Rare sample compound fault diagnosis method based on multi-source semantic combination guided diffusion

Abstract

The application discloses a kind of based on multi-source semantic combination guide diffusion's scarce sample compound fault diagnosis method, it is related to fault detection technical field.The method includes: obtaining the feature expression of each single fault under multiple sensor channels, constructs its continuous semantic vector;Target compound fault is disassembled into single fault set, and the compound fault semantic vector is obtained by the combination of corresponding single fault semantic vector;Fusion sensor space position coding and category attribute vector, form spatial perception continuous semantic representation;By wide receptive field one-dimensional convolutional neural network mapping to latent space, obtain compound fault latent feature;The latent feature is input into the multi-source conditional combination guide diffusion model based on bayesian inference, generates guide noise and carries out reverse denoising sampling, outputs compound fault pseudo-feature;Training set real feature and pseudo-feature are merged into balanced dataset, for neural network training.The application can effectively improve the diagnosis performance of compound fault under very small sample.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically, to a method for diagnosing composite faults in scarce samples based on multi-source semantic combination-guided diffusion. Background Technology

[0002] Composite fault diagnosis of key transmission components in rotating machinery is a core technological aspect for ensuring the safety and service life of highly reliable equipment systems in rail transit, aerospace, and energy equipment. In practical engineering, multi-source sensors are widely deployed at different measurement points on equipment to capture the differences in fault signals propagating in space. Existing composite fault diagnosis methods mainly follow two technical routes: one is fault feature extraction methods based on signal processing and mechanism analysis, which achieve fault identification by combining time-domain, frequency-domain, or time-frequency-domain statistical features with shallow classifiers; the other is end-to-end diagnostic methods based on deep learning, which automatically extract nonlinear fault features using convolutional neural networks, attention mechanisms, etc. In recent years, to address the problem of scarce fault samples, researchers have further introduced generative methods such as generative adversarial networks, variational autoencoders, and diffusion models to augment the data.

[0003] However, the existing solutions still have many limitations. First, most existing multi-source information fusion methods use simple feature splicing or black-box network joint modeling, lacking effective modeling of the independent semantic structure between multi-sensor signals, which easily introduces redundant information and causes feature submersion. Second, traditional methods usually treat composite faults as independent categories for overall modeling, making it difficult to characterize the coupling relationship between multiple single fault components in composite faults and the continuous evolution law of fault intensity. Furthermore, generative adversarial network-based methods suffer from inherent defects such as training instability and pattern collapse under small sample conditions; while existing diffusion models can generate high-quality samples, their condition guidance methods are mostly single conditions or simple condition splicing, lacking fine-grained combination and collaborative modeling of multi-source semantic conditions. Therefore, it is necessary to study a technology that can achieve fine-grained combination guidance of multi-source semantics and stably generate high-quality composite fault features under extremely scarce sample conditions. Summary of the Invention

[0004] In view of this, the present invention proposes a method for diagnosing composite faults of scarce samples based on multi-source semantic combination-guided diffusion, in order to solve the above problems.

[0005] Firstly, this application provides a method for diagnosing composite faults in scarce samples based on multi-source semantic combination-guided diffusion, including: S1: Obtain the feature performance of all single faults under different sensor channels, and construct a continuous semantic vector corresponding to each single fault based on the feature performance of the single faults under different sensor channels; S2: Obtain the continuous semantic vector corresponding to each composite fault, wherein, for the target composite fault, the target composite fault is decomposed to obtain the set of single faults corresponding to the target composite fault, and the continuous semantic vector corresponding to the target composite fault is obtained based on the continuous semantic vectors corresponding to the target single faults in the set of single faults. S3: Introduce sensor spatial location encoding vector and category attribute vector, and fuse the sensor spatial location encoding vector and the category attribute vector with the continuous semantic vector corresponding to the composite fault category to obtain the spatial perception continuous semantic representation corresponding to the composite fault category. S4: Input the spatially perceived continuous semantic representation corresponding to the composite fault category into a one-dimensional convolutional neural network with a wide receptive field to map the spatially perceived continuous semantic representation corresponding to the composite fault category to the latent space, so as to obtain the latent features corresponding to the composite fault category. S5: Input the latent features corresponding to the composite fault category into the multi-source condition combination guided diffusion model based on Bayesian inference to output the guided noise corresponding to each composite fault category, and perform reverse denoising sampling to output composite fault pseudo features; S6: The acquired training set features are merged with the composite fault pseudo-features to generate a balanced dataset, which is then input into the neural network for training.

[0006] One possible approach is to obtain the feature representation of all single faults under different sensor channels, and construct a continuous semantic vector corresponding to each single fault based on the feature representation of the single faults under different sensor channels. The characteristics of a single fault under different sensor channels include: the time-domain characteristics and frequency-domain characteristics of a single fault under different sensors; The continuous semantic vector corresponding to the single fault is constructed in the following manner: Regarding the sensor in the first The first channel Each statistical indicator has a continuous semantic vector representation as follows: ; in, For the k-th sensor channel, single fault category The continuous semantic value corresponding to the j-th statistical indicator; —Temperature coefficient ; —On the k-th sensor channel, single fault category The standardized value of the j-th statistical indicator; in, ; —No. The first channel The mean of a statistical indicator across all single fault categories; —No. The first channel The standard deviation of each statistical indicator across all single fault categories; —On the k-th sensor channel, single fault category The value of the j-th statistical indicator.

[0007] One possible approach is to obtain the continuous semantic vector corresponding to each composite fault in the step of obtaining the continuous semantic vector corresponding to the composite fault in the following manner: For the target composite fault in the first The continuous semantic vectors under each channel, based on the target single fault in the single fault set at the first... The maximum value of the continuous semantic vector under each channel is determined.

[0008] One possible approach is that, in the step of introducing a sensor spatial location encoding vector and an input category attribute vector, and fusing the sensor spatial location encoding vector and the category attribute vector with the continuous semantic vector corresponding to the composite fault category to obtain the spatial perception continuous semantic representation corresponding to the composite fault category, the spatial perception continuous semantic representation corresponding to the composite fault category is expressed as follows: ; —The spatially perceived continuous semantic representation of the composite fault category on the kth sensor channel; —A scalar quantity, positively correlated with the degree of influence of differences in fault categories; —A scalar quantity, positively correlated with the degree of influence of sensor position information; —Category attribute vector; —Fault category; — Fault Category Continuous semantic vectors under the k-th sensor channel; — Concatenation operator.

[0009] One possible approach is that, prior to the step of inputting the spatially perceived continuous semantic representation corresponding to the composite fault category into a one-dimensional convolutional neural network with a wide receptive field to map the spatially perceived continuous semantic representation corresponding to the composite fault category to the latent space to obtain the latent features corresponding to the composite fault category, the method further includes: The one-dimensional convolutional neural network is trained using a weighted sum of cross-entropy loss and center loss as the joint optimization objective.

[0010] One possible approach is that, in the step of inputting the latent features corresponding to the composite fault categories into a multi-source conditional combination guided diffusion model based on Bayesian inference to output the guided noise corresponding to each composite fault category, and performing inverse denoising sampling to output the pseudo-features of the composite faults, the corresponding pseudo-features are output in the following manner for any composite fault: At each time step, perform the following operations until the time step length is reached: Obtain the first noisy latent feature and input the first noisy latent feature into the multi-source condition combination guided diffusion model, and output the current guided noise; A second noisy latent feature is generated based on the first noisy latent feature and the current guiding noise, and the first noisy latent feature is updated based on the second noisy latent feature; The composite fault pseudo-feature is the difference between the noisy latent feature input to the multi-source condition combined guided diffusion model and the guided noise output by the multi-source condition combined guided diffusion model when the time step is reached.

[0011] One possible approach is to determine the guiding noise using the following formula: ; —A continuous semantic vector for the k-th sensor channel in a complex fault; —Number of sensor channels; —Weight of the kth sensor channel; —Noisy latent features; —Empty condition; —Time step; —Noise in the continuous semantic vector of the k-th sensor channel in the composite fault at the current time step; —Noise under idle conditions in the k-th sensor channel of the composite fault at the current time step; —The guiding noise of the continuous semantic vector of the k-th sensor channel in the composite fault at the current time step.

[0012] One possible approach is that the multi-source condition combination guided diffusion model uses the following formula as the optimization objective: ; in, —Used for backpropagation to update network parameters Optimization target value —The multi-source conditional combined guided diffusion model predicts noise based on the current input; —Hidden features of the output of a one-dimensional convolutional neural network; —A continuous semantic vector for the k-th sensor channel in a complex fault; —Regarding the loss function within the parentheses in the variables , , and Take the batch average on the joint distribution; —During the positive noise addition process, the system randomly generates and actually injects noise. The noise tensor in; —Cumulative signal retention coefficient; -norm.

[0013] One possible approach is to combine the acquired training set features with the composite fault pseudo-features and input them into the neural network for training. The predicted probabilities of each category are output by the Softmax activation function.

[0014] This invention relates to a method for diagnosing complex faults in scarce samples based on multi-source semantic combination guided diffusion, aiming to solve problems such as the difficulty of multi-source information coordination, the extreme scarcity of complex fault samples, and the difficulty in accurately characterizing the continuous evolution of fault intensity in complex fault diagnosis. By constructing a spatially aware continuous semantic representation, prior knowledge from sufficient single faults is transferred to scarce complex faults, effectively improving the accuracy of complex fault diagnosis. Simultaneously, the WideCNN1D feature extraction network is jointly optimized using cross-entropy loss and center loss to construct a compact intra-class and separable inter-class continuous feature latent space, providing a stable distribution basis for subsequent pseudo-feature generation. Finally, a multi-source conditional combination guided diffusion model based on Bayesian inference is designed. Through multi-source conditional independent inference and combination guidance mechanisms, high-quality generation of pseudo-features for complex faults is achieved. The pseudo-features are then combined with training neural networks to effectively improve the performance of complex fault diagnosis under extremely small sample conditions. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a method for diagnosing composite faults in scarce samples based on multi-source semantic combination guided diffusion provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for diagnosing composite faults in scarce samples based on multi-source semantic combination-guided diffusion, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the acquisition of spatially perceived continuous semantic representation in a rare sample composite fault diagnosis method based on multi-source semantic combination guided diffusion, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the acquisition of guiding noise in a rare sample composite fault diagnosis method based on multi-source semantic combination guided diffusion, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of a one-dimensional convolutional neural network training method provided in an embodiment of the present invention. Figure 6 A comparison chart of the accuracy of fault diagnosis in a dataset-comparison experiment in the rare sample composite fault diagnosis method based on multi-source semantic combination guided diffusion provided in the embodiments of the present invention. Figure 7 This is a comparison chart of the accuracy of fault diagnosis in another dataset-comparison experiment in the rare sample composite fault diagnosis method based on multi-source semantic combination guided diffusion provided in the embodiments of the present invention. Figure 8The fault diagnosis confusion matrix diagram of different models in the rare sample composite fault diagnosis method based on multi-source semantic combination guided diffusion provided in the embodiments of the present invention; Figure 9 The t-SNE visualization distribution of features extracted by different models in the rare sample composite fault diagnosis method based on multi-source semantic combination guided diffusion provided in the embodiments of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] Composite fault diagnosis of key transmission components in rotating machinery is a core technological aspect for ensuring the safety and service life of highly reliable equipment systems in rail transit, aerospace, and energy equipment. In practical engineering, multi-source sensors are widely deployed at different measurement points on equipment to capture the differences in fault signals propagating in space. Existing composite fault diagnosis methods mainly follow two technical routes: one is fault feature extraction methods based on signal processing and mechanism analysis, which achieve fault identification by combining time-domain, frequency-domain, or time-frequency-domain statistical features with shallow classifiers; the other is end-to-end diagnostic methods based on deep learning, which automatically extract nonlinear fault features using convolutional neural networks, attention mechanisms, etc. In recent years, to address the problem of scarce fault samples, researchers have further introduced generative methods such as generative adversarial networks, variational autoencoders, and diffusion models to augment the data.

[0018] However, the existing solutions still have many limitations. First, most existing multi-source information fusion methods use simple feature splicing or black-box network joint modeling, lacking effective modeling of the independent semantic structure between multi-sensor signals, which easily introduces redundant information and causes feature submersion. Second, traditional methods usually treat composite faults as independent categories for overall modeling, making it difficult to characterize the coupling relationship between multiple single fault components in composite faults and the continuous evolution law of fault intensity. Furthermore, generative adversarial network-based methods suffer from inherent defects such as training instability and pattern collapse under small sample conditions; while existing diffusion models can generate high-quality samples, their condition guidance methods are mostly single conditions or simple condition splicing, lacking fine-grained combination and collaborative modeling of multi-source semantic conditions. Therefore, it is necessary to study a technology that can achieve fine-grained combination guidance of multi-source semantics and stably generate high-quality composite fault features under extremely scarce sample conditions.

[0019] Reference Figure 1 and Figure 2 In view of this, this application provides a method for diagnosing composite faults in scarce samples based on multi-source semantic combination-guided diffusion, including: S1: Obtain the feature performance of all single faults under different sensor channels, and construct a continuous semantic vector corresponding to each single fault based on the feature performance of the single faults under different sensor channels.

[0020] The characteristics of a single fault under different sensor channels include: the time-domain characteristics and frequency-domain characteristics of a single fault under different sensors.

[0021] Specifically, in statistical feature extraction, vibration signals are the most direct physical carrier reflecting the operating state of mechanical equipment. Faults often alter the statistical characteristics of vibration signals; local defects generate transient impacts, changing the amplitude and energy distribution of the signal. Wear or degradation changes the frequency structure of the signal. Therefore, extracting statistical features from both the time and frequency domains allows for a more nuanced portrayal of the physical nature of faults.

[0022] In the time-domain statistical characteristics, the vibration signal sequence for each sensor channel (Length L), this embodiment extracts the following time-domain statistical indicators: Amplitude-related indicators: mean The absolute mean is used to reflect the average amplitude level of a signal. The absolute mean reflects the absolute magnitude of the signal amplitude.

[0023] Root mean square value The effective energy of the signal is reflected, and it is sensitive to global wear-related faults. The root square amplitude, on the other hand, is more sensitive to small amplitude fluctuations. Peak-related metrics: Peak value and peak value It directly reflects the maximum impact amplitude of the signal and is extremely sensitive to transient impacts caused by local defects; Dispersion index: variance and standard deviation Reflects the degree to which the signal fluctuates around the mean. Waveform factor Peak factor Pulse factor Margin factor Dimensionless indicators are insensitive to changes in operating conditions and can stably reflect the presence of faults. Distribution morphology indicators: skewness reflects the asymmetry of signal distribution, kurtosis... Extremely sensitive to impact pulses, it is a classic indicator for measuring the severity of faults. Let L be the mean value and L be the length of the vibration signal. It is a vibration signal sequence. The root mean square value, Peak value, Peak-to-peak value Let Variance be the variance.

[0024] When extracting frequency domain features, the amplitude spectrum is obtained by performing a Fast Fourier Transform on the time domain signal. (corresponding frequency) (Total number of spectral lines is M), and based on this, frequency domain feature statistical indicators are extracted: Spectral energy and centroid: spectral mean Spectral energy reflects the average amplitude level of the spectrum. It reflects the total frequency domain energy of the signal.

[0025] Center of gravity frequency It reflects the concentration position of the signal power spectrum. When a fault causes a shift in the frequency structure, the center frequency will change accordingly.

[0026] Frequency dispersion: spectral variance and frequency standard deviation It reflects the degree of dispersion of spectral energy across the frequency band.

[0027] Spectral morphology indicators include spectral skewness and spectral kurtosis, which reflect the degree of skewness and steepness of the spectral distribution.

[0028] Indicators such as root mean square frequency, frequency-to-mean ratio, and frequency-to-mean square ratio describe the clustering characteristics of the spectrum from different perspectives.

[0029] Characteristic frequency index: spectral peak frequency Directly locating the frequency components with the highest energy in the spectrum is extremely sensitive to the appearance of fault characteristic frequencies. Assuming a multi-source dataset contains K different sensor measurement points (e.g., horizontal, vertical, axial, etc.), given a single fault category... The original multi-channel vibration signal, for any single fault category On the k-th channel, the contents of the extraction can be extracted. The feature vector of a statistical indicator. Let all categories be denoted as . The mean of the j-th statistical indicator extracted from the k-th channel of the sample is In the above formula, Let be the amplitude of the j-th sampling point in the vibration signal sequence acquired by the sensor. Peak value, For waveform factor, As the peak factor, For pulse factor, As a margin factor, The mean of the spectrum. For amplitude spectrum, For spectral energy, For the center of gravity frequency, The standard deviation of the frequency. For spectral variance, For the corresponding frequency.

[0030] In this step, for the sensor in the first... The first channel Each statistical indicator has a continuous semantic vector representation as follows: ; in, For the k-th sensor channel, single fault category The continuous semantic value corresponding to the j-th statistical indicator; —Temperature coefficient ; —On the k-th sensor channel, single fault category Standardized value of the j-th statistical indicator when When the value is large (e.g., 3.0), it tends to become a step function, exhibiting a sharp jump near the zero point, approaching traditional hard thresholding binarization, and weakening the advantages of continuous mapping. When the value is small (e.g., 0.5), the semantic values ​​are concentrated around 0.5, resulting in insufficient differentiation between categories and a decrease in the discriminative power of the semantic vector. When the value is 1.0, the mapping curve achieves the best balance between preserving gradient information and amplifying differences between classes.

[0031] Understandably, by using a sigmoid function with a temperature coefficient instead of the traditional hard thresholding binarization method, the standardized statistical indicators are smoothly mapped to a continuous value range. The optimal value of the temperature coefficient ensures that the mapping curve maintains a moderate steepness near the zero point. Compared to hard thresholding, this avoids the information saturation problem caused by compressing different severity levels of the same positive or negative values ​​into the same discrete label. It allows the evolution of the same fault intensity from mild to severe to be fully preserved in the semantic space as a smoothly increasing continuous value, thereby improving the accuracy and interpretability of the semantic description of the fault evolution trend. Furthermore, it avoids the non-physical phenomenon of discrete label abrupt changes caused by small fluctuations in statistics near the zero point, making the semantic mapping insensitive to small perturbations in the data and enhancing the stability and robustness of the semantic representation. Meanwhile, compared to situations where an excessively large temperature coefficient leads to the curve approximating a step function and losing the advantage of continuous mapping, or an excessively small temperature coefficient results in the semantic value being squeezed into a narrow range and losing inter-class distinguishability, the optimal value of this invention achieves the best balance between preserving the continuous and gradual information of intra-class fault intensity and maintaining the ability to distinguish inter-class fault modes, thus improving the guiding effectiveness of continuous semantic vectors as conditional inputs to the generative model. Furthermore, this invention determines the optimal value of the temperature coefficient through a phased grid search method, using diagnostic accuracy as the primary evaluation index and cosine similarity between semantic vectors as an auxiliary monitoring index. This ensures that the selection of the temperature coefficient has clear engineering operability and verification basis, guaranteeing the adaptability and stability of semantic mapping parameters in tasks with different scarcity levels.

[0032] in, ; —No. The first channel The mean of a statistical indicator across all single fault categories; —No. The first channel The standard deviation of each statistical indicator across all single fault categories; —On the k-th sensor channel, single fault category The j-th statistical indicator.

[0033] The above method can eliminate the dimensional differences between different physical indicators, while preserving the continuous physical strength relationship between different fault types in the feature distribution. It is necessary to perform independent standardization and nonlinear mapping on the extracted indicator matrix.

[0034] S2: Obtain the continuous semantic vector corresponding to each composite fault.

[0035] Specifically, for a target composite fault, the target composite fault is decomposed to obtain a set of single faults corresponding to the target composite fault, and based on the continuous semantic vectors corresponding to the target single faults in the set of single faults, the continuous semantic vectors corresponding to the target composite fault are obtained.

[0036] Specifically, in this step, we assume a composite fault category. A set of single faults Combining, here, compound fault categories The continuous semantic vector corresponding to the j-th statistical index on the k-th sensor channel is as follows: ; in, —Combined fault category The continuous semantic vector corresponding to the j-th statistical index on the k-th sensor channel; —Combined fault categories; For a single target fault A continuous semantic vector in the k-th channel and the j-th statistical index.

[0037] In this application, the feature strength of faults on specific statistical indicators is quantified using each dimension. For composite faults, their performance on this dimension is usually dominated by the most significant of the component faults; therefore, using element-wise maximum values ​​as a semantic synthesis method is a reasonable engineering approximation. This approximation, as prior knowledge, can effectively compensate for statistical distribution bias caused by small sample data.

[0038] S3: Introduce sensor spatial location encoding vector and category attribute vector, and fuse the sensor spatial location encoding vector and the category attribute vector with the continuous semantic vector corresponding to the composite fault category to obtain the spatial perception continuous semantic representation corresponding to the composite fault category.

[0039] Combination Figure 3 Considering the significant spatial heterogeneity of the coupling interference behavior of composite faults at different sensor measurement points (such as horizontal and vertical directions), this application introduces a sensor spatial position encoding vector. The channel to which the current feature belongs is identified by a unique hot code. A strong correlation between "statistical characteristics" and "physical location" is established. Simultaneously, category attribute vectors are introduced. A unique hot code is used to identify the current category in order to enhance the category differentiation of different failure modes.

[0040] The spatially-aware continuous semantic representation corresponding to the composite fault category is as follows: ; —The spatially perceived continuous semantic representation of the composite fault category on the kth sensor channel; —A scalar quantity, positively correlated with the degree of influence of differences in fault categories; —A scalar quantity, positively correlated with the degree of influence of sensor position information; —Category attribute vector; —Fault category; — Fault Category Continuous semantic vectors under the k-th sensor channel; — Concatenation operator.

[0041] Understandably, by extracting multi-dimensional statistical indicators, including amplitude, energy, dispersion, distribution pattern, and spectral features, from both the time and frequency domains, the physical information of the fault contained in the vibration signal is fully explored. This allows the semantic description to comprehensively characterize the different physical aspects of the fault, such as amplitude anomalies, impact responses, and frequency structure shifts, thus improving the completeness of feature representation. By introducing dimensionless indicators such as waveform factors and peak factors, the interference of operating condition changes on feature distribution is effectively reduced, making the semantic description insensitive to changes in operating conditions such as speed and load, thereby improving feature stability across different operating conditions. By performing cross-category standardization on the statistical indicators of each channel, the dimensional differences between different physical indicators are eliminated, while preserving the relative strength ranking of different fault categories on the same statistical indicator. This provides an accurate numerical basis for subsequent semantic mapping and avoids the feature overload problem caused by differences in magnitude. By employing a sigmoid function with a temperature coefficient, standardized statistical indicators are nonlinearly mapped to a continuous value range. The optimal temperature coefficient strikes a balance between preserving gradient information and amplifying inter-class differences in the mapping curve. Compared to traditional hard-threshold binarization, this avoids information saturation caused by compressing different severity levels (both positive or negative) into the same label, and also prevents label abrupt changes caused by minor variations in statistics near zero. This fully preserves the continuous evolution gradient information of faults from mild to severe, improving the accuracy and physical interpretability of semantic descriptions. For scenarios with scarce composite fault samples, the semantics of the composite fault are derived by aggregating the element-wise maximum values ​​of the continuous semantic vectors of each individual fault. The physical prior of the most significant fault dominating the statistical dimension compensates for statistical distribution bias caused by small sample data. This allows composite faults to obtain stable and reliable semantic constraints even in the absence of sufficient real samples, improving semantic usability in zero-sample and extremely small-sample scenarios. By weighted fusion of continuous semantic vectors from each channel with sensor spatial location encoding and category attribute encoding, the semantic description explicitly embeds spatial heterogeneity information of fault features at different measurement points and identity information of fault modes, enhancing the semantic description's ability to distinguish multi-source coupled faults. At the same time, by flexibly adjusting the contribution of the two types of auxiliary information through weight coefficients, the auxiliary information is prevented from excessively dominating the semantic direction, thus improving the discriminativeness and engineering adjustability of the semantic vector.

[0042] S4: Input the spatially perceived continuous semantic representation corresponding to the composite fault category into a one-dimensional convolutional neural network with a wide receptive field to map the spatially perceived continuous semantic representation corresponding to the composite fault category to the latent space, so as to obtain the latent features corresponding to the composite fault category.

[0043] In order to map the high-dimensional original multi-source vibration signal to a low-dimensional latent space with strong discriminativeness, this application adopts a one-dimensional convolutional neural network (Wide CNN1D) based on a wide receptive field. This network uses large-size convolutional kernels to extract features of multiple vibration cycles at the same time, thereby enabling more accurate capture of global impact features caused by mechanical failure.

[0044] As one possible approach, the specific network structure of WideCNN1D in this embodiment is as follows: The first layer is the input layer, which receives the preprocessed frequency domain amplitude spectrum features. The input size is 1×512. This embodiment chooses the frequency domain amplitude spectrum instead of the original time domain signal as the input because the frequency domain representation can explicitly present the frequency components of the fault features on a specific frequency band, reducing the difficulty for the network to learn frequency decomposition from the original time domain signal and making feature extraction more efficient.

[0045] The second layer is a one-dimensional convolutional layer (Conv1d_1) with a kernel size of 64, a stride of 2, padding of 3, and 16 output channels. This layer is the core of WideCNN1D's wide receptive field; the large 64-kernel size covers a long interval in the input sequence, enabling it to perceive the relationships between multiple frequency components at once. This is followed by a batch normalization (BN) layer and a ReLU activation function, and then a max pooling layer (MaxPool1d_1) with a pooling window size of 2 and a stride of 2.

[0046] The third layer is a one-dimensional convolutional layer (Conv1d_2) with a kernel size of 15, a stride of 1, padding of 7, and an output channel count expanded to 32. This layer further extracts higher-level abstract features from the mid-level features, and the kernel size of 15 still maintains a large receptive field in deep networks. It also follows batch normalization, ReLU activation, and max pooling.

[0047] The fourth layer is a one-dimensional convolutional layer (Conv1d_3) with a kernel size of 3, a stride of 1, padding of 1, and an output channel count expanded to 64. This layer uses small-sized convolutional kernels for fine feature extraction, performing local refinement in the high-level feature space.

[0048] The fifth layer is the flatten layer, which unfolds the multi-channel temporal features extracted by the first three convolutional layers into a one-dimensional vector with a dimension of 4096.

[0049] The sixth layer is a fully connected feature layer (FC_feat), which compresses the 4096-dimensional features to 128 dimensions. This 128-dimensional output is defined as a continuous feature latent space. This layer is not only the direct domain of subsequent center loss for performing compact clustering of similar features, but also the space in which the generative model learns multi-source semantic mapping relationships and generates composite fault pseudo-features.

[0050] The seventh layer is a fully connected classification layer (FC_class), which maps the 128-dimensional hidden layer features to all fault categories and outputs the prediction score for each category.

[0051] S5: Input the latent features corresponding to the composite fault category into the multi-source condition combination guided diffusion model based on Bayesian inference to output the guided noise corresponding to each composite fault category, and perform reverse denoising sampling to output composite fault pseudo features.

[0052] Combination Figure 4 Faced with the multi-source concurrent nature of complex faults, traditional early fusion methods often directly input multi-sensor features into complex deep networks for black-box processing. This heuristic approach lacks rigorous theoretical basis and cannot achieve precise coordinated control over the evolution of features from each sensor channel. As a result, when faced with complex faults with extremely scarce samples, the model often struggles to achieve high-quality generalization and feature expansion without retraining the joint network.

[0053] To address the aforementioned issues, this application employs a multi-source conditional combination guided diffusion model based on Bayesian inference to achieve high-quality generalization and feature expansion.

[0054] In this step, at each time step, the following operation is performed until the time step length is reached: First, a first noisy latent feature is obtained and input into the multi-source conditional combination guided diffusion model to output the current guiding noise. Then, a second noisy latent feature is generated based on the first noisy latent feature and the current guiding noise, and the first noisy latent feature is updated based on the second noisy latent feature.

[0055] It should be noted that the first noisy latent feature is the noisy latent feature before the update, while the second noisy latent feature is the noisy latent feature after the update.

[0056] Understandably, the composite fault pseudo-feature is the difference between the noisy latent feature input to the multi-source condition combination guided diffusion model and the guided noise output by the multi-source condition combination guided diffusion model when the time step is reached.

[0057] Specifically, given a certain composite fault, Semantic condition set of each sensor channel The common characteristic is that the formula for the guided noise is as follows: ; —A continuous semantic vector for the k-th sensor channel in a complex fault; —Number of sensor channels; —Weight of the kth sensor channel; —Noisy latent features; —Empty condition; —Time step; —Noise in the continuous semantic vector of the k-th sensor channel in the composite fault at the current time step; —Noise under idle conditions in the k-th sensor channel of the composite fault at the current time step; —The guiding noise of the continuous semantic vector of the k-th sensor channel in the composite fault at the current time step.

[0058] S6: The acquired training set features are merged with the composite fault pseudo-features to generate a balanced dataset, which is then input into the neural network for training.

[0059] Here, when the number of composite fault samples is extremely limited, directly training a deep classification network using scarce real data is prone to severe overfitting, and the model will lose its ability to identify the minority class (composite faults) under the suppression of the majority class (single faults). In order to break the barrier of data imbalance, a trained multi-source combined guided diffusion model (MCCD) is used to generate high-quality composite fault pseudo-features in the latent space, and a balanced dataset is constructed to jointly train the all-class classifier.

[0060] Specifically, for any composite fault category, its corresponding continuous semantic vector is extracted and input into a multi-source combined guided diffusion model with frozen weights to perform inverse denoising sampling. To eliminate class imbalance, the model generates a sufficient number of pseudo-feature samples for each composite fault category. Subsequently, the training set features and the generated composite fault pseudo-features are merged to construct a joint training dataset covering all fault categories. This hybrid dataset effectively corrects the bias in the distribution of composite faults in the original data space, providing sufficient data support for the classifier to delineate the decision boundary.

[0061] The multi-source combined guided diffusion model provided in this application replaces the traditional black-box processing process with rigorous probabilistic derivation, fundamentally overcoming the defects of mutual interference and information overload caused by independent semantics from multiple sources. Based on this derivation paradigm, the joint distribution of multi-source features obtains a clear gradient guidance direction, prompting the generation process to steadily converge to the intersection distribution region that satisfies all channel constraints. This method achieves high-quality generation of composite fault pseudo-features through explicit conditional guidance, laying a solid data foundation for subsequent training of neural networks using real scarce samples.

[0062] The following will describe the training process of the multi-source conditional combined guided diffusion model: Here, for a one-dimensional convolutional neural network, the one-dimensional convolutional neural network is trained with a weighted sum of cross-entropy loss and center loss as the joint optimization objective.

[0063] Reference Figure 5 Specifically, the joint optimization objective function is: ,in, For joint losses, For cross-entropy loss, Loss at the center The hyperparameter is used to balance the weights of the two loss terms.

[0064] ; in, The batch size for model training. Indicates the first The learnable feature centers of each category in the latent space This refers to the hidden layer feature vectors output by the network.

[0065] Here, cross-entropy loss is responsible for "pushing apart" feature clusters of different categories, while centering loss is responsible for "tightening up" feature clusters of the same category. The synergistic effect of the two results in an ideal distribution of features in the latent space that is compact within each category and clear between categories, laying a solid feature space foundation for subsequent diffusion generation and classification diagnosis.

[0066] For the multi-source condition-guided diffusion model, the following formula is used as the optimization objective: ; in, —Used for backpropagation to update network parameters The optimization target value; —The multi-source conditional combined guided diffusion model predicts noise based on the current input; —Hidden features of the output of a one-dimensional convolutional neural network; —A continuous semantic vector for the k-th sensor channel in a complex fault; —Regarding the loss function within the parentheses in the variables , , and Take the batch average on the joint distribution; —During the positive noise addition process, the system randomly generates and actually injects noise. The noise tensor in; —Cumulative signal retention coefficient; -norm.

[0067] This application has at least the following beneficial effects: 1. A framework for diagnosing composite faults in scarce samples based on multi-source semantic combination-guided diffusion is proposed. Addressing the problems of extremely scarce composite fault samples and severe class imbalance, a diffusion model is used to generate high-quality pseudo-features for scarce composite faults at the feature level, providing effective data support for composite fault identification under extremely small sample conditions.

[0068] 2. A spatially perceptual continuous semantic representation and feature compact clustering strategy is proposed. By using continuous semantic modeling, the smoothing of information on continuous changes in fault intensity by traditional discrete semantics is avoided. Sensor spatial location encoding is introduced to establish the association between "statistical features - spatial orientation - fault mode". At the same time, the feature space is optimized by combining cross-entropy loss and center loss, which enhances intra-class compactness and inter-class separability.

[0069] 3. A multi-source conditional combination-based fault diagnosis diffusion model based on Bayesian inference is proposed. This model theoretically characterizes the joint distribution generation mechanism under independent conditions of multiple source sensors. It achieves collaborative generation of multi-source features through the combination of noise predictions from each independent channel, thereby effectively alleviating the problems of semantic interference and information loss in traditional multi-source fusion.

[0070] The following is a specific example: The experiment used two sets of bearing datasets. One set of datasets came from the bearing fault experimental platform, with a sampling frequency of 10kHz, and included normal, inner ring fault, outer ring fault, rolling element fault, cage fault, and various compound faults. The specific categories are shown in Table 1.

[0071] Table 1

[0072] Another dataset was collected using the SpectraQuest mechanical fault simulator at a sampling frequency of 25.6kHz. It includes normal conditions, inner race faults of varying severity, outer race faults, rolling element faults, and compound faults, as shown in Table 2.

[0073] Table 2

[0074] To address the scarcity of composite fault samples, this paper constructs three training tasks with different levels of scarcity, and the specific sample size settings are shown in Table 3. Specifically, the single-fault training samples for Tasks A, B, and C are all set to 3200 per class, while the composite fault training samples are set to 500, 50, and 10 per class, respectively, to simulate different scenarios ranging from relatively abundant to extremely scarce.

[0075] Table 3

[0076] Both datasets employ a consistent task partitioning principle. To ensure consistency in experimental procedures across different datasets, the same data preprocessing strategy is used in both applications. First, the original vibration signals are sampled using a sliding window with a window length of 1024 and a coverage rate of 50%. Then, a Fast Fourier Transform (FFT) is performed on each sample to extract the first half of the amplitude spectrum's frequency components, which are then normalized and scaled to 512 dimensions before being used as network input. To mitigate the impact of insufficient composite fault samples on feature learning stability under extremely small sample conditions, this application standardizes the composite fault samples in each task before training by adding noise and perturbation, ensuring they remain at a similar order of magnitude during training. This reduces the interference of differences in basic sample size on performance evaluation.

[0077] During model training, both the feature extraction network and the diffusion network use the Adam optimizer, with learning rates set to [values ​​to be filled in]. and The batch size was set to 64, and the learning rate remained fixed during training. The feature extraction network was trained for 30 epochs, and the diffusion network for 500 epochs, with a diffusion time step of 1000. Subsequently, 500 pseudo-feature samples were generated for each type of composite fault, and a full-class classifier was trained for 30 epochs based on these samples. To reduce the influence of random factors, all comparative and ablation experiments were repeated 5 times, and the average accuracy was used as the evaluation metric. To fully verify the effectiveness and advancement of the proposed method MSCGD in scenarios with scarce composite fault samples, this application set up three diagnostic tasks with different levels of sample scarcity based on two composite fault datasets and conducted a systematic comparison with existing advanced intelligent diagnostic methods. Specifically, the number of composite fault samples in the training set was gradually reduced from a relatively balanced 500 to 50 and 10 to construct scenarios with different levels of sample scarcity. The specific task settings are shown in Table 5. To objectively evaluate the comprehensive diagnostic performance of each method, this paper uses the average accuracy of 5 repeated experiments as the core evaluation metric.

[0078] This application compares MAACCN, DACN, LCM-GCN, MHTA-DDPM, and DCGAN. MAACCN combines one-dimensional convolution, an efficient channel attention mechanism, and capsule networks to achieve multi-sensor feature fusion and deep feature extraction; DACN achieves decoupled fault diagnosis for multi-domain generalization tasks through collaborative modeling of feature extractors, decoupled classifiers, and multi-domain classifiers; LCM-GCN uses static and dynamic graph convolutional networks to characterize label correlations and combines a multi-head attention mechanism to achieve composite fault identification; MHTA-DDPM enhances scarce samples by improving the denoising diffusion probability model and improves generation quality using a multi-head transposed attention mechanism; and DCGAN uses continuous wavelet transform, deep convolutional generative adversarial networks, and the large-kernel convolutional network RepLKNet to achieve composite fault sample enhancement and high-precision classification.

[0079] The diagnostic results for the first set of data are shown in Table 4: Table 4

[0080] Combine Table 4 and Figure 6 In Task A (a scenario with sufficient composite fault samples, i.e., a relatively ideal balanced data scenario), due to the training data volume being able to fully cover the feature distribution characteristics of various faults, all models except DACN (which only achieved 93.90%) demonstrated excellent diagnostic capabilities, with accuracy generally exceeding 99%. Among them, the proposed MSCGD method achieved a complete recognition accuracy of 100.00%, proving its effectiveness under normal conditions.

[0081] However, with the sharp decrease in the number of composite fault samples, especially under the most demanding task C, the performance of both traditional discriminative models and conventional generative models has declined to varying degrees.

[0082] The proposed MSCGD method still achieves an outstanding diagnostic accuracy of 98.75% in the most extreme task C, significantly outperforming all comparative models and fully demonstrating its significant advantages in scarce and complex fault data scenarios.

[0083] The diagnostic results for the second set of data are shown in Table 5: Table 5

[0084] Observe Table 5 and Figure 7 As can be seen, the overall accuracy of the three models, DACN, DCGAN, and MAACCN, mostly hovers between 70% and 80% across various tasks, yet they still maintain a diagnostic accuracy of 97.67%, demonstrating excellent small-sample diagnostic capabilities and cross-device generalization ability. This significant performance advantage directly confirms the effectiveness of this application in dealing with scarce and complex fault samples.

[0085] Reference Figure 8 The confusion matrix given in the image shows the model's predicted class on the horizontal axis and the true class of the sample on the vertical axis. The values ​​in the matrix represent the proportion of samples of the corresponding class. Figure 8 As shown in (a), while the CNN baseline model can accurately identify normal states and various single faults, its ability to identify compound faults is weak. For example, 78.5% and 83.8% of samples with the true categories of "inner ring + outer ring" and "inner ring + rolling element" were misclassified as outer ring faults, respectively. The "outer ring + rolling element" fault was also mainly misclassified as inner ring faults and outer ring faults. This indicates that without sample expansion, the classification model is easily dominated by the majority of single fault samples and it is difficult to form an effective decision boundary for compound faults. This demonstrates the superiority of our method (MSCGD) in distinguishing various faults (especially compound faults).

[0086] from Figure 9 As can be seen in (a), although the CNN baseline model can form relatively independent clusters for some single faults, the features of multiple composite faults overlap significantly in local areas, and the inter-class boundaries are blurred. This indicates that under extreme imbalanced sample conditions, the model struggles to effectively decouple the coupling features of different composite faults. Figure 9As can be seen in (b), after sample expansion, MHTA-DDPM significantly improves the separation of various features, and composite faults begin to form relatively independent clusters. However, some feature clusters still exhibit a relatively obvious elongated distribution, with significant intra-class dispersion and a small amount of class overlap, indicating that conventional diffusion generation methods are still insufficient in characterizing the distribution of complex composite faults. In contrast, Figure 9 As shown in (c), the different categorical features extracted by MSCGD have higher intra-class compactness and inter-class separation, forming clear distribution boundaries between each fault category. Composite fault features can also stably cluster in the corresponding category region. By comparing the clustering effects of each model, the superiority of the proposed method (MSCGD) in the feature space is intuitively demonstrated.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for diagnosing composite faults in scarce samples based on multi-source semantic combination-guided diffusion, characterized in that, include: S1: Obtain the feature performance of all single faults under different sensor channels, and construct a continuous semantic vector corresponding to each single fault based on the feature performance of the single faults under different sensor channels; S2: Obtain the continuous semantic vector corresponding to each composite fault, wherein, for the target composite fault, the target composite fault is decomposed to obtain the set of single faults corresponding to the target composite fault, and the continuous semantic vector corresponding to the target composite fault is obtained based on the continuous semantic vectors corresponding to the target single faults in the set of single faults. S3: Introduce sensor spatial location encoding vector and category attribute vector, and fuse the sensor spatial location encoding vector and the category attribute vector with the continuous semantic vector corresponding to the composite fault category to obtain the spatial perception continuous semantic representation corresponding to the composite fault category. S4: Input the spatially perceived continuous semantic representation corresponding to the composite fault category into a one-dimensional convolutional neural network with a wide receptive field to map the spatially perceived continuous semantic representation corresponding to the composite fault category to the latent space, so as to obtain the latent features corresponding to the composite fault category. S5: Input the latent features corresponding to the composite fault category into the multi-source condition combination guided diffusion model based on Bayesian inference to output the guided noise corresponding to each composite fault category, and perform reverse denoising sampling to output composite fault pseudo features; S6: Merge the acquired training set features with the composite fault pseudo-features to generate a balanced dataset, and input it into the neural network for training. The guiding noise is determined using the following formula: ; —A continuous semantic vector for the k-th sensor channel in a complex fault; —Number of sensor channels; —Weight of the kth sensor channel; —Noisy latent features; —Empty condition; —Time step; —Noise in the continuous semantic vector of the k-th sensor channel in the composite fault at the current time step; —Noise under idle conditions in the k-th sensor channel of the composite fault at the current time step; —The guiding noise of the continuous semantic vector of the k-th sensor channel in the composite fault at the current time step.

2. The method according to claim 1, characterized in that, In the step of obtaining the feature performance of all single faults under different sensor channels, and constructing a continuous semantic vector corresponding to each single fault based on the feature performance of the single faults under different sensor channels; The characteristics of a single fault under different sensor channels include: the time-domain characteristics and frequency-domain characteristics of a single fault under different sensors; The continuous semantic vector corresponding to the single fault is constructed in the following manner: Regarding the sensor in the first The first channel Each statistical indicator has a continuous semantic vector representation as follows: ； in, For the k-th sensor channel, single fault category The continuous semantic value corresponding to the j-th statistical indicator; —Temperature coefficient ; —On the k-th sensor channel, single fault category The standardized value of the j-th statistical indicator; in, ; —No. The first channel The mean of a statistical indicator across all single fault categories; —No. The first channel The standard deviation of each statistical indicator across all single fault categories; —On the k-th sensor channel, single fault category The j-th statistical indicator.

3. The method according to claim 1, characterized in that, In the step of obtaining the continuous semantic vector corresponding to each composite fault, the continuous semantic vector corresponding to the target composite fault is obtained in the following manner: For the target composite fault in the first The continuous semantic vectors under each channel, based on the target single fault in the single fault set at the first... The maximum value of the continuous semantic vector under each channel is determined.

4. The method according to claim 1, characterized in that, In the step of introducing a sensor spatial location encoding vector and an input category attribute vector, and fusing the sensor spatial location encoding vector and the category attribute vector with the continuous semantic vector corresponding to the composite fault category to obtain the spatial perception continuous semantic representation corresponding to the composite fault category... The spatially-aware continuous semantic representation corresponding to the composite fault category is as follows: ; —The spatially perceived continuous semantic representation of the composite fault category on the kth sensor channel; —A scalar quantity, positively correlated with the degree of influence of differences in fault categories; —A scalar quantity, positively correlated with the degree of influence of sensor position information; —Category attribute vector; —Fault category; — Fault Category Continuous semantic vectors under the k-th sensor channel; — Concatenation operator.

5. The method according to claim 4, characterized in that, Before the step of inputting the spatially perceived continuous semantic representation corresponding to the composite fault category into a one-dimensional convolutional neural network with a wide receptive field to map the spatially perceived continuous semantic representation corresponding to the composite fault category to the latent space to obtain the latent features corresponding to the composite fault category, the method further includes: The one-dimensional convolutional neural network is trained using a weighted sum of cross-entropy loss and center loss as the joint optimization objective.

6. The method according to claim 5, characterized in that, In the step of inputting the latent features corresponding to the composite fault categories into the multi-source conditional combination guided diffusion model based on Bayesian inference to output the guided noise corresponding to each composite fault category, and performing inverse denoising sampling to output the pseudo-features of the composite faults, the corresponding pseudo-features are output in the following manner for any composite fault: At each time step, perform the following operations until the time step length is reached: Obtain the first noisy latent feature and input the first noisy latent feature into the multi-source condition combination guided diffusion model, and output the current guided noise; A second noisy latent feature is generated based on the first noisy latent feature and the current guiding noise, and the first noisy latent feature is updated based on the second noisy latent feature; The composite fault pseudo-feature is the difference between the noisy latent feature input to the multi-source condition combined guided diffusion model and the guided noise output by the multi-source condition combined guided diffusion model when the time step is reached.

7. The method according to claim 6, characterized in that, The multi-source condition combination guided diffusion model uses the following formula as its optimization objective: ; in, —Used for backpropagation to update network parameters The optimization target value; —The multi-source conditional combined guided diffusion model predicts noise based on the current input; —Hidden features of the output of a one-dimensional convolutional neural network; —A continuous semantic vector for the k-th sensor channel in a complex fault; —Regarding the loss function within the parentheses in the variables , , and Take the batch average on the joint distribution; —During the positive noise addition process, the system randomly generates and actually injects noise. The noise tensor in; —Cumulative signal retention coefficient; -norm.

8. The method according to claim 7, characterized in that, In the step of merging the acquired training set features with the composite fault pseudo-features and inputting them into the neural network for training, The predicted probabilities of each category are output by the Softmax activation function.

Images