Diffusion model based device failure data augmentation and diagnosis method and system
By generating high-fidelity and diverse pseudo-samples through a diffusion model, the problem of scarce and imbalanced equipment fault data is solved, improving the accuracy and robustness of the fault diagnosis model. It is particularly suitable for fault diagnosis of motor vibration signals.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGSHA UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-02
AI Technical Summary
In industrial settings, equipment failure data samples are scarce and unevenly distributed, making it difficult for existing diagnostic models to accurately identify rare faults. Traditional methods, such as GANs, generate samples of poor quality, which cannot meet the needs for accurate early warning.
A data augmentation method based on a diffusion model is adopted. Pseudo-samples are generated through a conditional denoising diffusion model. Using a U-Net network and a noise-aware classifier, combined with the regional importance weight matrix of the bispectral map, high-fidelity and diverse synthetic fault data are generated.
In situations where fault samples are scarce, a large number of high-quality pseudo samples are effectively generated, which improves the accuracy and generalization ability of the fault diagnosis model, solves the problems of data scarcity and imbalanced distribution, and improves the accuracy and robustness of the diagnosis model.
Smart Images
Figure CN122135146A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of electrical equipment fault diagnosis and artificial intelligence, and relates to a method and system for enhancing and diagnosing equipment fault data based on a diffusion model. Background Technology
[0002] With the rapid development of the Industrial Internet and intelligent manufacturing, data-driven intelligent fault diagnosis technology has become a core means to ensure the safe and stable operation of critical equipment. Deep learning-based diagnostic models can automatically learn features from massive amounts of monitoring data to achieve high-precision condition identification. However, the performance of such methods heavily relies on a large amount of high-quality, class-balanced training data.
[0003] In real-world industrial scenarios, equipment is in a healthy state most of the time, with extremely low probability of specific faults occurring, often accompanied by high data collection costs. This results in scarce fault data samples available for training, and an imbalance in fault samples between categories. Directly training a diagnostic model on such an imbalanced dataset will cause the model to be severely biased towards the majority class, while the recognition rate for the minority class will be low, making it difficult to meet the urgent need for accurate early warning of rare faults in practical engineering.
[0004] To alleviate the data imbalance problem, traditional methods mainly focus on the algorithmic and data levels. Algorithmic approaches primarily involve designing cost-sensitive loss functions, but their effectiveness is limited. Data augmentation techniques are mainly employed at the data level. Traditional signal augmentation methods, such as adding noise, time shifting, and scaling, can only produce limited diversity. Although Synthetic Minority Oversampling (SMOTE) and its variants are widely used, for high-dimensional, nonlinear, and non-stationary mechanical vibration signals, the generated samples often suffer from feature space distortion. In recent years, deep generative models such as Generative Adversarial Networks (GANs) have been introduced into the field of fault diagnosis to generate synthetic fault data. However, GANs suffer from pattern collapse, training instability, and difficulty in convergence during training, often resulting in insufficient diversity or poor quality of generated samples. Therefore, the problem of data imbalance remains to be solved. Summary of the Invention
[0005] This invention aims to address the core problems of scarce fault samples and imbalanced data distribution. To this end, it provides a method and system for equipment fault data enhancement and diagnosis based on a diffusion model. This method focuses on the integration of data enhancement and diagnosis in equipment fault diagnosis. By generating data through a diffusion model, the training set is significantly expanded, thereby improving the accuracy and generalization ability of subsequent diagnostic models. The application scenarios are more focused on predictive maintenance and intelligent diagnosis. Specifically, the technical solution of this invention introduces a conditional denoising diffusion model, using time steps and fault category labels as conditional information to accurately guide the denoising direction. Simultaneously, by adding a noise-aware classifier, its gradient is used to impose conditional constraints on the pre-trained diffusion model, further guiding the diffusion model to generate the required category images. In particular, the technical solution is further optimized by using the symmetry of the bispectral image to introduce a regional importance weight matrix during the forward diffusion denoising process, enabling the model to focus on important fault feature regions and suppress redundant information regions. In the field of sample amplification technology, this invention provides a novel technical approach to solve the problem of data imbalance. It can efficiently generate a large amount of synthetic fault data with high fidelity and good diversity when fault samples are scarce, thus solving the problem of scarce and unbalanced fault data in industrial scenarios. Ultimately, in practical applications, it enables efficient and accurate fault diagnosis in scenarios with scarce and imbalanced data.
[0006] Therefore, the present invention provides the following technical solution:
[0007] On the one hand, the present invention provides a device fault data augmentation method based on a diffusion model, comprising the following steps:
[0008] Step 1: Collect and map the equipment fault timing signals under different conditions to generate equipment fault feature maps for training samples;
[0009] Step 2: Construct a conditional denoising diffusion model, which includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information consists of time steps and fault category labels.
[0010] The training process of the conditional denoising diffusion model is divided into forward diffusion denoising and reverse denoising. The equipment fault feature map of the training sample is generated into a feature map carrying noise through forward diffusion denoising. Based on the feature map, the U-Net network and the noise perception classifier perform reverse denoising to obtain the predicted noise and fault category probability distribution.
[0011] Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples for the specified fault category to enhance and expand the sample data;
[0012] At any given time step, the gradient of the fault category probability distribution of the noise-aware classifier corrects the current predicted noise output by the U-Net network. Then, the corrected predicted noise and the feature map of the current time step are used to calculate the feature map of the next time step. Through iterative calculation, the final output feature map is regarded as a pseudo-sample of the specified fault category.
[0013] Optionally, the U-Net network and the noise perception classifier share an encoder or use an encoder with the same structure. If the encoder is shared, the encoder output is divided into two branches: one branch is input to the decoder of the U-Net network, and the other branch is input to the classification head of the noise perception classifier.
[0014] The encoder consists of several downsampling layers connected in sequence. Each downsampling layer has a residual block and a downsampling convolution in sequence. The residual block has the following in sequence: a first convolutional layer, a first adaptive group normalization layer (AdaGN), an activation function layer, a second convolutional layer, and a second adaptive group normalization layer (AdaGN).
[0015] The input to the adaptive group normalization layer AdaGN, in addition to the output of the previous convolutional layer, also incorporates the conditional information, specifically:
[0016] Generate modulation vector parameters, that is, first pass the time step through a linear coding layer. and fault category labels Encoding is performed to obtain vectors of the same dimension, which are then superimposed. Finally, a fully connected layer is used to generate two modulation vectors, i.e., the scaling parameters. and bias parameters :
[0017] Feature modulation involves first normalizing the output of the previous convolutional layer, and then applying scaling parameters. and bias parameters The output after group normalization is scaled and translated.
[0018] Optionally, the U-Net network further includes a decoder and a bottleneck layer; wherein, the decoder consists of several upsampling layers connected in sequence, and each upsampling layer includes upsampling convolution, skip connections, and residual blocks; the bottleneck layer is the bottom layer of the U-Net network, consisting of several residual blocks, used to connect the encoder and the decoder;
[0019] The residual blocks in the decoder and the bottleneck layer have the same structure as the residual blocks in the encoder.
[0020] Optionally, the equipment fault feature map is a bispectral map. Correspondingly, during the forward diffusion noise addition process, a regional importance weight matrix is constructed based on the frequency resolution and feature point positions of the bispectral map. This matrix is used for anisotropic noise addition through the regional importance weights, and the regional importance weight matrix is a diagonal covariance matrix. diagonal elements satisfy:
[0021]
[0022] In the formula, the coordinates of the feature points in the bispectral image are: The regional importance weighting function is:
[0023]
[0024] in, Indicates the location of feature points in the bispectral image The corresponding regional importance weights, where i,j are the row and column markers in the bispectral plot. For the frequency resolution of the bispectral, It is an inhibitory factor;
[0025] Based on the region importance weight matrix, the iterative model for forward diffusion with noise addition is expressed as:
[0026] ,in, All of these are intermediate values from the forward diffusion process, obtained by forward diffusion and noise addition to the original bispectral image, corresponding to the feature maps obtained after the t-th and t-1-th forward diffusion and noise addition. Represents a normal distribution; It is random noise; The coefficients are retained for single-step signals.
[0027] Optionally, during the training of the conditional denoising diffusion model, the U-Net network uses the feature map to predict noise under the guidance of conditional information; the noise-aware classifier uses the feature map to obtain the fault classification prediction result under the guidance of conditional information; and then in step 2, the model is trained based on the predicted noise and the noise added by the forward diffusion noise addition, as well as the contrast loss between the fault classification prediction result and the real fault category.
[0028] Specifically, a diffusion loss incorporating the diagonal covariance matrix is proposed based on a comparison between the predicted noise and the noise added during forward diffusion. , is represented as:
[0029] ;
[0030] In the formula, For expectation operators; The noise is randomly sampled Gaussian noise; The U-Net network is based on parameters Predictive noise; This is the initial feature map, i.e., the device fault feature map under one fault category; y represents the fault category. Represents the diagonal covariance matrix -1 / 2.
[0031] It also proposes a classification loss based on the comparison between fault classification prediction results and actual fault categories. , is represented as:
[0032] ;
[0033] In the formula, For expectation operators; For noise perception classifier based Obtain the probability distribution of fault category y. This is the feature map for time step t.
[0034] The loss function described in this invention incorporates conditional information, making the generation of pseudo-samples more directional and controllable.
[0035] Optionally, the process of generating pseudo-samples in step 3 is as follows:
[0036] Step 3.1: Sample pure noise from pure Gaussian noise And then proceed to the calculations at each time step in sequence; among them, such as , Representing a standard normal distribution; or introducing a diagonal covariance matrix, pure noise. , Represents a normal distribution;
[0037] Step 3.2: For each time step The feature map of time step t Time step and target fault category Input into the U-Net network to predict noise ;
[0038] Step 3.3: Calculate the gradient of the noise-aware classifier; transfer the feature map... Each time step and target fault category The probability distribution is obtained by inputting it into a noise-sensing classifier. For feature maps gradient Represented as: ;
[0039] Step 3.4: Utilizing gradients Correcting the noise in the prediction: ,in, It guides the scale and controls the strength of the gradient's influence. For the predicted noise, The corrected noise. Preservation coefficients for single-step signals The average value, The single-step signal retention coefficient corresponding to time step t satisfy: , These are noise scheduling parameters;
[0040] Step 3.5: Use the corrected noise Calculate the feature map at the next time step :
[0041]
[0042] Step 3.6: Repeat steps 3.2-3.5 sequentially to obtain the final feature map. , i.e., target fault category Feature map;
[0043] In this process, the number of repetitions and the required fault categories are set, and steps 3.1-3.6 above are repeated to obtain the feature maps of each fault category.
[0044] Optionally, the equipment fault timing signals collected in step 1 under different states are: motor vibration timing signals under fault and healthy states;
[0045] Then, the vibration time-series signals under various motor conditions are divided into equal intervals to obtain several samples. The bispectrum of each sample is then calculated and a bispectral map is generated. The bispectral map serves as a fault feature map of the equipment.
[0046] The bispectral pattern is at two frequencies and Using the amplitude as the coordinate axis The thermal values were obtained.
[0047] It should be understood that, in terms of feature extraction, traditional analysis methods based on the time or frequency domain of vibration signals cannot effectively capture the nonlinearity and phase coupling phenomena caused by faults. Higher-order spectral analysis, especially bispectral analysis, can preserve the phase information of the signal and more effectively reveal the essential characteristics of the fault, providing a superior feature representation for fault diagnosis. This invention combines this powerful feature representation with an advanced generative model through the above technical solution to achieve more accurate and controllable fault data enhancement.
[0048] Secondly, the present invention provides a device fault diagnosis method based on a diffusion model, comprising:
[0049] Based on the above diffusion model-based device fault data augmentation method, augmented pseudo-samples are obtained.
[0050] Enhanced pseudo-samples are added to the training set to train a neural network-based fault diagnosis classifier.
[0051] Among them, the trained fault diagnosis classifier is used for equipment fault category diagnosis.
[0052] In three aspects, the system provided by the present invention, based on the above-mentioned diffusion-based device fault data enhancement method, includes at least:
[0053] The signal acquisition and mapping module is used to acquire and map equipment fault timing signals under different states, and generate equipment fault feature maps of training samples.
[0054] The conditional denoising diffusion model construction and training module is used to construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step t and the fault category label y.
[0055] The sample enhancement module is used to generate pseudo samples of a specified fault category by the conditional denoising diffusion model, thereby enhancing and expanding the sample data.
[0056] In four aspects, the present invention provides a computer device comprising at least: one or more processors and a memory storing one or more computer programs;
[0057] The processor invokes a computer program to implement: the method of the above-mentioned device fault data enhancement method or the steps of the above-mentioned fault diagnosis method.
[0058] In five aspects, the present invention provides a computer-readable storage medium storing a computer program, which is invoked by a processor to implement:
[0059] The method of the above-mentioned equipment fault data enhancement method or the steps of the above-mentioned fault diagnosis method.
[0060] Compared with the prior art, the present invention achieves the following progress and beneficial effects:
[0061] 1. The technical solution of this invention utilizes a conditional denoising diffusion model for data augmentation, enabling the efficient generation of a large amount of high-fidelity and diverse synthetic fault data even when fault samples are scarce. This solves the problem of data scarcity and imbalanced distribution in industrial scenarios. Specifically, through regional importance weighting, the model focuses on the fault-sensitive main triangular region during generation, while appropriately ignoring redundant symmetrical regions. The generated pseudo-samples have their key fault features enhanced, while noise and redundant features are suppressed.
[0062] 2. The technical solution of the present invention further targets the motor vibration signal by selecting bispectral analysis to preprocess the original vibration signal, effectively extracting the nonlinear coupling features that can characterize the essence of the fault, thereby constructing a time-frequency spectrum with purer information and more prominent features.
[0063] 3. The technical solution of this invention uses an enhanced balanced dataset for training, which significantly improves the accuracy and robustness of the final intelligent diagnostic model. Attached Figure Description
[0064] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0065] Figure 1 This is a flowchart illustrating a device fault data enhancement and diagnosis method based on a diffusion model in one embodiment;
[0066] Figure 2 Here is a diagram of the residual block structure in one embodiment;
[0067] Figure 3 This is an improved U-Net structure in one embodiment;
[0068] Figure 4 This is a noise-sensing classifier structure in one embodiment;
[0069] Figure 5 This is a flowchart of data augmentation in one embodiment. Detailed Implementation
[0070] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other. It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0071] To address the problem of data imbalance in the field, particularly in motor fault diagnosis where data imbalance significantly limits the accuracy of diagnosis, this invention proposes a data augmentation and diagnosis method for equipment faults based on a diffusion model. The core of this invention is a novel approach to data augmentation: utilizing a limited existing dataset, a conditional denoising diffusion model is designed and trained. This trained model then generates pseudo-samples for any specified fault category, thereby expanding the dataset.
[0072] This invention employs a novel conditional denoising diffusion model, which performs multi-level deep optimization based on forward diffusion noise addition and U-Net network noise prediction. This includes: proposing conditional information containing time steps and fault category labels, and incorporating it into the residual blocks of the U-Net network and the noise-aware classifier. This guides the AdaGN layer in the residual blocks to dynamically adjust the distribution and scale of features, thereby precisely guiding the denoising direction. Utilizing the symmetry of bispectral maps, regional importance weighting is applied during forward diffusion noise addition to ensure the model focuses on key feature regions and suppresses redundant feature regions. Furthermore, the model training approach is optimized by using a comparative loss based on predicted noise, noise added during forward diffusion noise addition, fault classification prediction results, and the actual fault category. This allows for direct use of the pre-trained U-Net network and noise-aware classifier during the sample expansion stage to obtain the feature map of the target fault category from pure noise. Furthermore, the technical solution of this invention is aimed at the field of motor fault diagnosis. Based on the characteristics of motor vibration signals, it proposes to use bispectral graphs as equipment fault feature maps, effectively capturing the nonlinearity and phase coupling phenomena caused by motor faults. The bispectral graphs fully express the fault characteristics of motor vibration signals, and combined with the symmetry of the bispectral graphs, the forward diffusion noise addition process is further optimized.
[0073] The technical solution of the present invention will be described in detail below with reference to specific embodiments. The embodiments below will take motors as the object and motor fault diagnosis as the application object for illustration. However, it should be understood that, according to the technical concept of the present invention, the present invention is also applicable to the fault diagnosis of other devices.
[0074] Example 1
[0075] An embodiment of the present invention provides a device fault data augmentation method based on a diffusion model, comprising the following steps:
[0076] Step 1: Collect equipment fault timing signals under different conditions, then construct a graph of the equipment fault timing signals to generate equipment fault feature maps for training samples.
[0077] This embodiment addresses motor malfunctions:
[0078] S1.1: Acquire the original vibration timing signals of the motor under different conditions using vibration sensors. The motor status includes the healthy status and various fault statuses under different operating conditions. Operating conditions include, but are not limited to, speed and load. Fault categories include, but are not limited to, rotor bar breakage faults, bow-shaped faults, bearing faults, rotor misalignment faults, mass imbalance faults, single-phase open circuit faults, and winding short circuit faults.
[0079] S1.2: The vibration signal sequences under various motor conditions are divided into equal intervals to obtain several samples. The bispectral density of each sample is then calculated, and a bispectral graph is generated. Each segmented time sequence is considered a sample, and a bispectral graph for that sample is constructed. Subsequent sample expansion creates pseudo-samples, which are also bispectral graphs of a sequence segment of the same length. It should be understood that using the bispectral graph as a fault feature map is a preferred technique in this embodiment, but not the only feasible technique.
[0080] In some embodiments, the specific process of step S1.2 is as follows:
[0081] Step 1.2.1: Divide the vibration signal of the motor in a healthy state into equal parts. Each serving contains Points; divide the motor vibration signals of various fault states into equal parts. ( ) portions, each containing There are 1 point, where C represents the number of fault categories. The number of segments corresponding to different fault categories.
[0082] Step 1.2.2: Calculate the bispectral density of each motor vibration signal.
[0083] Here, the bispectrum is the two-dimensional Fourier transform of the third-order cumulant, defined as:
[0084]
[0085] in, It is a double spectrum. It is a third-order cumulant. All are time-lag variables. These are two independent frequency variables, whose values must conform to the Nyquist frequency requirement. Third-order cumulants. Represented as: , The expectation operator is a commonly used notation in the field of statistics. This is the original vibration signal.
[0086] Generally, bispectral It can be intuitively understood as a measure of the second-order nonlinear phase coupling in a signal, and its direct estimation formula is expressed as:
[0087]
[0088] in, Represents Fourier transform, denoted as complex conjugate, f is the frequency. It can be represented as: This indicates that the bispectral measure is frequency. , and its frequency and Phase relationship between them:
[0089]
[0090] In the formula, This represents the phase angle of the bispectral spectrum. , , Both are phase angles, representing the Fourier components respectively. , , The phase.
[0091] If there is a second phase coupling, i.e. Then double spectrum amplitude A peak will occur.
[0092] Step 1.2.3: Construct a bispectral graph based on the calculated bispectral data. The bispectral data... It is a two-dimensional complex array. We are usually more interested in its magnitude. It reflects the different frequencies The strength of phase coupling between them. This invention uses two frequencies. and Using the amplitude as the coordinate axis A two-dimensional thermogram was plotted as a bispectral graph to represent the thermal values.
[0093] It should be understood that significant peaks in the bispectral graphs represent core diagnostic information. If at a point... The appearance of a peak at a certain point indicates the frequency at that point. and There is a second phase coupling.
[0094] Step 1.2.4: Repeat steps 1.2.2 and 1.2.3 to obtain bispectral images of all samples in all categories. Place all bispectral images and their corresponding labels into the training set for later use. In this embodiment, the label is the motor state, i.e., the fault category, which includes different fault types and health status.
[0095] This invention focuses on how to extract effective features from images, and is applicable to situations where fault data is scarce and categories are severely imbalanced. Therefore, it chooses to directly calculate the bispectrum based on the original signal.
[0096] Step 2: Construct and train a conditional denoising diffusion model. The model includes a U-Net network with conditional information (time step and fault category label). The model training process is divided into forward diffusion denoising and backward denoising. Specifically, the equipment fault feature map is first forward diffused to generate a noisy feature map, which is then input into the U-Net network of the conditional denoising diffusion model. Guided by the conditional information, the U-Net network uses the feature map to predict noise. The noise-sensing classifier, also guided by the conditional information, uses the feature map to obtain the fault classification prediction result. The model is then trained based on the predicted noise, the noise added during forward diffusion denoising, and the contrast loss between the fault classification prediction result and the actual fault category.
[0097] The forward diffusion noise addition process is also iterated at set time steps. Whether a conventional iterative process is used, or the importance weight matrix is defined based on the symmetry of the bispectral graph in this embodiment, both are applicable to the technical concept of this invention and fall within the protection scope of this invention.
[0098] In this embodiment, the coordinates of the feature points in the bispectral image are set as follows: The regional importance weighting function is defined as follows:
[0099]
[0100] in, Indicates position The corresponding regional importance weights, where i,j are the row and column markers in the bispectral plot. For the frequency resolution of the bispectral, The suppression factor can be optimized using a grid search method. This regional importance function utilizes the symmetry of the bispectral image to divide it into primary and redundant regions, thereby performing anisotropic noise addition based on the importance of different regions. The diagonal elements of the diagonal covariance matrix are defined. for:
[0101]
[0102] The iterative process at each time step of the forward process is defined as follows: ,in, For probability distribution, These are the intermediate values of the forward diffusion process, obtained by forward diffusion and noise addition to the original bispectral image; Indicates a normal distribution. The mean, This is the variance, i.e., the noise scheduling parameter. In this embodiment, it is... Replaced , This is the identity matrix. In some embodiments, the random sampling time step is... , The total step size, To ensure uniform distribution, this embodiment employs linear scheduling, i.e., intermediate... Calculated through linear interpolation. For example... , ,middle Linear interpolation is obtained. Therefore, we have:
[0103]
[0104] in, It is random noise, and its distribution conforms to The difference between this embodiment and the standard diffusion model is that the noise matrix here no longer conforms to a standard normal distribution. The diagonal elements represent the region importance weights, thus achieving anisotropy. Standard-intensity noise is added to the primary region to fully learn the complex fault characteristics of that region, while it is attenuated in redundant regions. Parameters are defined. Then we have:
[0105]
[0106] By iteratively solving the above formula, we can obtain:
[0107]
[0108] Among them, the average value .
[0109] In summary, the embodiments of the present invention obtain a noisy image through a forward diffusion noise addition process. Feature maps, also known as feature maps, aim to make the image as pleasing as possible. This becomes Gaussian noise. Therefore, besides the above implementation, other existing techniques can obtain a noisy image from the original bispectral image through iterative operations at each time step. Existing technologies that ultimately tend towards Gaussian noise can all be applied to the technical solutions of this invention and fall within the protection scope of this invention.
[0110] Regarding the inverse denoising process, this process aims to obtain... The reverse process This process cannot be directly calculated. The technical solution of this invention uses a neural network to fit this distribution, that is, to use a neural network with parameters to calculate... To address this, the constructed conditional denoising diffusion model comprises an improved U-Net network and a noise-aware classifier. The improved U-Net network is used to predict noise, while the noise-aware classifier is used to predict the probability of fault categories, guiding the generation of images of the corresponding categories.
[0111] It should be understood that the core improvement of the conditional denoising diffusion model provided by this invention is the introduction of a region importance weight matrix. During the forward diffusion denoising process, the symmetry of the bispectral image is used to weight different regions based on their importance, and importance weights are introduced into the loss, making the model focus more on important feature regions. Secondly, conditional information is also introduced, which improves the U-Net network.
[0112] like Figure 3 As shown, the improved U-Net in this embodiment includes an encoder, a bottleneck layer, and a decoder. The core improvement of the U-Net network lies in more effectively injecting conditional information (time steps) into the encoder. and fault category Category labels Enhance the ability to express features.
[0113] The encoder consists of several downsampling layers connected sequentially, each downsampling layer containing a residual block and a downsampling convolution; the decoder consists of several upsampling layers connected sequentially, each upsampling layer containing an upsampling convolution, skip connections, and a residual block. For example... Figure 3 In one embodiment shown, the encoder has four downsampling layers and the decoder has four upsampling layers, and thus the encoder and decoder each have a corresponding residual block 1, residual block 2, residual block 3 and residual block 4.
[0114] like Figure 2 As shown, the residual block contains, in sequence: a first convolutional layer, a first adaptive group normalization layer (AdaGN), an activation function layer, a second convolutional layer, and a second adaptive group normalization layer (AdaGN). One of the core optimizations of the U-Net network is the introduction of time steps in both adaptive group normalization layers (AdaGN). and fault category .
[0115] The convolutional layer within the residual block performs basic feature extraction. It aggregates local pixel information and extracts features such as edges and textures by using a sliding window mechanism with the convolutional kernel. This embodiment uses a single 3x3 convolutional layer; in other feasible embodiments, the size of the convolutional layer can be adaptively adjusted according to the application, environment, and accuracy requirements.
[0116] In addition to the output h of the previous convolutional layer, the input of the Adaptive Group Normalization (AdaGN) layer also incorporates conditional information, namely the time step. and fault category (All are integers), thus controlling the entire generation process. The AdaGN layer mainly consists of two steps:
[0117] Generate modulation vector parameters, that is, first pass the time step through a linear coding layer. and fault category labels Encoding is performed to obtain vectors of the same dimension, which are then superimposed. Finally, a fully connected layer is used to generate two modulation vectors, i.e., the scaling parameters. and bias parameters ,like ,in, It is a scaling parameter. It is a bias parameter. This represents a linear coding layer that converts discrete... and Encoding is done as a vector of the same dimension. One possible method is sinusoidal positional encoding. Other feasible encoding methods also fall within the protection scope of this invention. This indicates a fully connected layer.
[0118] Feature modulation involves first normalizing the output h of the previous convolutional layer, and then applying scaling parameters. and bias parameters The normalized output h is scaled and translated, as follows: , The result of characteristic modulation, Normalize the group.
[0119] This invention introduces conditional information into the AdaGN layer, in which time steps are included. Controlling the generation phase enables the model to adaptively adjust its behavior based on the generation phase, including fault category labels. The AdaGN layer controls the generated content, enabling directional and controllable generation of pseudo-samples of a specified category. Furthermore, since time step t is the index of the forward diffusion noise addition process, a larger t indicates that the noise addition process occurs later. The AdaGN layer dynamically adjusts the distribution and scale of features, thereby precisely guiding the denoising direction.
[0120] The activation function layer in the residual block is used to introduce nonlinear transformations, enabling the network to learn and simulate more complex functional relationships, greatly enhancing the model's expressive power. In some embodiments, the SiLU function is used, which can be expressed as: , This is the result for the activation function layer. It should be understood that this invention is not limited to the SiLU activation function.
[0121] In the residual block, the output of the activation function layer After passing through the same convolutional layer and AdaGN layer again, we get The processing of features, the second convolutional layer and the second adaptive group normalization layer AdaGN are as described above, so they will not be explained in detail here.
[0122] It should be understood that the encoder's downsampling convolution and the decoder's upsampling convolution are both existing in the field, and therefore will not be described in detail.
[0123] The bottleneck layer is a series of layers at the bottom of the network architecture used to connect the encoder and decoder, performing deep integration and modulation of global information at the lowest resolution. In a preferred embodiment of the invention, the bottleneck layer is composed of several identical residual blocks stacked together. These residual blocks have the same structure as the residual blocks in the encoder, and therefore will not be described further. Figure 3 In one embodiment shown, the bottleneck layer has two residual blocks.
[0124] The decoder includes a series of upsampling layers, each upsampling layer comprising: upsampling convolution, skip connections, and residual blocks. Preferably, in this embodiment of the invention, the residual block structure of the decoder is identical to that of the residual block structure in the encoder. Other structures of the encoder can refer to existing technologies.
[0125] In some embodiments, the decoder's upsampling convolution gradually restores the spatial resolution of the image through transposed convolution.
[0126] Skip connections are the core idea of the U-Net architecture. They fuse feature maps of the same resolution from the encoder path with upsampled feature maps to provide details lost during encoding, thus helping the decoder reconstruct the image more accurately. Specifically, this involves fusing skip connection features from the encoder... Features after upsampling Then, the parts are assembled.
[0127] Encoder features contain more local details, while upsampled features contain more global semantic information. Fusing the two allows for a complementary balance of detail and semantics. In summary, after multiple upsampling layers in the decoder, the final output is predicted noise. The goal of the U-Net network is to accurately predict the noise added during the forward diffusion noise step.
[0128] Therefore, in order to achieve this objective, this embodiment designs a diffusion loss. In other feasible embodiments, if a diagonal covariance matrix is not introduced, it can also be defined as: .
[0129] Regarding the noise-aware classifier, this invention proposes to impose conditional constraints on the pre-trained diffusion model through the gradient of the noise-aware classifier, thereby further guiding the diffusion model to generate the required category images.
[0130] The noise-aware classifier is a functionally independent architecture used to predict the class of an image, but it is structurally coupled with an improved U-Net, such as... Figure 4 As shown, the preferred noise perception classifier in this embodiment includes: an encoder, an attention pooling layer, and a linear layer. The attention pooling layer and the linear layer constitute the classification head. It should be noted that in some embodiments, such as... Figure 5 As shown, the U-Net network and the noise-aware classifier share an encoder. The encoder output is divided into two branches: one branch is input to the U-Net network's decoder, and the other branch is input to the noise-aware classifier's classification head. In other embodiments, the U-Net network and the noise-aware classifier use the same structure but independent encoders.
[0131] In this embodiment, a shared encoder output feature map is defined. The output of the attention pooling layer of the noise-aware classifier Represented as:
[0132]
[0133] in, , and For learnable weights, is the scaling factor, and T is the matrix transpose symbol.
[0134] The linear layer of the noise-aware classifier consists of two fully connected layers, and the final output is the probability distribution of the predicted image. .
[0135] It should be understood that the technical solution of the present invention limits the noise perception classifier to the function of predicting the fault category, that is, to obtain the fault category classification result, but does not restrict the network architecture of the above classification head to be the only feasible way.
[0136] Based on the classification structure of the noise-sensing classifier, a classification loss is set. .
[0137] During training, this invention preferably first trains an improved U-Net network, with the optimization objective being... Then, fix the improved U-Net architecture, train the classification head again, and optimize the objective as follows: .
[0138] Therefore, step 2 involves training a conditional denoising diffusion model, which includes an improved U-Net network and a noise-aware classifier. .
[0139] Step 3: Utilize the trained conditional denoising diffusion model to generate pseudo-samples for the specified fault category, thereby augmenting the sample data. This aims to transform pure noise into pseudo-samples. To begin, generate an image belonging to a specific fault category. Its core method lies in using the gradient provided by the noise-aware classifier to guide the denoising process.
[0140] In this embodiment, step 3 is implemented as follows:
[0141] Step 3.1: Sample the initial state from pure Gaussian noise Pure noise Then, the calculations for each time step are performed sequentially.
[0142] Step 3.2: For each time step The feature map of the current time step Time step and target fault category The input is fed into the improved U-Net network to obtain the predicted noise. .
[0143] Step 3.3: Calculate the gradient of the noise-aware classifier. (This involves processing the noisy image.) Each time step and target fault category The probability distribution is obtained by inputting it into a noise-sensing classifier. ,for The gradient can be expressed as: .
[0144] Step 3.4: Use gradient g to guide the correction of noisy predictions. The gradient calculated in Step 3.3 is used to correct the noise in the U-Net predictions. The noise obtained is guided and corrected. ,in, It is a guiding scale that controls the intensity of the gradient's influence.
[0145] Step 3.5: Perform the denoising step. Use the corrected noise prediction. To calculate the feature map of the next time step The calculation formula is: .
[0146] Step 3.6: Repeat steps 3.2-3.5 sequentially to obtain the final feature map. That is, the generated fault category conforms to the target fault category. The high-quality image, in this embodiment, is a bispectral image.
[0147] Step 3.7: If steps 3.1-3.6 are repeated P times, P images will be obtained; after the P images are included in the training set, the sample size of each category can be kept consistent with the sample size of the healthy state.
[0148] It should be understood that this embodiment constructs a bispectral image for the vibration time-series signal of the motor, and the high-quality image obtained by sample augmentation is also a bispectral image. In other feasible embodiments, if other types of feature maps are selected, the high-quality image obtained by sample augmentation will be a feature map of the corresponding type.
[0149] Example 2:
[0150] This embodiment of the invention implements a fault diagnosis method based on Embodiment 1. Therefore, the equipment fault diagnosis method based on a diffusion model provided in this embodiment includes the following steps:
[0151] Step 1: Collect equipment fault time-series signals under different conditions, then construct a graph of the equipment fault time-series signals to generate equipment fault feature maps for training samples;
[0152] Step 2: Construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step t and the fault category label y.
[0153] Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples for the specified fault category to enhance and expand the sample data.
[0154] Step 4: Add the augmented pseudo-samples from Step 3 to the training set to train a neural network-based fault diagnosis classifier.
[0155] Step 5: Use the trained fault diagnosis classifier to diagnose the equipment fault category.
[0156] In some embodiments, ResNet18 is selected as the classification framework in step 5, and the device fault feature map is used as the model input. It should be understood that ResNet18 is only an example, and in other feasible embodiments, other existing classification networks can be selected according to application requirements and accuracy. This invention does not impose specific limitations on this.
[0157] Similarly, such as Figure 1 The illustrated process uses motor fault diagnosis as an example. It should be understood that this fault diagnosis method, in addition to its application in motor fault diagnosis, can also be applied to the fault diagnosis of other types of equipment, depending on its effectiveness.
[0158] Example 3
[0159] Based on Embodiment 1, this invention provides a system based on a device fault data augmentation method, which includes at least a signal acquisition and mapping module, a conditional denoising diffusion model construction and training module, and a sample augmentation module that are connected sequentially or interconnected.
[0160] Among them, the signal acquisition and mapping module is used to acquire equipment fault time-series signals under different states, and then to map the equipment fault time-series signals to generate equipment fault feature maps of training samples.
[0161] The conditional denoising diffusion model construction and training module is used to construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step t and the fault category label y.
[0162] The sample augmentation module is used to generate pseudo samples of a specified fault category using the trained conditional denoising diffusion model, thereby enhancing and expanding the sample data.
[0163] In other embodiments, the system also implements fault diagnosis, that is, it also includes a training module and a diagnosis module. The training module is used to add augmented pseudo-samples to the training set, thereby training a neural network-based fault diagnosis classifier; the diagnosis module is used to use the trained fault diagnosis classifier to diagnose the equipment fault category.
[0164] The specific application process of each module shall be implemented in accordance with the foregoing description, and the combination and splitting of modules are also applicable to this invention. Furthermore, the merging and integration, disassembly and separation of modules can also be achieved through other methods. It should be understood that the specific implementation method described is only a preferred embodiment of this invention. Any modifications or substitutions made within the scope of knowledge possessed by those skilled in the art, without departing from the spirit and scope of this invention, are also included within the protection scope of this invention.
[0165] Example 4
[0166] This invention provides a computer device comprising at least one or more processors and a memory storing one or more computer programs; wherein the processor invokes the computer programs to implement the method of the above-described device fault data enhancement method or the steps of the above-described fault diagnosis method.
[0167] For specific implementation:
[0168] Step 1: Collect equipment fault timing signals under different conditions, then construct a graph of the equipment fault timing signals to generate equipment fault feature maps for training samples.
[0169] Step 2: Construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step and the fault category label.
[0170] Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples of the specified fault category to enhance and expand the sample data.
[0171] Or execute:
[0172] Step 1: Collect equipment fault timing signals under different conditions, then construct a graph of the equipment fault timing signals to generate equipment fault feature maps for training samples.
[0173] Step 2: Construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step and the fault category label.
[0174] Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples of the specified fault category to enhance and expand the sample data.
[0175] Step 4: Add the augmented pseudo-samples from Step 3 to the training set to train a neural network-based fault diagnosis classifier.
[0176] Step 5: Use the trained fault diagnosis classifier to diagnose the equipment fault category.
[0177] Please refer to the explanation of the method above for the specific implementation process of each step.
[0178] The controller can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The controller is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.
[0179] The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory, and the controller calls and executes the algorithm program of the data enhancement or diagnostic method of the embodiments of this invention.
[0180] Example 5
[0181] This invention provides a computer-readable storage medium storing a computer program that is invoked by a processor to implement the method of the above-described device fault data enhancement method or the steps of the above-described fault diagnosis method.
[0182] For specific implementation:
[0183] Step 1: Collect equipment fault timing signals under different conditions, then construct a graph of the equipment fault timing signals to generate equipment fault feature maps for training samples.
[0184] Step 2: Construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step and the fault category label.
[0185] Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples of the specified fault category to enhance and expand the sample data.
[0186] Or execute:
[0187] Step 1: Collect equipment fault timing signals under different conditions, then construct a graph of the equipment fault timing signals to generate equipment fault feature maps for training samples.
[0188] Step 2: Construct a conditional denoising diffusion model and train the model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step and the fault category label.
[0189] Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples of the specified fault category to enhance and expand the sample data.
[0190] Step 4: Add the augmented pseudo-samples from Step 3 to the training set to train a neural network-based fault diagnosis classifier.
[0191] Step 5: Use the trained fault diagnosis classifier to diagnose the equipment fault category.
[0192] Please refer to the explanation of the method above for the specific implementation process of each step.
[0193] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the foregoing embodiments, such as the controller's hard drive or memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both the controller's internal storage unit and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.
[0194] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0195] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.
Claims
1. A method for enhancing equipment fault data based on a diffusion model, characterized in that: Includes the following steps: Step 1: Collect and map the equipment fault timing signals under different conditions to generate equipment fault feature maps for training samples; Step 2: Construct a conditional denoising diffusion model, which includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information consists of time steps and fault category labels. The training process of the conditional denoising diffusion model is divided into forward diffusion denoising and reverse denoising. The equipment fault feature map of the training sample is generated into a feature map carrying noise through forward diffusion denoising. Based on the feature map, the U-Net network and the noise perception classifier perform reverse denoising to obtain the predicted noise and fault category probability distribution. Step 3: Use the trained conditional denoising diffusion model to generate pseudo-samples for the specified fault category to enhance and expand the sample data; At any given time step, the gradient of the fault category probability distribution of the noise-aware classifier corrects the current predicted noise output by the U-Net network. Then, the corrected predicted noise and the feature map of the current time step are used to calculate the feature map of the next time step. Through iterative calculation, the final output feature map is regarded as a pseudo-sample of the specified fault category.
2. The equipment fault data augmentation method based on the diffusion model according to claim 1, characterized in that: The U-Net network and the noise perception classifier share an encoder or use an encoder with the same structure. If they share an encoder, the encoder output is divided into two branches: one branch is input to the decoder of the U-Net network, and the other branch is input to the classification head of the noise perception classifier. The encoder consists of several downsampling layers connected in sequence. Each downsampling layer has a residual block and a downsampling convolution in sequence. The residual block has the following in sequence: a first convolutional layer, a first adaptive group normalization layer (AdaGN), an activation function layer, a second convolutional layer, and a second adaptive group normalization layer (AdaGN). In addition to the output of the previous convolutional layer, the input to the adaptive group normalization layer AdaGN also incorporates the aforementioned conditional information, specifically: Generate modulation vector parameters, that is, first pass the time step through a linear coding layer. and fault category labels Encoding is performed to obtain vectors of the same dimension, which are then superimposed. Finally, a fully connected layer is used to generate two modulation vectors, i.e., the scaling parameters. and bias parameters : Feature modulation involves first normalizing the output of the previous convolutional layer, and then applying scaling parameters. and bias parameters The output after group normalization is scaled and translated.
3. The device fault data augmentation method based on the diffusion model according to claim 1, characterized in that: The equipment fault feature map is a bispectral map. Correspondingly, during the forward diffusion noise addition process, a regional importance weight matrix is constructed based on the frequency resolution and feature point positions of the bispectral map. This matrix is used for anisotropic noise addition through regional importance weights. The regional importance weight matrix is a diagonal covariance matrix. diagonal elements satisfy: ； In the formula, the coordinates of the feature points in the bispectral image are: The regional importance weighting function is: ； in, Indicates the location of feature points in the bispectral image The corresponding regional importance weights, where i,j are the row and column markers in the bispectral graph. For the frequency resolution of the bispectral, It is an inhibitory factor; Based on the region importance weight matrix, the iterative model for forward diffusion with noise addition is expressed as: ,in, All of these are intermediate values from the forward diffusion process, obtained by forward diffusion and noise addition to the original bispectral image, corresponding to the feature maps obtained after the t-th and t-1-th forward diffusion and noise addition. Represents a normal distribution; It is random noise; The coefficients are retained for single-step signals.
4. The device fault data augmentation method based on the diffusion model according to claim 3, characterized in that: During the training of the conditional denoising diffusion model, the U-Net network uses the feature map to predict noise under the guidance of conditional information; the noise-aware classifier uses the feature map to obtain the fault classification prediction result under the guidance of conditional information; and then in step 2, the model is trained based on the predicted noise and the noise added by the forward diffusion noise addition, as well as the contrast loss between the fault classification prediction result and the real fault category. Specifically, a diffusion loss incorporating the diagonal covariance matrix is proposed based on a comparison between the predicted noise and the noise added during forward diffusion. , is represented as: ； In the formula, For expectation operators; The noise is randomly sampled Gaussian noise; For the U-Net network based on parameters Predictive noise; This is the initial feature map, i.e., the device fault feature map under one fault category; y represents the fault category. Represents the diagonal covariance matrix -1 / 2.
5. The equipment fault data augmentation method based on the diffusion model according to claim 1, characterized in that: The process of generating pseudo-samples in step 3 is as follows: Step 3.1: Sample pure noise from pure Gaussian noise And then proceed to the calculations at each time step in sequence; Step 3.2: For each time step The feature map of time step t Time step and target fault category Input into the U-Net network to predict noise ; Step 3.3: Calculate the gradient of the noise-aware classifier; transfer the feature map... Each time step and target fault category The probability distribution is obtained by inputting it into a noise-sensing classifier. For feature maps gradient Represented as: ; Step 3.4: Utilizing gradients Correcting the noise in the prediction: ,in, It guides the scale and controls the strength of the gradient's influence. For the predicted noise, The corrected noise. Preservation coefficients for single-step signals The average value, The single-step signal retention coefficient corresponding to time step t satisfy: , These are noise scheduling parameters; Step 3.5: Use the corrected noise Calculate the feature map at the next time step : Step 3.6: Repeat steps 3.2-3.5 sequentially to obtain the final feature map. , i.e., target fault category Feature map; In this process, the number of repetitions and the required fault categories are set, and steps 3.1-3.6 above are repeated to obtain the feature maps of each fault category.
6. The device fault data augmentation method based on the diffusion model according to claim 1, characterized in that: The equipment fault timing signals collected in step 1 under different states are: motor vibration timing signals under fault state and healthy state; Then, the vibration time-series signals under various motor conditions are divided into equal intervals to obtain several samples. The bispectrum of each sample is then calculated and a bispectral map is generated. The bispectral map serves as a fault feature map of the equipment. The bispectral pattern is at two frequencies and Using the amplitude as the coordinate axis The thermal values were obtained.
7. A device fault diagnosis method based on a diffusion model, characterized in that: include: The enhanced pseudo-samples are obtained based on the device fault data enhancement method based on the diffusion model according to any one of claims 1-6; Enhanced pseudo-samples are added to the training set to train a neural network-based fault diagnosis classifier. Among them, the trained fault diagnosis classifier is used for equipment fault category diagnosis.
8. A system based on the device fault data augmentation method based on the diffusion model as described in any one of claims 1-6, characterized in that: At least including: The signal acquisition and mapping module is used to acquire and map equipment fault timing signals under different conditions, and generate equipment fault feature maps of training samples. The conditional denoising diffusion model construction and training module is used to construct and train the conditional denoising diffusion model. The conditional denoising diffusion model includes a U-Net network that introduces conditional information and a noise perception classifier. The conditional information is the time step t and the fault category label y. The sample enhancement module is used to generate pseudo samples of a specified fault category by the conditional denoising diffusion model, thereby enhancing and expanding the sample data.
9. A computer device, characterized in that: At least includes: One or more processors; A memory that stores one or more computer programs; The processor invokes a computer program to achieve the following: The steps of the device fault data enhancement method based on the diffusion model according to any one of claims 1-6 or the device fault diagnosis method based on the diffusion model according to claim 7.
10. A computer-readable storage medium, characterized in that: The computer program is stored and is invoked by the processor to implement: The steps of the device fault data enhancement method based on the diffusion model according to any one of claims 1-6 or the device fault diagnosis method based on the diffusion model according to claim 7.