A pure hydrogen shaft furnace ironmaking process fault diagnosis method based on a generative adversarial network

Abstract

The application discloses a pure hydrogen shaft furnace ironmaking process fault diagnosis method based on a generative adversarial network, which comprises adaptive dynamic space-time feature regularization, slow feature optimization generative adversarial diagnosis model construction, offline fault sample generation and diagnosis model training, and pure hydrogen shaft furnace ironmaking process online fault diagnosis. In the actual pure hydrogen shaft furnace ironmaking process, process data often has the characteristics of few fault samples and non-stationarity. The application ingeniously refers to the similar double-input structure of the generative adversarial network and deep slow feature analysis, proposes a new model, and solves the problems of few fault samples and process non-stationarity inherent in the pure hydrogen shaft furnace ironmaking fault process. And aiming at the pain point of difficult training of the new model, a new loss function is designed to further improve the diagnosis accuracy and model robustness, thereby effectively guaranteeing the safety and high efficiency of the production process.

Description

Technical Field

[0001] This invention relates to a fault diagnosis method for pure hydrogen shaft furnace ironmaking process based on generative adversarial networks. Background Technology

[0002] In the green transformation of the steel industry, pure hydrogen shaft furnace ironmaking, as a highly promising low-carbon ironmaking technology, has attracted much attention. This technology uses hydrogen as a reducing agent, significantly reducing carbon dioxide emissions compared to traditional blast furnace ironmaking, aligning with the global trend of sustainable development. However, the pure hydrogen shaft furnace ironmaking process involves complex physicochemical changes, and the raw materials are flammable and explosive, posing numerous challenges in actual production. Real-time fault diagnosis is a key factor in ensuring its stable and efficient operation.

[0003] Currently, diagnostic methods for ironmaking processes mainly include mechanistic modeling, knowledge modeling, and data-driven methods. Mechanismic modeling, based on physicochemical principles and system structure, constructs mathematical models that can theoretically reveal the relationship between faults and parameters. However, the complexity of hydrogen reactions and the uncertainty of gas-solid flow within the furnace during pure hydrogen shaft furnace ironmaking make accurate physical model construction extremely difficult, resulting in low fault diagnosis accuracy. Knowledge modeling relies on expert experience; in practical applications, obtaining comprehensive and accurate rules and rich historical fault information is difficult and costly. Furthermore, with constantly changing production conditions, knowledge updates are not timely, making it difficult to adapt to complex and ever-changing fault diagnosis needs. For pure hydrogen shaft furnace ironmaking, as a relatively new ironmaking technology, the accumulation of expert knowledge still requires time.

[0004] With the development of sensor and computer technology, data-driven methods have been applied to fault diagnosis in pure hydrogen shaft furnace ironmaking. However, the data characteristics of this process pose significant obstacles to fault diagnosis. On the one hand, as a relatively new type of ironmaking equipment, the pure hydrogen shaft furnace itself has relatively little historical fault data. Furthermore, labeling fault data in pure hydrogen shaft furnace ironmaking is difficult. Due to the timeliness of fault handling, production needs to be quickly restored after a fault occurs, resulting in a limited number of available labeled fault samples. This limited sample size leads to insufficient model training, decreased classification accuracy, and difficulty in accurately identifying various faults. On the other hand, the pure hydrogen shaft furnace ironmaking process exhibits significant non-stationarity. Key parameters such as hydrogen flow rate, pressure, and furnace temperature are affected by factors such as fluctuations in raw material characteristics, changes in equipment operating status, and external environmental interference, causing their time series mean and variance to change continuously over time. This non-stationarity results in the intertwining of changes in normal production status and changes caused by faults, making it difficult for existing methods to effectively extract fault-related features and increasing the difficulty of fault diagnosis. Therefore, developing a fault diagnosis method for pure hydrogen shaft furnace ironmaking that can effectively solve the above problems is urgently needed. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention aims to provide a fault diagnosis method for the pure hydrogen shaft furnace ironmaking process based on generative adversarial networks, comprising the following steps: adaptive dynamic spatiotemporal feature regularization, slow feature optimization for generative adversarial diagnostic model construction, offline fault sample generation and diagnostic model training, and online fault diagnosis of the pure hydrogen shaft furnace ironmaking process.

[0006] This invention specifically studies fault diagnosis in the pure hydrogen shaft furnace ironmaking process. Therefore, this invention proposes a fault diagnosis method for the pure hydrogen shaft furnace ironmaking process based on a generative adversarial framework to address the characteristics of limited fault sample size and non-stationarity in the pure hydrogen shaft furnace ironmaking process data. First, adaptive dynamic spatiotemporal feature regularization is used to preprocess the data, integrating dynamic features and unifying the scale. Next, a slow feature-optimized generative adversarial diagnostic model is constructed, which can extract the non-stationarity of the data, generate high-quality fault samples, and improve the model training stability. Then, the generated high-quality fault samples are used to expand the dataset and train the discriminator of the proposed model to achieve more accurate classification performance. Finally, the online fault diagnosis stage of the pure hydrogen shaft furnace ironmaking process is entered. Simply input the real-time collected samples into the discriminator to quickly and accurately determine the current operating condition.

[0007] A fault diagnosis method for pure hydrogen shaft furnace ironmaking process based on generative adversarial networks includes the following steps: adaptive dynamic spatiotemporal feature regularization, slow feature optimization generative adversarial diagnosis model construction, offline fault sample generation and diagnosis model training, and online fault diagnosis of pure hydrogen shaft furnace ironmaking process.

[0008] First, adaptive dynamic spatiotemporal feature regularization is used to preprocess the data, integrating dynamic features and unifying the scale. Next, a slow feature optimization and adversarial diagnostic model is constructed. Slow features refer to those features that change slowly during data transformation and reflect long-term trends and inherent characteristics. By introducing slow feature constraints and a least-squares structure, the model focuses on slow features during generation, eliminating the influence of noise and non-stationarity, thereby generating high-quality fault samples and improving model training stability. Then, the generated high-quality fault samples are used to expand the dataset and train the model's discriminator, enabling the discriminator to achieve more accurate classification performance. Finally, the online fault diagnosis stage of the pure hydrogen vertical shaft furnace ironmaking process is entered, where real-time collected samples are input into the discriminator to determine the current operating condition.

[0009] The adaptive dynamic spatiotemporal feature normalization includes the following steps:

[0010] (1.1) Data sample collection: Based on the fault occurrence time recorded in the worker's duty log, normal sample x and fault sample x' of the pure hydrogen vertical furnace ironmaking process are collected by sensors, including variables: hydrogen circulation humidity, hydrogen circulation pressure, hydrogen circulation temperature, vertical furnace mass flow rate, metallization rate in the vertical furnace, temperature in the vertical furnace, and gas flow rate in the vertical furnace.

[0011] (1.2) Integrating dynamic characteristics: The collected pure hydrogen shaft furnace ironmaking data is processed using sliding window technology. The sliding window step size l and size ω are set to reshape the sample into two-dimensional data.

[0012] (1.3) Unify data scale: The data after sliding window processing is standardized by using the z-score standardization method to make the scale of vertical furnace data of different feature dimensions consistent.

[0013] The construction of the slow feature optimization generation adversarial diagnostic model includes the following steps:

[0014] The model encompasses a generator, a discriminator, a slow feature constraint module, and an improved least-squares structure.

[0015] (2.1) The generator consists of fully connected layers and deconvolutional layers, and generates fault samples of pure hydrogen shaft furnaces with random vector z as input.

[0016] (2.2) The discriminator consists of a convolutional layer and a fully connected layer, used to distinguish between real and fake samples and determine the fault category c of the pure hydrogen vertical furnace;

[0017] (2.3) Introduce slow feature constraints into the discriminator objective function, and extract slow features through the redefined objective function;

[0018] (2.4) By drawing on the structure of the least squares loss function, the objective function is further optimized to improve the sample generation quality and training stability of the proposed model.

[0019] The offline fault sample generation and diagnostic model training include the following steps:

[0020] (3.1) Collect pure hydrogen vertical furnace process data under different working conditions as training dataset, construct training samples using adaptive dynamic spatiotemporal feature regularization, and initialize model weights and biases;

[0021] (3.2) Improve the generator’s fault sample generation performance by adversarial training between the generator and the discriminator. That is, after training the discriminator for d rounds with the generator weight fixed, train the generator with the discriminator weight fixed, and then use gradient descent to update K times alternately.

[0022] (3.3) The generated samples are put into the original dataset, and the diagnostic performance of the discriminator is trained with the enhanced dataset to make the discriminator more accurate in classification.

[0023] The online fault diagnosis of the pure hydrogen shaft furnace ironmaking process includes the following steps:

[0024] (4.1) Real-time acquisition of pure hydrogen vertical furnace measurement data, and construction of test samples through adaptive dynamic spatiotemporal feature regularization;

[0025] (4.2) Input the test sample into the trained discriminator for fault diagnosis. The discriminator determines whether the sample is a fault sample and the type of fault it belongs to.

[0026] The beneficial effects of this invention are:

[0027] This invention overcomes the shortcomings of current fault diagnosis methods in pure hydrogen shaft furnace ironmaking processes by proposing a fault diagnosis method based on generative adversarial networks (GANs). This method, referencing the adversarial training of GANs, uses a generator to generate fault samples, addressing the problem of insufficient fault samples. Simultaneously, a slow feature constraint module is introduced to extract slow features from the data, effectively handling the non-stationarity of pure hydrogen shaft furnace ironmaking data. This enables the model to better identify changes in normal conditions and changes caused by faults, improving the accuracy of fault diagnosis. Attached Figure Description

[0028] Figure 1 The present invention relates to a fault diagnosis method for pure hydrogen shaft furnace ironmaking process based on generative adversarial networks.

[0029] Figure 2 This is a schematic diagram of the slow feature optimization generation and adversarial diagnostic model for the pure hydrogen shaft furnace ironmaking process according to the present invention.

[0030] Figure 3 This is a schematic diagram of the fault diagnosis process for the pure hydrogen vertical shaft furnace ironmaking process of the present invention. Detailed Implementation

[0031] The following uses a fault in a pure hydrogen shaft furnace ironmaking process as an example to introduce a fault diagnosis method based on generative adversarial networks. Figure 1 As shown, it includes the following steps:

[0032] (1) Data Sample Collection: A steel plant using a pure hydrogen shaft furnace ironmaking process was selected as the research object, and the pure hydrogen shaft furnace of the plant was used as the data collection source. Data was collected at fixed time intervals (e.g., 10 seconds) over a period of time, covering data under normal operation and various fault conditions of the pure hydrogen shaft furnace. Among them, the number of samples for hydrogen leakage, pipeline failure, and furnace charge blockage failure were 126, 968, and 333, respectively. Based on the on-site operation logs and the judgment of engineers, multiple subsets were extracted from the entire dataset, including training set and test set, covering key variables such as hydrogen circulation humidity, hydrogen circulation pressure, hydrogen circulation temperature, shaft furnace mass flow rate, metallization rate in the shaft furnace, temperature in the shaft furnace, and gas flow rate in the shaft furnace for monitoring.

[0033] (2) Adaptive dynamic spatiotemporal feature normalization: Based on the number of key variables selected, the step size of the sliding window is adaptively set to l=1 and the size ω=7. This allows the model to fully capture the dynamic features of the data while avoiding information loss or redundancy caused by excessively large or small windows. After integrating the samples into two-dimensional data using the sliding window technique, the z-score normalization method is used to ensure that the process data of different feature dimensions are scaled uniformly.

[0034] (3) Slow feature optimization generates adversarial diagnostic model construction: such as Figure 2 As shown, the generator consists of fully connected layers, two-dimensional deconvolutional layers, etc. The discriminator consists of two-dimensional convolutional layers, fully connected layers, etc. A slow feature constraint module and a least squares improved structure are added to the discriminator. The network structures of the generator and discriminator are shown in Table 1 below.

[0035]

[0036] The objective function of the slow feature constraint is to minimize Its symbol

[0037] After network layer computation, one-dimensional batch normalization and modified linear unit activation functions are used to improve model performance. Specific network structure parameters can be adjusted through parameter tuning and based on the characteristics of pure hydrogen shaft furnace ironmaking data. These structural designs fully leverage the feature extraction capabilities of deep learning models to meet the needs of fault diagnosis. Based on the number of fault categories, the discriminator iteration count K is set to 3. The number of training epochs is set to 100, and due to limited training data, the batch size is set to 50. These parameters were determined through multiple experiments and optimizations, ensuring both effective model training and improved training efficiency. RMSProp is chosen as the model's optimizer. Momentum-based optimization algorithms may cause the discriminator gradient to become negative, leading to unstable model training. RMSProp, with its iterative decay mechanism, converges faster than SGD and is more suitable for the training requirements of this model.

[0038] (4) Offline fault sample generation and diagnostic model training: such as Figure 3 As shown, during model training, the discriminator is first trained 3 times with fixed generator weights, then the generator is trained with fixed discriminator weights, alternating for 100 rounds. During training, the model weights gradually converge, continuously generating high-quality fault samples. Ultimately, the generator and discriminator reach a Nash equilibrium during training. The generator synthesizes various fault samples to augment the training dataset, thereby improving the diagnostic performance of the discriminator.

[0039] (5) Online fault diagnosis of pure hydrogen shaft furnace ironmaking process: Real-time acquisition of pure hydrogen shaft furnace data, after sliding window processing and standardization, is input into the discriminator for fault diagnosis. The discriminator outputs the fault type judgment result of the sample. If the sample is judged to be faulty, an alarm is issued in time to prompt the operator to take appropriate measures, such as checking whether there is a leak in the hydrogen delivery pipeline and adjusting the furnace temperature control parameters. In order to better reflect the superiority of the method, three indicators are defined in the diagnosis process: accuracy, precision, and recall.

[0040]

[0041] Where TP represents true positives, i.e., the number of samples that are actually positive and are predicted as positive; TN represents true negatives, i.e., the number of samples that are actually negative and are predicted as negative; FP represents false positives, i.e., the number of samples that are actually negative but are predicted as positive; and FN represents false negatives, i.e., the number of samples that are actually positive but are predicted as negative. Using 50 test samples of each type of fault, and the remainder as training samples, the results compared with existing methods are shown in Table 2 below.

[0042] Multilayer perceptron 80.85％ 85.81％ 80.85％ Convolutional Neural Networks 81.95％ 87.03％ 81.95％ Auxiliary classifier generative adversarial network 93.25％ 94.21％ 93.25％ The method 95.75％ 95.95％ 95.75％

[0043] It can be seen that convolutional neural networks are more suitable for extracting multi-dimensional fault features of vertical shaft furnaces compared to multi-layer perceptrons. The auxiliary classifier, generative adversarial network (GAN), can solve the few-sample problem by generating fault samples, thus significantly improving model performance. Furthermore, the method described above considers both the few-sample problem and the non-stationarity of the process, thereby generating higher-quality samples and achieving better accuracy. Similarly, its accuracy and recall also achieve optimal performance.

[0044] Through the above specific implementation methods, the fault diagnosis method for pure hydrogen shaft furnace ironmaking based on generative adversarial networks of the present invention can be effectively applied to the actual pure hydrogen shaft furnace ironmaking process. It achieves high-precision fault diagnosis for three types of faults: hydrogen leakage, pipeline malfunction, and charge blockage, providing a reliable guarantee for the stable operation of the pure hydrogen shaft furnace. In practical applications, the model parameters and structure can be further optimized and adjusted according to the actual operating conditions and data characteristics of the pure hydrogen shaft furnace to adapt to different production needs.

[0045] The embodiments described above can be further combined or replaced, and these embodiments are merely descriptions of preferred embodiments of the present invention, not limitations on the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalent technical solutions.

Claims

1. A fault diagnosis method for a pure hydrogen shaft furnace ironmaking process based on generative adversarial networks, characterized in that, The process includes the following steps: adaptive dynamic spatiotemporal feature regularization, slow feature optimization to generate adversarial diagnostic model construction, offline fault sample generation and diagnostic model training, and online fault diagnosis of pure hydrogen vertical shaft furnace ironmaking process. First, adaptive dynamic spatiotemporal feature regularization is used to preprocess the data, integrate dynamic features and unify the scale; then, a slow feature optimization and adversarial diagnostic model is constructed. Slow features refer to those features that change relatively slowly during the data change process and can reflect the long-term trend and inherent characteristics of the data; by introducing slow feature constraints and least squares structure, the model focuses on slow features during the generation process, eliminates the influence of noise and non-stationarity, thereby generating high-quality fault samples and improving the model training stability. Then, the dataset is expanded using the generated high-quality fault samples, and the discriminator of the model is trained to enable the discriminator to obtain more accurate classification performance. Finally, the online fault diagnosis stage of the pure hydrogen shaft furnace ironmaking process is entered, and the real-time collected samples are input into the discriminator to determine the current working condition. The adaptive dynamic spatiotemporal feature normalization includes the following steps: (1.1) Data sample collection: Based on the fault occurrence time recorded in the worker's shift log, normal samples of the pure hydrogen shaft furnace ironmaking process were collected using sensors. x With fault samples x' The variables include: hydrogen circulation humidity, hydrogen circulation pressure, hydrogen circulation temperature, shaft furnace mass flow rate, metallization rate in the shaft furnace, temperature in the shaft furnace, and gas flow rate in the shaft furnace. (1.2) Integrating dynamic characteristics: The collected pure hydrogen shaft furnace ironmaking data is processed using sliding window technology, and the sliding window step size is set. l And the size ω, reshape the sample into two-dimensional data; (1.3) Unify data scale: Use the z-score standardization method to standardize the data after sliding window processing to make the scale of vertical furnace data of different feature dimensions consistent; The construction of the slow feature optimization generation adversarial diagnostic model includes the following steps: The model encompasses a generator, a discriminator, a slow feature constraint module, and an improved least-squares structure. (2.1) The generator consists of fully connected layers and deconvolutional layers, and generates fault samples of pure hydrogen shaft furnace with random vector z as input. ; (2.2) The discriminator consists of a convolutional layer and a fully connected layer, used to distinguish between real and fake samples and determine the fault category c of the pure hydrogen shaft furnace; (2.3) Introduce slow feature constraints into the discriminator objective function, and extract slow features through the redefined objective function; (2.4) By drawing on the structure of the least squares loss function, the objective function is further optimized to improve the sample generation quality and training stability of the proposed model; The offline fault sample generation and diagnostic model training include the following steps: (3.1) Collect pure hydrogen vertical furnace process data under different working conditions as training dataset, construct training samples using adaptive dynamic spatiotemporal feature regularization, and initialize model weights and biases; (3.2) Improve the generator’s fault sample generation performance by adversarial training between the generator and the discriminator. That is, after training the discriminator for d rounds with the generator weight fixed, train the generator with the discriminator weight fixed, and then use gradient descent to update it K times alternately. (3.3) The generated samples are put into the original dataset, and the diagnostic performance of the discriminator is trained with the enhanced dataset to make the discriminator more accurate in classification.

2. The method according to claim 1, characterized in that, The online fault diagnosis of the pure hydrogen shaft furnace ironmaking process includes the following steps: (4.1) Real-time acquisition of pure hydrogen vertical furnace measurement data, and construction of test samples through adaptive dynamic spatiotemporal feature regularization; (4.2) Input the test sample into the trained discriminator for fault diagnosis. The discriminator determines whether the sample is a fault sample and the type of fault it belongs to.

Images