A small sample condition under the die forging press fault diagnosis method

Abstract

The present application relates to the field of die forging press fault diagnosis, and more particularly to a die forging press fault diagnosis method under small sample conditions, comprising: S1, obtaining fault window samples under various types of typical faults, and setting fault category labels; S2, for each type of typical fault, respectively constructing a conditional diffusion generation network and training to generate corresponding augmented fault window samples; S3, obtaining an augmented sample set; S4, pre-training an abnormality detection model by performing abnormality detection pre-training on a preset time sequence convolution network; S5, inputting each sample in the augmented sample set into the pre-trained abnormality detection model to extract a high-dimensional transfer feature vector; and S6, training a random forest classifier using the high-dimensional transfer feature vector as training data to obtain a fault diagnosis model. The present application can realize high-quality generation of die forging press fault samples and effective transfer of multi-dimensional features under small sample and class imbalance conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of forging presses, and more particularly to a fault diagnosis method for forging presses under small sample conditions. Background Technology

[0002] The 800MN die forging press, as a core piece of equipment in the aerospace and heavy equipment manufacturing fields, directly affects production safety and product quality. However, in actual industrial settings, this type of large equipment operates under normal conditions for extended periods, with extremely low failure rates. This results in a severe scarcity of tagged fault samples, and a significant imbalance in the number of typical faults. This small sample size and class imbalance present a significant challenge to fault diagnosis. Traditional fault diagnosis methods often rely on discriminative models such as support vector machines and neural networks. The performance of these models is largely limited by the scale and quality of the labeled data, making them prone to overfitting when the sample size is insufficient, and hindering their ability to maintain stable generalization.

[0003] To alleviate the problem of data scarcity, existing technologies often employ data augmentation or generative models to expand samples. However, for the nonlinear, strongly coupled, and highly dynamic characteristics of multivariate time-series signals from forging presses, existing sample expansion methods still have significant shortcomings: On the one hand, traditional augmentation methods based on geometric transformations or signal processing can only change the surface morphology of the signal and cannot learn and characterize the complex probability distribution inherent in the fault data; on the other hand, while deep generative models such as Generative Adversarial Networks (GANs) possess certain distribution learning capabilities, they suffer from poor training stability under small sample conditions, easily experiencing pattern collapse or gradient vanishing, resulting in insufficient diversity and poor physical consistency of generated samples, and even introducing false features that mislead subsequent classifier training. Furthermore, even if some expanded samples are obtained, how to fully extract prior knowledge from normal data under small sample conditions and effectively transfer it to the fault identification task to compensate for the sparsity of the fault feature space is also a challenge currently faced by the technology. Due to the lack of an effective feature transfer mechanism, existing methods often treat sample generation and fault classification separately, failing to form end-to-end collaborative optimization and limiting the improvement of overall diagnostic accuracy.

[0004] In summary, when dealing with small-sample fault diagnosis of 800MN die forging presses, existing technologies are unable to generate high-quality fault samples that conform to real physical laws, nor can they make full use of the rich data under normal operating conditions for knowledge transfer. As a result, the robustness and accuracy of the diagnostic model in scenarios with few samples and imbalance cannot meet engineering requirements.

[0005] Therefore, how to achieve high-quality generation of fault samples of forging presses and effective transfer of multi-dimensional features under conditions of small sample size and class imbalance, so as to significantly improve the generalization performance and recognition accuracy of fault diagnosis models, has become an urgent problem to be solved. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a fault diagnosis method for forging presses under small sample conditions. This method aims to achieve high-quality generation of fault samples and effective transfer of multi-dimensional features of forging presses under small sample and class imbalance conditions, thereby significantly improving the generalization performance and recognition accuracy of the fault diagnosis model.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A fault diagnosis method for a forging press under small sample conditions includes the following steps:

[0009] S1. Obtain fault window samples of the forging press under various typical faults, and set corresponding fault category labels for each fault window sample;

[0010] S2. For each type of typical fault, a dedicated conditional diffusion generation network is constructed and trained using the fault window samples of the typical fault obtained in S1. The trained conditional diffusion generation networks are used to generate expanded fault window samples of the corresponding typical faults.

[0011] The structure of each conditional diffusion generation network is consistent, with Diffusion Transformer as the backbone. The network internally uses an adaptive layer normalization module to embed the corresponding fault category label as a condition into each layer of the network. During training, a cosine noise scheduling strategy is used to control the noise addition and denoising process. When generating the expanded fault window samples, a classifier-free guidance mechanism is used to enhance the class consistency of the generated samples.

[0012] S3. Merge the expanded fault window samples generated in S2 with the fault window samples obtained in S1 to obtain the expanded sample set.

[0013] S4. Obtain the multivariate time-series signal of the forging press under normal working conditions, perform anomaly detection pre-training on the preset time-series convolutional network, and obtain a pre-trained anomaly detection model containing the backbone of the time-series convolutional network and a dual-track pooling layer.

[0014] S5. After freezing the backbone parameters of the temporal convolutional network of the pre-trained anomaly detection model in S4, input each sample from the expanded sample set in S3 into the pre-trained anomaly detection model and extract the high-dimensional transfer feature vector output by the dual-track pooling layer of each sample.

[0015] S6. Using the high-dimensional transfer feature vectors of each sample extracted in S5 and the fault category labels of each sample as training data, train the preset random forest classifier to obtain the fault diagnosis model.

[0016] S7. Input the window sample of the forging press to be diagnosed into the pre-trained anomaly detection model in S4, extract its high-dimensional transfer feature vector and input it into the fault diagnosis model in S6 to obtain the fault type identification result.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. Improved quality and stability of fault data generation under small sample conditions. In existing technologies, Generative Adversarial Networks (GANs) are prone to pattern collapse during small sample training, while traditional data augmentation methods struggle to capture the complex distribution of multivariate time-series signals. This solution employs a conditional diffusion generative network with Diffusion Transformer as its backbone, combined with a cosine noise scheduling strategy. Compared to GANs, this results in a more stable training process and superior distribution coverage. Simultaneously, by deeply embedding fault category labels into each layer of the network through an adaptive layer normalization module, and coupled with a classifier-free guidance mechanism, the consistency of generated sample categories and physical realism are effectively enhanced, avoiding the generation of false or ambiguous features.

[0019] 2. This approach effectively transfers prior knowledge of normal operating conditions to fault diagnosis tasks. Existing methods typically separate sample generation from feature extraction, and directly training classifiers with small sample sizes easily leads to overfitting. This scheme utilizes a large amount of readily available normal operating condition signals to pre-train a temporal convolutional network for anomaly detection, enabling the model to learn the general temporal patterns and dynamic characteristics of equipment operation in advance. Subsequently, the backbone parameters are frozen, and only dual-track pooling layers are used to extract high-dimensional transfer features from the augmented samples. This method fully leverages the rich information in normal data to compensate for the feature sparsity of fault samples. Compared to training a fault classification model directly from scratch, it significantly reduces the model's dependence on labeled data and enhances the discriminative power and robustness of the features.

[0020] 3. Enhanced generalization ability of the fault diagnosis model in imbalanced scenarios. Traditional discriminative models often favor the majority class when samples are extremely imbalanced, resulting in low identification rates for minority class faults. This solution fundamentally alleviates the class imbalance problem by constructing a dedicated conditional diffusion generative network for each typical fault class. Simultaneously, it employs a random forest as the downstream classifier, combined with high-quality transfer features for training, which exhibits stronger anti-overfitting ability and greater tolerance for small sample classes compared to a single neural network classifier, thus ensuring balanced and stable performance across various fault identification types.

[0021] In summary, this method can achieve high-quality generation of fault samples from forging presses and effective transfer of multi-dimensional features under conditions of small sample size and class imbalance, thereby significantly improving the generalization performance and recognition accuracy of the fault diagnosis model.

[0022] Preferably, in S2, the forward diffusion process of the conditional diffusion generation network gradually moves towards the original fault window samples within T steps. By injecting Gaussian noise and utilizing reparameterization techniques, the noisy result at any step can be directly obtained from the original fault window samples. The closed form is:

[0023] ;

[0024] ;

[0025] in, For cumulative signal-to-noise ratio; The noise intensity at step s is given by the cosine noise scheduling strategy; For noise variables; Representing noise variables It follows a standard multivariate Gaussian distribution with a mean of 0 and a covariance matrix of identity; t represents the current time step in the forward diffusion process.

[0026] This setup significantly improves the training stability and sample fidelity of the diffusion model in generating small-sample time-series signals. Compared to traditional diffusion models that directly discretely sample the noise variance or use fixed scheduling, this scheme explicitly models the noise addition process as an analytical combination of the signal-to-noise ratio accumulation function and standard Gaussian noise through reparameterization, avoiding training instability caused by numerical integration errors and random sampling fluctuations; simultaneously, because It can change smoothly over time, making the noise path more consistent with the physical signal degradation law. Thus, under the constraint of limited fault samples, it can still generate extended samples with coherent temporal structure, reasonable amplitude-frequency characteristics and clear category semantics, laying a reliable data foundation for subsequent high-precision fault identification.

[0027] Preferably, in S2, the cosine noise scheduling strategy defines the cumulative signal-to-noise ratio using a cosine function. Attenuation curve:

[0028] ;

[0029] in, To prevent exist The offset constant is too small; This represents the total number of steps in the diffusion process.

[0030] This setup achieves a smooth, non-linear decay of the signal-to-noise ratio during the noise addition process. It avoids premature loss of the original signal information due to a sudden drop in the signal-to-noise ratio in the initial stage, while ensuring sufficient perturbation capability in the later high-noise stage. This allows for more accurate modeling of the multi-scale degradation characteristics of the forging press fault time-series signal under small sample conditions. Compared to traditional strategies such as linear or square root scheduling, cosine scheduling... In the early stages, changes are slow, preserving more details of the original fault characteristics; in the middle stages, the decay accelerates, promoting pattern discovery; and in the final stages, it approaches zero, ensuring noise dominance. This dynamic characteristic of slow-to-rapid decay aligns better with the physical law of gradual damage leading to sudden failure in the failure evolution of mechanical systems; combined with the offset constant... The introduction of [something] can effectively prevent [something]. hour The problem of initial gradient explosion or generation distortion caused by approaching 0 significantly improves the training convergence and generation fidelity of the diffusion model with a very small number of fault samples.

[0031] Preferably, in S2, when training the conditional diffusion generation network, fault category labels are used. The training objective is to minimize the conditional noise prediction error on various fault samples, and the loss function is:

[0032] ;

[0033] ;

[0034] in, Gaussian noise added during the forward diffusion process; The noise predicted by the denoising network; For tags The actual data distribution corresponding to the fault categories.

[0035] This setup, by explicitly guiding the noise prediction task with category labels as conditions, enables the diffusion generation network to learn category-specific temporal degradation patterns even with small sample sizes, significantly improving the semantic consistency and physical plausibility of generated samples. Compared to unconditional diffusion or weakly conditional modeling methods that only embed labels in the input layer, this scheme deeply couples category information to the noise prediction process, forcing the network to adjust its denoising strategy based on the fault type at each diffusion step. This fine-grained conditional constraint effectively alleviates the category confusion problem caused by data sparsity in small sample scenarios, avoids generating cross-class mixed samples, and thus provides high-fidelity, highly discriminative extended data support for subsequent classification tasks.

[0036] Preferably, in S2, the conditional diffusion generation network uses a Diffusion Transformer as its backbone. The Diffusion Transformer is composed of multiple DiT Blocks connected in series. Each DiT Block includes a self-attention submodule, a feedforward neural network submodule, and two adaptive layer normalization modules.

[0037] The adaptive layer normalization module embeds the corresponding fault category label as a condition into each layer of the network, including:

[0038] The current denoising step is converted into a temporal embedding vector using sine-cosine position encoding. The target fault category label is obtained as a category embedding vector by looking up a table using a learnable embedding matrix. The two are concatenated and input into a multilayer perceptron to obtain a conditional vector. :

[0039] ;

[0040] in, Standard sine-cosine position encoding; The fault category embedding vector is obtained by looking up a table of the learnable matrix; the MLP is a two-layer nonlinear unfolded network; the subscript t represents the current time step in the forward diffusion process;

[0041] The condition vector The scaling, translation, and gating parameters are generated through a zero-initialization linear transformation layer.

[0042] In the adaptive layer normalization module of each DiT Block, the input features are first normalized using a standard layer, then the normalized features are scaled element-wise using the scaling parameter, translated element-wise using the translation parameter, and the overall contribution of the DiT Block is adjusted using the gating parameter to obtain the modulated feature output.

[0043] This setup enables the joint integration of fault category semantics and diffusion time steps, along with fine-grained feature control. This allows the network to adaptively strengthen or suppress key temporal patterns related to the current fault type under varying noise levels, maintaining category specificity and physical consistency in the generation process even with minimal sample constraints. Compared to traditional methods that simply concatenate labels to the input or use simple attention guidance, this scheme injects conditional information into the normalization stage of each DiT Block. This avoids the weakening of high-level semantics caused by shallow embedding and overcomes the limitation of global conditional vectors being difficult to adapt to multi-scale features. Especially in highly nonlinear systems like forging presses, different faults should exhibit differentiated frequency domain responses and transient characteristics under the same noise intensity. AdaLN, by dynamically adjusting the mean and variance of features in each layer, can accurately guide the network to focus on the class discriminative subspace, significantly reducing the risk of inter-class confusion in generated samples and enhancing the effective support capability of expanded data for downstream classification tasks.

[0044] Preferably, in S2, the classifier-free guidance mechanism used in generating expanded fault window samples has the following noise prediction calculation formula during the inference phase:

[0045] ;

[0046] in, For guiding strength; This is the prediction result for unconditional noise; This is the result of conditional noise prediction;

[0047] During the training phase, fault categories are randomly labeled with preset probabilities. Replace it with the null flag to enable the network to estimate noise with both conditional and unconditional conditions.

[0048] This setup, without introducing an additional classifier, dynamically enhances the semantic focus of the generation process on fault categories through linear interpolation, significantly improving the class fidelity and detail clarity of generated samples with small sample sizes, while avoiding common problems such as unstable classifier training and gradient conflicts.

[0049] Preferably, in S2, the sampling process for generating expanded fault window samples is accelerated by using the denoising diffusion implicit model DDIM, and the number of sampling steps is shortened to 100 steps by constructing a deterministic inverse Markov chain.

[0050] This setup increases the inference sampling speed by more than an order of magnitude without sacrificing the quality of the generated data, making the diffusion model feasible for engineering implementation in industrial fault data expansion scenarios with high real-time requirements.

[0051] Preferably, in S4, the method for obtaining the pre-trained anomaly detection model is as follows: obtain the multivariate time-series signal of the forging press under normal working conditions, and obtain the normal working condition window sample after sliding window segmentation; use the normal working condition window sample as the training set to perform anomaly detection pre-training on the preset temporal convolutional network, and after training, only retain the parameters of the backbone of the temporal convolutional network and the dual-track pooling layer to obtain the pre-trained anomaly detection model;

[0052] The dual-track pooling layer includes a global max pooling branch and a global average pooling branch set in parallel. The high-dimensional transfer feature vector is a 512-dimensional feature vector formed by concatenating the peak response feature extracted by the global max pooling branch and the global statistical mean feature extracted by the global average pooling branch.

[0053] This setup, by using only normal data for pre-training and freezing the backbone network parameters, achieves efficient representation learning of implicit degradation patterns in industrial equipment, significantly improving the sensitivity and generalization ability of anomaly detection with small or zero samples, while avoiding the risk of model bias and overfitting caused by introducing abnormal samples.

[0054] Preferably, in S6, when training the random forest classifier, the class weights are allocated according to the reciprocal of the number of samples in each fault category, so that the minority category samples obtain an equal contribution to the impurity calculation when splitting nodes.

[0055] This setup effectively alleviates the typical long-tail distribution problem in industrial fault data. Without relying on data-level operations such as oversampling or undersampling, it significantly improves the recall rate of minority class faults through a loss function-level class reweighting mechanism, while maintaining the stability of overall classification accuracy.

[0056] Preferably, in S1, the fault window sample is a multivariate time series signal segmented by a sliding window; the window length of the sliding window is consistent with the input dimension of the conditional diffusion generation network and the input dimension of the pre-trained anomaly detection model.

[0057] The typical fault types include pressure / leakage faults, speed / servo faults, and tilt / motion faults.

[0058] This setup, by unifying the input dimensions and sample construction paradigm, eliminates information distortion and model degradation caused by scale or structural mismatches between modules, enabling the generated expanded samples to be seamlessly embedded into the downstream detection process while maintaining semantic consistency. This significantly improves the end-to-end effectiveness and reliability of the entire intelligent diagnostic system in small-sample fault scenarios. Attached Figure Description

[0059] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0060] Figure 1 This is a flowchart of the method;

[0061] Figure 2 This is a schematic diagram of the conditional diffusion generation network in Example 1;

[0062] Figure 3 This is an expanded flowchart of the fault sample of the forging press in Example 1;

[0063] Figure 4 This is a schematic diagram comparing the TimeGAN data augmentation waveforms in Example 2;

[0064] Figure 5 A schematic diagram of the t-SNE projection of TimeGAN data augmentation quality in Example 2;

[0065] Figure 6 This is a schematic diagram comparing the augmented waveforms of the conditional DDPM data in Example 2;

[0066] Figure 7 Enhance the quality t-SNE projection of the conditional DDPM data in Example 2;

[0067] Figure 8 This is a schematic diagram illustrating the accuracy of different fault diagnosis methods in Example 2;

[0068] Figure 9 This is a Macro-F1 schematic diagram of different fault diagnosis methods in Example 2;

[0069] Figure 10 This is a comparative schematic diagram of the t-SNE projection diagram of the diffusion and expansion fault diagnosis method in Example 2;

[0070] Figure 11 This is a schematic diagram comparing the 5-fold CV confusion matrix of the diffusion expansion fault diagnosis method in Example 2. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0073] Example 1

[0074] like Figure 1 As shown, this invention provides a fault diagnosis method for forging presses under small sample conditions, comprising the following steps:

[0075] S1. Obtain fault window samples of the forging press under various typical faults, and set corresponding fault category labels for each fault window sample;

[0076] The typical fault types include pressure / leakage faults, speed / servo faults, and tilt / motion faults; the fault window samples are multivariable time-series signals segmented by a sliding window.

[0077] To facilitate better understanding, the following explanation is provided.

[0078] In this method, based on actual production site maintenance records, forging press faults are categorized into three types: pressure / leakage faults, speed / servo faults, and tilt / motion faults. These three types of faults exhibit distinct observable characteristics on the core variable set obtained after two-stage variable reduction.

[0079] (1) Pressure / Leakage Faults: The significant signal characteristic of this type of fault appears in the pressure of the four master cylinders. Under normal operating conditions, the pressures of the four master cylinders are highly correlated, with similar amplitudes and almost zero difference between variables; when this type of fault occurs, the pressures of the four corners become unbalanced, and the difference between variables increases significantly from a near-zero state, forming a separation structure of amplitude between variables that does not exist under normal operating conditions. At the same time, the balance cylinder pressure also shows amplitude change response in this type of fault.

[0080] (2) Speed / Servo-type faults: The significant signal characteristics of this type of fault appear in the suppression speed and the target speed. Under normal operating conditions, the actual suppression speed closely follows the suppression target speed, and the deviation between the two is small; when this type of fault occurs, the ability of the actual suppression speed to follow the target suppression speed decreases, the control deviation between the two variables continues to widen, and may be accompanied by oscillations in the actual suppression speed. The amplitude changes of other variables in this type of fault are relatively limited.

[0081] (3) Tilt / Motion Faults: The significant signal characteristics of this type of fault are manifested in the tilt angle variable, with an observable and continuous deflection of the tilt angle in the AD or BC direction; at the same time, the pressure of the four corner master cylinders also shows an asymmetrical distribution response, manifested as an increase in the pressure difference between the diagonal variables. This type of fault generates responses on multiple variable groups simultaneously, and the signal pattern differs from that of pressure / leakage faults in terms of variable manifestation: the former is dominated by the continuous deflection of the tilt angle variable, while the latter is dominated by a single-corner pressure drop.

[0082] Although the three types of faults exhibit overlapping signal responses on some variables, the principal discriminant variable sets and their signal change patterns differ for each type of fault, making the three types of samples separable in the feature space. Therefore, if samples from different fault categories are indiscriminately expanded, the generative model will be forced to fit a mixed distribution of multiple fault types, failing to accurately capture the independent statistical characteristics of each category, and even generating overlapping samples across categories. Therefore, fault sample expansion must use fault type labels as explicit conditions to ensure that each synthetic sample strictly conforms to the true signal distribution of the corresponding fault category.

[0083] S2. For each type of typical fault, a dedicated conditional diffusion generation network is constructed and trained using the fault window samples of the typical fault obtained in S1. The trained conditional diffusion generation networks are used to generate expanded fault window samples of the corresponding typical faults.

[0084] The structure of each conditional diffusion generation network is consistent, with Diffusion Transformer as the backbone. The network internally uses an adaptive layer normalization module to embed the corresponding fault category label as a condition into each layer of the network. During training, a cosine noise scheduling strategy is used to control the noise addition and denoising process. When generating the expanded fault window samples, a classifier-free guidance mechanism is used to enhance the class consistency of the generated samples.

[0085] The window length of the sliding window is consistent with the input dimension of the conditional diffusion generation network and the input dimension of the pre-trained anomaly detection model.

[0086] Conditional Diffusion Generator Network (i.e., Conditional DDPM) consists of two stages: a forward diffusion process and a backward denoising process. The forward process gradually destroys the fault window into Gaussian noise, while the backward process learns the distribution of the fault signal from the noise based on the fault category label.

[0087] In practice, the forward diffusion process of the conditional diffusion generation network gradually moves towards the original fault window samples within T steps. By injecting Gaussian noise and utilizing reparameterization techniques, the noisy result at any step can be directly obtained from the original fault window samples. The closed form is:

[0088] ;

[0089] ;

[0090] in, For cumulative signal-to-noise ratio; The noise intensity at step s is given by the cosine noise scheduling strategy; For noise variables; Representing noise variables It follows a standard multivariate Gaussian distribution with a mean of 0 and a covariance matrix of identity; t represents the current time step in the forward diffusion process.

[0091] In specific implementation, settings can be made This refers to the 128 time steps and 12-dimensional variable fault window samples after sliding window segmentation. The 12-dimensional variables include the main cylinder pressure at angles A / B / C / D, the pressing speed, the pressing target speed, the AD / BC tilt angle, the C balance cylinder pressure, and the switching variables M3.2 / M3.3 / M3.5.

[0092] Among them, the switch variables M3.2, M3.3, and M3.5 are binary logic signals representing the states of three key process nodes in the pressing cycle of the die forging press. These signals are acquired in real time by the PLC control system, and their state reversal times are temporally correlated with the abnormal characteristics of continuous variables such as pressure and speed, thus aiding in the differentiation of fault types.

[0093] M3.2 corresponds to the status indicator of the end of the pressing process;

[0094] M3.3 corresponds to the status indicator indicating the end of the pressure holding phase;

[0095] M3.5 corresponds to the status indicator of the slider return start phase.

[0096] The above signals are all dimensionless binary quantities (0 / 1). Their specific physical triggering conditions depend on the device control logic, but functionally they mark three discrete event nodes in the suppression cycle.

[0097] Compared to traditional diffusion models that directly discretely sample the noise variance or use fixed scheduling, this scheme explicitly models the noise addition process as an analytical combination of the signal-to-noise ratio accumulation function and standard Gaussian noise through reparameterization, avoiding training instability caused by numerical integration errors and random sampling fluctuations; simultaneously, because It can change smoothly over time, making the noise path more consistent with the physical signal degradation law. Thus, under the constraint of limited fault samples, it can still generate extended samples with coherent temporal structure, reasonable amplitude-frequency characteristics and clear category semantics, laying a reliable data foundation for subsequent high-precision fault identification.

[0098] In the direction Noise scheduling strategy during Gaussian noise injection The design directly affects the rate of signal-to-noise ratio (SNR) decay during diffusion, thus impacting the quality of sample augmentation. The standard linear scheduling method results in a slow SNR decrease in the early stages of diffusion followed by a sharp collapse in the later stages, leading to uneven distribution of effective information in intermediate time steps. For the timing signal of a forging press, key fault characteristics such as the step abrupt change at the moment of forging contact and the high-frequency oscillation of the balance cylinder pressure need to be fully learned by the denoising network under moderate noise levels. If the SNR collapses sharply in the later stages, these characteristics will be excessively submerged in the intermediate time steps, preventing the denoising network from effectively learning its timing structure.

[0099] Therefore, in specific implementation, the cosine scheduling strategy proposed by Nichol and Dhariwal is adopted. This cosine noise scheduling strategy defines the cumulative signal-to-noise ratio through a cosine function. Attenuation curve:

[0100] ;

[0101] in, To prevent exist The offset constant is too small; This represents the total number of steps in the diffusion process.

[0102] This setup achieves a smooth, non-linear decay of the signal-to-noise ratio during the noise addition process. It avoids premature loss of the original signal information due to a sudden drop in the signal-to-noise ratio in the initial stage, while ensuring sufficient perturbation capability in the later high-noise stage. This allows for more accurate modeling of the multi-scale degradation characteristics of the forging press fault time-series signal under small sample conditions. Compared to traditional strategies such as linear or square root scheduling, cosine scheduling... In the early stages, changes are slow, preserving more details of the original fault characteristics; in the middle stages, the decay accelerates, promoting pattern discovery; and in the final stages, it approaches zero, ensuring noise dominance. This dynamic characteristic of slow-to-rapid decay aligns better with the physical law of gradual damage leading to sudden failure in the failure evolution of mechanical systems; combined with the offset constant... The introduction of [something] can effectively prevent [something]. hour The problem of initial gradient explosion or generation distortion caused by approaching 0 significantly improves the training convergence and generation fidelity of the diffusion model with a very small number of fault samples.

[0103] In practice, the denoising network is trained through a reverse process. Labeled by fault category As a condition, from the noise-increasing fault window The noise added during prediction. The training objective is to minimize the conditional noise prediction error on various fault samples, and the loss function is:

[0104] ;

[0105] ;

[0106] in, Gaussian noise added during the forward diffusion process; The noise predicted by the denoising network; For tags The actual data distribution corresponding to the fault categories.

[0107] The essence of this loss function is to encode fault category information into the noise prediction task, that is, to perceive which type of fault the current window belongs to when predicting noise, because different fault categories have different signal structures and corresponding noise fields.

[0108] By explicitly guiding the noise prediction task with category labels as conditions, the diffusion generation network can still learn category-specific temporal degradation patterns even with small samples, significantly improving the semantic consistency and physical plausibility of the generated samples.

[0109] In specific implementation, the conditional diffusion generation network uses Diffusion Transformer as its backbone. Diffusion Transformer is composed of multiple DiT Blocks connected in series. Each DiT Block contains a self-attention submodule, a feedforward neural network submodule, and two adaptive layer normalization modules.

[0110] To facilitate better understanding, the following explanation is provided.

[0111] For fault category labels to truly influence the signal structure of generated samples, the labels need to permeate the feature transformations of every layer of the network to continuously constrain the signal reconstruction direction at each stage of denoising. The global self-attention mechanism of the Transformer can globally model long-range dependencies across 128 time steps, making it naturally suitable for capturing cross-variable coupling relationships between multi-source signals from a forging press (such as the joint response of tilt angle and corner pressure variables). More importantly, the adaptive layer normalization interface built into each block provides structural support for the layer-by-layer deep injection of fault category conditions. Therefore, this method uses the Diffusion Transformer (DiT) as its backbone architecture.

[0112] Noisy fault window at the network input. The feature sequence is obtained by linearly projecting the image onto the hidden feature space and then superimposing the learnable positional encoding. Preserve the sequence position information at each time step.

[0113] In specific implementation, the adaptive layer normalization module embeds the corresponding fault category label as a condition into each layer of the network, including:

[0114] The current denoising step is converted into a temporal embedding vector using sine-cosine position encoding. The target fault category label is obtained as a category embedding vector by looking up a table using a learnable embedding matrix. The two are concatenated and input into a multilayer perceptron to obtain a conditional vector. :

[0115] ;

[0116] in, Standard sine-cosine position encoding; The fault category embedding vector is obtained by looking up the learnable matrix; the MLP is a two-layer nonlinear unfolded network; the subscript t represents the current time step in the forward propagation process. The fault category embedding is obtained by looking up the learnable matrix, and each of the three fault categories has an independent 256-dimensional embedding vector, which is optimized in the feature direction of its respective fault distribution during training. The two are added together to obtain a unified conditional vector. This serves as the modulation signal for each subsequent DiTBlock.

[0117] The condition vector The scaling, translation, and gating parameters are generated through a zero-initialization linear transformation layer.

[0118] In the adaptive layer normalization module of each DiT Block, the input features are first normalized using a standard layer, then the normalized features are scaled element-wise using the scaling parameter, translated element-wise using the translation parameter, and the overall contribution of the DiT Block is adjusted using the gating parameter to obtain the modulated feature output.

[0119] In practice, the Diffusion Transformer consists of six sequentially connected DiT Blocks. Within these six DiT Blocks, each attention submodule and feedforward network submodule uses adaptive layer normalization (AdaLayerNorm) instead of fixed layer normalization: conditional vectors... By using zero-initialization of linear layer projections as three sets of modulation parameters—scale, shift, and gate—variable-wise adaptive modulation is applied to the normalized feature distribution. Scale / shift grants the conditional vectors fine-grained control over the activation patterns of each layer's features: when inputting pressure / leakage labels, the network activates more strongly in feature dimensions related to the magnitude separation structure of the relevant variable dimensions; when inputting tilt / motion labels, it shifts towards the tilt angle feature direction. The gate dynamically adjusts the overall contribution of each block, allowing the network to adaptively allocate expression weights to each layer under different denoising stages and different fault categories. This layer-wise conditional modulation mechanism ensures that the constraint of the fault category on the denoising direction persists throughout the entire forward propagation process, rather than only acting at the network entry point. Zero-initialization of the projection layer weights for all modulation parameters ensures that the network approaches an identity mapping in the early stages of training, effectively avoiding gradient instability in the early training phase under small sample conditions. After 6 DiT Block layers, the final features are mapped back to 12 dimensions through the output layer, yielding noise predictions corresponding to 12-dimensional variables and 128 time steps. The complete structure is shown in Figure 2.

[0120] This setup enables the joint integration of fault category semantics and diffusion time steps, along with fine-grained feature control. This allows the network to adaptively strengthen or suppress key temporal patterns related to the current fault type under varying noise levels, maintaining category specificity and physical consistency in the generation process even with minimal sample constraints. Compared to traditional methods that simply concatenate labels to the input or use simple attention guidance, this scheme injects conditional information into the normalization stage of each DiT Block. This avoids the weakening of high-level semantics caused by shallow embedding and overcomes the limitation of global conditional vectors being difficult to adapt to multi-scale features. Especially in highly nonlinear systems like forging presses, different faults should exhibit differentiated frequency domain responses and transient characteristics under the same noise intensity. AdaLN, by dynamically adjusting the mean and variance of features in each layer, can accurately guide the network to focus on the class discriminative subspace, significantly reducing the risk of inter-class confusion in generated samples and enhancing the effective support capability of expanded data for downstream classification tasks.

[0121] In practice, even with the conditional DDPM framework, the expressive power of the class embedding vectors may still be insufficient under small sample conditions, causing the generated samples to drift near the inter-class boundaries. Therefore, this method introduces a classifier-free guidance (CFG) mechanism.

[0122] The classifier-free guidance mechanism used in generating expanded fault window samples has the following noise prediction calculation formula during the inference phase:

[0123] ;

[0124] in, For guiding strength; This is the prediction result for unconditional noise; This is the result of conditional noise prediction;

[0125] During the training phase, fault labels are randomly assigned with a 15% probability. Replacing it with a null token forces the network to have both conditional and unconditional noise estimation capabilities.

[0126] In this way, without introducing an additional classifier, the semantic focus of the generation process on the fault category is dynamically enhanced by linear interpolation, which significantly improves the class fidelity and detail clarity of the generated samples under small sample conditions, while avoiding common problems such as unstable classifier training and gradient conflicts.

[0127] In practice, after training is completed, the generation phase needs to start from pure Gaussian noise. Starting from the beginning, a clean fault timing window is recovered through iterative denoising. Standard DDPM requires a network forward propagation at each step, resulting in extremely high computational overhead. Therefore, this method employs Denoising Diffusion Implicit Models (DDIM) to accelerate the sampling strategy. DDIM reparameterizes the Markov denoising chain into a non-Markov form, rapidly converging to the target distribution with an approximately deterministic trajectory. During inference, labels are assigned to the three types of faults. And booted with CFG Drive 100 steps of DDIM sampling, generating 600 expanded windows for each category. The complete process for expanding the fault samples of a forging press is as follows: Figure 3 As shown

[0128] S3. Merge the expanded fault window samples generated in S2 with the fault window samples obtained in S1 to obtain the expanded sample set.

[0129] S4. Obtain the multivariate time-series signal of the forging press under normal operating conditions, and perform anomaly detection pre-training on the preset temporal convolutional network to obtain a pre-trained anomaly detection model containing the backbone of the temporal convolutional network and a dual-track pooling layer.

[0130] In specific implementation, the method for obtaining the pre-trained anomaly detection model is as follows: obtain the multivariate time series signal of the forging press under normal working conditions, and obtain the normal working condition window sample after sliding window segmentation; use the normal working condition window sample as the training set to perform anomaly detection pre-training on the preset temporal convolutional network, and after training, only retain the parameters of the backbone of the temporal convolutional network and the dual-track pooling layer to obtain the pre-trained anomaly detection model;

[0131] The dual-track pooling layer includes a global max pooling branch and a global average pooling branch set in parallel. The high-dimensional transfer feature vector is a 512-dimensional feature vector formed by concatenating the peak response feature extracted by the global max pooling branch and the global statistical mean feature extracted by the global average pooling branch.

[0132] S5. After freezing the backbone parameters of the temporal convolutional network of the pre-trained anomaly detection model in S4, input each sample from the expanded sample set in S3 into the pre-trained anomaly detection model and extract the high-dimensional transfer feature vector output by the dual-track pooling layer of each sample.

[0133] To facilitate better understanding, the following explanation is provided.

[0134] In practical implementation, the pre-trained anomaly detection model DPTCN-SVDD in this method consists of three parts: a TCN backbone, a dual-track pooling layer, and a fully connected projection layer. The fully connected layer projects the pooling features onto the low-dimensional representation space (64-dimensional) constrained by the hypersphere of SVDD. Its parameters are driven by the hypersphere objective function during training. Compressing the representations of each sample towards the hypersphere center is necessary in anomaly detection tasks, but in fault diagnosis tasks, it will destroy the discriminative distance between different fault types, causing the representations of different faults to overlap after projection. Therefore, the fully connected layer is skipped, and the output of its upstream dual-track pooling layer is selected as the transfer feature.

[0135] The dual-track pooling layer extracts features in parallel using Global Max Pooling (GMP) and Global Average Pooling (GAP), then concatenates them to obtain a 512-dimensional transfer feature vector. GMP extracts the peak response of each variable within the entire window, making it sensitive to impact fault features such as sudden pressure drops and abrupt velocity changes; GAP extracts the global statistical mean of each variable, which is more effective for time-domain cumulative fault features such as amplified velocity tracking deviation and persistent tilt deflection. The concatenation of the two forms a complementary 512-dimensional representation, which, compared to a single pooling method, can simultaneously retain the characterization information of both impactful and cumulative fault signals.

[0136] S6. Using the high-dimensional transfer feature vectors of each sample extracted in S5 and the fault category labels of each sample as training data, train the preset random forest classifier to obtain the fault diagnosis model.

[0137] In practice, when training the random forest classifier, class weights are assigned based on the reciprocal of the number of samples in each fault category. This ensures that minority class samples receive an equal contribution to the impurity calculation during node splitting as the majority class. This effectively alleviates the typical long-tail distribution problem in industrial fault data. Without relying on data-level operations such as oversampling or undersampling, a loss function-level class reweighting mechanism significantly improves the recall rate of minority class faults while maintaining overall classification accuracy.

[0138] S7. Input the window sample of the forging press to be diagnosed into the pre-trained anomaly detection model in S4, extract its high-dimensional transfer feature vector and input it into the fault diagnosis model in S6 to obtain the fault type identification result.

[0139] In existing technologies, Generative Adversarial Networks (GANs) are prone to pattern collapse during small-sample training, while traditional data augmentation methods struggle to capture the complex distribution of multivariate temporal signals. This scheme employs a conditional diffusion generative network with a Diffusion Transformer backbone, combined with a cosine noise scheduling strategy, resulting in a more stable training process and superior distribution coverage compared to GANs. Simultaneously, an adaptive layer normalization module deeply embeds fault category labels into each network layer, coupled with a classifier-free guidance mechanism, effectively enhancing the class consistency and physical realism of generated samples, avoiding the generation of false or ambiguous features. Furthermore, this scheme utilizes a large amount of readily available normal operating condition signals for anomaly detection pre-training of the temporal convolutional network, enabling the model to pre-learn the general temporal patterns and dynamic characteristics of equipment operation. Subsequently, the backbone parameters are frozen, and only dual-track pooling layers are used to extract high-dimensional transfer features from the augmented samples. This approach fully leverages the rich information of normal data to compensate for the feature sparsity of fault samples. Compared to training a fault classification model directly from scratch, it significantly reduces the model's dependence on labeled data and enhances the discriminative power and robustness of features. Furthermore, this scheme fundamentally alleviates the class imbalance problem by constructing a dedicated conditional diffusion generation network for each type of typical fault. At the same time, it uses random forest as the downstream classifier and combines it with high-quality transfer features for training. Compared with a single neural network classifier, it has stronger anti-overfitting ability and inclusiveness for small sample classes, thus ensuring the balance and stability of the recognition performance of various faults.

[0140] This method can achieve high-quality generation of fault samples of forging presses and effective transfer of multi-dimensional features under conditions of small sample size and class imbalance, thereby significantly improving the generalization performance and recognition accuracy of the fault diagnosis model.

[0141] Example 2

[0142] To better illustrate the effectiveness of this method, the following experimental verification was conducted.

[0143] The fault data used in this experiment comes from abnormal production cycles of the die forging production line. After screening by the anomaly detection module, labeled window sets of three typical faults are obtained. All sliding windows are grouped by forging cycle number before splitting. The five-fold division is performed at the cycle level rather than the window level. All windows of the same cycle can only appear in the same fold. There are no overlapping windows from the same cycle between the training set and the test set, strictly avoiding cross-set data information leakage. The experiment conducts a systematic comparison from the following three dimensions:

[0144] (1) Feature extraction method: manual statistical features and pre-trained TCN deep features to verify the effectiveness of transfer learning.

[0145] (2) Classifier selection: SVM, RandomForest, TCN+Softmax end-to-end, and combination with pre-trained TCN features, to analyze the adaptability of different classifiers in small sample high-dimensional scenarios.

[0146] (3) Fault sample augmentation strategies: no augmentation, simple transformation augmentation, conditional DDPM augmentation, quantifying the gain of different augmentation methods on diagnostic performance.

[0147] To comprehensively evaluate the impact of different fault sample augmentation strategies on fault diagnosis performance, two baseline methods are introduced for comparison: signal transformation augmentation methods and deep generation methods based on the generative adversarial network TimeGAN. Signal transformation augmentation methods include the following approaches:

[0148] Gaussian jittering: Adding independent Gaussian white noise to each data point within the window.

[0149] ;

[0150] in, To simulate the random fluctuations in sensor measurement noise, this method introduces small-amplitude random perturbations without altering the macroscopic trend of the signal, which is equivalent to applying regularized noise to the original sample.

[0151] Scaling: Multiply each feature variable by an independent random scaling factor.

[0152] ;

[0153] in, To simulate small deviations in sensor gain drift or operating conditions, the scaling factor is sampled independently along the variable dimension, preserving the temporal synchronization relationship between variables.

[0154] Time warping: Applying local stretching and compression to the time axis through non-linear time mapping. Specifically, it involves applying stretching and compression to the time axis. Upper equidistant setting Control Nodes Apply a Gaussian offset to each internal node as follows:

[0155] ;

[0156] In the formula, The torsional strength coefficient controls the deformation amplitude of the time axis. Using the offset control node as the anchor point, a continuous time mapping function is constructed through cubic spline interpolation to non-uniformly remap the original time axis to the new time axis, simulating the small fluctuations in the timing of the forging process steps.

[0157] The expansion strategy of the above method is to randomly combine 1 to 2 transformations for each original window to generate 4 times the expanded sample.

[0158] Data augmentation quality analysis

[0159] First, we independently evaluate the generation quality of two deep generative augmentation methods, TimeGAN and Conditional DDPM (i.e., the Conditional Diffusion Generative Network of our method), to support the theoretical analysis of the limitations of the TimeGAN method.

[0160] (1) TimeGAN generation quality verification

[0161] TimeGAN uses discriminator accuracy as a quantitative measure of generation quality. A discriminator accuracy closer to 0.5 indicates greater difficulty in distinguishing generated samples from real samples, signifying higher generation quality; an accuracy closer to 1.0 indicates a significant difference between the generated and real distributions. On the three types of forging press fault data, TimeGAN's discriminator accuracy reached 1.0, indicating that the discriminator, after training convergence, can distinguish real samples from generated samples with 100% accuracy, resulting in extremely poor generation quality. Figure 4 The waveforms of a certain variable signal and the signal generated by TimeGAN are compared.

[0162] As can be seen, real fault signals exhibit typical non-stationary dynamic characteristics, ranging from a step rise to a peak and then a sharp decline. However, the samples generated by TimeGAN degenerate into monotonically smooth curves, losing the physical structure of the peak and rapid decline, and failing to reproduce the characteristics of the fault occurrence. The generated samples deviate from the distribution of real samples. The step-change characteristics of the forging press fault signal make it easy for the discriminator to detect shortcut differences between real and generated samples, leading to an imbalance in adversarial training. The generator cannot effectively optimize due to gradient saturation, ultimately degenerating into an approximation of the mean of the training samples. The TimeGAN data augmentation t-SNE effect intuitively reflects the deviation of the generated data, such as... Figure 5As shown.

[0163] (2) Conditional DDPM generation quality verification

[0164] Compared to TimeGAN, Conditional DDPM shows waveform comparisons for the same fault category and variable, for example... Figure 6 As shown, the samples generated by DDPM maintain a consistent macroscopic trend with the real signal. The low-amplitude oscillations in the first half, the continuous rise in the middle, the high-amplitude response in the peak region, and the decay process in the second half are all effectively reproduced. The DTW distance between the generated curve and the real curve is significantly lower than that of TimeGAN.

[0165] Quantitatively, the accuracy of the logistic regression discriminator in evaluating DDPM-generated samples is close to 0.65, significantly closer to the ideal value of 0.5 compared to TimeGAN's 1.000, indicating that the statistical distribution of DDPM-generated samples is highly similar to that of real fault samples. Although conditionally generated DDPM samples still differ from real signals in local high-frequency details, fault discrimination in forging presses mainly relies on multi-scale trend features (amplitude separation, tracking bias, tilt angle deflection) rather than point-by-point precise values. The multi-layer dilated convolution of the pre-trained TCN is naturally robust to local high-frequency noise, so this difference has a limited impact on downstream classification tasks. The t-SNE effect of conditionally augmented DDPM data is as follows: Figure 7 As shown.

[0166] Based on the above analysis, TimeGAN fails in this scenario due to the instability of adversarial training, and will not be included in subsequent classification experiments for TimeGAN augmentation conditions; the sample quality generated by Conditional DDPM (i.e., the Conditional Diffusion Generation Network of this method) meets the requirements of downstream classification augmentation, and will be used as the deep augmentation method proposed in this method for subsequent comparison.

[0167] Comparison and verification

[0168] The 5-fold cross-validation performance of all methods under three conditions—no augmentation, signal transformation augmentation, and diffusion augmentation—is summarized. The key parameter configurations for all methods are shown in Table 1, and the results are shown in Table 2.

[0169] Table 1

[0170]

[0171] Table 2

[0172]

[0173] Figure 8 and Figure 9 The effects of Accuracy and Macro-F1 on different fault diagnosis methods were plotted.

[0174] As can be seen from the comparison charts above, under the same classifier conditions, the pre-trained TCN features outperformed manually generated statistical features by 3.3 percentage points on the same classifier. This difference is statistically significant in small sample scenarios. Manually generated statistical features only perform global statistical aggregation in the time domain, resulting in irreversible loss of information about the temporal structure of the signal. In contrast, the pre-trained TCN, through 6 layers of dilated causal convolutional residual blocks, performs hierarchical representation of multi-scale temporal dynamics within a receptive field of 253 steps. The extracted 512-dimensional features contain the common encoding of local impact waveforms and global trend shifts, providing a more discriminative feature foundation for downstream classification.

[0175] The unexpanded Macro-F1 score of TCN+Softmax is 0.8999, with a standard deviation as high as 0.0913, exhibiting the most significant instability among all methods. This confirms the analysis: with an insufficient annotation window and approximately one million TCN parameters, end-to-end training inevitably leads to overfitting. Notably, while TCN+Softmax shows some performance improvement after signal transformation expansion, diffusion expansion does not bring further improvement (F1=0.9068). This is because the distributional discrepancy between DDPM-generated samples and real samples is amplified by the network parameters during end-to-end training, introducing additional distributional confusion noise. In contrast, pre-trained TCN+RF is unaffected by this, as the frozen TCN encoder is highly robust to small differences in the input distribution, and Random Forest only needs to handle small domain shifts at the 512-dimensional feature level.

[0176] Furthermore, pre-trained TCN+SVM performs worse than Random Forest with equivalent features, by a margin of 4 percentage points. This is attributed to the bandwidth parameter of the RBF kernel. Setting this is extremely difficult with a 512-dimensional input; the gamma=scale strategy sets the bandwidth to... In high-dimensional scenarios, the kernel matrix tends to approach the identity matrix, causing the kernel method to degenerate into an approximately linear classification, resulting in the loss of nonlinear discriminative information in the feature space. Random forests, by randomly sampling each split node... The random subspace mechanism is used to evaluate the features, which is equivalent to implicit regularization in high-dimensional space. It effectively suppresses overfitting to noisy features and shows stronger high-dimensional adaptability.

[0177] Diffusion augmentation significantly improves traditional feature extraction methods, increasing the SVM+statistical feature score from 0.8716 to 0.9122. This indicates that when feature extraction capabilities are limited, the data diversity brought by generated samples can effectively compensate for the insufficient discriminative power of statistical features. However, for our method, which already achieved an F1 score of 0.9580 without augmentation thanks to the high-quality representation of transfer features, the marginal gain from diffusion augmentation approaches zero. This phenomenon suggests that when the discriminative power of pre-trained deep features is sufficiently strong, the marginal contribution of data augmentation to classification performance tends to saturate, and the upper limit of model performance is mainly constrained by feature quality rather than the number of samples.

[0178] To visually demonstrate the classification behavior of this method on the test set, t-SNE was used to perform two-dimensional projection visualization of the 512-dimensional feature vectors of each fold of the test set in the 5-fold cross-validation of the diffusion-amplified fault diagnosis method. The results are as follows: Figure 10 As shown.

[0179] The t-SNE projection diagram reveals a clear cluster separation of the three fault types in the t-SNE two-dimensional space. Tilt / motion faults cluster on the right side of the projection space, with almost no overlap with the other two types. Pressure / leakage faults and speed / servo faults are distributed in the left region, maintaining a separable vertical distribution in the second dimension. This separation structure confirms the essential differences in the physical mechanisms of the three fault types at the feature space level. Tilt / motion faults primarily affect the continuous shift of the BC Tilt variable, fundamentally different from the sensor activation modes of the other two fault types, thus forming the most independent cluster in the feature space. Pressure / leakage and speed / servo faults both involve abnormal pressure or flow in the hydraulic system and share some similarities in variable characteristics, thus they are relatively close in the projection space. The intra-class clustering of various test samples is high, with the class center located near the geometric center of the corresponding cluster. The cluster of tilt / motion faults splits into two sub-clusters in the t-SNE projection. Based on the feedback waveform, it can be seen that the tilt angle shifts in the AD and BC directions in this type of fault correspond to different hydraulic misalignment modes physically. The two have directional differences in the waveform morphology of variables 7 and 8. The pre-trained TCN remains sensitive to this directional difference, thus forming two independent sub-clusters in the feature space. However, the sub-clusters are independent of the distribution areas of the other two classes and do not affect the overall classification performance.

[0180] Diffusion-based fault diagnosis methods, such as 5-fold CV confusion matrix comparisons. Figure 11As shown, this further quantitatively reflects the above conclusions. The proposed method achieves zero missed detections for pressure / leakage faults, a recall rate of 0.90 for speed / servo faults, and a recall rate of 0.95 for tilt / motion faults. There were no cross-class misclassifications for any of the three fault types, and no confusion exists between speed / servo and tilt / motion faults. This experimentally verifies that there is indeed feature overlap between the two types of faults with similar physical mechanisms presented by t-SNE projection. Furthermore, the misclassification of tilt / motion faults, which have the greatest difference in physical mechanism, only stems from a few boundary windows and does not affect the overall reliability of the diagnosis.

[0181] To analyze the independent contribution of each component in this method to the model's diagnostic performance, the following ablation experiment was designed, and the results are shown in Table 3.

[0182] Table 3

[0183]

[0184] The ablation experiments show that the Macro-F1 of randomly initialized TCN+RF drops to 0.8834, indicating that without pre-training knowledge, the multi-layer convolutional stacking of TCN actually impairs classification performance due to the randomness of features introduced by random initialization. Pre-training is the fundamental guarantee for the effectiveness of feature transfer. The Macro-F1 of the 64-dimensional output after using the SVDD fully connected layer, i.e., the Hard-SVDD optimized target-driven latent space, is only 0.8752, showing the largest decrease. This verifies the analysis: the SVDD training target compresses normal samples towards a single center, and its 64-dimensional FC layer learns a mapping that smooths out differences, rather than a mapping that preserves discriminative differences; the landing points of faulty samples in the 64-dimensional SVDD latent space do not possess inter-class separability, and direct classification in this space is worse than statistical features. The 512-dimensional dual-track pooling feature layer, located upstream of the SVDD fully connected layer, preserves the multi-scale temporal discriminative representation of TCN convolutional stacking and is a key node for achieving effective transfer.

[0185] Experimental results show that the proposed method achieves good diagnostic performance on the actual measured dataset of forging press operation status, with a Macro-F1 score of 0.9580 and an accuracy of 0.9667, both superior to the comparative methods. Ablation experiments further demonstrate the positive role of pre-trained weights and frozen feature extraction layers in performance improvement. The t-SNE visualization results show a clear distribution structure of different fault types in the feature space, proving the discriminative ability of the proposed method in small-sample typical fault diagnosis scenarios of forging presses.

[0186] 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 the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for fault diagnosis of a forging press under small sample conditions, characterized in that, Includes the following steps: S1. Obtain fault window samples of the forging press under various typical faults, and set corresponding fault category labels for each fault window sample; S2. For each type of typical fault, a dedicated conditional diffusion generation network is constructed and trained using the fault window samples of the typical fault obtained in S1. The trained conditional diffusion generative networks are used to generate expanded fault window samples corresponding to typical faults. The structure of each conditional diffusion generation network is consistent, with Diffusion Transformer as the backbone. The network internally uses an adaptive layer normalization module to embed the corresponding fault category label as a condition into each layer of the network. During training, a cosine noise scheduling strategy is used to control the noise addition and denoising process. When generating the expanded fault window samples, a classifier-free guidance mechanism is used to enhance the class consistency of the generated samples. S3. Merge the expanded fault window samples generated in S2 with the fault window samples obtained in S1 to obtain the expanded sample set. S4. Obtain the multivariate time-series signal of the forging press under normal working conditions, perform anomaly detection pre-training on the preset time-series convolutional network, and obtain a pre-trained anomaly detection model containing the backbone of the time-series convolutional network and a dual-track pooling layer. S5. After freezing the backbone parameters of the temporal convolutional network of the pre-trained anomaly detection model in S4, input each sample from the expanded sample set in S3 into the pre-trained anomaly detection model and extract the high-dimensional transfer feature vector output by the dual-track pooling layer of each sample. S6. Using the high-dimensional transfer feature vectors of each sample extracted in S5 and the fault category labels of each sample as training data, train the preset random forest classifier to obtain the fault diagnosis model. S7. Input the window sample of the forging press to be diagnosed into the pre-trained anomaly detection model in S4, extract its high-dimensional transfer feature vector and input it into the fault diagnosis model in S6 to obtain the fault type identification result.

2. The method for fault diagnosis of a forging press under small sample conditions according to claim 1, characterized in that, In S2, the forward diffusion process of the conditional diffusion generation network gradually moves towards the original fault window sample within T steps. By injecting Gaussian noise and utilizing reparameterization techniques, the noisy result at any step can be directly obtained from the original fault window samples. The closed form is: ； ； in, For cumulative signal-to-noise ratio; The noise intensity at step s is given by the cosine noise scheduling strategy; For noise variables; Representing noise variables It follows a standard multivariate Gaussian distribution with a mean of 0 and a covariance matrix of identity; t represents the current time step in the forward diffusion process.

3. The method for fault diagnosis of a forging press under small sample conditions according to claim 2, characterized in that, In S2, the cosine noise scheduling strategy defines the cumulative signal-to-noise ratio using a cosine function. Attenuation curve: ； in, To prevent exist The offset constant is too small; This represents the total number of steps in the diffusion process.

4. The method for fault diagnosis of a forging press under small sample conditions according to claim 3, characterized in that, In S2, when training the conditional diffusion generation network, fault category labels are used. The training objective is to minimize the conditional noise prediction error on various fault samples, and the loss function is: ； ； in, Gaussian noise added during the forward diffusion process; The noise predicted by the denoising network; For tags The actual data distribution corresponding to the fault categories.

5. The method for fault diagnosis of a forging press under small sample conditions according to claim 4, characterized in that, In S2, the conditional diffusion generation network uses Diffusion Transformer as its backbone. Diffusion Transformer is composed of multiple DiT Blocks connected in series. Each DiT Block contains a self-attention submodule, a feedforward neural network submodule, and two adaptive layer normalization modules. The adaptive layer normalization module embeds the corresponding fault category label as a condition into each layer of the network, including: The current denoising step is converted into a temporal embedding vector using sine-cosine position encoding. The target fault category label is obtained as a category embedding vector by looking up a table using a learnable embedding matrix. The two are concatenated and input into a multilayer perceptron to obtain a conditional vector. : ； in, Standard sine-cosine position encoding; The fault category embedding vector is obtained by looking up a table of the learnable matrix; the MLP is a two-layer nonlinear unfolded network; the subscript t represents the current time step in the forward diffusion process; The condition vector The scaling, translation, and gating parameters are generated through a zero-initialization linear transformation layer. In the adaptive layer normalization module of each DiT Block, the input features are first normalized using a standard layer, then the normalized features are scaled element-wise using the scaling parameter, translated element-wise using the translation parameter, and the overall contribution of the DiT Block is adjusted using the gating parameter to obtain the modulated feature output.

6. The method for fault diagnosis of a forging press under small sample conditions according to claim 5, characterized in that, In S2, the classifier-free guidance mechanism used in generating expanded fault window samples has the following noise prediction calculation formula during the inference phase: ； in, For guiding strength; This is the prediction result for unconditional noise; This is the result of conditional noise prediction; During the training phase, fault categories are randomly labeled with preset probabilities. Replace it with the null flag to enable the network to estimate noise with both conditional and unconditional conditions.

7. The method for fault diagnosis of a forging press under small sample conditions according to claim 6, characterized in that, In S2, the denoising diffusion implicit model DDIM is used to accelerate the sampling process of generating expanded fault window samples, and the number of sampling steps is shortened to 100 steps by constructing a deterministic inverse Markov chain.

8. The method for fault diagnosis of a forging press under small sample conditions according to claim 1, characterized in that, In S4, the method for obtaining the pre-trained anomaly detection model is as follows: the multivariate time-series signal of the forging press under normal working conditions is obtained, and the normal working condition window sample is obtained after sliding window segmentation; the normal working condition window sample is used as the training set to perform anomaly detection pre-training on the preset temporal convolutional network; after training, only the parameters of the backbone of the temporal convolutional network and the dual-track pooling layer are retained to obtain the pre-trained anomaly detection model. The dual-track pooling layer includes a global max pooling branch and a global average pooling branch set in parallel. The high-dimensional transfer feature vector is a 512-dimensional feature vector formed by concatenating the peak response feature extracted by the global max pooling branch and the global statistical mean feature extracted by the global average pooling branch.

9. The method for fault diagnosis of a forging press under small sample conditions according to claim 1, characterized in that, In S6, when training the random forest classifier, class weights are assigned based on the reciprocal of the number of samples in each fault category, so that minority category samples receive an equal contribution to the impurity calculation during node splitting as the majority category.

10. The method for fault diagnosis of a forging press under small sample conditions according to claim 1, characterized in that, In S1, the fault window sample is a multivariate time series signal segmented by a sliding window; the window length of the sliding window is consistent with the input dimension of the conditional diffusion generation network and the input dimension of the pre-trained anomaly detection model. The typical fault types include pressure / leakage faults, speed / servo faults, and tilt / motion faults.

Images