Stamping machine bearing fault diagnosis method and system based on conditional generative adversarial network

Abstract

This invention belongs to the technical field of bearing fault diagnosis. To address the inaccuracy of existing bearing fault diagnosis methods, this invention proposes a fault diagnosis method and system for press bearings based on conditional generative adversarial networks (GANs). A generator is trained using operating condition vectors and Gaussian noise. A discriminator is trained using the health state features generated by the generator and the actual health state features. The generator and discriminator are then subjected to adversarial training to prevent the discriminator from distinguishing between the health state features generated by the generator and the actual health state features, as well as from distinguishing the matching between the health state features generated by the generator and the corresponding operating condition vectors. This results in a well-trained health state model. The health state model is then used to generate a dynamic health baseline for the press shaft to be diagnosed, thereby obtaining the fault diagnosis result for the press shaft and achieving high-precision fault diagnosis and early warning.

Description

Technical Field

[0001] This invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a method and system for diagnosing faults in stamping press bearings based on conditional generative adversarial networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the deepening of Industry 4.0 and intelligent manufacturing, the continuous and stable operation of large-scale, high-value industrial equipment (such as large stamping presses, CNC machine tools, and aerospace power equipment) is of decisive significance to production efficiency and economic benefits. Unplanned equipment downtime will lead to huge economic losses and the risk of production chain disruption. Therefore, fault diagnosis, as a core component of PHM (Problem Diagnosis and Management) technology, has become a key supporting technology for ensuring industrial production safety and improving the utilization rate of equipment throughout its entire life cycle.

[0004] Currently, data-driven fault diagnosis is the mainstream research direction. Its core idea is to analyze monitoring data generated during equipment operation (such as vibration, temperature, pressure, acoustic signals, etc.) to determine whether the equipment is in a healthy state and to issue alarms and diagnoses when abnormalities occur. A key step is establishing a health status baseline, that is, defining the data pattern of the equipment during normal operation. When real-time monitoring data deviates from this baseline to a certain extent, the system determines that a fault may have occurred. However, existing technologies face a serious challenge in establishing a health baseline: the variability of operating conditions.

[0005] In typical scenarios involving the stamping mechanism of large-scale stamping presses, existing health baseline technologies face severe challenges due to the variability of operating conditions. Taking the J31-315 closed-type single-point stamping press as an example, its stamping mechanism bears the critical support function for the reciprocating motion of the slide block. Failure of the core rotating component, the 6206 deep groove ball bearing, will lead to unplanned downtime, with each hour of downtime causing tens of thousands of yuan in economic losses. The operating conditions of this type of equipment exhibit significant characteristics of strong coupling of multiple parameters and dynamic drift throughout its entire life cycle: the dynamic range of stamping force is 100-450kN, the spindle speed fluctuates between 600-1400rpm, the bearing temperature varies with load between 30-75℃, and the cumulative number of stamping cycles reaches 8×10 5 Subsequently, due to lubrication loss and mechanical fatigue, natural aging will occur, ultimately leading to the monitoring data exhibiting complex characteristics of high impact, multi-dimensional coupling, and life-cycle drift.

[0006] Currently used methods for constructing health baselines, such as the static threshold method, the multi-model method, and traditional machine learning methods, each have their own drawbacks. The static threshold method sets fixed alarm thresholds for each monitoring indicator, completely ignoring the impact of changes in operating conditions on the equipment's operating status. The direct consequence is a persistently high false alarm rate and missed alarm rate. Normal data fluctuations caused by changes in operating conditions are misjudged as faults, while weak signals of early faults are masked by operating noise, failing to meet the high reliability maintenance requirements of stamping machines.

[0007] The multi-model approach establishes independent health models for different operating conditions. For example, a diagnostic model is trained for each typical combination of operating conditions, such as high-speed-heavy load, high-speed-light load, and low-speed-heavy load. The disadvantages of this approach are: actual operating conditions are often continuously changing, making it difficult to clearly divide them into a limited number of discrete modes, leading to difficulties in operating condition classification; each operating condition requires a sufficient amount of health data for model training, resulting in high data acquisition costs; a large number of models need to be maintained, and the switching logic of the models may introduce new problems when operating conditions change; this method cannot effectively handle intermediate or new operating conditions not present in the training set, exhibiting poor generalization ability.

[0008] Traditional machine learning methods rely on massive amounts of labeled fault data, while fault samples in industrial settings typically account for less than 1% of the total, and they do not explicitly incorporate operating condition constraints. When operating conditions change, the dynamic shift in feature distribution leads to a significant increase in model generalization error. Although some methods attempt to incorporate operating condition parameters as input features into the model (such as SVM and traditional ANN), they essentially learn a black-box mapping from vibration features + operating condition features to health and fault conditions. They do not explicitly model the typical vibration patterns of healthy equipment under specific operating conditions, resulting in insufficient sensitivity to early, weak fault signals and poor model interpretability.

[0009] In summary, the existing fault diagnosis of stamping bearings suffers from problems such as frequent false alarms, missed early faults, and failure to adapt to aging throughout the entire life cycle, which are problems that urgently need to be solved. Summary of the Invention

[0010] To overcome the shortcomings of the prior art, this invention provides a fault diagnosis method and system for stamping press bearings based on conditional generative adversarial networks. By adapting a dynamic health baseline to the high impact, multi-condition, and full life-cycle aging characteristics of stamping presses, it achieves high-precision fault diagnosis and early warning, which has urgent practical significance for reducing equipment downtime losses and ensuring the safe and stable operation of intelligent manufacturing systems.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks, including: Obtain the training dataset for the health status model; wherein, the training dataset includes vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions; the health status model is constructed based on a conditional generative adversarial network; A generator is trained using a working condition vector and Gaussian noise. A discriminator is trained using the health status features generated by the generator and the real health status features. The generator and the discriminator are then subjected to adversarial training so that the discriminator cannot distinguish between the health status features generated by the generator and the real health status features, nor can it distinguish between the matching of the health status features generated by the generator and the corresponding working condition vector, thus obtaining the trained health status model. Real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed are collected. The health status characteristics of the press shaft to be diagnosed are generated using the health status model. A dynamic health baseline is determined based on the generated health status characteristics. The dynamic health baseline is compared with the health status characteristics to be diagnosed to obtain the fault diagnosis result of the press shaft to be diagnosed.

[0012] Secondly, the present invention provides a fault diagnosis system for stamping press bearings based on conditional generative adversarial networks, comprising: The acquisition module is configured to: acquire the training dataset of the health status model; wherein, the training dataset includes vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions; the health status model is constructed based on a conditional generative adversarial network; The training module is configured to: train a generator using a working condition vector and Gaussian noise; train a discriminator using health state features generated by the generator and real health state features; and perform adversarial training on the generator and the discriminator so that the discriminator cannot distinguish between the health state features generated by the generator and the real health state features, nor can it distinguish the matching between the health state features generated by the generator and the corresponding working condition vector, thereby obtaining the trained health state model. The diagnostic module is configured to: acquire real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed; generate health status characteristics of the press shaft using the health status model; determine a dynamic health baseline based on the generated health status characteristics; and compare the dynamic health baseline with the health status characteristics to obtain the fault diagnosis result of the press shaft. Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0013] The above one or more technical solutions have the following beneficial effects: This invention extracts health status characteristics of stamping press bearings under different operating conditions, uses conditional generative adversarial networks to establish a nonlinear mapping model between operating condition parameters and health status vectors, and achieves dynamic health assessment of stamping press bearings under multiple operating conditions by matching the real-time status of stamping press bearings under different operating conditions with the corresponding dynamic health baseline.

[0014] This invention constructs an initial feature pool by extracting time-domain, frequency-domain, time-frequency-domain, and entropy features from vibration signals. After redundancy removal through Pearson correlation analysis, weight sorting using the ReliefF algorithm, and importance screening using random forest, health status features are obtained. This can comprehensively capture the feature evolution of bearings from healthy to early failure and then to severe failure, improve the accuracy of subsequent classification, and provide sufficient feature support for early failure identification.

[0015] In this invention, the generator utilizes 1D-CNN to effectively preserve the core impact information in the vibration signal, and uses dual LSTM to accurately capture the periodic repetition pattern of the impact signal, which is more suitable for the temporal characteristics of the vibration signal, so that the generated health features are consistent with the dimensions of the real health features; the discriminator simultaneously verifies the authenticity of the features and the adaptability of the working conditions, avoiding interference from invalid samples with real features but mismatched working conditions, and improving the discrimination accuracy.

[0016] In this invention, a three-stage training method is adopted for the health status model. Through progressive training, the shallow network first learns the basic mapping relationship between working conditions and health features, and then LSTM is gradually introduced to avoid the gradient vanishing problem caused by the network being too deep. At the same time, the feature matching loss is strengthened in stages to ensure that the low-level features are aligned with the real distribution first, laying the foundation for the learning of high-level temporal features, and ultimately improving the convergence speed and generation accuracy of the health status model.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is an overall flowchart of the stamping press bearing fault diagnosis method based on conditional generative adversarial networks in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the construction of a health status vector in an embodiment of the present invention. Figure 3 This is a schematic diagram of the network structure of the conditional adversarial network in an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] Example 1 This embodiment discloses a fault diagnosis method for stamping press bearings based on conditional generative adversarial networks, including: Obtain the training dataset for the health status model; the training dataset includes vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions; the health status model is constructed based on a conditional generative adversarial network. A generator is trained using a working condition vector and Gaussian noise. A discriminator is trained using the health status features generated by the generator and the real health status features. The generator and discriminator are then subjected to adversarial training so that the discriminator cannot distinguish between the health status features generated by the generator and the real health status features, nor can it distinguish the matching between the health status features generated by the generator and the corresponding working condition vector, thus obtaining a well-trained health status model. Real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed are collected. The health status characteristics of the press shaft to be diagnosed are generated using a health status model. A dynamic health baseline is determined based on the generated health status characteristics. The dynamic health baseline is compared with the health status characteristics to be diagnosed to obtain the fault diagnosis result of the press shaft to be diagnosed.

[0024] This embodiment extracts health status features of the press bearing under different operating conditions, uses conditional generative adversarial network to establish a nonlinear mapping model between operating condition parameters and health status vector, and achieves dynamic health assessment of the press bearing under multiple operating conditions by matching the real-time status of the press bearing under different operating conditions with the corresponding dynamic health baseline.

[0025] The following is combined with Figure 1 The fault diagnosis method for press bearings based on conditional generative adversarial networks proposed in this embodiment is described in detail below: Step 1: Obtain the training dataset for the health status model; the training dataset includes vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions.

[0026] In this embodiment, the operating parameters and corresponding vibration signals of the press bearing are collected; correlation analysis is performed on the operating parameters to screen out the core operating parameters whose absolute values ​​of correlation coefficients are greater than a set value; differential normalization is applied to the core operating parameters to obtain the corresponding operating condition vector; time domain, frequency domain, time-frequency domain and entropy features of the vibration signal are collected to construct an initial feature pool; redundant features are removed from the initial feature pool using the Pearson correlation coefficient; feature weights are sorted using the ReliefF algorithm; the importance of features is verified and evaluated using random forest to obtain health status features.

[0027] As a specific implementation method, taking the deep groove ball bearing of the stamping mechanism of a stamping press as a typical example, vibration signals and operating parameters are synchronously collected by multi-dimensional sensors. After impact-resistant preprocessing, a 32-dimensional multi-domain health feature vector and a 4-dimensional key operating condition vector are extracted. The two vector pairs are used as input to improve the conditions to generate an adversarial network, generating health benchmark features and confidence intervals that match the current operating conditions. Fault classification diagnosis is achieved through cosine similarity and trend analysis. Healthy samples are periodically screened at the edge and uploaded to the cloud to drive model iteration and update, forming a closed-loop system of perception-modeling-diagnosis-iteration.

[0028] Collect 10,000 sets of sample pairs of operating parameters and vibration characteristics covering all operating conditions. Select core indicators of vibration characteristics, such as kurtosis, center of gravity frequency, and sample entropy.

[0029] Spearman correlation analysis was performed on eight types of operating parameters involved in the operation of the stamping press, namely, stamping force, spindle speed, bearing temperature, ambient temperature, grease quantity, cumulative stamping times, slide stroke, and die weight. Four core parameters with an absolute correlation coefficient greater than the set value (e.g., 0.3) were selected. The core parameters are stamping force, spindle speed, bearing temperature, and cumulative stamping times. Weakly correlated operating parameters such as ambient temperature (which is less affected by the insulation layer) and grease quantity (which cannot be collected in real time) were excluded.

[0030] Based on the linear relationship model between measured operating parameters and vibration characteristics, such as an increase of 0.2g in peak-to-peak vibration for every 100kN of impact pressure; an increase of 0.3kHz in outer ring fault characteristic frequency for every 200rpm increase in rotational speed; and an increase of 0.05g in vibration noise for every 15℃ increase in temperature, linear normalization is applied to impact pressure F (100-450kN), rotational speed n (600-1400rpm), and bearing temperature T (30-75℃), with errors controlled within ±2%.

[0031]

[0032]

[0033] in, This represents the minimum value of the impact force. This represents the maximum value of the impact force. This is the normalized impact force value; This represents the minimum value of the rotational speed; This is the normalized rotational speed value; This represents the maximum value of the rotational speed; This is the normalized bearing temperature; This represents the minimum value of the bearing temperature. This represents the maximum value of the bearing temperature.

[0034] Design cumulative stamping times Maximum lifespan Next, establish the relationship between the cumulative number of stampings and the kurtosis K, initially with the cumulative number of stampings being 0-2×10⁻⁶. 5 At this time, the kurtosis K increased rapidly from 3.0 to 3.3; in the later stage, the cumulative number of stamping times reached 6×10. 5 -8×10 5 Next, the kurtosis K increases more slowly from 3.6 to 3.8; the fitted equation is: , This indicates a non-linear relationship between the cumulative number of stamping operations and health characteristics. The cumulative number of stamping operations is normalized using the following formula: .

[0035] After the above steps, the final output is a 4D working condition vector. The values ​​of each dimension are all in the range of [0,1], and the influence of the parameters on health characteristics is accurately preserved.

[0036] As a specific implementation method, such as Figure 2 As shown, the core objective of the health status feature vector is to comprehensively capture the feature evolution pattern of the device from health to early failure to severe failure, while taking into account both feature effectiveness and computational efficiency.

[0037] This embodiment breaks through the limitations of traditional single-domain features + manual selection, and constructs a three-order framework of initial feature pool → accurate extraction of four-dimensional features → multi-step redundancy optimization → feature vector output. The specific process is as follows: First, health features are extracted, covering four dimensions: time domain, frequency domain, time-frequency domain, and entropy, to ensure a comprehensive characterization of the device's health status. These features include: The 18-dimensional time-domain features include basic indicators and impact indicators. Basic indicators include mean, peak value, peak-to-peak value, standard deviation, kurtosis, skewness, root mean square (RMS), and absolute mean. Impact indicators include impulse indicators, margin factor, maximum probability (i.e., the proportion of sample points with amplitude > 0.5g), waveform factor, peak factor, variance, mean square value, peak offset, waveform slope, and steady-state volatility.

[0038] 12-dimensional frequency domain characteristics: including centroid frequency, spectral kurtosis, harmonic ratio (i.e., the ratio of fundamental frequency to harmonics), spectral energy, peak frequency, spectral flatness, spectral peak density, low-frequency bandwidth ratio (i.e., the proportion of 0-5kHz), mid-frequency bandwidth ratio (i.e., the proportion of 5-15kHz), high-frequency bandwidth ratio (i.e., the proportion of 15-20kHz, which increases with early wear), third harmonic ratio (i.e., the ratio of fundamental frequency to third harmonics), and spectral roll-off rate.

[0039] 20-dimensional time-frequency domain features: A 3-level decomposition using the db5 wavelet basis is selected, covering the entire frequency band from 0 to 20 kHz. The proportion of energy in each sub-band to the total energy is calculated. For the 8 sub-bands after the 3-level wavelet packet decomposition, one sub-band energy proportion (i.e., sub-band energy / total energy) is extracted from each sub-band, directly reflecting the energy distribution of different frequency bands. Under healthy conditions, the proportion of low and mid frequencies is high, while the proportion of high frequencies increases during faults. To capture local features within the sub-bands, 1-2 detailed indicators are further derived for each sub-band, resulting in a total of 12 dimensions for the 8 sub-bands, including "sub-band kurtosis" and "sub-band entropy".

[0040] 10-dimensional entropy features: quantifying the complexity and regularity of signals, and adapting to the enhanced signal regularity phenomenon caused by faults. These mainly include: sample entropy, approximate entropy, permutation entropy, wavelet packet entropy, spectral entropy, etc., and optimizing the sample entropy parameter with an embedding dimension of 2 and a similarity tolerance of 0.2σ.

[0041] Then, the initially extracted 60-dimensional features are optimized, specifically as follows: The first step is to perform Pearson correlation analysis to remove redundancy.

[0042] Calculate the Pearson correlation coefficient r between any two health features. If |r|>0.9, the feature is considered redundant, and the health feature with clearer physical meaning or higher computational efficiency is retained.

[0043] The second step is to sort the weights using the ReliefF algorithm.

[0044] The top 40 features with the highest weights are selected by calculating the contribution of each feature to the classification of the samples. The number of neighbors is k=10. Experiments have verified that when k=10, the weight calculation error is ≤5%, and the number of iterations equals the number of samples (10,000 times).

[0045] The feature weight update formula is:

[0046] In the formula, represents the weight of health feature A; m is the total number of samples; Let i be the set of nearest neighbors of the same kind as sample i; Let i be the set of nearest neighbors of sample i in category c (different class); Let be the prior probability of category c; Let health feature A be the difference between sample i and its nearest neighbor j, and continuous feature be... The discrete characteristic is 0 or 1.

[0047] The above measures the feature's ability to maintain similarity among similar samples and its ability to distinguish between different samples, thus quantifying the feature's contribution. A higher weight indicates that the feature is more important for classifying health status.

[0048] The third step is to validate using a random forest.

[0049] A random forest of 100 decision trees was constructed, and the Gini coefficient (the reduction in node impurity for each feature) was calculated. The top 32 features by importance were selected. The average importance of these 32 features was >0.03, resulting in an 8% improvement in classification accuracy and a 47% improvement in computational efficiency compared to the initial 60-dimensional feature set.

[0050] The reduction in node impurity = parent node Gini index - weighted average child node Gini index.

[0051] The formula for calculating the Gini index is as follows:

[0052] In the formula: The value represents the proportion of samples of the k-th class in a node. In this embodiment, "class" refers to different operating condition clusters under healthy conditions, and K is the total number of classes. The parent node's Gini index represents the impurity of the node before the split, and the weighted average Gini index of the child nodes is the weighted sum of the Gini indices of each child node after the split, based on the proportion of sample size. The larger this value, the greater the contribution of the feature to reducing node impurity, and the stronger the feature's ability to distinguish healthy states under different operating conditions.

[0053] After the above steps, a 32-dimensional health status feature vector is finally output. The values ​​of each dimension are normalized to the interval [0,1].

[0054] Step 2: Train the generator using the condition vector and Gaussian noise, and train the discriminator using the health status features generated by the generator and the real health status features. Perform adversarial training on the generator and the discriminator so that the discriminator cannot distinguish between the health status features generated by the generator and the real health status features, nor can it distinguish the matching between the health status features generated by the generator and the corresponding condition vector. This results in a well-trained health status model, which is built based on a conditional generative adversarial network.

[0055] Traditional conditional generative adversarial networks suffer from three fatal flaws in generating health benchmarks for industrial equipment: structural mismatch, poor generation quality, and unstable training. These flaws directly lead to large deviations between the benchmark and the actual health state, and failure to adapt to multiple operating conditions.

[0056] This embodiment focuses on the core requirements of vibration signal temporality, operating condition coupling, and health feature diversity, and innovates through structural reconstruction, loss enhancement, and training optimization.

[0057] like Figure 3 As shown, the generator input is a 100-dimensional noise vector z + a 4-dimensional condition vector C; the output is a 32-dimensional health state feature X. gen .

[0058] The generator includes: Working Condition-Noise Deep Coupling Layer: The 4-dimensional working condition vector C is mapped to the same dimension as the noise through a fully connected layer. The working condition constraints are embedded in the initial generation stage and concatenated with 100-dimensional Gaussian noise z to form a 200-dimensional vector, achieving pixel-level coupling between working conditions and noise, thus solving the problem of feature generation being disconnected from the working conditions under multiple conditions. Here, z ~ N (0,1); the fully connected layer has an input dimension of 4 and an output dimension of 100, using the LeakyReLU activation function with a slope of 0.2.

[0059] Local Impact Feature Extraction Layer: A 200-dimensional vector is reshaped into a (32,8) matrix by a fully connected layer, input to a 1D-CNN, and then reduced to (16,32) dimensionality through max pooling, preserving core impact information while reducing computational cost. The fully connected layer has an input dimension of 200 and an output dimension of 256, employing batch normalization (BN) to suppress distribution shifts caused by high-impact signals. The 1D-CNN uses 64 filters with a kernel size of 3, a stride of 1, and BN. The max pooling layer is 2×2 with a stride of 2.

[0060] Cross-cycle temporal modeling layer: A dual LSTM stacked design is used. The first layer has 128 units, returning the sequence with dropout=0.2; the second layer has 64 units and does not return the sequence. The 128 units adapt to the temporal memory of 3 rotation cycles, and dropout=0.2 suppresses workshop noise interference. The second 64 units compress redundant temporal information, ultimately outputting a 64-dimensional temporal feature vector. This is a relatively pure CNN structure, capable of accurately capturing the periodic repetition patterns of impact signals.

[0061] Feature Output Layer: Normalized features are output via a fully connected layer (64→32, Tanh activation). This is consistent with the 32-dimensional feature vector of the real health status, ensuring compatibility with subsequent benchmark generation.

[0062] Discriminator input: 32-dimensional health status features X + 4-dimensional working condition vector C; output: true probability P∈[0,1].

[0063] The discriminator includes: Working condition-feature strong coupling layer: After the 4-dimensional working condition vector is mapped by a fully connected layer (4→32, LeakyReLU), it is concatenated with the 32-dimensional health status features to form a 64-dimensional vector with a temporal format of (64,1). This forces the discriminator to verify the working condition-feature matching degree while discriminating features. Traditional cGANs only discriminate the authenticity of features and ignore the working condition adaptability.

[0064] Multi-scale difference extraction layer: Dual 1D-CNN stacked, the first layer has 64 filters, kernel size 3, Batch Normalization (BN), and dropout=0.3; the second layer has 128 filters, kernel size 3, and BN. The first 1D-CNN captures low-frequency operating condition-related features, such as frequency shifts corresponding to rotational speed; the second 1D-CNN captures high-frequency impact difference features, such as high-frequency energy changes in micro-wear, with dropout=0.3 to suppress discrimination bias caused by noise. After max pooling, the result is a 4096-dimensional multi-scale feature vector.

[0065] Probability Output and Feature Feedback Layer: Outputs deep features (feat.) via a fully connected layer (4096→64). The true probability is then output through a fully connected layer (64→1, Sigmoid activation).

[0066] The discrimination accuracy is 18% higher than that of traditional pure CNN discriminators. It can accurately identify invalid samples with real features but mismatched operating conditions, such as generating high-speed features under low-speed operating conditions.

[0067] Traditional cGANs rely solely on adversarial loss, leading to a disconnect between generated features and the real distribution, and a high rate of sample duplication. This embodiment constructs a triple constraint of adversarial loss + feature matching loss + consistency loss, enabling the discriminator to accurately distinguish between the real health status features with conditional vector C and the generated health status features with conditional vector C, maximizing the probability of real samples being classified as true and the probability of generated samples being classified as false.

[0068] Construct the discriminator loss function : ; Where D represents the discriminator; C represents the actual health status characteristics; E represents the working condition vector; and E represents the expectation operator.

[0069] The generator generates features that the discriminator misclassifies as real, thus enabling the adversarial loss function. :

[0070] Traditional adversarial loss only constrains surface probabilities. Therefore, we designed feature matching loss to constrain the spatial distance of deep features, which solves the problem that generated features look similar but are semantically inconsistent, such as kurtosis values ​​being close but not matching rotational speed.

[0071] Feature matching loss function as follows:

[0072] Traditional cGANs, focusing solely on the probability of true detection, are prone to generating duplicate samples, resulting in a mode collapse rate ≥15%. Therefore, a consistency impairment function is designed to suppress mode collapse. This function calculates the variance of feature samples generated from 10 independently generated noise samples under the same operating condition vector C. Maximizing the variance ensures that the generated samples cover the normal fluctuation range of healthy conditions, such as slight changes in bearing lubrication. The impairment function is as follows:

[0073] Therefore, the total generator loss function is:

[0074] in, , C represents the optimal weights; C is the working condition vector. It is Gaussian noise; D represents the generator, E represents the discriminator, and E represents the expectation operator. Characteristics of a true state of health; Indicates deep intermediate features; L2 norm is represented; var represents the variance operator; This represents the samples generated by generator G under a given vector of operating conditions. Regarding noise The variance; , These are the weighting coefficients.

[0075] In this embodiment, a progressive training method is adopted for the health status model. Specifically, the first stage of training is performed on the 1D-CNN of the generator and the first layer of the 1D-CNN of the discriminator, so that the shallow network learns the basic mapping relationship between working conditions and health status features; the second stage of training is performed on the 1D-CNN and the first layer of the LSTM of the generator, as well as the two-layer 1D-CNN of the discriminator, to model cross-cycle temporal correlation; and the third stage of training is performed on the complete generator and the complete discriminator to optimize the matching accuracy between deep features and working conditions.

[0076] As a specific implementation method, in the first stage, the first 200 epochs (training rounds) use a learning rate of lr=0.0002 and a batch size of 32 to train both the generator's 1D-CNN and the first layer of the discriminator's 1D-CNN, employing the same loss function architecture as the complete network, i.e., the generator loss is adversarial loss + feature matching loss. λ 1=10) + Consistency loss (λ2=5), the discriminator loss is the adversarial loss.

[0077] In the second stage, epochs range from 201 to 500, the learning rate is lr = 0.0001, and cosine annealing is halved. The 1D-CNN for the generator and the first-layer LSTM, as well as the two-layer 1D-CNN for the discriminator, are trained using the same loss function architecture as the complete network, i.e., the generator loss is adversarial loss + feature matching loss. λ 1=10) + Consistency loss (λ2=5), the discriminator loss is the adversarial loss.

[0078] In the third stage, epochs are set from 501 to 1000, the learning rate is lr=0.00005, and the training terminates when the similarity is ≥0.95. This completes the training of the generator and the discriminator.

[0079] Through progressive training, the shallow network first learns the basic mapping relationship between working conditions and health characteristics, such as the local amplitude and frequency characteristics of vibration signals; then LSTM is gradually introduced to avoid the gradient vanishing problem caused by the network being too deep; at the same time, the feature matching loss is strengthened in stages to ensure that the low-level features are aligned with the real distribution first, laying the foundation for the learning of high-level temporal features, and ultimately improving the model's convergence speed and generation accuracy.

[0080] The Adam optimizer is used, with β1=0.5 to adapt to the non-normal distribution of industrial data and β2=0.999 to improve gradient stability; the cosine annealing learning rate is halved every 100 epochs to avoid training oscillations in the later stages and to adapt to the slow changes in healthy features; the batch size is 32 to balance memory usage and gradient estimation accuracy.

[0081] Step 3: Collect real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed, generate health status characteristics of the press shaft to be diagnosed using the health status model, determine the dynamic health baseline based on the generated health status characteristics, and compare the dynamic health baseline with the health status characteristics to be diagnosed to obtain the fault diagnosis result of the press shaft to be diagnosed.

[0082] In this embodiment, the degree of matching between the health status characteristics to be diagnosed and the dynamic health baseline is calculated by cosine similarity; the fault diagnosis result of the press shaft to be diagnosed is determined based on the calculation result.

[0083] As a specific implementation method, the real-time status of the online operating equipment is synchronously collected, including four key operating parameters: punching force F, spindle speed n, bearing temperature T, and cumulative punching times N. These parameters are then converted into operating condition vectors using a differentiated normalization strategy. Ensure that the matching error between the input conditions and the current operating status of the device is ≤2%.

[0084] Ten independent 100-dimensional Gaussian noise vectors are generated, each following an N(0,1) distribution. This noise diversity ensures the comprehensiveness of the generated samples. The noise vectors are then compared with... The trained health status model is input synchronously, and 10 32-dimensional generated health status features are output. , … Multiple sets of samples are generated from 10 independent noise sources to avoid baseline deviation caused by a single noise source. The generation process takes ≤0.25s, meeting industrial real-time requirements, and the total diagnostic process takes ≤0.5s.

[0085] Based on 10 generated health status features, a center-interval-structure dynamic baseline is constructed using statistical modeling methods. The mean of the 10 generated health status features is taken as the baseline center. ,in, The calculation error for core features (such as kurtosis and centroid frequency) is ≤ ±3%. Construct a 95% confidence interval, where, The standard deviation of the 10 generated health status features.

[0086] Using 10 5 The dual-trigger update strategy, implemented every 3 months for each stamping cycle, ensures that the baseline drifts synchronously with equipment aging. This update occurs when the equipment's cumulative stamping count reaches 10. 5 The update process is initiated either once or after 3 months of operation, whichever comes first; 3000 new healthy samples are collected, and outliers are removed using the 3σ criterion, with a proportion ≤2%, to avoid contamination of aging data; the new samples are merged with the original training set, the earliest 3000 samples are removed to maintain the timeliness of the dataset, and the conditional adversarial network is rapidly retrained using the 3rd stage training parameters, epoch=200, lr=0.00005, to generate a new baseline.

[0087] A dual criterion of cosine similarity and feature interval is employed to achieve accurate fault classification and targeted handling. 32-dimensional health status features of the equipment in real-time are extracted. The cosine similarity metric is used to measure its relationship with the baseline center, i.e., the dynamic health baseline. The degree of matching is calculated using the following formula:

[0088] Where S∈[0,1], the closer it is to 1, the better the real-time status matches the health baseline. The health status levels are shown in Table 1.

[0089] Table 1:

[0090] This embodiment addresses the shortcomings of existing industrial equipment fault diagnosis systems, which typically rely on models trained under fixed baselines or single operating conditions. These systems fail to accurately depict the dynamic changes in equipment health characteristics under different operating conditions, leading to high false alarm rates, high false negative rates, and poor generalization ability in health diagnosis under multiple operating conditions. This problem is particularly pronounced in high-impact, high-noise scenarios such as stamping presses. This embodiment proposes a multi-domain fusion health feature extraction mechanism that integrates time-domain, frequency-domain, time-frequency-domain, and entropy features to comprehensively capture the details of equipment health status. It designs nonlinearly quantized key operating condition vectors and achieves deep coupling between operating conditions and health features through correlation screening and differential normalization. An improved adversarial network with an innovative CNN-LSTM fusion structure is used, introducing dual-auxiliary loss to address mode collapse. Finally, a quantitative health scoring system is implemented to achieve health assessment under multiple operating conditions.

[0091] This embodiment breaks through the limitations of traditional single feature + static model, and constructs an innovative end-to-end solution for data acquisition, preprocessing, condition vector construction, benchmark generation, edge-cloud collaboration, and diagnostic applications. The core architecture has three major linkage characteristics: Dual-vector collaborative representation: By precisely coupling the working condition vector and the health feature vector, the nonlinear mapping model of working condition parameters and health status is realized, which solves the distortion problem of traditional single-vector representation. Model Deep Adaptation: To address the temporal and impact characteristics of industrial vibration signals, a CNN-LSTM fusion conditional adversarial network is designed, combining dual-auxiliary loss and progressive training to overcome the collapse and accuracy bottleneck of traditional adversarial network models. Edge-cloud dynamic iteration: Real-time inference at the edge and model updates in the cloud work together to achieve full lifecycle self-adaptation of the baseline and solve the problem of adapting to equipment aging and operating condition drift.

[0092] Example 2 The purpose of this embodiment is to provide a fault diagnosis system for stamping press bearings based on conditional generative adversarial networks, including: The acquisition module is configured to: acquire the training dataset of the health status model; wherein, the training dataset includes the condition vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions; the health status model is constructed based on a conditional generative adversarial network; The training module is configured to: train a generator using the working condition vector and Gaussian noise; train a discriminator using the health status features generated by the generator and the real health status features; and perform adversarial training on the generator and the discriminator so that the discriminator cannot distinguish between the health status features generated by the generator and the real health status features, nor can it distinguish the matching between the health status features generated by the generator and the corresponding working condition vector, thus obtaining a trained health status model. The diagnostic module is configured to: collect real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed; generate health status characteristics of the press shaft to be diagnosed using a health status model; determine a dynamic health baseline based on the generated health status characteristics; and compare the dynamic health baseline with the health status characteristics to be diagnosed to obtain the fault diagnosis result of the press shaft to be diagnosed.

[0093] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0094] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0095] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0096] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0097] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0098] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0099] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0100] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0101] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0102] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0103] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks, characterized in that, include: Obtain the training dataset for the health status model; wherein, the training dataset includes vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions; the health status model is constructed based on a conditional generative adversarial network; A generator is trained using a working condition vector and Gaussian noise. A discriminator is trained using the health status features generated by the generator and the real health status features. The generator and the discriminator are then subjected to adversarial training so that the discriminator cannot distinguish between the health status features generated by the generator and the real health status features, nor can it distinguish between the matching of the health status features generated by the generator and the corresponding working condition vector, thus obtaining the trained health status model. Real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed are collected. The health status characteristics of the press shaft to be diagnosed are generated using the health status model. A dynamic health baseline is determined based on the generated health status characteristics. The dynamic health baseline is compared with the health status characteristics to be diagnosed to obtain the fault diagnosis result of the press shaft to be diagnosed.

2. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 1, characterized in that, The generator takes the working condition vector and Gaussian noise as input, uses 1D-CNN to extract local impact features, and uses a two-layer LSTM to capture cross-cycle temporal characteristics to generate health status features. The discriminator takes the health status features generated by the generator and the real health status features as input, and uses a two-layer 1D-CNN and a fully connected layer to obtain the authenticity of the health status features generated by the generator, as well as the comprehensive probability value of the matching between the health status features generated by the generator and the corresponding working condition vector.

3. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 1 or 2, characterized in that, The loss function of the generator is: ； ； ； ； Where C is the working condition vector; It is Gaussian noise; D represents the generator, E represents the discriminator, and E represents the expectation operator. Characteristics of a true state of health; Indicates deep intermediate features; L2 norm is represented; var represents the variance operator; This represents the samples generated by generator G under a given vector of operating conditions. Regarding noise The variance; , These are the weighting coefficients. The loss function of the discriminator is: ； Where D represents the discriminator; C represents the actual health status characteristics; E represents the working condition vector; and E represents the expectation operator.

4. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 1, characterized in that, The acquisition of the operating condition vector of the press bearing and the corresponding health status characteristics under the operating conditions is as follows: Collect the operating parameters and corresponding vibration signals of the stamping press bearings; Correlation analysis is performed on the operating parameters to screen out the core operating parameters whose absolute values ​​of correlation coefficients are greater than the set values. Differential normalization is applied to the core operating parameters to obtain the corresponding operating condition vector. An initial feature pool is constructed by collecting time-domain, frequency-domain, time-frequency-domain, and entropy features of the vibration signal; Redundant features are optimized in the initial feature pool to obtain health status features.

5. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 1, characterized in that, By comparing the dynamic health baseline with the characteristics of the health status to be diagnosed, the fault diagnosis results of the press shaft to be diagnosed are obtained, specifically: The degree of matching between the characteristics of the health status to be diagnosed and the dynamic health baseline is calculated using cosine similarity. The fault diagnosis result of the press shaft to be diagnosed is determined based on the calculation results.

6. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 1, characterized in that, The health status model is used to generate health status features of the press shaft to be diagnosed, and a dynamic health baseline is determined based on the generated health status features, specifically: Multiple independent Gaussian noises and the real-time operating condition vector of the press shaft to be diagnosed are input into the health state model to obtain multiple health state vectors of the press shaft to be diagnosed. The average of multiple health state vectors of the press shaft to be diagnosed is taken as the dynamic health baseline.

7. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 1, characterized in that, The health status model is trained using a progressive method, specifically as follows: The generator's 1D-CNN and the discriminator's first-layer 1D-CNN are trained in the first stage, enabling the shallow network to learn the basic mapping relationship between working conditions and health status features. The generator's 1D-CNN and first-layer LSTM, as well as the discriminator's two-layer 1D-CNN, are trained in a second stage to model cross-cycle temporal correlations. A third stage of training is performed on the complete generator and the complete discriminator to optimize the matching accuracy between deep features and working conditions.

8. The method for fault diagnosis of stamping press bearings based on conditional generative adversarial networks as described in claim 4, characterized in that, Redundant features are optimized in the initial feature pool to obtain health status features. Specifically, redundant features are removed using the Pearson correlation coefficient, feature weights are sorted using the ReliefF algorithm, and the importance of features is verified and evaluated using a random forest to obtain health status features.

9. A fault diagnosis system for stamping press bearings based on conditional generative adversarial networks, characterized in that, include: The acquisition module is configured to: acquire the training dataset of the health status model; wherein, the training dataset includes vectors of different working conditions of the press bearing and the corresponding health status features under the working conditions; the health status model is constructed based on a conditional generative adversarial network; The training module is configured to: train a generator using a working condition vector and Gaussian noise; train a discriminator using health state features generated by the generator and real health state features; and perform adversarial training on the generator and the discriminator so that the discriminator cannot distinguish between the health state features generated by the generator and the real health state features, nor can it distinguish the matching between the health state features generated by the generator and the corresponding working condition vector, thereby obtaining the trained health state model. The diagnostic module is configured to: collect real-time operating parameters and corresponding vibration signals of the bearing of the press to be diagnosed; generate health status characteristics of the press shaft to be diagnosed using the health status model; determine a dynamic health baseline based on the generated health status characteristics; and compare the dynamic health baseline with the health status characteristics to be diagnosed to obtain the fault diagnosis result of the press shaft to be diagnosed.

10. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-8.

Images