A method for information-poor fault prediction based on multi-domain feature fusion and interpretable adversarial convolutional coding networks

Abstract

This invention relates to the field of intelligent fault diagnosis technology for rotating machinery, and in particular to a method for fault prediction with limited information based on multi-domain feature fusion and an interpretable adversarial convolutional coding network. The method includes: acquiring vibration signals of the target rotating machinery; inputting the vibration signals into a preset fault prediction model; and outputting fault categories, wherein the fault categories include known fault categories and new fault categories. The fault prediction model is trained based on a training set, which includes fault samples of several rotating machinery under various operating conditions. The fault prediction model is constructed by combining multi-domain feature fusion and an interpretable adversarial convolutional coding network. This invention focuses on new fault types that may occur in rotating machinery under complex industrial environments, improves the reliability of fault diagnosis under small sample data conditions, and comprehensively coordinates and optimizes the fault diagnosis process and model performance.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for rotating machinery, and in particular to a method for predicting information-poor faults based on multi-domain feature fusion and interpretable adversarial convolutional coding networks. Background Technology

[0002] With the rapid development of modern industrial technology, the automation, intelligence, and precision levels of mechanical equipment are constantly improving, and the structure of rotating machinery is becoming increasingly complex, with significantly increased operating speeds and load capacities. However, during long-term operation, the components of rotating machinery are highly susceptible to damage from factors such as friction, impact, and fatigue, resulting in wear, cracks, and deformation. If this damage is not detected and addressed in a timely manner, it may lead to serious equipment failures or even safety accidents. Therefore, to ensure the stable and efficient operation of industrial equipment, accurately identify potential faults, and take maintenance measures, higher demands are placed on intelligent fault diagnosis technology and health status monitoring of mechanical systems.

[0003] Traditional fault diagnosis methods rely on signal processing, feature extraction, and pattern recognition techniques, requiring in-depth mechanistic analysis and depending on expert experience and prior knowledge. With the development of artificial intelligence (AI) technology, intelligent fault diagnosis based on deep learning has become a research hotspot, demonstrating significant advantages. Deep learning is an AI method based on multi-layered neural networks. Its core idea is to simulate the hierarchical structure of the human brain, automatically learning features and hidden patterns from data through models such as deep neural networks, convolutional neural networks, and recurrent neural networks. Compared to traditional methods, deep learning can directly extract effective information from raw data without complex signal mechanism analysis, thus reducing human intervention and improving diagnostic accuracy and generalization ability. In recent years, intelligent fault diagnosis methods based on deep learning have been widely applied to the condition monitoring and fault prediction of rotating machinery, achieving remarkable results.

[0004] However, in real-world industrial environments, the complex operating conditions of rotating machinery result in diverse and uncertain fault types. While some fault modes are likely to occur, their rarity often prevents them from appearing in or being recorded in historical data. This uneven data distribution and the unpredictability of new faults mean that supervised learning methods, relying on large amounts of labeled historical data, lack the adaptability to effectively detect unfamiliar faults. Furthermore, current deep learning-based methods such as supervised learning, transfer learning, and few-shot learning, while improving generalization capabilities to some extent, primarily focus on enhancing the diagnostic accuracy of known fault types. They tend to overlook objective limitations and practical needs, leading to less than ideal results when diagnosing new fault types, and thus exhibiting certain limitations. Summary of the Invention

[0005] The purpose of this invention is to provide a fault prediction method based on multi-domain feature fusion and interpretable adversarial convolutional coding network, which focuses on new fault types that may occur in rotating machinery in complex industrial environments, improves the reliability of fault diagnosis under small sample data conditions, and comprehensively coordinates and optimizes the fault diagnosis process and model performance.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for predicting information-poor faults based on multi-domain feature fusion and interpretable adversarial convolutional coding networks includes:

[0008] Collect vibration signals from the target rotating machinery;

[0009] The vibration signal is input into a preset fault prediction model, which outputs a fault category. The fault category includes known fault categories and new fault categories. The fault prediction model is trained based on a training set, which includes fault samples of several rotating machines under various operating conditions. The fault prediction model is constructed by combining multi-domain feature fusion and an interpretable adversarial convolutional coding network.

[0010] Optionally, before training the fault prediction model based on the training set, the method further includes preprocessing the training set, wherein the preprocessing includes:

[0011] The collected fault samples are segmented and normalized.

[0012] Optionally, the fault prediction model includes:

[0013] The new fault detection module is used to perform multi-scale feature fusion on input data based on an autoencoder, obtain encoded data and input it into a self-decoder, and output decoded data. The autoencoder is constructed using wavelet scattering transform and convolution modules of different scales, and the self-decoder is constructed using deconvolution modules of different scales.

[0014] A known fault diagnosis module is used to input the encoded data into a classifier and output a fault category prediction vector, wherein the classifier is constructed using a fully connected neural network;

[0015] The collaborative adversarial module is used to input the encoded and decoded data into the discriminator and optimize the model using an adversarial training strategy.

[0016] The fault output module is used to calculate the reconstruction error between the input data and the decoded data, and to determine the fault category based on the reconstruction error. If the reconstruction error reaches a preset threshold, it is determined to be a new fault category. If the reconstruction error does not reach the preset threshold, it is the known fault category corresponding to the largest component in the fault category prediction vector.

[0017] Optionally, the encoded data is:

[0018]

[0019] Where E represents the autoencoder, x represents the input data, and E(x) represents the encoded data. and Let J represent the wavelet scattering transform and convolution module with the j-th scale information, respectively. E This represents the number of scales contained in the autoencoder.

[0020] Optionally, the decoded data is:

[0021]

[0022] Where D represents the self-decoder, and D(E(x)) represents the decoded data. J represents a deconvolution module with information at the k-th scale. D This represents the number of scales contained in the self-decoder.

[0023] Optionally, the objective function for model optimization using an adversarial training strategy is:

[0024]

[0025] Among them, E, D, P E P D These represent autoencoder, autodecoder, encoder discriminator, and decoder discriminator, respectively. J represents a deconvolution module with information at the k-th scale. D L represents the number of scales contained in the self-decoder. ae L is the mean square error loss function used by the autoencoder. dsc The binary cross-entropy loss function used by the discriminator, x i Let m represent the i-th input data, and E(x) represent the total number of input data. i ), Let D(E(x) represent the encoded data and pseudo-encoded data of the i-th input data, respectively. i )) These represent the decoded data and pseudo-decoded data of the i-th input data, respectively.

[0026] Optionally, the autoencoder can be optimized as follows:

[0027]

[0028] Where, θ ae b ae These represent the weights and bias parameters involved in the autoencoder, respectively.

[0029] Optionally, the discriminator can be optimized as follows:

[0030]

[0031] Where, θ dsc b dsc These are the weights and bias parameters included in the discriminator.

[0032] Optionally, the classifier can be optimized as follows:

[0033]

[0034] Where C represents the classifier, θ cls b cls These are the weights and bias parameters included in the classifier, L. cls The cross-entropy loss function used by the classifier, C(E(x) i )) represents the predicted category of the i-th input data, y i This represents the true category of the i-th input data.

[0035] The beneficial effects of this invention are as follows:

[0036] This invention focuses on new types of faults that may occur in rotating machinery under complex industrial environments, and takes the accurate identification of new fault types as the basis for improving fault diagnosis technology. This invention improves the reliability of fault diagnosis under small sample data conditions, and takes enhancing the model's learning ability on small sample new fault data as one of the key directions of technology optimization. This is because it is difficult to obtain a large number of fault sample data in actual industrial scenarios, and small sample diagnosis capability is crucial. This invention comprehensively coordinates and optimizes the fault diagnosis process and model performance, including improving diagnosis efficiency, reducing the false diagnosis rate, and enhancing the model's generalization ability. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a method for predicting information-poor faults based on multi-domain feature fusion and interpretable adversarial convolutional coding networks according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the fault prediction model structure according to an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the parameters of the autoencoder and autodecoder in an embodiment of the present invention;

[0041] Figure 4 This diagram illustrates the test accuracy obtained by the method and the comparative method under different working conditions in this embodiment of the invention.

[0042] Figure 5 This is a schematic diagram of the ablation experiment results of the method and the comparative method in this embodiment of the invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] This embodiment provides a method for predicting information-poor faults based on multi-domain feature fusion and interpretable adversarial convolutional coding networks, including:

[0046] Collect vibration signals from the target rotating machinery;

[0047] The vibration signal is input into a preset fault prediction model, which outputs a fault category. The fault category includes known fault categories and new fault categories. The fault prediction model is trained based on a training set, which includes fault samples of several rotating machines under various operating conditions. The fault prediction model is constructed by combining multi-domain feature fusion and an interpretable adversarial convolutional coding network.

[0048] Prior to training the fault prediction model based on the training set, the training set is preprocessed, and the preprocessing includes:

[0049] The collected fault samples are segmented and normalized.

[0050] Furthermore, the fault prediction model includes:

[0051] The new fault detection module is used to perform multi-scale feature fusion on input data based on an autoencoder, obtain encoded data and input it into a self-decoder, and output decoded data. The autoencoder is constructed using wavelet scattering transform and convolution modules of different scales, and the self-decoder is constructed using deconvolution modules of different scales.

[0052] A known fault diagnosis module is used to input the encoded data into a classifier and output a fault category prediction vector, wherein the classifier is constructed using a fully connected neural network;

[0053] The collaborative adversarial module is used to input the encoded and decoded data into the discriminator and optimize the model using an adversarial training strategy.

[0054] The fault output module is used to calculate the reconstruction error between the input data and the decoded data, and to determine the fault category based on the reconstruction error. If the reconstruction error reaches a preset threshold, it is determined to be a new fault category. If the reconstruction error does not reach the preset threshold, it is the known fault category corresponding to the largest component in the fault category prediction vector.

[0055] Specifically, this embodiment focuses on new fault types that may occur in rotating machinery under complex industrial environments, and takes the accurate identification of new fault types as the basis for improving fault diagnosis technology. This embodiment improves the reliability of fault diagnosis under small sample data conditions, and takes enhancing the model's learning ability on small sample new fault data as one of the key directions of technology optimization. This is because it is difficult to obtain a large number of fault sample data in actual industrial scenarios, and small sample diagnosis capability is crucial. This embodiment comprehensively coordinates and optimizes the fault diagnosis process and model performance, including improving diagnosis efficiency, reducing the misdiagnosis rate, and enhancing the model's generalization ability.

[0056] The following is combined with Figure 1 The fault prediction model constructed in this embodiment is described in detail. The model construction, training optimization, and testing application specifically include the following steps:

[0057] Step S1: Collect the original vibration signals of the rotating machinery under specific experimental conditions, and label the fault samples obtained under different operating conditions after normalization to construct a complete dataset.

[0058] Step S2: Construct a new information-poor fault diagnosis model (ME-DACAE) based on multi-domain feature fusion and interpretable adversarial convolutional coding networks for new fault detection with small sample sizes, such as... Figure 2As shown, the network framework includes: a novel fault detection module based on a multi-scale autoencoder, a cooperative adversarial module based on a dual discriminator, a known fault diagnosis module based on a fully connected classifier, and a fault output module.

[0059] Step S3: During the training phase, wavelet scattering transform with different scale parameters is used to perform multi-scale data augmentation on the training samples. The augmented training samples at each scale are then further extracted using a convolution module. At the same time, the average value of the extracted features at each scale is calculated to achieve feature fusion.

[0060] Step S4: The fused features and randomly generated pseudo-coded samples are input into the multi-scale deconvolution modules of the encoder discriminator, classifier, and decoder. The average value of the decoding reconstruction of each scale deconvolution module is calculated to obtain the decoded samples and pseudo-decoded samples, which are then input into the decoder discriminator.

[0061] Step S5: In the optimization phase, calculate the loss of the sample on the autoencoder, classifier, encoder-discriminator and decoder-discriminator, use the gradient descent algorithm and combine it with the adversarial training strategy to alternately optimize the model parameters, and complete the training of the overall model after multiple iterations.

[0062] Step S6: During the testing phase, calculate the reconstruction error between the test sample and the decoded sample. If the error is greater than or equal to the set threshold, it is determined to be a new fault category; otherwise, it is the known category corresponding to the largest component in the prediction vector.

[0063] Furthermore, in step S1, the dataset is sourced from a laboratory experimental platform or an actual industrial environment equipment monitoring platform.

[0064] When collecting data under various working conditions, the process also includes: preprocessing the collected data, that is, segmenting the collected dataset and normalizing the resulting sample set to the range [-1,1], which is suitable for input to the network model.

[0065] Furthermore, the new fault detection module constructed in step S2 adopts an autoencoder and self-decoder structure. The convolutional module in the autoencoder consists of convolutional layers, activation layers, and pooling layers stacked sequentially, while the deconvolutional module in the self-decoder consists of deconvolutional layers, activation layers, and unpooling layers stacked sequentially. The cooperative adversarial module constructs an encoder discriminator and a decoder discriminator to enhance the training effect of the convolutional autoencoder. The known fault diagnosis module is constructed using a fully connected neural network, with the Softmax activation function used in its last layer.

[0066] Furthermore, step S3 specifically includes:

[0067] For the original fault signal data x received in step S1, multiple scale parameters j = 1, 2, ..., J are set to perform a multi-scale wavelet scattering transform on the original sample x. When performing this transform on the sample, a scale function φ and multiple sets of wavelet functions are required. Where ψ λ This represents the set of wavelet functions obtained by applying a series of scaling transformations to the original mother wavelet. The wavelet functions are first used at level zero. Performing a wavelet transform on sample x yields wavelet coefficients, which, when moduloed, form the first layer of the wavelet scattering transform, represented as:

[0068]

[0069] Continue using the first-level wavelet coefficients Perform convolution operations and take the modulus of the newly obtained wavelet coefficients to obtain the second layer, which is represented as follows:

[0070]

[0071] Following this pattern, after m layers, the final output of the wavelet scattering transform can be obtained. This output is called the scattering coefficient, and its expression is as follows:

[0072]

[0073] Here, the zeroth layer L0 represents the original sample, and the scattering coefficient S is the depth feature extracted after performing wavelet scattering transform on sample x. The scaling function used has an adjustable scaling parameter, and multi-scale data augmentation can be achieved by setting multiple scaling parameters. Finally, the scattering coefficients are concatenated to obtain the scattering sample.

[0074] Multi-layer convolutional modules were designed for scattering samples at different scales to further extract depth features while ensuring that the same feature dimension is obtained after passing through multiple convolutional modules. Then, the average value of the extracted features at each scale was calculated to achieve multi-scale feature fusion, and the fused features were used as the encoded features of the original sample (encoded sample). For the original sample x, J... E At this scale, the encoded features of sample x can be expressed as:

[0075]

[0076] Where E represents an autoencoder, and These represent wavelet scattering transform and convolution modules with information at the j-th scale, respectively.

[0077] Furthermore, in step S4, multiple deconvolution modules with different convolution parameters are designed in the self-decoder to decode the encoded features separately to reconstruct the original sample. The final decoded sample is obtained by calculating the average value of each reconstructed sample, which can be represented as:

[0078]

[0079] Where D represents the self-decoder, This represents a deconvolution module with information at the k-th scale.

[0080] An adversarial training strategy was introduced to enhance the training effect of the autoencoder. Two discriminators based on fully connected neural networks were constructed, namely an encoding discriminator and a decoding discriminator, which were used to assist the autoencoder in achieving adversarial training.

[0081] The input to the encoding discriminator includes encoded samples from the training samples and some randomly generated auxiliary samples (called pseudo-encoded samples). The features of the pseudo-encoded samples follow a normal distribution with values ​​ranging from [-1, 1]. On the other hand, the input to the decoding discriminator includes decoded samples from the training samples and the output obtained by decoding the pseudo-encoded samples (called pseudo-decoded samples). For a dataset A containing m samples of known classes, Where x i and y i Let and represent the i-th sample in the dataset and its corresponding true class label, respectively. The discriminant label for encoded samples is set to 1; while the label for pseudo-encoded samples, obtained through random generation, is set to 0. The objective function during adversarial training can be defined as:

[0082]

[0083] Among them, E, D, P E P D , J represents the autoencoder, decoder, encoder discriminator, autodecoder discriminator, and the deconvolution module with the k-th scale information in the autodecoder, respectively. D L represents the number of scales contained in the self-decoder. ae L is the mean squared error (MSE) loss function used by the autoencoder, while L dsc The binary cross-entropy loss function used by the two discriminators, E(x) i ), Let D(E(x) represent the encoded sample and the pseudo-encoded sample, respectively. i )) These represent the decoded sample and the pseudo-decoded sample, respectively.

[0084] Furthermore, in step S5, the optimization stage, the output of the auto-decoder in S4 should reproduce the original sample as much as possible. The objective function of the autoencoder can be defined as:

[0085]

[0086] Where, x i and m represent the i-th sample and the total number of samples in the training set, respectively. Then it means that sample x i The decoded sample obtained by reconstructing at the k-th scale, L ae Let MSE be the objective function. The autoencoder is trained by minimizing this objective function; this process can be represented as:

[0087]

[0088] Where, θ ae b ae These represent the various convolution weights and bias parameters involved in the autoencoder.

[0089] The autoencoder and the two discriminators should be trained and optimized alternately. When training the autoencoder, the parameters of the discriminators need to be fixed, and the parameters of the autoencoder are optimized by minimizing the objective function, i.e., formula (6). This process can be expressed as:

[0090]

[0091] Where, θ ae b ae These represent the weights and bias parameters involved in the autoencoder, respectively. Conversely, when training the discriminator, the parameters of the autoencoder need to be fixed, and the parameters of the discriminator are updated by maximizing the objective function (6), as shown in the following equation:

[0092]

[0093] Where, θ dsc b dsc These represent the weights and bias parameters included in the discriminator. The alternating optimization process described above is repeated multiple times to complete adversarial training, effectively improving the novel fault detection performance of the proposed method.

[0094] For optimizing the classifier, given a dataset A containing m samples of known classes, Where x i and y i Let and represent the i-th sample in the dataset and its corresponding class label, respectively. During classifier training, the objective function can be defined as:

[0095]

[0096] Where E and C represent the encoder and classifier, respectively, and L cls Let be the cross-entropy loss function of the classifier. The classifier is trained by minimizing this objective function; this process can be represented as:

[0097]

[0098] Where, θ cls b cls These represent the weights and bias parameters involved in the classifier, respectively. Furthermore, by simultaneously solving equations (10), (11), and (12), the training optimization expression for the entire model can be obtained:

[0099]

[0100] Furthermore, in step S6, during the testing phase, after the proposed method has been trained, the autoencoder and classifier will output a decoded sample and a predicted vector for each test sample, respectively. Then, the reconstruction error δ between the test sample and the decoded sample is calculated using the MSE function and compared with the set detection threshold φ. If δ ≥ φ, the test sample is considered a new fault category; if δ < φ, the test sample is considered a known category (healthy or known fault category), and the category corresponding to the largest component in the predicted vector is its specific category. This achieves small-sample new fault detection and known fault diagnosis for rotating machinery.

[0101] The method proposed in this embodiment is verified by comparative experiments below:

[0102] 1. Experiment Introduction:

[0103] Experiments were designed using the Five-Speed ​​Transmission (FST) dataset, which was collected from an automotive transmission system test bench. Accelerometers were mounted on the five-speed transmission on the test bench, and vibration signal data were collected at a sampling frequency of 24 kHz under three operating conditions: a constant transmission load of 50 Nm and input shaft speeds of 750, 1000, and 1250 rpm. The FST dataset includes data on bearing faults, gear faults, and combined faults of both. The transmission conditions covered include four types: healthy, 0.2 mm inner ring (IR) cracking, single gear tooth (GT) fracture, and compound fault (CF), denoted by categories 1 to 4. Healthy, IR, and GT faults are considered known categories on the transmission, while CF faults are considered as a new, unknown fault category. The fault was implanted by wire cutting. A 0.2mm crack was machined on the inner ring of the bearing to simulate bearing cracking, and the gear teeth were cut to simulate gear tooth breakage.

[0104] Samples were collected using a non-overlapping sliding window on the time-domain vibration signals provided by the FST dataset. 450 samples (150 for each operating condition) with a length of 4096 were collected for each data class. Small-sample novel fault detection tasks with 1-shot, 2-shot, and 5-shot samples were designed for the three operating conditions. One, two, and five samples from each of the known categories were randomly selected for training. Simultaneously, 50 samples from each of the known categories (excluding the training samples) were selected to form the test set along with all samples from the novel fault category. Each task's test set contained 150 samples from the known category and 150 samples from the novel fault category. The goal of the task was to accurately distinguish between these two categories.

[0105] 2. Method Validation:

[0106] The parameters of the constructed multi-domain feature fusion and interpretable adversarial convolutional coding network are shown in Table 1 and Figure 3 As shown:

[0107] Table 1

[0108]

[0109] (1) Comparison of the method of the present invention with other fault diagnosis methods:

[0110] In practical engineering, models are required not only to accurately detect new fault categories but also to further differentiate known fault categories. This prevents confusion between healthy and known fault categories and allows for rapid location of known faults when the model identifies them. This experiment selected four comparison methods: Multi-Task Autoencoder (MTAE), Multi-Task Convolutional Autoencoder (MTCAE), Multi-Task Autoencoder with Fine-Tuning (MTAE-FT), and Multi-Task Convolutional Autoencoder with Fine-Tuning (MTCAE-FT). All are multi-task models capable of simultaneously detecting new faults and diagnosing known faults. Furthermore, healthy, IR, and GT faults were considered known categories, while CF faults were considered as unknown new fault categories. Then, small-sample tasks for new fault detection and known fault diagnosis were designed for three operating conditions: 1-shot, 2-shot, and 5-shot. One, two, and five samples from each of the known categories were randomly selected for training, while the remaining samples, along with all samples from the new fault category, formed the test set. For transfer learning-based methods such as MTAE-FT and MTCAE-FT, additional source domain data is provided to assist in their training. Since the FST dataset covers data from three different operating conditions, when data from one operating condition is designated as the target domain, data from the other two operating conditions will be used to construct the source domain. Based on the above settings, the number of samples used for each method in a diagnostic task is shown in Table 2.

[0111] Table 2

[0112]

[0113] The test accuracy achieved by each method is as follows: Figure 4As shown in the figure. Experimental results show that MTAE achieves an accuracy of only about 50% in small-sample variable-condition fault diagnosis tasks, and the improved version MTAE-FT also fails to significantly improve this (below 55%). Although MTAE has a high F1 score in the new fault detection task, it performs poorly in the detailed differentiation of known categories, mainly because its encoder based on fully connected layers has limited feature extraction capabilities, making it difficult to effectively identify fault features and affecting classification performance. In contrast, MTCAE and MTCAE-FT, which use convolutional layer encoders, have significantly better test accuracies than the MTAE series. In the 1-shot task at 750 rpm, MTCAE-FT achieves a test accuracy of 90.16%, an improvement of 41.18% over MTAE-FT, highlighting the feature extraction advantage of convolutional layers. Furthermore, the ME-DACAE model achieves the highest accuracy in most tasks, only slightly lagging behind MTCAE-FT in the 5-shot task at 750 rpm. However, MTCAE-FT relies on a large amount of source domain data for training, while ME-DACAE only requires scarce target data, avoiding additional costs and combining economy and effectiveness, making it more suitable for small-sample fault diagnosis problems in engineering practice.

[0114] (2) Ablation experiment using the method of the present invention:

[0115] Ablation experiments were designed using the FST dataset to investigate the impact of multi-scale encoding, multi-scale decoding, and adversarial training strategies on model performance. Four ablation models were designed based on ME-DACAE, labeled AM1 to AM4. AM1 contains only a single-scale encoder and decoder; AM2 combines a multi-scale encoder (covering five scales) with a single-scale decoder; AM3, in contrast to AM2, uses a single-scale encoder and a multi-scale decoder (covering five scales); AM4 integrates both a multi-scale encoder and decoder; and ME-DACAE adds a cooperative adversarial module to AM4. Figure 5 The test accuracy of various ablation models and ME-DACAE in a 1-shot task at 1000 rpm is presented. Furthermore, the experimental results of MTCAE are also presented. Figure 4 (As shown) was also added Figure 5A comparison was conducted. Experimental results show that both multi-scale encoding and decoding strategies contribute to improving the detection and diagnosis performance of the model. The test accuracy of AM2 and AM3 is improved by 2.74% and 4.55% respectively compared to AM1, while AM4 further improves upon this, indicating that using both strategies simultaneously is more effective than using them alone. The multi-scale encoder enhances the data through feature fusion, while the multi-scale decoder integrates information from various scales to generate better decoding samples. Compared to AM4, ME-DACAE's accuracy is improved by 0.62%, verifying the effectiveness of the adversarial module in new fault detection and known fault diagnosis. Furthermore, AM1's accuracy is significantly higher than MTCAE, indicating that wavelet scattering transform has a positive effect on model performance under small sample conditions.

[0116] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting information-poor faults based on multi-domain feature fusion and interpretable adversarial convolutional coding networks, characterized in that, include: Collect vibration signals from the target rotating machinery; The vibration signal is input into a preset fault prediction model, and the fault category is output. The fault category includes known fault categories and new fault categories. The fault prediction model is trained based on a training set, which includes fault samples of several rotating machines under various working conditions. The fault prediction model is constructed by combining multi-domain feature fusion and an interpretable adversarial convolutional coding network. The fault prediction model includes: The new fault detection module is used to perform multi-scale feature fusion on input data based on an autoencoder, obtain encoded data and input it into a self-decoder, and output decoded data. The autoencoder is constructed using wavelet scattering transform and convolution modules of different scales, and the self-decoder is constructed using deconvolution modules of different scales. A known fault diagnosis module is used to input the encoded data into a classifier and output a fault category prediction vector, wherein the classifier is constructed using a fully connected neural network; The cooperative adversarial module is used to input the encoded data into the encoding discriminator and the decoded data into the decoding discriminator, and to optimize the model using an adversarial training strategy. The fault output module is used to calculate the reconstruction error between the input data and the decoded data, and determine the fault category based on the reconstruction error. If the reconstruction error reaches a preset threshold, it is determined to be a new fault category. If the reconstruction error does not reach the preset threshold, it is the known fault category corresponding to the largest component in the fault category prediction vector. The encoded data is: ； Where E represents the autoencoder, x represents the input data, and E(x) represents the encoded data. and Let J represent the wavelet scattering transform and convolution module with the j-th scale information, respectively. E This represents the number of scales contained in the autoencoder; The decoded data is: ； Where D represents the self-decoder, and D(E(x)) represents the decoded data. J represents a deconvolution module with information at the k-th scale. D This represents the number of scales contained in the self-decoder; The objective function for model optimization using an adversarial training strategy is: ； Among them, E, D, , These represent autoencoder, autodecoder, encoder discriminator, and decoder discriminator, respectively. This represents a deconvolution module with information at the k-th scale. This indicates the number of scales contained in the self-decoder. The mean squared error loss function used by the autoencoder. The binary cross-entropy loss function used by the encoder-discriminator or decoder-discriminator, x i Let m represent the i-th input data, and m represent the total number of input data. , These represent the encoded data and pseudo-encoded data of the i-th input data, respectively. , These represent the decoded data and pseudo-decoded data of the i-th input data, respectively.

2. The information-poor fault prediction method based on multi-domain feature fusion and interpretable adversarial convolutional coding network according to claim 1, characterized in that, Before training the fault prediction model based on the training set, the training set is preprocessed, and the preprocessing includes: The collected fault samples are segmented and normalized.

3. The information-poor fault prediction method based on multi-domain feature fusion and interpretable adversarial convolutional coding network according to claim 1, characterized in that, The optimized representation of the autoencoder is as follows: ； in, , These represent the weights and bias parameters involved in the autoencoder, respectively.

4. The information-poor fault prediction method based on multi-domain feature fusion and interpretable adversarial convolutional coding network according to claim 1, characterized in that, The optimized representation of the encoder discriminator or decoder is as follows: ； in, , These are the weights and bias parameters included in the encoder discriminator or decoder discriminator, respectively.

5. The information-poor fault prediction method based on multi-domain feature fusion and interpretable adversarial convolutional coding network according to claim 1, characterized in that, The optimized classifier is represented as follows: ； Where C represents the classifier, , These represent the weights and bias parameters of each element in the classifier. The cross-entropy loss function used by the classifier, C(E(x) i )) represents the predicted category of the i-th input data, y i This represents the true category of the i-th input data.

Images