Steel production equipment vibration abnormity detection method based on DCGAN
By combining DCGAN with FFT transform, the problem of small sample size in the detection of vibration anomalies in steel production equipment using traditional methods is solved, achieving efficient real-time fault early warning and condition monitoring, and improving the accuracy and early detection.
Patent Information
- Application Number
- CN202511402898.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-28
- Publication Date
- 2026-02-03
AI Technical Summary
Traditional supervised learning methods for detecting vibration anomalies in steel production equipment suffer from poor generalization ability of models trained on small abnormal datasets, resulting in low recognition rates, weak detection capabilities, and difficulty in identifying novel anomaly patterns. Furthermore, existing methods are slow to respond to progressive faults.
A DCGAN-based approach is adopted, combining Fast Fourier Transform and Generative Adversarial Network. Through adversarial training of the generator and discriminator, the features of vibration signals are extracted. Data is processed using sliding window and wavelet transform. Asymmetric training rounds and label smoothing techniques are employed to improve the model's training stability and generalization ability, thereby achieving dynamic anomaly detection.
It enables efficient and real-time anomaly detection of steel production equipment, reduces manual inspection costs, and improves the accuracy and early warning of faults. It is suitable for condition monitoring of rotating machinery equipment.
Smart Images
Figure CN121456435A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial artificial intelligence and intelligent equipment operation and maintenance technology, and in particular to a method for detecting abnormal vibrations in steel production equipment based on DCGAN. Background Technology
[0002] Deep Convolutional Generative Adversarial Networks (DCGANs) learn the essential feature distribution of vibration signals through adversarial training between the generator and discriminator, effectively identifying abnormal states that deviate from normal patterns. This technology directly extracts deep features from raw vibration data, avoiding the limitations of traditional methods that rely on expert experience to design features, significantly improving the accuracy and automation level of anomaly detection. In other words, DCGAN technology directly extracts various features from vibration data, saving on manual inspection costs.
[0003] High-frequency vibration signal analysis is a key technology for industrial equipment condition monitoring. It involves converting mechanical vibrations into electrical signals using acquisition devices such as accelerometers, and then extracting feature information using signal processing techniques. In industrial equipment condition monitoring, the collected vibration data exhibits a significant imbalance: the sample size of vibration data under normal operating conditions is sufficient, while vibration data under abnormal conditions (especially before and after a fault) is extremely scarce. This data distribution characteristic poses a serious challenge to traditional supervised learning methods—when training models using small-sample abnormal datasets, the models often exhibit poor generalization ability and detection accuracy due to insufficient learning of abnormal patterns. Specifically, this manifests as: 1) low recognition rate of known abnormal types; 2) weak detection ability of novel abnormal patterns; and 3) a tendency to misclassify marginal normal states as abnormal. This small-sample learning dilemma severely restricts the practical application effect of data-driven fault early warning systems.
[0004] Taking a gearbox as an example, the fault records of this equipment from March to July 2024 show that before March 2024, the equipment was in normal operation, and the vibration signal characteristic parameters were stable within the standard threshold range. However, after the first fault occurred in March 2024, without maintenance intervention, the equipment continued to exhibit abnormal operating characteristics, specifically manifested in significant morphological changes in its Fast Fourier Transform (FFT) spectrum. This complete "normal-abnormal" state evolution process provides a typical data sample for studying the gradual development law of mechanical faults. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a method for detecting vibration anomalies in steel production equipment based on DCGAN. This method achieves intelligent feature extraction and dynamic anomaly detection of high-frequency vibration signals by integrating fast Fourier transform frequency domain analysis and generative adversarial network feature learning. It is particularly suitable for real-time status monitoring and early fault warning of rotating machinery equipment in the steel industry.
[0006] To solve the above-mentioned technical problems, the present invention provides a DCGAN-based method for detecting abnormal vibrations in steel production equipment, comprising the following steps: Step 1: Data Acquisition and Processing. High-frequency acceleration data of the device is acquired through the server and formed into a CSV file. The original vibration signal and timestamp data are loaded from the CSV file. The vibration signal in string format is cleaned and converted. After unifying the data length, the signal is divided into 1024 data segments using a sliding window. Step 2: Extract the frequency domain features of the vibration signal through FFT transformation, screen samples with qualified spectral peaks, obtain time-frequency features through wavelet transform, normalize the data and divide it into training set and validation set, store the processed spectrum data and corresponding timestamps in NPZ format, and automatically generate spectrum visualization charts for quality monitoring. Step 3: Using the DCGAN architecture, the generator generates simulated data through transposed convolution, and the discriminator uses a convolutional network to distinguish the authenticity of the data. An asymmetric training ratio of three training rounds for the discriminator and two training rounds for the generator is adopted to train the discriminator model. Step 4: Set the real sample label in the discriminator model to 0.9, and keep the generated sample label at 0.0. This prevents the discriminator from being overconfident in its predictions of real samples, improves the gradient vanishing problem, and enhances training stability. Step 5: The discriminator model divides the dataset into a training set and a test set ratio of 1:1. During training, the model performance is evaluated by monitoring the discriminator's ability to distinguish between real and generated data. After training, the training data itself is used for testing to calculate the discriminator's acceptance rate and anomaly detection rate on real data. Step 6: Load the trained discriminator model, calculate the discrimination score of the real-time vibration signal data, and trigger an alarm when the discrimination score is lower than the threshold.
[0007] Furthermore, during the training of the discriminator model, the vibration signal is superimposed with two Gaussian noises of different scales. By combining the diversity of noise, the distribution coverage of the generated samples is enhanced, thus preventing the generator from getting trapped in local optima.
[0008] Furthermore, in step four, the BCELoss loss function is used to indirectly improve gradient behavior through label smoothing, making the training process smoother and suitable for generating high-dimensional time series data.
[0009] The present invention, a DCGAN-based method for detecting vibration anomalies in steel production equipment, employs the aforementioned technical solution. Specifically, this method acquires high-frequency acceleration data from the server, performs FFT transformation to convert the time-domain signal into frequency-domain features, and generates a 1024-point data segment through a sliding window. It utilizes a DCGAN architecture, where the generator generates simulated data through transposed convolution, and the discriminator uses a convolutional network to determine the authenticity of the data. An asymmetric training ratio of three training rounds for the discriminator and two for the generator is employed. The trained discriminator model is loaded, and the discrimination score of the real-time vibration signal data is calculated. An alarm is triggered when the discrimination score falls below a threshold. This method, by integrating Fast Fourier Transform frequency domain analysis with generative adversarial network feature learning, achieves intelligent feature extraction and dynamic anomaly detection of high-frequency vibration signals, making it particularly suitable for real-time status monitoring and early fault warning of rotating machinery in the steel industry. Attached Figure Description
[0010] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the vibration anomaly detection method for steel production equipment based on DCGAN of the present invention. Detailed Implementation
[0011] Implementation, for example Figure 1 As shown, the DCGAN-based method for detecting abnormal vibrations in steel production equipment according to this invention includes the following steps: Step 1: Data Acquisition and Processing. High-frequency acceleration data of the device is acquired through the server and formed into a CSV file. The original vibration signal and timestamp data are loaded from the CSV file. The vibration signal in string format is cleaned and converted. After unifying the data length, the signal is divided into 1024 data segments using a sliding window. Step 2: Extract the frequency domain features of the vibration signal through FFT transformation, screen samples with qualified spectral peaks, obtain time-frequency features through wavelet transform, normalize the data and divide it into training set and validation set, store the processed spectrum data and corresponding timestamps in NPZ format, and automatically generate spectrum visualization charts for quality monitoring. Step 3: Adopting the DCGAN architecture, the generator generates simulated data through transposed convolution, and the discriminator uses a convolutional network to distinguish the authenticity of the data. An asymmetric training ratio of three training rounds for the discriminator and two training rounds for the generator is adopted to train the discriminator model. By frequently updating the discriminator, its discrimination ability is improved, and the generator is prevented from dominating the training process too early, thereby alleviating the mode collapse problem. Step 4: Set the real sample label in the discriminator model to 0.9, and keep the generated sample label at 0.0. This prevents the discriminator from being overconfident in its predictions of real samples, improves the gradient vanishing problem, and enhances training stability. Step 5: The discriminator model divides the dataset into a training set:test set ratio of 1:1 (without the traditional validation set partitioning). During training, the model performance is evaluated by monitoring the discriminator's ability to distinguish between real and generated data. After training, the training data itself is used for testing to calculate the discriminator's acceptance rate and anomaly detection rate on real data. Step 6: Load the trained discriminator model and calculate the discrimination score of the real-time vibration signal data. When the discrimination score is lower than the threshold, an alarm is triggered. The discrimination score is the model's real-time confidence output for a single vibration signal sample, which is equivalent to the acceptance rate of a single sample.
[0012] Preferably, during discriminator model training, the vibration signal is superimposed with two Gaussian noises of different scales. By mixing the diversity of noise, the distribution coverage of the generated samples is enhanced, and the generator is prevented from getting trapped in local optima.
[0013] Preferably, in step four, the BCELoss loss function is used to indirectly improve gradient behavior through label smoothing, making the training process smoother and suitable for generating high-dimensional time series data.
[0014] This method addresses the issue of vibration signal discrimination in steel production equipment. It solves problems such as insufficient training of supervised learning models due to the large amount of normal equipment data but scarce fault samples, and the lag in response of existing threshold alarm methods to progressive faults. This enables real-time early warning of equipment faults and reduces the cost of manual inspection. Through a hybrid learning framework combining adversarial training and autoencoders, it achieves high-precision equipment anomaly detection under unsupervised conditions, overcoming the dependence of traditional supervised learning on labeled data. The method employs a DCGAN architecture combined with FFT transform to ensure efficient feature extraction of vibration signals, making it more suitable for time-series signal analysis than traditional 2D convolution methods. Furthermore, an adversarial generative method based on small-sample data augmentation ensures improved model generalization ability even with scarce training samples.
[0015] Taking a gearbox as an example, 677 days of high-frequency acceleration signal data from a measuring point, collected from May 11, 2023 to March 7, 2025, were used for training. The original dataset was divided into training and testing sets using a 1:1 dataset partitioning strategy. Data with an accuracy above 0.5 was classified as normal operating conditions, while data with an accuracy below 0.5 was classified as abnormal operating conditions. The matching rate was used as the evaluation metric, i.e., whether the model's operating condition identification accuracy matched the fault records. On the validation set, the matching rate was 99.70%, indicating that this method can fully realize the detection and diagnosis of vibration anomalies in steel production equipment.
Claims
1. A method for detecting abnormal vibration in steel production equipment based on DCGAN, characterized in that... This method includes the following steps: Step 1: Data Acquisition and Processing. High-frequency acceleration data of the device is acquired through the server and formed into a CSV file. The original vibration signal and timestamp data are loaded from the CSV file. The vibration signal in string format is cleaned and converted. After unifying the data length, the signal is divided into 1024 data segments using a sliding window. Step 2: Extract the frequency domain features of the vibration signal through FFT transformation, screen samples with qualified spectral peaks, obtain time-frequency features through wavelet transform, normalize the data and divide it into training set and validation set, store the processed spectrum data and corresponding timestamps in NPZ format, and automatically generate spectrum visualization charts for quality monitoring. Step 3: Using the DCGAN architecture, the generator generates simulated data through transposed convolution, and the discriminator uses a convolutional network to distinguish the authenticity of the data. An asymmetric training ratio of three training rounds for the discriminator and two training rounds for the generator is adopted to train the discriminator model. Step 4: Set the real sample label in the discriminator model to 0.9, and keep the generated sample label at 0.
0. This prevents the discriminator from being overconfident in its predictions of real samples, improves the gradient vanishing problem, and enhances training stability. Step 5: The discriminator model divides the dataset into a training set and a test set ratio of 1:
1. During training, the model performance is evaluated by monitoring the discriminator's ability to distinguish between real and generated data. After training, the training data itself is used for testing to calculate the discriminator's acceptance rate and anomaly detection rate on real data. Step 6: Load the trained discriminator model, calculate the discrimination score of the real-time vibration signal data, and trigger an alarm when the discrimination score is lower than the threshold.
2. The method for detecting abnormal vibrations in steel production equipment based on DCGAN according to claim 1, characterized in that: During discriminator model training, the vibration signal is superimposed with two Gaussian noises of different scales. By mixing the noise, the distribution coverage of the generated samples is enhanced, thus preventing the generator from getting stuck in local optima.
3. The method for detecting abnormal vibrations in steel production equipment based on DCGAN according to claim 1, characterized in that: In step four, the BCELoss loss function is used to indirectly improve gradient behavior through label smoothing, making the training process smoother and suitable for generating high-dimensional time series data.