to obtain a trained variational autoencoder model; The detection module is used to input the sample to be tested into the trained variational autoencoder model, calculate the reconstruction error value, and compare the reconstruction error value with a preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the sample to be tested is judged to be an unknown fault category. Otherwise, the distance between the sample to be tested and the prototype center is measured to determine whether the sample to be tested belongs to the known fault category corresponding to the nearest prototype center.

[0029] The third aspect: The present invention provides a device for open-set diagnosis of rotating equipment faults based on voiceprints, comprising a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the open-set diagnosis method for rotating equipment faults based on voiceprints.

[0030] Compared with the prior art, the present invention has the following significant advantages: (1) Using the voiceprint signal of rotating equipment as a feature for fault identification, as a non-invasive detection method, can reduce the cost of installation and maintenance. At the same time, the detection method based on deep learning can reduce the labor cost.

[0031] (2) Voiceprint signals can effectively extract tiny operating signals during the operation of rotating equipment. Compared with traditional vibration signals, they are more sensitive and can detect early faults of rotating equipment.

[0032] (3) To address the problem of scarce samples in the real world, the variational autoencoder was improved to reconstruct the latent feature space of the voiceprint signal, thus realizing the detection of unknown fault samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a method for open-set diagnosis of rotating equipment faults based on voiceprints according to the present invention; Figure 2 This is a comparison chart of the effects of the present invention and other methods. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0035] Example 1: Figure 1 As shown, the present invention provides a method for open-set diagnosis of rotating equipment faults based on voiceprints, which is applied to a rotating equipment fault diagnosis system. Taking motor fault diagnosis as an example, the method specifically includes the following steps: Step 1: Using a non-contact acoustic sensor (such as a high-sensitivity condenser microphone), real-time sampling of operating soundprint signals from various types (powers) of permanent magnet synchronous motors (PMSMs) is performed in a soundproof room for motor fault detection. These signals, including both faulty and non-faulty types, are used as training and test datasets. In this embodiment, the samples collected include PMSMs with powers of 0.4 kW, 1 kW, 1.2 kW, 1.5 kW, and 1.8 kW, respectively. The 0.4 kW, 1.5 kW, and 1.8 kW PMSMs contain both faulty and non-faulty types, the 1 kW PMSM only has faulty type samples, and the 1.2 kW PMSM only has non-faulty type samples. Each sample consists of 6 seconds of operating data, and the specific sampling frequency is set to 22.05 kHz. By processing samples of various power levels, the model aims to improve its generalization and accuracy for customized PMSMs of multiple specifications.

[0036] The collected data set is randomly divided into 80% for training and 20% for testing. The training data and labels are denoted as and , the test data and labels are denoted as and The sample labels of one of the power categories that includes both faulty and non-faulty samples are hidden and put into the test dataset as an unknown class. The training dataset does not contain samples of this power category.

[0037] Step 2: Collect the original voiceprint signal Perform pre-emphasis processing: ,in, Represents the index of the global sampling point, express The emphasis signal at the sampling point, express The original signal at the sampling point, express The original signal at the sampling point, Indicates the pre-emphasis coefficient. In this embodiment Take 0.97.

[0038] The pre-emphasized signal is divided into frames with a window length of 1024 points, and the frames overlap by 50%. The total number of frames can be obtained =257, and use the Hamming window function to perform windowing processing on each frame of data. The Hamming window function is expressed as follows:

[0039] in, Indicates the length of the window, which is 1024 in this embodiment. It is the sample index within the single frame signal after framing. In this embodiment, .

[0040] Therefore, the voiceprint signal of each frame after framing It can be expressed as:

[0041] in, Indicates the Intraframe The windowed overlapping signal of sampling points, Indicates the The emphasized signal of each sampling point.

[0042] The signal spectrum is then obtained by performing a fast Fourier transform (FFT) on the windowed signal frame. , and then converted into the corresponding power spectrum :

[0043]

[0044] in, Indicates the Frame The frequency domain signal of the frequency components, represents the index of the frequency component, represents an imaginary number, Indicates the Frame The power spectrum of the frequency components is obtained by applying a Mel filter bank to the power spectrum, taking the logarithm of the output of each filter, and applying a discrete cosine transform (DCT) to obtain the Mel frequency cepstrum (MFCC) coefficients.

[0045]

[0046]

[0047] in, Indicates the Frame in The output energy of the Mel filter, Indicates the The Mel filter is The response of the frequency components, Indicates the MFCC coefficients of the frame, is the number of cepstral coefficients retained, which is 13 in this example, is the total number of Mel filters, The index of the Mel filter.

[0048] The sound spectrum centroid feature of the original signal is obtained according to the calculation formula of the sound spectrum centroid feature :

[0049] in, Indicates the The spectral centroid feature of the frame, It is The frequency of the frequency component, The spectrum is The amplitude of the frequency component, is the dimension of the spectrum, that is, the total number of frequency components in a frame of signal.

[0050] Finally, the two signal features of MFCC coefficient and spectrum centroid feature are fused to obtain the fusion feature , formally expressed as , used for the subsequent modeling process of latent space disentanglement.

[0051] Step 3: This embodiment uses a variational autoencoder model based on the ResNet18 structure. In the specific network structure, the encoder part contains multiple residual convolution units, each of which includes a convolution layer, batch normalization, and a Leaky ReLU activation function; the decoder part mirrors the encoder structure and uses inverse convolution to achieve signal reconstruction. Specific method of model training: fusion features Input into the encoder, the encoder outputs two vectors, which are categorical variables and noise variables , expressed as:

[0052]

[0053] in, Indicates the The fusion features of the frame, Indicates that the weight parameter is The encoder pair fusion feature The latent space representation after nonlinear mapping, is the noise sampled from the standard normal distribution. Then the final latent variable is generated by fusion through reparameterization techniques :

[0054] in, is the weight adjustment value, which is selected as 0.5 in this example.

[0055] Step 4: Use the decoder to calculate latent variables Perform signal reconstruction to reconstruct the error and KL divergence To optimize the goal, we need to achieve accurate and stable reconstruction of the voiceprint signal, thereby improving the generalization ability of the model. and KL divergence They are defined as:

[0056]

[0057] in, represents the reconstruction error, Represents KL divergence, which is used to measure the posterior distribution Relative to the prior distribution The difference, Indicates the The fusion features of the frame, Indicates the Reconstructed features of the frame, represents the total number of dimensions of the latent space, Represents the latent space The mean of the dimensions, Represents the latent space The standard deviation of each dimension.

[0058] Therefore, the loss function of this link can be expressed as:

[0059] in, Represents the unknown category detection loss function.

[0060] Step 5: This embodiment uses prototype representation to achieve classification diagnosis of fault categories. During the training process, the prototype center (category prototype point) of each known fault category is calculated. A set of training samples, represented by the classification variables of each sample , calculate the center of the class in the latent space , which can be expressed as:

[0061] in, Indicates the fault type The center of the prototype in the latent space, Indicates the fault type The number of samples, .

[0062] This calculation provides a representative point in the latent space for each known class. The prototype center of each fault class is then calculated Euclidean distance and dot product distance The method is to measure the similarity between the input sample and the prototype center of the known fault category. Specifically expressed as:

[0063]

[0064]

[0065] in, represents the categorical variable of the current sample, Represents the classification variable and fault category of the current sample The Euclidean distance between the centers of the prototypes in the latent space, Represents the classification variable and fault category of the current sample The dot product distance between prototype centers in latent space, Represents the classification variable and fault category of the current sample The integrated distance between prototype centers in the latent space, Represents the total number of dimensions of the latent space.

[0066] By using this distance measurement method, the model can effectively identify those samples that may be close to a known fault category in terms of geometric distance. In the prototype network, this embodiment hopes that the feature vector of each sample can be as close as possible to the prototype center of its corresponding fault category, while maintaining a certain distance from the prototype centers of other fault categories. Therefore, the loss function of this link can be expressed as:

[0067] in, represents the known category classification loss function, Indicates the total number of samples in the current training batch, , Indicates the fault category to which the current sample belongs, Represents the comprehensive distance between the classification variable of the current sample and the prototype center of the fault category to which the current sample belongs in the latent space, represents the activation function, Represents a hyperparameter, which is used to control the weight of pushing away the negative class part. It represents the preset distance threshold to encourage the sample to maintain a sufficient distance from other prototype centers. Its value is often determined based on experience.

[0068] Step 6: Combine the two loss functions to get the total loss function .in This is a hyperparameter that controls the model's performance for identifying known and unknown fault categories. Based on the total loss function, the training dataset is input and the model parameters are trained. In this example, the Adam optimizer is used for optimization with a learning rate of 0.001, 200 training rounds, and a batch size of 128.

[0069] Step 7: Set the unknown class threshold (preset threshold). When the reconstruction error value of the sample to be tested after passing the trained model is greater than or equal to the preset threshold, the sample can be identified as an unknown fault category. Otherwise, the distance between the sample to be tested and the center of the prototype of the known category is measured by Euclidean distance and dot product similarity, and the sample to be tested is judged to belong to the known category closest to the prototype center to obtain the voiceprint fault diagnosis result of the motor.

[0070] This method is compared with other open set recognition methods on the test data set described in step 1. The comparison methods include: SoftMAX (Soft Maximum, normalized exponential function), OpenGAN (Open Generative Adversarial Network, open adversarial generative network), CGDL (Conditional Gaussian Distribution Learning, conditional Gaussian distribution learning), ARPL (Adversarial Reciprocal Points Learning, adversarial exchange point learning), PROSER (PlaceholdeRs for Open-SEt Recognition, learning placeholders for open set recognition). Among them, LSDPR is the method described in this case. OACC is the open set recognition rate, CACC is the closed set recognition rate, and AUROC is the area under the ROC curve, which is used to measure the stability of the classification performance of the model under all possible thresholds. The average value of the evaluation data of each method is shown in the following figure. Figure 2 The specific data are shown in Table 1 and Table 1. TaskA, TaskB, TaskC, and TaskD correspond to multiple rotating groups of different unknown categories, namely, the 1.8 kW group in TaskA is the unknown class, the 1 kW group in TaskB is the unknown class, the 1.5 kW group in TaskC is the unknown class, and the 0.4 kW group in TaskD is the unknown class.

[0071] Table 1 Evaluation data of each dataset measured using different methods

[0072] Table 1 Evaluation data of each dataset measured using different methods

[0073] Example 2: This embodiment provides a voiceprint-based open-set diagnosis system for rotating equipment faults, including: a feature extraction module, a decoupling and reconstruction module, a prototype center calculation module, an optimization training module, and a detection module, wherein: The feature extraction module is used to obtain the voiceprint signal of the rotating device; preprocess the voiceprint signal, extract the Mel frequency cepstral coefficient and the sound spectrum centroid feature based on the preprocessed voiceprint signal, and fuse them to obtain a fusion feature; The decoupling and reconstruction module is used to construct a variational autoencoder model, map the fused features to the latent space through the encoder in the variational autoencoder model, decouple them into classification variables and noise variables, and reconstruct the fused features through the decoder in the variational autoencoder model to obtain reconstructed features; The prototype center calculation module is used to calculate the prototype center of each known fault category based on the classification variables; The optimization training module is used to construct an unknown category detection loss function with reconstruction error and KL divergence as optimization targets, construct a known category classification loss function based on the prototype center, and optimize the parameters of the variational autoencoder model based on the unknown category detection loss function and the known category classification loss function to obtain a trained variational autoencoder model; The detection module is used to input the sample to be tested into the trained variational autoencoder model, calculate the reconstruction error value, and compare the reconstruction error value with a preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the sample to be tested is judged to be an unknown fault category. Otherwise, the distance between the sample to be tested and the prototype center is measured to determine whether the sample to be tested belongs to the known fault category corresponding to the nearest prototype center.

[0074] The specific limitations of the voiceprint-based open-set diagnosis system for rotating equipment faults can be found in the above limitations of the voiceprint-based open-set diagnosis method for rotating equipment faults, which will not be repeated here.

[0075] Example 3: This embodiment provides a device for open-set diagnosis of rotating equipment faults based on voiceprints, comprising a processor and a memory storing a plurality of computer instructions. When the computer instructions are executed by the processor, the steps of the open-set diagnosis method for rotating equipment faults based on voiceprints are implemented.

[0076] The specific limitations of the open-set diagnostic device for rotating equipment faults based on voiceprints can be found in the above limitations of the open-set diagnostic method for rotating equipment faults based on voiceprints, which will not be repeated here.

[0077] The memory and processor are electrically connected, directly or indirectly, to enable data transmission or interaction. For example, these components may be electrically connected via one or more communication buses or signal lines. The memory stores a computer program executable on the processor, and the processor implements the method of the present invention by executing the computer program stored in the memory.

[0078] The memory may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The memory is used to store a program, and the processor executes the program after receiving an execution instruction.

[0079] The processor may be an integrated circuit chip with data processing capabilities. The processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor may be a microprocessor or any conventional processor.

[0080] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A method for open set diagnosis of rotating equipment faults based on voiceprint, characterized in that: The open-set diagnosis method for rotating equipment faults based on voiceprints includes: Acquire a voiceprint signal of the rotating equipment, wherein the voiceprint signal includes a fault signal and a non-fault signal, wherein the fault signal includes multiple known fault categories; Preprocess the voiceprint signal, extract the Mel-frequency cepstral coefficient and the sound spectrum centroid features based on the preprocessed voiceprint signal, and fuse them to obtain the fusion feature; Construct a variational autoencoder model, map the fused features to the latent space through the encoder in the variational autoencoder model, decouple them into classification variables and noise variables, and reconstruct the fused features through the decoder in the variational autoencoder model to obtain reconstructed features; Calculate the prototype center for each known fault category based on the categorical variables; An unknown category detection loss function is constructed with reconstruction error and KL divergence as optimization targets, and a known category classification loss function is constructed based on the prototype center. The parameters of the variational autoencoder model are optimized based on the unknown category detection loss function and the known category classification loss function to obtain a trained variational autoencoder model. The sample to be tested is input into the trained variational autoencoder model, the reconstruction error value is calculated, and the reconstruction error value is compared with the preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the sample to be tested is judged to be an unknown fault category. Otherwise, the distance between the sample to be tested and the prototype center is measured to judge that the sample to be tested belongs to the known fault category corresponding to the nearest prototype center.

2. The method for open set diagnosis of rotating equipment faults based on voiceprint according to claim 1 is characterized in that: The preprocessing of the voiceprint signal includes: The pre-emphasis processing of the voiceprint signal is expressed as follows: ,in, Represents the index of the global sampling point, express The emphasized signal at the sampling point, express The original signal at the sampling point, express The original signal at the sampling point, Indicates the pre-emphasis coefficient; Divide the pre-emphasized voiceprint signal into short time frames; Apply the Hamming window function to each frame for windowing.

3. The method for open set diagnosis of rotating equipment faults based on voiceprint according to claim 1, characterized in that: The variational autoencoder model is based on the ResNet18 structure. The encoder contains multiple residual convolution units, each of which includes a convolution layer, a batch normalization layer and a leaky ReLU activation function. The structure of the decoder mirrors that of the encoder.

4. The method for open-set diagnosis of rotating equipment faults based on voiceprint according to claim 1, characterized in that: The categorical variables and noise variables are fused through the reparameterization technique to generate latent variables, which are input into the decoder and expressed as follows: ； ； ； in, represents a categorical variable, represents the noise variable, represents the latent variable, Indicates the total number of frames. Indicates the The fusion features of the frame, Indicates that the weight parameter is The encoder pair fusion feature The latent space representation after nonlinear mapping, is the weight adjustment value, represents noise sampled from a standard normal distribution.

5. The method for open-set diagnosis of rotating equipment faults based on voiceprint according to claim 1, characterized in that: The prototype center of each known fault category is calculated based on the classification variable and is expressed as follows: ； in, Indicates known fault categories The center of the prototype in the latent space, Indicates known fault categories The number of samples, , represents a categorical variable.

6. The method for open set diagnosis of rotating equipment faults based on voiceprint according to claim 1, characterized in that: The unknown category detection loss function is constructed with reconstruction error and KL divergence as optimization targets, and is expressed as follows: ； ； ； in, represents the reconstruction error, represents the KL divergence, represents the unknown category detection loss function, Indicates the total number of frames. Indicates the The fusion features of the frame, Indicates the Reconstructed features of the frame, represents the total number of dimensions of the latent space, Represents the latent space The mean of the dimensions, Represents the latent space The standard deviation of each dimension.

7. The method for open-set diagnosis of rotating equipment faults based on voiceprint according to claim 1, characterized in that: The method of constructing a known category classification loss function based on the prototype center includes: ； ； ； ； in, represents the categorical variable of the current sample, Indicates known fault categories The center of the prototype in the latent space, represents the L1 norm, Represents the classification variable of the current sample and the known fault category The Euclidean distance between the centers of the prototypes in the latent space, Represents the classification variable of the current sample and the known fault category The dot product distance between prototype centers in latent space, Represents the classification variable of the current sample and the known fault category The integrated distance between prototype centers in the latent space, represents the total number of dimensions of the latent space, represents the known category classification loss function, Indicates the total number of samples in the current training batch, , Indicates the fault category to which the current sample belongs, Represents the comprehensive distance between the classification variable of the current sample and the prototype center of the fault category to which the current sample belongs in the latent space, represents the activation function, represents the hyperparameter, Indicates the preset distance threshold.

8. The method for open-set diagnosis of rotating equipment faults based on voiceprint according to claim 1, characterized in that: The optimization of the parameters of the variational autoencoder model based on the unknown category detection loss function and the known category classification loss function includes: The total loss function is obtained based on the unknown category detection loss function and the known category classification loss function. The formula is expressed as: ； in, represents the total loss function, represents the unknown category detection loss function, represents the known category classification loss function, represents a hyperparameter; Optimize the parameters of the variational autoencoder model based on the total loss function.

9. A rotating equipment fault open set diagnosis system based on voiceprint, characterized in that: include: Feature extraction module, decoupling and reconstruction module, prototype center calculation module, optimization training module and detection module, among which: The feature extraction module is used to obtain the voiceprint signal of the rotating device; preprocess the voiceprint signal, extract the Mel frequency cepstral coefficient and the sound spectrum centroid feature based on the preprocessed voiceprint signal, and fuse them to obtain a fusion feature; The decoupling and reconstruction module is used to construct a variational autoencoder model, map the fused features to the latent space through the encoder in the variational autoencoder model, decouple them into classification variables and noise variables, and reconstruct the fused features through the decoder in the variational autoencoder model to obtain reconstructed features; The prototype center calculation module is used to calculate the prototype center of each known fault category based on the classification variables; The optimization training module is used to construct an unknown category detection loss function with reconstruction error and KL divergence as optimization targets, construct a known category classification loss function based on the prototype center, and optimize the parameters of the variational autoencoder model based on the unknown category detection loss function and the known category classification loss function to obtain a trained variational autoencoder model; The detection module is used to input the sample to be tested into the trained variational autoencoder model, calculate the reconstruction error value, and compare the reconstruction error value with a preset threshold. When the reconstruction error value is greater than or equal to the preset threshold, the sample to be tested is judged to be an unknown fault category. Otherwise, the distance between the sample to be tested and the prototype center is measured to determine whether the sample to be tested belongs to the known fault category corresponding to the nearest prototype center.

10. A voiceprint-based open-set diagnostic device for rotating equipment faults, comprising a processor and a memory storing a plurality of computer instructions, characterized in that: When the computer instructions are executed by the processor, the steps of the open-set diagnosis method for rotating equipment faults based on voiceprints described in any one of claims 1 to 8 are implemented.

Images