Blast furnace ironmaking process fault diagnosis method based on robust semi-supervised classification framework
Through a robust semi-supervised classification framework combined with supervised and unsupervised learning, the training signal annealing and anti-perturbation calculation method is used to solve the problem of unbalanced fault data during blast furnace iron smelting, achieving efficient and robust fault diagnosis and improving The accuracy and real-time nature of fault detection.
Patent Information
- Application Number
- CN202410938311.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-07-10
- Filing Date
- 2024-07-13
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-07-13
AI Technical Summary
The failure data is unbalanced during blast furnace iron smelting. Traditional supervised learning methods are difficult to effectively learn and capture failure modes. The semi-supervised learning methods perform bottlenecks when processing labeled and unlabeled data imbalances, resulting in robustness and ability of diagnostic results. Inadequate explanation.
Using a robust semi-supervised classification framework (RSCF), combined with supervised and unsupervised parts, the signal annealing method is trained to prevent overfitting, the model smoothness is enhanced by using unlabeled data, and a power iteration method and finite difference method are designed. Anti-perturbation calculation method, integrates supervised and unsupervised loss training to achieve a robust fault diagnosis model.
Effectively utilize limited labeled samples and large amounts of unlabeled samples to prevent overfitting, enhance the generalization ability and robustness of the model, improve the accuracy and real-time of fault diagnosis, and be able to detect faults early and issue alarms to avoid failure deterioration.
Smart Images

Figure CN118916738A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of fault diagnosis of blast furnace ironmaking process, and in particular to a fault diagnosis method of blast furnace ironmaking process based on a robust semi-supervised classification framework. Background Art
[0002] The blast furnace ironmaking process is the core link of steel production. Through blast furnace ironmaking, iron ore reacts with coke at high temperature to produce pig iron, which is the basis of steel manufacturing. The efficiency and quality of blast furnace ironmaking directly affect the cost and performance of steel products, so its position in the steel industry is crucial. However, the blast furnace ironmaking process is extremely complex, involving a variety of physical and chemical reactions, and needs to be carried out under high temperature and high pressure. This process has strict requirements on raw material composition, temperature control, airflow distribution, equipment condition, etc. Problems in any link may cause failures and affect the continuity and stability of production. Effective fault diagnosis can detect potential problems in advance, avoid production interruptions, reduce equipment losses, and reduce economic losses.
[0003] Fault events are relatively rare in the blast furnace ironmaking process, resulting in very limited labeled fault data (i.e., instances of known faults), while data on normal operation is relatively abundant. This data imbalance poses a huge challenge to the training of fault diagnosis models. Traditional supervised learning methods rely on a large amount of labeled data. When the labeled data is insufficient, it is difficult for the model to effectively learn and capture fault modes, and overfitting problems are prone to occur. The fault modes in the blast furnace ironmaking process are diverse and complex, and the causes of the faults may involve the interaction of multiple parameters. To build an accurate fault diagnosis model, it is necessary to fully consider all possible fault modes, which places high demands on the design and training of the model. The blast furnace ironmaking process needs to run continuously, and any downtime will cause huge economic losses. Therefore, the fault diagnosis model must not only be accurate, but also real-time, and be able to detect and issue alarms in time at the early stages of the fault to prevent the fault from further deteriorating.
[0004] Traditional fault diagnosis mostly uses supervised learning methods, which require a large amount of labeled data for training, such as support vector machine (SVM), random forest (RF), artificial neural network (ANN), etc. However, in practical applications, the scarcity of labeled fault data makes it difficult to effectively apply these methods. In addition, supervised learning methods have poor robustness when facing changes in data distribution and are difficult to adapt to the complex and changeable working conditions in the blast furnace ironmaking process. Unsupervised learning methods use unlabeled data for training, such as cluster analysis, principal component analysis (PCA), isolation forest, etc. Although these methods can perform anomaly detection without labeled data, they lack clear fault pattern recognition capabilities, are difficult to provide specific fault causes and types, and have poor interpretability of diagnostic results. Semi-supervised learning methods attempt to combine the advantages of supervised and unsupervised learning, using a small amount of labeled data and a large amount of unlabeled data for model training. However, existing semi-supervised learning methods still have performance bottlenecks when dealing with imbalanced labeled and unlabeled data, and fail to fully utilize the potential information in unlabeled data to improve the diagnostic ability and robustness of the model. Summary of the invention
[0005] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a blast furnace ironmaking process fault diagnosis method based on a robust semi-supervised classification framework (RSCF).
[0006] The technical solution of the present invention to achieve its technical purpose is as follows:
[0007] A fault diagnosis method for blast furnace ironmaking process based on robust semi-supervised classification framework (RSCF) is proposed. The method combines supervised and unsupervised parts to effectively utilize limited labeled samples of blast furnace and a large number of unlabeled samples. In the supervised part, a training technique based on training signal annealing (TSA) is adopted to prevent overfitting in view of the gap between labeled data and unlabeled data. In addition, the unsupervised loss part enhances the smoothness of the model by utilizing unlabeled data, which is achieved by making the output results of the model consistent before and after adding adversarial interference. Then, an adversarial disturbance calculation method is designed based on power iteration method and finite difference method. Finally, a robust fault diagnosis model is realized through overall training of supervised loss and unsupervised loss.
[0008] The supervised part of the training signal annealing method is:
[0009] First, x1 and y represent the labeled input variable and output variable respectively, then the supervised cross entropy loss is established using the labeled data part as follows:
[0010]
[0011] Where G(·) represents the neural network fault classification model, is the cross entropy function, n l Represents the number of labeled samples; based on the training signal annealing method, when the model is trained on more and more unlabeled samples, the training signal from the labeled samples will be gradually released without overfitting, which is defined as follows:
[0012]
[0013] Among them, when the conditions in brackets are met, τ t Represents the threshold relative to the training step t. In order to control the ratio of unlabeled data to labeled data as needed, τ t Using exponential form, with K as the number of categories, then:
[0014]
[0015] The unsupervised loss part is:
[0016] x u The input variable represents the unlabeled sample, and the unsupervised loss is defined using the Kullback-Leibler (KL) divergence as follows:
[0017]
[0018] Among them, n u represents the number of unlabeled samples, ψ represents the model parameters, r represents the perturbation, and the adversarial perturbation is defined as follows:
[0019] The countermeasure disturbance calculation method based on the power iteration method and the finite difference method is:
[0020]
[0021] Among them, H(x u ,ψ) is defined as the Hessian matrix, given by the formula: According to this approximation, r is H(x u ,ψ) the first principal eigenvector u at amplitude ξ;
[0022]
[0023] in represents the unit vector with the same direction as the maximum eigenvector; assuming that d0 is the initial unit vector randomly sampled, as long as d0 is consistent with the main eigenvector v max is not vertical, the following iterative calculation will make d k Converges to v max ;
[0024]
[0025] Hd i The finite difference method is used to approximate the following:
[0026]
[0027] Update d by repeatedly applying the formula:
[0028]
[0029] After k iterations, we finally get an approximate estimate of r:
[0030]
[0031] The overall loss of the RSCF model is:
[0032] The overall robust semi-supervised classification loss of the RSCF model combines the labeled cross entropy loss and the unlabeled loss through a weight factor μ (0 < μ ≤ 1).
[0033]
[0034] The present invention has the beneficial effects:
[0035] 1. By combining supervised loss and unsupervised loss, a fault diagnosis method for blast furnace ironmaking process based on robust semi-supervised classification framework (RSCF) is proposed, aiming to solve the obvious imbalance problem between fault labeled dataset and unlabeled dataset in blast furnace ironmaking process.
[0036] 2. A supervised loss based on the training signal annealing method (TSA) is proposed to effectively alleviate the huge gap between labeled data and unlabeled data and prevent the model from overfitting to limited labeled data in the early training stage.
[0037] 3. Unsupervised loss is proposed by utilizing unlabeled data to enhance the smoothness of the model. The specific method is to make the output of the model consistent before and after adding adversarial interference, which effectively enhances the generalization ability and robustness of the diagnostic model.
[0038] 4. In order to calculate the adversarial perturbation, an efficient calculation method based on the power iteration method and the finite difference method is designed. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 Shown is an overall structural diagram of a blast furnace ironmaking process fault diagnosis method based on a robust semi-supervised classification framework (RSCF) of the present invention.
[0040] Figure 2The fault diagnosis results of RSCF on the actual data of blast furnace ironmaking process. DETAILED DESCRIPTION
[0041] In order to more clearly describe the technical content of the present invention, the specific implementation method is further explained below in conjunction with the accompanying drawings.
[0042] like Figure 1 As shown, it is an overall structure diagram of the blast furnace ironmaking process fault diagnosis method based on the robust semi-supervised classification framework (RSCF) of the present invention, and its steps are as follows:
[0043] A fault diagnosis method for blast furnace ironmaking process based on a robust semi-supervised classification framework (RSCF), characterized in that the present invention combines supervised and unsupervised parts, and can effectively utilize limited blast furnace labeled samples and a large number of unlabeled samples. In the supervised part, in order to solve the huge gap between labeled data and unlabeled data, a new training technique called training signal annealing method (TSA) is adopted to prevent overfitting. In addition, the unsupervised loss part enhances the smoothness of the model by utilizing unlabeled data, which is achieved by causing the output results of the model to be consistent before and after the addition of adversarial interference. Then, an efficient adversarial disturbance calculation method is designed based on the power iteration method and the finite difference method. Finally, a more robust fault diagnosis model is achieved through the overall training of supervised loss and unsupervised loss.
[0044] The supervised part of the training signal annealing method is:
[0045] First, x1 and y represent the labeled input variable and output variable respectively, then the supervised cross entropy loss is established using the labeled data part as follows:
[0046]
[0047] Where G(·) represents the neural network fault classification model, is the cross entropy function, n l Represents the number of labeled samples. However, in the blast furnace ironmaking process, since it is much easier to obtain unlabeled data than labeled data, there is a large gap between the amount of unlabeled data and the amount of labeled data. In order for our model to make use of unlabeled data as much as possible, a sufficiently large model is usually required, but a large model can easily overfit to supervised data of limited size. To solve this problem, we adopted a new training technique called training signal annealing method (TSA). Specifically, as the model is trained on more and more unlabeled samples, the training signal from the labeled samples is gradually released without overfitting. It is defined as follows:
[0048]
[0049] Among them, when the conditions in brackets are met, τ t Represents the threshold relative to the training step t. In order to control the ratio of unlabeled data to labeled data according to our needs, τ t Using exponential form, with K as the number of categories, then:
[0050]
[0051] The unsupervised loss part is:
[0052] Unlabeled data is used to enhance the smoothness of the model. Specifically, this method makes the prediction results of unlabeled data consistent before and after perturbation, so that the classification boundary is located in the low-density area. Therefore, x u The input variable represents the unlabeled sample, and the unsupervised loss is defined using the Kullback-Leibler (KL) divergence as follows:
[0053]
[0054] Among them, n u Represents the number of unlabeled samples, ψ represents the model parameters, and r represents the perturbation. Generally speaking, the perturbation vector r is set to obey a specific distribution, such as Gaussian distribution, uniform distribution, etc. However, the criteria for selecting a suitable distribution are not uniform, and the effect of different distributions on improving the model is unknown. Referring to the adversarial training method, consider applying perturbations in the direction where the model changes the steepest. The adversarial perturbation is defined as follows:
[0055]
[0056] The countermeasure disturbance calculation method based on the power iteration method and the finite difference method is:
[0057] The perturbation vector r has no obvious closed form. Therefore, the present invention proposes an efficient calculation method. When the perturbation vector r is 0, KL(r,x u ,ψ) reaches its minimum value. According to the properties of calculus, at this time, the first-order derivative KL(r,x u ,ψ)=0. Therefore, the present invention uses KL(r,x u ,ψ) is approximately calculated by the second-order Taylor expansion.
[0058]
[0059] Among them, H(x u ,ψ) is defined as the Hessian matrix, which can be given by the formula: According to this approximation, r is H(x u,ψ) is the first principal eigenvector u at amplitude ξ.
[0060]
[0061] in represents the unit vector with the same direction as the largest eigenvector. In practice, it is very challenging to directly calculate the eigenvalue, especially for high-order matrices. The power iteration method is an effective method to estimate the principal eigenvalue. Assume that d0 is a randomly sampled initial unit vector. As long as d0 is consistent with the principal eigenvector v max is not vertical, the following iterative calculation will make d k Converges to v max .
[0062]
[0063] Then, for the convenience and efficiency of calculation, Hd i The finite difference method can be used to approximate the following:
[0064]
[0065] In summary, d can be updated by repeatedly applying the formula:
[0066]
[0067] After k iterations, we can finally get an approximate estimate of r:
[0068]
[0069] The overall loss of the RSCF model is:
[0070] The overall robust semi-supervised classification loss of the RSCF model combines the labeled cross entropy loss in claim 2 and the unlabeled loss in claim 3 through a weight factor μ (0<μ≤1).
[0071]
[0072] The above is the goal of the robust semi-supervised classification framework based on adversarial perturbations. The framework is good at handling scenarios with a limited number of labeled fault samples and provides a robust solution for fault diagnosis in blast furnace ironmaking process.
[0073] The present invention was verified by substituting the actual data of blast furnace ironmaking process into the actual data of a large ironmaking plant in South China with a volume of 2650m 3The data sample of No. 2 blast furnace in 2023 contains 14 parameters and the sampling rate is 10s. The training set includes labeled and unlabeled blast furnace ironmaking process fault data. Specifically, the labeled part includes 2000 normal operation samples, 816 hanging material faults, 1636 pipeline faults and 1100 tuyere small set faults, as well as 5000 unlabeled samples. The test set is designed to evaluate the performance of the model, which includes 600 normal operation samples, 255 hanging material fault samples, 531 pipeline fault samples, and 250 tuyere small set fault samples. Figure 2 The figure shows the confusion matrix result of the method of the present invention on the actual blast furnace ironmaking process data. From the fault diagnosis effect, it can be seen that the model effect is very good.
[0074] The technical features of the above-described embodiments can be further combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0075] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, all of which belong to the protection scope of the present invention. The protection scope of the present invention is given by the attached claims and any equivalent technical solutions thereof.
Claims
1. A method for fault diagnosis of blast furnace ironmaking process based on robust semi-supervised classification framework (RSCF), characterized in that: The supervised and unsupervised parts are combined to effectively utilize the limited blast furnace labeled samples and a large number of unlabeled samples; in the supervised part, a training technique based on the training signal annealing method (TSA) is used to prevent overfitting in view of the gap between labeled data and unlabeled data; in addition, the unsupervised loss part enhances the smoothness of the model by utilizing unlabeled data, which is achieved by forcing the output results of the model to be consistent before and after the addition of adversarial interference; then, an adversarial perturbation calculation method is designed based on the power iteration method and the finite difference method; finally, a robust fault diagnosis model is realized through the overall training of supervised loss and unsupervised loss.
2. The method according to claim 1, characterized in that The supervised part of the training signal annealing method is: First, x1 and y represent the labeled input variable and output variable respectively, then the supervised cross entropy loss is established using the labeled data part as follows: Where G(·) represents the neural network fault classification model, is the cross entropy function, n l Represents the number of labeled samples; based on the training signal annealing method, when the model is trained on more and more unlabeled samples, the training signal from the labeled samples will be gradually released without overfitting, which is defined as follows: Among them, when the conditions in brackets are met, τ t Represents the threshold relative to the training step t. In order to control the ratio of unlabeled data to labeled data as needed, τ t Using exponential form, with K as the number of categories, then:
3. The method according to claim 2, characterized in that The unsupervised loss part is: x u The input variable represents the unlabeled sample, and the unsupervised loss is defined using the Kullback-Leibler (KL) divergence as follows: Among them, n u represents the number of unlabeled samples, ψ represents the model parameters, r represents the perturbation, and the adversarial perturbation is defined as follows:
4. The method according to claim 3, characterized in that The countermeasure disturbance calculation method based on the power iteration method and the finite difference method is: Among them, H(x u ,ψ) is defined as the Hessian matrix, given by the formula: According to this approximation, r is H(x u ,ψ) the first principal eigenvector u at amplitude ξ; in represents the unit vector with the same direction as the maximum eigenvector; assuming that d0 is the initial unit vector randomly sampled, as long as d0 is consistent with the main eigenvector v max is not vertical, the following iterative calculation will make d k Converges to v max ; Hd i The finite difference method is used to approximate the following: Update d by repeatedly applying the formula: After k iterations, we finally get an approximate estimate of r:
5. The method according to claim 4, characterized in that The overall loss of the RSCF model is: The overall robust semi-supervised classification loss of the RSCF model combines the labeled cross entropy loss and the unlabeled loss through a weight factor μ (0 < μ ≤ 1).
Citation Information
Patent Citations
Water chilling unit unknown fault diagnosis method based on semi-supervised adversarial variational automatic coding
CN116089894A
Data processing method and device, electronic equipment and storage medium
CN117035048A
