A method and apparatus for diagnosing rotating machinery faults based on audible sound signals

Abstract

This invention discloses a method and apparatus for fault diagnosis of rotating machinery based on audible sound signals. The method includes: preprocessing the acquired raw audible sound signals of the rotating machinery using a sliding window method to obtain a raw audible sound signal sample set; assigning a corresponding label to each sample in the raw audible sound signal sample set using a multi-label encoding method; performing Mel-spectrum processing on the raw audible sound signal sample set to generate an input dataset; performing sample data augmentation on the input dataset to obtain a balanced dataset; training the balanced dataset and its corresponding labels according to a pre-constructed neural network model to obtain a fault classification model, wherein the neural network model consists of multiple convolutional layers and capsule network layers; and using the fault classification model to identify faults in the audible sound signal data of the rotating machinery to be identified, thereby determining the fault classification result of the rotating machinery to be identified.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for rotating machinery, and more specifically, to a method and apparatus for fault diagnosis of rotating machinery based on audible sound signals. Background Technology

[0002] In the context of smart industry, modern rotating machinery is evolving towards greater complexity, precision, efficiency, and intelligence. Therefore, research on intelligent fault diagnosis (IFD) has become crucial for ensuring the safe, efficient, and reliable operation of such machinery. In recent years, data-driven methods utilizing deep learning have attracted widespread attention for developing effective intelligent monitoring and fault diagnosis models; however, these methods primarily rely on vibration analysis for contact measurements. Compared to traditional vibration-based fault diagnosis methods, acoustic-based intelligent fault diagnosis (AIFD) for rotating machinery, utilizing microphones to capture acoustic signals, offers several distinct advantages. First, AIFD enables non-contact and non-invasive measurements without requiring equipment disassembly. Second, it exhibits strong adaptability to extreme environments, such as high-temperature or corrosive conditions. Third, the signal acquisition process of AIFD is unaffected by the surface condition of the equipment, capturing acoustic signals even on oily, smooth, or geometrically irregular surfaces. Finally, AIFD allows for continuous diagnosis without interrupting machinery operation, as the installation and diagnostic processes do not interfere with normal operation. Furthermore, as a key technological driver of the Fourth Industrial Revolution, AIFD has emerged as a very promising fault diagnosis technology for rotating machinery in recent years. It is important to note that most existing literature on AIFD (Autonomous Asynchronous Failure) focuses primarily on the occurrence of single failures in rotating machinery. However, in real-world industrial environments, the internal structures of rotating machinery are typically complex, with tightly coupled components. This complexity often leads to multiple failures occurring within a single component or failures distributed across multiple components. Research on AIFD addressing such complex failures is relatively limited and requires further exploration.

[0003] Strictly speaking, the challenges of acoustic-based intelligent composite fault diagnosis can be categorized as a small-sample problem. Composite fault samples are typically difficult to obtain and costly, resulting in datasets dominated by normal operation and single fault categories, with a limited number of composite fault samples. This extreme data imbalance causes classification models to ignore a minority of classes, leading to poor generalization performance for these classes. Data augmentation techniques are considered a powerful solution to address the limited sample problem in fault diagnosis. Generative Adversarial Networks (GANs) provide an effective data augmentation method, using adversarial training to learn the underlying data distribution from noise, thereby generating more realistic and representative samples and overcoming the limitations of traditional sampling methods. Current research highlights the extensive exploration of various GAN architectures in mechanical fault diagnosis, particularly addressing the challenges posed by small-sample vibration signal scenarios. However, in the context of AIFD (Autonomous Alternating Current Fault Diagnosis), a comprehensive methodological framework is lacking to address the severe challenges posed by the extreme scarcity of composite fault samples. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and apparatus for diagnosing rotating machinery faults based on audible sound signals.

[0005] According to one aspect of the present invention, a method for diagnosing rotating machinery faults based on audible sound signals is provided, comprising:

[0006] The original audible sound signals of the acquired rotating machinery are preprocessed using a sliding window method to obtain an original audible sound signal sample set, which includes single fault data and compound fault data.

[0007] A multi-label encoding method is used to label each sample in the original audible sound signal sample set.

[0008] The original audible sound signal sample set is processed using Mel spectrograms to generate the input dataset.

[0009] Augment the input dataset to obtain a balanced dataset;

[0010] A fault classification model is obtained by training a balanced dataset and its corresponding labels based on a pre-built neural network model. The neural network model consists of multiple convolutional layers and capsule network layers.

[0011] A fault classification model is used to identify faults in the audible signal data of the rotating machinery to be identified, and the fault classification result of the rotating machinery to be identified is determined.

[0012] Optionally, the original audible sound signal sample set is subjected to Mel spectrogram processing to generate an input dataset, including:

[0013] The original audible sound signal sample set is processed into frames, and a preset window function is added to each frame signal to generate framed audible sound signals.

[0014] Perform Discrete Fourier Transform on the framed audible sound signal to obtain discrete signals;

[0015] Calculate the power spectrum of a discrete signal;

[0016] A set of triangular filters is uniformly distributed across the linear frequency using the Mel filter bank.

[0017] The input dataset is generated by multiplying the power spectrum by the triangular filter and taking the logarithm.

[0018] Optionally, the truncated audible sound signal f m The formula for calculating (n) is:

[0019] f m (n)=f(n+m·h)·ω(n)

[0020] In the formula, m represents the frame index, n represents the frame size, h represents the frame shift, ω(n) represents the window weight of the frame, and f m (n) represents the frame after the window function is added;

[0021] The formula for calculating discrete signals is:

[0022]

[0023] In the formula, k is the index of the frequency point, K is the number of points in the Fourier transform, and f m (k) is the discrete signal of the discrete Fourier transform corresponding to the frequency components of the frame;

[0024] The formula for calculating the power spectrum is:

[0025] P m (k)=|f m (k)| 2

[0026] The Mel frequency range of the triangular filter is:

[0027]

[0028] The scaling relationship between the Mel frequency and the sampling frequency is as follows: f is the frequency of the signal, F mel (f) is the Mel frequency, H i (k) represents the frequency response of the Mel filter, i is the filter index, M is the number of filters, and f(i) is the center frequency of the filter on the Mel frequency scale.

[0029] Optionally, the input dataset is augmented with sample data to obtain a balanced dataset, including:

[0030] Construct orthogonalized latent space vectors that include implicit label information;

[0031] A generative adversarial model is constructed that combines latent space control with local perception mechanism. The generative adversarial model includes a generator and a discriminator, where the input of the generator is the latent space vector.

[0032] A generative adversarial model is used to augment the input dataset to obtain a balanced dataset.

[0033] Optionally, the generator includes five deconvolutional layers, each followed by batch normalization and nonlinearity introduced using the ReLU activation function, and residual blocks are added to the first two layers. The discriminator includes five convolutional layers and introduces the LeakyReLU activation function and batch normalization. The first layer of the discriminator embeds label information, and the output layer uses a local discriminative perceptual mechanism for discrimination.

[0034] Optionally, the loss function of the discriminator is:

[0035]

[0036] In the formula, The interpolated samples are the real data and the generated data, where ε is the sampled data from a uniform distribution. Let D be the gradient of the output, and c be the class label vector. This indicates the ability of the measurement discriminator to identify real data as real, while The ability of a measurement discriminator to identify generated data as fake, λ gp It is the penalty coefficient that controls the weights of the gradient penalty term;

[0037] The loss function of the generator is:

[0038]

[0039] In the formula, φ(·) represents the feature extraction performed by the neural network, x and y are the input real image and the generated image, respectively, and λ p It is a weighting factor used to balance the contributions of adversarial loss and perceived loss.

[0040] Optionally, the capsule network layer uses a dynamic routing algorithm to determine the coefficients for each capsule to pass information to the next layer. The coefficient update formula is as follows:

[0041]

[0042] In the formula, b ijrepresents the initial weight between capsule i and capsule j, and C is the total number of capsules in the next layer;

[0043] The nonlinear compression function of the capsule is:

[0044]

[0045] in, K is the predicted vector from capsule i to capsule j. l Let d be the total number of capsules i. j This represents the total weighted input from all capsules i to j in the previous layer, v j The output of each capsule is used to train a fault classification model on a balanced dataset and its corresponding labels based on a pre-built neural network model, including:

[0046] Optionally, the loss function of the neural network model is:

[0047] L k =T k ·max(0,m + -||v k ||) 2 +λ·(1-T k )·max(0,||v k ||-m - ) 2

[0048] In the formula, T k A binary indicator representing the label; ||v k || represents the length of the output vector of the k-th capsule; m + It is the boundary value of the positive class; m - λ is the boundary value of the negative class; λ is the balance coefficient.

[0049] According to another aspect of the present invention, a rotating machinery fault diagnosis device based on audible sound signals is provided, comprising:

[0050] The preprocessing module is used to preprocess the acquired raw audible sound signals of the rotating machinery using a sliding window method to obtain a raw audible sound signal sample set, which includes single fault data and compound fault data.

[0051] The tag encoding module is used to assign a corresponding tag to each sample in the original audible sound signal sample set using a multi-tag encoding method;

[0052] The signal characterization module is used to perform Mel spectrogram processing on the original audible sound signal sample set to generate the input dataset.

[0053] The data augmentation module is used to augment the input dataset with sample data to obtain a balanced dataset;

[0054] The model training module is used to train a balanced dataset and its corresponding labels based on a pre-built neural network model to obtain a fault classification model. The neural network model consists of multiple convolutional layers and capsule network layers.

[0055] The fault identification module is used to identify faults in the audible signal data of the rotating machinery to be identified using a fault classification model, and to determine the fault classification result of the rotating machinery to be identified.

[0056] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing a computer program for performing the methods described in any of the above aspects of the present invention.

[0057] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the method described in any of the preceding aspects of the present invention.

[0058] Therefore, this invention proposes a novel method for diagnosing complex faults in rotating machinery based on audible sound signals with limited samples. First, the original audible sound signals of the rotating machinery are acquired, and a sample set of original audible sound signals containing a small amount of complex fault data is constructed. Then, Mel spectrogram preprocessing is performed on these signals to obtain a training set, which is used as input for training a generative adversarial model (GAP). An LP-LSGAN GAP is constructed, combining latent space control and local perception mechanisms, to control the generation of Mel spectrograms for complex faults, thereby enhancing the complex fault samples in the training set and obtaining a balanced training set. A CapNet model is then constructed, and the balanced dataset is input into the CapNet model for training. Multi-label decoupling classification of fault types is performed to achieve accurate prediction of fault states. This invention, based on audible sound signals, proposes a novel intelligent fault diagnosis method that effectively solves the problem of extremely scarce complex fault samples by expanding the complex fault types in the training set through the generation of high-resolution Mel spectrograms. Attached Figure Description

[0059] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures:

[0060] Figure 1 This is a schematic flowchart of a rotating machinery fault diagnosis method based on audible sound signals provided in an exemplary embodiment of the present invention;

[0061] Figure 2This is another schematic flowchart of a rotating machinery fault diagnosis method based on audible sound signals provided in an exemplary embodiment of the present invention;

[0062] Figure 3 This is a flowchart of the Mel spectrum acquisition process provided by an exemplary embodiment of the present invention;

[0063] Figure 4 This is a detailed structural diagram of the LP-LSGAN generative adversarial model provided in an exemplary embodiment of the present invention;

[0064] Figure 5 This is a flowchart of fault diagnosis for multi-label classification using the CapNet model provided in an exemplary embodiment of the present invention;

[0065] Figure 6 This is a schematic diagram of the structure of a rotating machinery fault diagnosis device based on audible sound signals provided in an exemplary embodiment of the present invention;

[0066] Figure 7 This is the structure of an electronic device provided in an exemplary embodiment of the present invention. Detailed Implementation

[0067] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0068] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0069] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of the present invention are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0070] It should also be understood that in the embodiments of the present invention, "multiple" can refer to two or more, and "at least one" can refer to one, two or more.

[0071] It should also be understood that any component, data or structure mentioned in the embodiments of the present invention can generally be understood as one or more unless explicitly defined or given contrary instructions in the context.

[0072] Furthermore, the term "and / or" in this invention is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this invention generally indicates that the preceding and following related objects have an "or" relationship.

[0073] It should also be understood that the description of the various embodiments in this invention emphasizes the differences between the various embodiments, and the similarities or similarities can be referred to each other. For the sake of brevity, they will not be described in detail.

[0074] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0075] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0076] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0077] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0078] The embodiments of this invention can be applied to electronic devices such as terminal devices, computer systems, and servers, and can operate together with a wide range of other general-purpose or special-purpose computing system environments or configurations. Well-known examples of terminal devices, computing systems, environments, and / or configurations suitable for use with electronic devices such as terminal devices, computer systems, and servers include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments including any of the above systems, etc.

[0079] Electronic devices such as terminal devices, computer systems, and servers can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in distributed cloud computing environments, where tasks are executed by remote processing devices linked through communication networks. In distributed cloud computing environments, program modules can reside on local or remote computing system storage media, including storage devices.

[0080] Exemplary methods

[0081] Figure 1 This is a schematic flowchart of a rotating machinery fault diagnosis method based on audible sound signals provided in an exemplary embodiment of the present invention. This embodiment can be applied to electronic devices, such as... Figure 1 As shown, the rotating machinery fault diagnosis method 100 based on audible sound signals includes the following steps:

[0082] Step 101: Use a sliding window method to preprocess the acquired raw audible sound signals of the rotating machinery to obtain a raw audible sound signal sample set, which includes single fault data and compound fault data.

[0083] Step 102: Use a multi-label encoding method to assign a corresponding label to each sample in the original audible sound signal sample set;

[0084] Step 103: Perform Mel spectrogram processing on the original audible sound signal sample set to generate the input dataset;

[0085] Step 104: Perform sample data augmentation on the input dataset to obtain a balanced dataset;

[0086] Step 105: Train the balanced dataset and its corresponding labels according to the pre-built neural network model to obtain the fault classification model. The neural network model consists of multiple convolutional layers and capsule network layers.

[0087] Step 106: Use a fault classification model to identify faults in the audible signal data of the rotating machinery to be identified, and determine the fault classification result of the rotating machinery to be identified.

[0088] Specifically, this invention discloses a novel method for diagnosing complex faults in rotating machinery based on audible acoustic signals and limited samples. It includes Mel spectrograms for feature representation, a Latent Spatial Controlled Generative Adversarial Network (LP-LSGAN) with a local perception mechanism for generating high-quality complex fault samples, and a CapNet-based multi-label classification model. This invention addresses the accuracy bias caused by the imbalance of training samples due to the scarcity of complex faults, providing an effective intelligent fault diagnosis scheme.

[0089] The technical solution adopted in this invention is as follows:

[0090] Step 1): Use the envelope surface measurement method to obtain the audible sound signals of multiple preset components of rotating machinery under preset working conditions, organize the audible sound signal datasets of single faults and compound faults, and assign corresponding multi-labels to each obtained sample.

[0091] Step 2): Perform non-overlapping sampling on the above signals and convert them into two-dimensional grayscale images using Mel spectrograms to enhance the representation of audible sound signal fault features. Then, normalize these images and use a large number of single faults and a small number of compound faults as training samples, and a sufficient number of single faults and compound faults as test samples.

[0092] Step 3): Construct an orthogonalized latent space vector containing implicit label information; and develop an LP-LSGAN generative adversarial model that combines latent space control with a local perception mechanism. Input the training samples obtained in Step 2) into the LP-LSGAN for adversarial training, and use the LP-LSGAN to precisely control the output of the composite fault Mel spectrogram.

[0093] Step 4): Use the LP-LSGAN generative adversarial model trained in Step 3) to generate a large number of composite fault samples, expanding the limited training sample set. Use the generated samples to assist in training the CapNet deep learning model for accurate classification and decoupling of composite fault types.

[0094] It should be further explained that in step 1), a multi-label classification method is used to encode the fault types. The multi-label encoding method is similar to binary encoding, where each bit represents the presence (1) or absence (0) of a specific single fault. For example, if fault1 and fault2 exist, their labels are [1,1,0]; if fault1 and fault3 exist, their labels are [1,0,1]. This encoding method effectively captures the relationship between single faults and compound faults through encoding.

[0095] It should be further explained that in step 2), the Mel spectrogram conversion process uses the Mel scale to convert the audible sound signal from the time domain to a more suitable time-frequency domain form for analysis. It exhibits significant sensitivity and rapid response in the low-frequency spectrum, very close to the capabilities of human hearing. Conversely, its sensitivity decreases at higher frequencies, which also aligns with human auditory perception. Given these advantages, the Mel scale can analyze and process acoustic signals more accurately and effectively. The process of converting the original time-domain signal to a Mel spectrogram using the Mel scale is as follows:

[0096] First, the audible sound signal f(c) is divided into overlapping frames, where c is the sampling point and c = 1, 2, ..., N. Then, frame shifting is used to control the degree of overlap between adjacent frames. A window function is added to each frame to obtain the framed audible sound signal f. m (n):

[0097] f m (n)=f(n+m·h)·ω(n)

[0098] In the formula, m represents the frame index, n represents the frame size, h represents the frame shift, ω(n) represents the window weight of the frame, and f m (n) represents the frame after the window function is added.

[0099] Next, the frames with the window function applied are processed by Discrete Fourier Transform (DFT):

[0100]

[0101] In the formula, k is the index of the frequency point, K is the number of points in the Fourier transform, and f m (k) is the result of the DFT corresponding to the frequency components of the frame.

[0102] Then, the power spectrum P is calculated based on the DFT results. m (k)=|f m (k)| 2 .

[0103] The scaling relationship between Mel frequency and sampling frequency is as follows:

[0104]

[0105] In the formula, f is the frequency of the signal, F mel (f) is the Mel frequency.

[0106] Based on the Mel frequency scale, a Mel filter bank is designed to correspond to a specific Mel frequency range. The calculation method is as follows:

[0107]

[0108] In the formula, H i (k) represents the frequency response of the Mel filter, i is the filter index, M is the number of filters, and f(i) is the center frequency of the filter on the Mel frequency scale.

[0109] The Mel spectrum is obtained by multiplying the power spectrum by the Mel filter bank and then taking the logarithm.

[0110]

[0111] In the formula, t represents the time frame and f represents the Mel frequency.

[0112] It should be further explained that in step 3), an orthogonalized latent space vector containing implicit label information needs to be constructed first. To represent different composite fault attributes, a d-dimensional orthogonalized latent space vector, following a Gaussian distribution, is designed. Assuming there are three single fault types, this vector is divided into four segments: the first part represents normal operation of the rotating machinery, ranging from 0 to d / 4; the second part represents the occurrence of single fault type fault1, ranging from d / 4 to d / 2; similarly, d / 2 to 3d / 4 represents the occurrence of fault2; and 3d / 4 to d represents the occurrence of fault3. This division of the vector attribute expression intervals ensures that the latent space expression vectors of the three single faults are completely orthogonal, relatively independent, and do not affect each other. Furthermore, unprecedented bearing composite fault categories can be generated through numerical operations on these intervals.

[0113] Further explanation is needed regarding step 3). An LP-LSGAN generative adversarial model combining latent space control and local perception mechanisms was developed. This model mainly consists of a generator and a discriminator. The generator has five deconvolutional layers, each followed by batch normalization (BatchNorm) to stabilize the training process, while ReLU activation is used to introduce non-linearity. Additionally, residual blocks are introduced in the first two deconvolutional layers. These blocks help preserve and transmit label information to deeper layers of the generator, reducing the risk of information loss or distortion during intermediate convolutional operations. The discriminator incorporates a local perception mechanism, judging local regions (patches) of the image patch-by-pattern to generate higher-resolution images. The discriminator has five convolutional layers and uses the LeakyReLU activation function to mitigate the potential gradient vanishing problem, while also applying BatchNorm to stabilize the training process and accelerate convergence. Furthermore, label information is embedded in the first layer of the discriminator to combine with category information, enabling the discriminator to make more accurate judgments.

[0114] It should be further noted that in step 3), during the training of the LP-LSGAN generative adversarial model, the Wasserstein distance is used as the loss function to iteratively measure the difference between the generated data and the real data, and a gradient penalty term (GP) is introduced to prevent gradient explosion. The final loss formula for the discriminator is calculated as follows:

[0115]

[0116] In the formula, It is an interpolated sample between real data and generated data, where ε is sampled from a uniform distribution. is the gradient of the output of D. Here, c is the class label vector. This indicates the ability of the measurement discriminator to identify real data as real, while The ability of a measurement discriminator to identify generated data as fake. λ gp It is the penalty coefficient that controls the weights of the gradient penalty term.

[0117] In addition to adversarial loss, perceptual loss can further improve the quality of generated images. Based on pre-trained CNNs, such as VGG, perceptual loss captures higher-level visual features by measuring the differences between generated and real images in the feature space. Therefore, the final generator loss formula is calculated as follows:

[0118]

[0119] In the formula, φ(·) represents the feature extraction performed by the VGG network, x and y are the input real image and the generated image, respectively, and λ p It is a weighting factor used to balance the contributions of adversarial loss and perceived loss.

[0120] It should be further noted that in step 4), the samples generated by LP-LSGAN are used to assist in training the CapNet deep learning model for accurate classification and decoupling of complex fault types. This classification model uses a dynamic routing algorithm to determine how much information each capsule should pass to the next layer. Initially, the coupling coefficients (representing the connection strength between one layer of capsules and the next) are randomly initialized. These coefficients are iteratively updated using the softmax function:

[0121]

[0122] In the formula, b ij This represents the initial weight between capsule i and capsule j, and the total number of capsules in the next layer is C.

[0123] The routing algorithm iteratively adjusts these coupling coefficients based on the consistency between the prediction of the current capsule and the output of the next layer. This consistency is achieved through a vector v.j and The scalar product between them is used to measure, where is the output vector of capsule j, and is... The predicted vector from capsule route i to capsule j.

[0124]

[0125] In the formula, d j This represents the total weighted input of capsule j in the current layer.

[0126] To ensure that the capsule output is within a reasonable range, a nonlinear compression function is applied to the capsule output vector.

[0127]

[0128] This function compresses short vectors to near-zero lengths and ensures longer vectors are close to 1, thus helping the network focus on the most relevant features. The final output of the decoupled classifier is converted into a prediction vector y. pred , where each element y i This represents the probability that the input belongs to class i. The classifier uses the L2 norm to refine the output tensor and uses a threshold. To determine the predicted category. If y i If the value exceeds a threshold γ, the sample is classified into that category; otherwise, it is not classified. This process ensures reliable predictions by distinguishing features that are clearly associated with each category.

[0129] It should be further noted that in step 4), the CapNet classification method uses a marginal loss function as the loss function for iterative calculation. The marginal loss function in Capsule Networks can effectively handle multi-label problems. For each label, the formula for the marginal loss function can be expressed as:

[0130] L k =T k ·max(0,m + -||v k ||) 2 +λ·(1-T k )·max(0,||v k ||-m - ) 2

[0131] In the formula, T k A binary indicator representing the label. If the true category is k, then T... k =1, otherwise T k =0. ||v k || represents the length (modulus) of the output vector of the k-th capsule. m +This is the boundary value for the positive class, indicating that correctly classified samples should have a larger output. - λ is the boundary value for the negative class, indicating that misclassified samples should have a smaller output. λ is a balancing coefficient used to control the weight of the negative class loss, and is usually taken as a small value.

[0132] In one embodiment of the present invention, the specific implementation process is as follows:

[0133] Step 1): Acoustic signals from multiple preset components of the rotating machinery under preset operating conditions are acquired using an envelope surface measurement method. The data is preprocessed using a sliding window. Non-overlapping segmentation is applied to the signals to obtain a raw audible signal sample set, including a large amount of single-fault data and a small amount of composite fault data. To illustrate the implementation of this embodiment, a planetary gearbox is used as the rotating machinery experimental object, and three different single fault types (cracks, pitting, and broken teeth) and four composite fault types are machined on the sun gear. The torque of the magnetic powder brake is set to 1 N·m, and the speed of the variable frequency motor is set to 1800 rpm. Then, an acoustic testing system is used to sample at a sampling frequency of 10 kHz to obtain the raw audible signals. By setting the length of the sliding window to 4096, non-overlapping segmentation is performed to obtain a sample set of 3000 samples for each fault type.

[0134] Step 2): Perform Mel spectrogram processing on the original audible sound signal sample set obtained in Step 1), such as... Figure 2 As shown, the process begins by loading faulty audio data from a file, then segmenting the data into frames and applying an audio window to each frame using a window function. Next, a Discrete Fourier Transform is performed on each frame to calculate its graph, followed by squaring the spectrum of each frame to obtain the power spectrum. Then, a set of triangular filters is uniformly distributed along the linear frequency band using a Mel filter bank. Finally, to simulate human sound perception, the energy of the Mel filter bank is logarithmically transformed to obtain the Mel spectrum. In the above embodiment, where the frame size is set to 1024, the frame shift is set to 64, and a Hanning window is used, a 64×64 Mel grayscale image is obtained after processing according to this Mel spectrum processing procedure. This image is used as the final input dataset for training the adversarial model.

[0135] In step 2), the 3000 Mel image samples obtained for each category are further divided into training and test sets. To evaluate the effectiveness of data-augmented fault diagnosis when the number of composite fault samples is very limited, only 5 Mel images are added to the training set for each composite fault type. 1500 Mel images are added to the training set for each single fault type and normal operation class. The test set includes 1500 samples for each of the normal operation, single fault, and composite fault types.

[0136] Step 3): An LP-LSGAN generative model combining latent space control and local perception mechanisms is used to control the generation of composite fault Mel spectrograms, thereby enhancing the composite fault samples in the training set. The generator and discriminator are the two core modules of LP-LSGAN, such as... Figure 3 The structure of these two core modules is described step by step according to this embodiment.

[0137] In step 3), the input to the generator module is a 256-dimensional orthogonalized random noise vector, representing the latent space vector. The design of the latent space vector involves first initializing a 256-dimensional random noise vector following a Gaussian distribution, then dividing this vector into four parts: the first part represents normal operation of the rotating machinery, ranging from 0 to 63; the second part represents the occurrence of a single fault type, fault1, ranging from 64 to 127; similarly, 128 to 191 represent the occurrence of fault2; and 192 to 255 represent the occurrence of fault3. If a fault type exists, the value of that part is retained; otherwise, the value of that part is set to zero. This allows the latent space vector to be obtained based on the fault type labels in the training set. For example, the multi-label type [1,1,0] represents a composite fault type where fault1 and fault2 coexist, and its corresponding latent space vector consists of random noise bits 0 to 191, with bits 192 to 256 being 0.

[0138] In step 3), the orthogonalized latent space vector is used as the input to the generator. Each layer of the generator consists of a convolutional transpose layer (ConvTranspose2d), batch normalization (BatchNorm), and an activation function (ReLU). This design progressively upsamples the latent space vector, converting it into the target Mel grayscale image obtained in step 2). Furthermore, residual blocks are added to the first two layers. These residual blocks contain two Conv2d operations with a kernel size of 3×3 and a stride of 1×1. Each convolutional operation is followed by a BatchNorm and ReLU activation function. The residual blocks preserve input information through skip connections and perform deep feature extraction. The specific details of the generator are shown in the table below:

[0139]

[0140] In step 3), the input image and labels are concatenated and used as input to the discriminator. The discriminator consists of multiple convolutional layers (Conv2d), with BatchNorm and LeakyReLU activation functions added after most convolutional layers to improve network stability and non-linear expressive power. This design extracts local features from the image and labels layer by layer and ultimately outputs a result representing the true / false judgment. Furthermore, the discriminator employs a local discrimination (Patch) approach, using a 1×1 convolutional layer in the output layer to obtain the discrimination results for local image regions, further enhancing the model's discriminative ability. Specific details of the discriminator are shown in the table below:

[0141]

[0142]

[0143] In step 3), the Wasserstein distance is used as the loss function to measure the difference between generated and real data, and a gradient penalty (GP) term is introduced to prevent gradient explosion. Furthermore, the generator's loss function consists of adversarial loss and perceptual loss. The perceptual loss calculates the differences between images in a higher-level feature space, implemented using a pre-trained model (VGG16). Where λ... p Set to 0.5. Finally, λ is the loss of the discriminator. gp Set it to 10.

[0144] In step 3), the training set obtained in step 2) is used as input to train the constructed LP-LSGAN generative adversarial model. The Adam optimizer is used, with a batch size of 512 for each experiment; the learning rates for both the generative model and the model are set to 0.0001, and training is performed for multiple epochs until Nash equilibrium is reached. Finally, the trained LP-LSGAN is used to precisely control the output of the composite fault Mel spectrogram, which is then added to the original training set as an augmentation dataset to balance the training data.

[0145] Step 4): Use the balanced training dataset obtained in Step 3) as input to the CapNet deep learning model, such as... Figure 4 The diagram shows the architectural details of the CapNet deep learning model. The model is trained and decoupled for multi-label classification of fault types on the test set to obtain the final fault diagnosis result. In this embodiment, the learning rate of the CapNet model is set to 0.0001 during training, and a step decay method is used to adjust the learning rate for 100 epochs.

[0146] In step 4), the CapNet model consists of multiple convolutional layers and capsule network layers, combining spatial information extraction with the feature processing capabilities of capsule layers, and further optimizing the classification results through a dynamic routing mechanism. First, the input image passes through a convolutional layer (Conv2d), which receives a single-channel input and outputs a feature map with 256 channels. The kernel size is 5×5, the stride is 2, and the padding is 1. A non-linear transformation is introduced through the ReLU activation function. Next is the PrimaryCaps layer, which outputs an 8×32 feature map (i.e., 256 capsules, 32 channels, each capsule containing 8 features) through a convolutional operation (Conv2d). The kernel size is 9×9, and the stride is 2. This layer transforms the convolutional feature map into capsule representations, allowing the capture of more complex patterns and spatial information. The DigitCaps layer consists of three linear layers, with the linear transformation in each capsule layer responsible for generating a 16-dimensional capsule feature vector. In this way, the model can capture different category features of the input image. The dynamic routing mechanism iteratively executes between capsules, ensuring that high-confidence outputs are passed to the target capsule layer. In this embodiment, since it is assumed that there are three single-fault types, the final capsule output vector, after flattening, yields three independent feature vectors. Each feature vector represents a classification label; therefore, this model is designed as a multi-label classification problem, where the model simultaneously predicts the class probabilities of three labels. A threshold of 0.5 is set here: when the predicted probability value is greater than 0.5, the label position is 1; when it is less than 0.5, the label position is 0.

[0147] In step 4), the CapNet model constructed above is used to train the balanced training dataset. During training, a marginal loss function is used to measure the difference between the model's output features and the true labels. The marginal loss function can effectively handle multi-label problems in capsule networks. In this embodiment, the boundary value of the positive class in the marginal loss function is set to 0.9, the boundary value of the negative class is set to 0.1, and the weight controlling the loss of the negative class is set to 0.25.

[0148] In summary, the novel method framework for finite-sample composite fault diagnosis of rotating machinery based on audible sound signals of this invention mainly consists of three parts: Mel spectrum processing, LP-LSGAN generative adversarial network generation of Mel spectrum corresponding to composite faults, and CapNet model for multi-label decoupling classification. By focusing on the representation of audible sound signals, balancing the dataset, and multi-label decoupling, corresponding processing methods are designed to improve the diagnostic accuracy of composite fault diagnosis based on finite samples of audible sound signals.

[0149] Compared with the prior art, the present invention has the following beneficial effects:

[0150] This invention introduces speech signal processing technology and uses Mel spectrogram mapping, which contains rich time-frequency information, as a quantitative representation of audible sound signals to effectively capture and characterize the features of different fault types.

[0151] This invention addresses the data imbalance problem caused by the limited number of composite fault samples by proposing an LP-LSGAN generative adversarial model with latent spatial control and a local awareness mechanism. This model can generate high-quality image data that preserves both local details and global features, effectively mitigating the limitations of dataset imbalance.

[0152] To decouple compound faults, this invention introduces the CapNet multi-label decoupling classifier. This classifier outputs multiple labels for samples representing compound faults, thereby overcoming the limitations of existing compound fault studies that typically treat single and compound faults as independent fault modes, neglecting the potential relationship between them, and enabling more detailed diagnosis.

[0153] Exemplary device

[0154] Figure 6 This is a schematic diagram of the structure of a rotating machinery fault diagnosis device based on audible sound signals provided in an exemplary embodiment of the present invention. Figure 6 As shown, the device 600 includes:

[0155] The preprocessing module 610 is used to preprocess the acquired raw audible sound signals of the rotating machinery using a sliding window method to obtain a raw audible sound signal sample set, wherein the raw audible sound signal sample set includes single fault data and compound fault data.

[0156] The tag encoding module 620 is used to assign a corresponding tag to each sample in the original audible sound signal sample set using a multi-tag encoding method;

[0157] The signal characterization module 630 is used to perform Mel spectrogram processing on the original audible sound signal sample set to generate the input dataset.

[0158] Data augmentation module 640 is used to augment the input dataset with sample data to obtain a balanced dataset;

[0159] The model training module 650 is used to train a balanced dataset and its corresponding labels based on a pre-built neural network model to obtain a fault classification model. The neural network model consists of multiple convolutional layers and capsule network layers.

[0160] The fault identification module 660 is used to identify faults in the audible signal data of the rotating machinery to be identified using a fault classification model, and to determine the fault classification result of the rotating machinery to be identified.

[0161] Optionally, the signal characterization module 630 includes:

[0162] The framing submodule is used to process the original audible sound signal sample set into frames and add a preset window function to each frame signal to generate framed audible sound signals.

[0163] The transformation submodule is used to perform discrete Fourier transform processing on the framed audible sound signal to obtain discrete signals;

[0164] The computation submodule is used to calculate the power spectrum of discrete signals;

[0165] The distribution submodule is used to uniformly distribute a set of triangular filters across the linear frequency using the Mel filter bank.

[0166] The generation submodule is used to multiply the power spectrum with the triangular filter and take the logarithm to generate the input dataset.

[0167] Optionally, the truncated audible sound signal f m The formula for calculating (n) is:

[0168] f m (n)=f(n+m·h)·ω(n)

[0169] In the formula, m represents the frame index, n represents the frame size, h represents the frame shift, ω(n) represents the window weight of the frame, and f m (n) represents the frame after the window function is added;

[0170] The formula for calculating discrete signals is:

[0171]

[0172] In the formula, k is the index of the frequency point, K is the number of points in the Fourier transform, and f m (k) is the discrete signal of the discrete Fourier transform corresponding to the frequency components of the frame;

[0173] The formula for calculating the power spectrum is:

[0174] P m (k)=|f m (k)| 2

[0175] The Mel frequency range of the triangular filter is:

[0176]

[0177] The scaling relationship between the Mel frequency and the sampling frequency is as follows: f is the frequency of the signal, F mel (f) is the Mel frequency, H i(k) represents the frequency response of the Mel filter, i is the filter index, M is the number of filters, and f(i) is the center frequency of the filter on the Mel frequency scale.

[0178] Optionally, the data augmentation module 640 includes:

[0179] The first construction submodule is used to construct orthogonalized latent space vectors containing implicit label information;

[0180] The second construction submodule is used to construct a generative adversarial model that combines latent space control with local perception mechanisms. The generative adversarial model includes a generator and a discriminator, where the input of the generator is a latent space vector.

[0181] The data augmentation submodule is used to augment the input dataset with sample data using a generative adversarial model to obtain a balanced dataset.

[0182] Optionally, the generator includes five deconvolutional layers, each followed by batch normalization and nonlinearity introduced using the ReLU activation function, and residual blocks are added to the first two layers. The discriminator includes five convolutional layers and introduces the LeakyReLU activation function and batch normalization. The first layer of the discriminator embeds label information, and the output layer uses a local discriminative perceptual mechanism for discrimination.

[0183] Optionally, the loss function of the discriminator is:

[0184]

[0185] In the formula, The interpolated samples are the real data and the generated data, where ε is the sampled data from a uniform distribution. Let D be the gradient of the output, and c be the class label vector. This indicates the ability of the measurement discriminator to identify real data as real, while The ability of a measurement discriminator to identify generated data as fake, λ gp It is the penalty coefficient that controls the weights of the gradient penalty term;

[0186] The loss function of the generator is:

[0187]

[0188] In the formula, φ(·) represents the feature extraction performed by the neural network, x and y are the input real image and the generated image, respectively, and λ p It is a weighting factor used to balance the contributions of adversarial loss and perceived loss.

[0189] Optionally, the capsule network layer uses a dynamic routing algorithm to determine the coefficients for each capsule to pass information to the next layer. The coefficient update formula is as follows:

[0190]

[0191] In the formula, b ij represents the initial weight between capsule i and capsule j, and C is the total number of capsules in the next layer;

[0192] The nonlinear compression function of the capsule is:

[0193]

[0194] in, K is the predicted vector from capsule i to capsule j. l Let d be the total number of capsules i. j This represents the total weighted input from all capsules i to j in the previous layer, v j The output of each capsule is used to train a fault classification model on a balanced dataset and its corresponding labels based on a pre-built neural network model, including:

[0195] Optionally, the loss function of the neural network model is:

[0196] L k =T k ·max(0,m + -||v k ||) 2 +λ·(1-T k )·max(0,||v k ||-m - ) 2

[0197] In the formula, T k A binary indicator representing the label; ||v k || represents the length of the output vector of the k-th capsule; m + It is the boundary value of the positive class; m - λ is the boundary value of the negative class; λ is the balance coefficient.

[0198] Exemplary electronic devices

[0199] Figure 7 This is the structure of an electronic device provided in an exemplary embodiment of the present invention. For example... Figure 7 As shown, the electronic device 70 includes one or more processors 71 and a memory 72.

[0200] The processor 71 may be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions.

[0201] The memory 72 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 71 may execute the program instructions to implement the methods of the software programs of the various embodiments of the present invention described above, and / or other desired functions. In one example, the electronic device may also include an input device 73 and an output device 74, these components being interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0202] In addition, the input device 73 may also include, for example, a keyboard, a mouse, etc.

[0203] The output device 74 can output various information to the outside. The output device 74 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0204] Of course, for the sake of simplicity, Figure 7 Only some of the components of this electronic device relevant to the present invention are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0205] Exemplary computer program products and computer-readable storage media

[0206] In addition to the methods and apparatus described above, embodiments of the present invention may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of the present invention described in the "Exemplary Methods" section above.

[0207] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of the present invention. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0208] Furthermore, embodiments of the present invention may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps of the methods according to various embodiments of the present invention described in the "Exemplary Methods" section above.

[0209] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0210] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the aforementioned specific details.

[0211] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0212] The block diagrams of devices, systems, devices, and systems involved in this invention are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, systems, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0213] The methods and systems of the present invention may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of the present invention are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, the present invention may also be implemented as a program recorded on a recording medium, the program comprising machine-readable instructions for implementing the methods according to the present invention. Thus, the present invention also covers recording media storing programs for performing the methods according to the present invention.

[0214] It should also be noted that in the systems, apparatus, and methods of the present invention, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered equivalents of the present invention. The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the invention. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the invention. Therefore, the invention is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0215] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the invention to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for diagnosing rotating machinery faults based on audible sound signals, characterized in that, include: The original audible sound signals of the acquired rotating machinery are preprocessed using a sliding window method to obtain an original audible sound signal sample set, wherein the original audible sound signal sample set includes single fault data and compound fault data. Each sample in the original audible sound signal sample set is labeled with a corresponding tag using a multi-tag encoding method. The original audible sound signal sample set is processed using Mel spectrograms to generate the input dataset; The input dataset is augmented with sample data to obtain a balanced dataset; A fault classification model is obtained by training the balanced dataset and its corresponding labels according to a pre-built neural network model, wherein the neural network model consists of multiple convolutional layers and capsule network layers. The fault classification model is used to identify faults in the audible sound signal data of the rotating mechanical equipment to be identified, and the fault classification result of the rotating mechanical equipment to be identified is determined. The original audible sound signal sample set is subjected to Mel spectrogram processing to generate an input dataset, including: The original audible sound signal sample set is processed into frames, and a preset window function is added to each frame signal to generate framed audible sound signals. Perform Discrete Fourier Transform on the framed audible sound signal to obtain discrete signals; Calculate the power spectrum of the discrete signal; A set of triangular filters is uniformly distributed across the linear frequency using the Mel filter bank. The power spectrum is multiplied by the triangular filter and the logarithm is taken to generate the input dataset; The framed audible sound signal f m The formula for calculating (n) is: f m (n)=f(n+m·h)·ω(n) In the formula, m represents the frame index, n represents the frame size, h represents the frame shift, ω(n) represents the window weight of the frame, and f m (n) represents the frame after the window function is added; The formula for calculating the discrete signal is: In the formula, k is the index of the frequency point, K is the number of points in the Fourier transform, and f m (k) is the discrete signal of the discrete Fourier transform corresponding to the frequency components of the frame; The formula for calculating the power spectrum is: P m (k)=|f m (k)| 2 The Mel frequency range of the triangular filter is: The scaling relationship between the Mel frequency and the sampling frequency is as follows: f is the frequency of the signal, f mel (f) is the Mel frequency, H i (k) represents the frequency response of the Mel filter, i is the filter index, M is the number of filters, and f(i) is the center frequency of the filter on the Mel frequency scale. The input dataset is augmented with sample data to obtain a balanced dataset, including: Construct orthogonalized latent space vectors that include implicit label information; A generative adversarial model is constructed that combines latent space control with local perception mechanism. The generative adversarial model includes a generator and a discriminator, wherein the input of the generator is the latent space vector. The input dataset is augmented using the generative adversarial model to obtain the balanced dataset.

2. The method according to claim 1, characterized in that, The generator comprises five deconvolutional layers, each followed by batch normalization and nonlinearity introduced using the ReLU activation function. Residual blocks are added to the first two layers. The discriminator comprises five convolutional layers and introduces the LeakyReLU activation function and batch normalization. The first layer of the discriminator embeds label information, and the output layer uses a local discriminative perceptual mechanism for discrimination.

3. The method according to claim 1, characterized in that, The loss function of the discriminator is: In the formula, The interpolated samples are the real data and the generated data, where ε is the sampled data from a uniform distribution. Let D be the gradient of the output, and c be the class label vector. This indicates the ability of the measurement discriminator to identify real data as real, while The ability of a measurement discriminator to identify generated data as fake, λ gp It is the penalty coefficient that controls the weights of the gradient penalty term; The loss function of the generator is: In the formula, φ(·) represents the feature extraction performed by the neural network, x and y are the input real image and the generated image, respectively, and λ p It is a weighting factor used to balance the contributions of adversarial loss and perceived loss.

4. The method according to claim 1, characterized in that, The capsule network layer uses a dynamic routing algorithm to determine the coefficients for each capsule to pass information to the next layer. The coefficient update formula is as follows: In the formula, b ij represents the initial weight between capsule i and capsule j, and C is the total number of capsules in the next layer; The nonlinear compression function of the capsule is: in, K is the predicted vector from capsule i to capsule j. l d represents the total number of capsules. j This represents the total weighted input from all capsules i to j in the previous layer, v j It is the output of each capsule.

5. The method according to claim 1, characterized in that, The loss function of the neural network model is: L k =T k ·max(0,m + -||v k ||) 2 +λ·(1-T k )·max(0,||v k ||-m - ) 2 In the formula, T k A binary indicator representing the label; ||v k || represents the length of the output vector of the k-th capsule; m + It is the boundary value of the positive class; m - λ is the boundary value of the negative class; λ is the balance coefficient.

6. A rotating machinery fault diagnosis device based on audible sound signals, the device being used to implement the method described in any one of claims 1-5, characterized in that, include: The preprocessing module is used to preprocess the acquired raw audible sound signals of the rotating machinery using a sliding window method to obtain a raw audible sound signal sample set, wherein the raw audible sound signal sample set includes single fault data and compound fault data. The tag encoding module is used to assign a corresponding tag to each sample in the original audible sound signal sample set using a multi-tag encoding method; The signal characterization module is used to perform Mel spectrogram processing on the original audible sound signal sample set to generate an input dataset; The data augmentation module is used to augment the input dataset with sample data to obtain a balanced dataset; The model training module is used to train the balanced dataset and its corresponding labels based on the pre-built neural network model to obtain a fault classification model. The neural network model consists of multiple convolutional layers and capsule network layers. The fault identification module is used to identify faults in the audible sound signal data of the rotating machinery to be identified using the fault classification model, and to determine the fault classification result of the rotating machinery to be identified.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program for performing the method described in any one of claims 1-5.

Images