Blast furnace abnormal working condition early warning method based on class condition manifold alignment

Abstract

The invention provides a blast furnace abnormal working condition early warning method based on class condition manifold alignment. The blast furnace abnormal working condition early warning method comprises the following steps: (1) representing a learning model based on a time sequence neighborhood comparison submerged space; (2) establishing a class condition manifold alignment model based on domain adaptation; and (3) a prototype dynamic updating mechanism based on confidence degree adjustment. Therefore, accurate early warning of the real abnormal working condition of the blast furnace is realized, the flexibility and the accuracy of early warning of the abnormal working condition are improved, meanwhile, combined characterization of stability and dynamic characteristics of the blast furnace in the operation process is realized, and the problem of knowledge migration caused by data distribution drift and working condition characteristic difference between different blast furnaces is solved.

Description

Technical Field

[0001] This invention relates to the technical field of abnormal operating condition early warning methods, specifically a blast furnace abnormal operating condition early warning method based on class conditional manifold alignment. Background Technology

[0002] As a core link in the steel industry, the stability of blast furnace ironmaking directly determines the production efficiency and molten iron quality of the blast furnace. Accurate early warning of abnormal blast furnace operating conditions can provide timely and accurate operating information to on-site operators, thereby effectively improving production efficiency and reducing production risks, and has important engineering application value.

[0003] Currently, early warning systems for abnormal blast furnace operating conditions still face multiple challenges. First, frequent fluctuations in operating conditions, inconsistent labeling rules, and delayed information feedback result in a scarcity of abnormal operating condition sample data, and the quality of the labeled data is low, making it difficult to utilize. Second, the integration and collaborative utilization of multi-blast furnace datasets faces technical bottlenecks, preventing the full exploitation of data value. Furthermore, existing early warning methods have insufficient warning duration to meet field requirements, involve significant manual intervention, and are highly subjective, thus limiting the objectivity and stability of the warnings. Therefore, effectively extracting high-quality data samples and fully utilizing multi-blast furnace datasets to accurately predict abnormal blast furnace operating conditions is a crucial prerequisite for subsequent process modeling, optimized control, and performance evaluation.

[0004] Existing early warning methods for abnormal blast furnace operating conditions are mainly divided into three categories: those based on expert experience, those based on mechanistic knowledge, and those based on data. Methods based on expert experience rely primarily on expert knowledge to predict operating conditions and adjust parameters, while methods based on mechanistic knowledge construct interpretable early warning models based on physicochemical principles. However, the knowledge systems of these two types of methods are relatively fixed and slow to update, making it difficult to effectively uncover the complex implicit correlations in the blast furnace operation process. This results in their early warning accuracy and foresight failing to meet the high-precision monitoring requirements of the big data era. In the context of big data, data-driven methods have emerged, demonstrating significant advantages in processing high-dimensional, nonlinear, and time-varying data. For example, machine learning methods can achieve fault early warning through feature extraction and algorithm fusion, while deep learning methods can effectively extract spatiotemporal features from process data using convolutional operations, graph neural networks, and other algorithms and structures, thereby enabling early anomaly detection.

[0005] A comprehensive analysis of the advantages and limitations of the various operational condition early warning methods reveals that deep learning-based methods have promising application prospects. However, when applied to early warning of abnormal blast furnace conditions, existing methods still face several challenges. First, current methods do not adequately model the short-term stability and physical continuity of blast furnace operation, thus affecting the accuracy of data representation and condition warnings. Second, when comprehensively utilizing data from multiple blast furnaces, existing cross-equipment data transfer methods typically rely on global feature alignment, failing to fully consider intra-class local manifold structures and the scarcity of target domain label information, thus limiting the effectiveness of cross-blast furnace knowledge transfer.

[0006] For example, patent application CN120744818A, entitled "A Method for Identifying Operating Conditions of Gas Turbines Based on Multi-Source Information Fusion," describes a method that uses speed signals as an intermediate benchmark to synchronously correlate vibration and thermal data from different control systems to construct a fused dataset. This dataset is then used as input from multiple sources and a back-propagation (BP) neural network to identify operating conditions, supporting overall assessment and fault warning. However, this method does not evaluate the reliability of abnormal operating condition warning results, leaving field workers without clear risk guidance for parameter adjustments and operational decisions.

[0007] For example, patent application CN120832577A, entitled "A Method for Identifying Operating Conditions and Providing Risk Warnings for High-Power Gearboxes," uses multiple sensors to collect operating data of the gearbox under various operating conditions. It constructs a multi-branch feature extraction model and an operating condition identification model to achieve multi-label operating condition identification and confidence assessment. Combined with a multi-level dynamic warning module, it ensures the safe and stable operation of the gearbox. However, the risk warning duration used in this method is short and fixed, making it difficult to adapt to complex scenarios with frequently changing operating conditions, thus affecting the accuracy and effectiveness of the warnings.

[0008] To address the aforementioned difficulties in existing methods, this invention proposes a blast furnace abnormal operation condition early warning method based on conditional manifold alignment, which can accurately predict abnormal blast furnace operating conditions, has high flexibility and accuracy, and provides early warning of operating conditions in process industry systems. This is the problem that this invention urgently needs to solve. Summary of the Invention

[0009] The purpose of this invention is to propose an early warning method for abnormal blast furnace operating conditions based on conditional manifold alignment, thereby enabling accurate early warning of abnormal blast furnace operating conditions.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for early warning of abnormal operating conditions in blast furnaces based on conditional manifold alignment, characterized by the following steps:

[0012] (1) Latent space representation learning model based on temporal neighborhood comparison. The latent space representation learning model based on temporal neighborhood comparison can make full use of the short-term stability and physical continuity of the blast furnace operation process, regard the temporal neighborhood samples as real positive samples to enhance the model's ability to capture the local temporal consistency of process data. At the same time, the future state prediction task and causal prediction coding loss function are designed in the latent space to enable the encoder to learn the predictable data feature evolution mode and realize the joint representation of the stability and dynamics of the blast furnace operation process.

[0013] (2) Domain-adaptive class-conditional manifold alignment model: By constructing a domain-adaptive class-conditional manifold alignment model, a loss function based on the maximum mean difference is designed to achieve accurate alignment of manifold structures within the class. While maintaining the separability between classes, it effectively reduces the distribution difference between the target domain and the source domain, and solves the problem of accurate knowledge transfer caused by significant data distribution drift and differences in operating conditions between different blast furnaces.

[0014] (3) A prototype dynamic update mechanism based on confidence adjustment is designed to adaptively adjust the prototype update rate in a confidence-weighted manner, and only perform prototype updates under high confidence conditions. At the same time, the warning window length is calculated based on the system's time constant, so that the warning is synchronized with the physical process of the blast furnace in the time dimension, thereby improving the flexibility and accuracy of the warning of abnormal blast furnace conditions.

[0015] As a further improvement, the present invention assumes It is an unlabeled dataset from the source domain, used for unsupervised representation learning. It is a labeled dataset from the source domain, containing rich label information. It is the target domain dataset, containing a small number of labeled samples. It is a prototype support set, usually from A labeled subset is extracted from the dataset and used to calculate the prototype. It is a labeled target domain time series dataset used for model testing.

[0016] As a further improvement, the latent space representation learning based on temporal neighborhood in step (1) specifically includes the following steps:

[0017] Step 1: Establish the mathematical expression of the physical continuity assumption, assuming that the blast furnace operation process satisfies local stationarity within a short time window:

[0018] ;

[0019] in, The "time constant" of the blast furnace operation system, i.e., the characteristic time scale of the system, represents the time required for significant changes to occur in the system. Represents time series observations, Is with Small correction terms of the same order, Indicates that in the known In the case of The conditional expectation states that, within a sufficiently short time window, a dynamically changing system can be approximated as a "stationary" or "predictable" system. This means that the state / time series data at adjacent times statistically belong to the same distribution family. Indicates time interval, Represents any time , This indicates that the time interval is less than the time required for a significant change to occur in the system. This means that when the time interval is less than the time required for a significant change in the system, for any given moment... All satisfied ;

[0020] For time Observations Its temporal neighborhood is defined as:

[0021] ;

[0022] in, Represents the neighborhood radius (in time steps). Represents the values ​​in the temporal neighborhood. and Indicates a time index. Represents the observed value The neighborhood radius is The set of observations.

[0023] To construct a latent space model that can characterize the evolution pattern of blast furnace time-series data, this invention designs a latent space representation learning model based on the fundamental principles of contrastive learning. Traditional methods construct positive samples through data augmentation. However, the time-series data of blast furnace operation is subject to causal constraints, which means that random disturbances may disrupt physical consistency. Simultaneously, by utilizing time-series neighborhoods, physically real positive and negative sample pairs are constructed, and a time-series neighborhood comparison design is implemented. The loss function is as follows:

[0024] ;

[0025] in, This represents the set of all possible time steps from the entire time series of the training data. That is , Represents L2 normalized cosine similarity. Indicates the anchor point. These represent the original data points. via encoder The resulting low-dimensional representation vector Indicates the batch size during training. This indicates that negative samples come from all other samples within the same batch, and a batch contains a total of One negative sample, one positive sample. Represents the set of time steps All moments Then for all those belonging to of Calculate mathematical expectation. These represent the original data points. via encoder The resulting low-dimensional representation vector;

[0026] Step 2: The evolution of the blast furnace operation process follows deterministic dynamics and low-dimensional stochastic dynamics, using the system's intrinsic dynamics as a guiding principle for learning a good representation:

[0027] ;

[0028] in, This represents the deterministic drift term, which is the inherent and decisive dynamic law of the system. This represents the random diffusion term, which is the random disturbance or noise experienced by the system. Indicates the intensity of noise. This represents white noise with a mean of zero.

[0029] That is, during the operation of the blast furnace, Representing the sum of the kinetic laws governing all physical and chemical processes within a blast furnace, it can be understood as the mathematical essence of the blast furnace's "physical laws" and "chemical reaction rules." It governs the entire transformation process of the furnace charge from its entry into the blast furnace to the final formation of molten iron and slag, including the stepwise reduction of iron oxides, carbon dissolution and loss reactions, and slagging reactions. It is the random diffusion term, representing the random disturbances or noise experienced by the system. The intensity of noise can depend on the system's state and time. White noise, typically represented by a zero mean, is uncorrelated at different time points. Low-dimensional stochastic dynamics implies that, despite the presence of noise, the core behavior of the system is constituted by a relatively low-dimensional, structured deterministic component. The dominant effect is minimal, with noise having a negligible impact.

[0030] Step 3: The trajectory of the blast furnace during operation in the latent space should be predictable, just like its trajectory in the original physical space. Dataset, defining the autoregressive prediction task:

[0031] ;

[0032] in, It is an autoregressive prediction model, specifically an LSTM model, with the following parameters: , To predict the step size, The predicted value of the feature. for Feature values ​​at and before time point;

[0033] However, learning dynamics solely through autoregressive prediction tasks faces a fundamental challenge. If the model directly minimizes the gap between the predicted value and the future true representation, it easily finds a shortcut that leads to representation collapse—that is, the encoder maps all inputs to the same output point, making the prediction task trivial and the representation ineffective. To avoid this risk, it is necessary to block the collapsing gradient transmitted to the encoder while introducing the prediction objective. This invention designs a causal prediction encoding loss. Its definition is:

[0034] ;

[0035] in, This indicates that the gradient has stopped, meaning that to prevent representation collapse, during forward propagation... The output value is It is exactly the same as without it; during backpropagation, when calculating the gradient, It is treated as a constant, its gradient is forced to zero, and it is not propagated back to its internal variables. Therefore, it will not be transmitted back to the source. encoder (i.e.) );

[0036] Gradient flow direction predictor Predictor It still receives the full gradient from the loss, and it is trained better and better to make its output... Approaching a "fixed" target At the same time, gradients do not flow to the encoder; the encoder will not adjust its future output to accommodate gradients. It is easier to predict and optimize, and it no longer has an incentive to collapse into a constant, because "keeping the target fixed" no longer provides gradient rewards for the encoder;

[0037] Step 4: The encoder generates an information-rich, non-collapsed representation space, while the predictor learns the dynamics of the system within this established space. The two loss functions work together to create a high-quality latent space that can distinguish different states and reflect the laws of state evolution.

[0038] The joint loss function is defined as follows:

[0039] ;

[0040] in, , is the regularization term. , and These are the weights of the three losses.

[0041] As a further improvement, the method for preparing domain migration based on class-conditional manifold alignment in (2) includes the following steps:

[0042] Step 1: We now possess a large amount of labeled data from the source domain (i.e., ) and a small amount of labeled data in the target domain (i.e. The core objective of domain adaptation is to enable the model to perform well in another target domain. To reduce the target domain error, the optimization direction is clearly to minimize the distribution difference between the target domain and the source domain while ensuring model performance.

[0043] Currently, the mainstream paradigm of domain alignment aims to achieve knowledge transfer by minimizing the global distribution difference between the target domain and the source domain. However, its limitations in complex scenarios are becoming increasingly apparent. It is difficult to avoid category confusion caused by local structural mismatch. For example, when the working condition type 1 in the source domain and the working condition type 2 in the target domain have similar geometric structures in the feature space, global alignment will indiscriminately bring all sample representations closer together, resulting in blurred decision boundaries and decreased semantic separability. This phenomenon is particularly prominent in practical scenarios such as blast furnace working condition transfer. Different categories of abnormal working conditions may have similar local manifold structures in the feature space. If only global alignment is performed, it is very easy to cause negative transfer.

[0044] To overcome this limitation, it is necessary to introduce class-supervised information to construct a more discriminative alignment method while maintaining the inherent manifold geometry of the data. Therefore, this invention proposes a class-conditional manifold alignment model. Its core idea is to align the conditional distributions of the same class in the target domain and the source domain in the embedding space, thereby narrowing the gap between domains while maintaining inter-class separability.

[0045] This invention designs a class-conditional alignment loss based on the maximum mean difference, for each class ( (where the total number of categories is 1), calculate the distribution distance between samples of that class in the target domain and the source domain respectively:

[0046] ;

[0047] in, This indicates that the actual label in the source domain is The sample distribution Indicates that the predicted category in the target domain is Regarding the sample distribution, in actual training, this invention uses the model's current prediction results to assign labels to the target domain samples. These labels change dynamically during the training process. This represents the maximum mean difference, used to calculate the distribution distance between various samples in the target domain and the source domain;

[0048] Step 2: The class-conditional alignment loss is combined by weighted summation of the contributions from each class.

[0049] ;

[0050] Among them, weight Based on the number of samples in each category Settings are configured to mitigate the impact of category imbalance on alignment.

[0051] As a further improvement, the method for preparing abnormal operating condition early warning based on prototype dynamic updates in step (3) includes the following steps:

[0052] The design incorporates a prototype dynamic update mechanism based on confidence level adjustment. This mechanism adaptively adjusts the prototype update rate using confidence level weighting, ensuring that updates are only performed when the confidence level is high. Simultaneously, it calculates the warning window length using the system time constant, synchronizing the warning with the physical processes of the blast furnace in the time dimension. This improves the accuracy and flexibility of abnormal operating condition warnings. The specific steps include:

[0053] Step 1: Based on the prototype support set For each work condition category, a probabilistic prototype is established, treating the prototype as a sufficient statistic for the category. The prototype of a class is typically defined as the mean vector of all samples in the feature space of that class, under the Gaussian assumption.

[0054] ;

[0055] ;

[0056] ;

[0057] in, Indicates the first Class prototype sample, for The Middle Latent space characteristics of a sample Class conditional probability, representing the probability given a known class. Under these conditions, the observed features The probability, The mean represents the category. The "central position" of the sample set in the feature space. This represents the covariance matrix, which describes the categories. The distribution of the sample set, i.e., the "size" and "shape" of the category, where "·" represents "T", indicating the transpose of the matrix;

[0058] Step 2: Classification decision based on Bayesian criteria:

[0059] ;

[0060] in, Let be the posterior probability, representing the probability given a feature vector. Under the condition that the sample belongs to category The goal of classification is to find the category that maximizes the probability of a given probability. Let be the prior probability, representing the category. The probability of occurrence in the entire sample set, compared with specific observations The fact that it is irrelevant reflects the prevalence of this category. Marginal probability represents the observed feature. The overall probability is independent of the category;

[0061] In uniform prior This is equivalent to maximum likelihood classification. In real-world scenarios, new data samples may contain noise or distribution drift. Directly updating the prototype will reduce robustness. This invention should place more trust in the model's predictions of samples with high confidence and reduce the impact of samples with low confidence. At the same time, the update magnitude should decay over time to ensure eventual convergence.

[0062] In summary, this Bayesian classification formula is based on the Bayesian classification framework, models the categories in the feature space as Gaussian distributions, derives the posterior probability through class conditional probability and class prior, and selects the category that maximizes the posterior probability as the classification result.

[0063] Step 3: To address noise interference in industrial data, a confidence-weighted exponential moving average strategy was designed for prototype updates.

[0064] ;

[0065] in, and They are respectively Time and The mean of the class prototype at time step. For the first One in Latent space characteristics at time step, adaptive learning rate Defined as:

[0066] ;

[0067] It's the base learning rate, which controls the overall update speed; it's a hyperparameter. The confidence level of the sample is determined based on the current sample. Adjust its influence based on the predictive certainty. The time decay factor is a constant less than 1, which increases with time steps. Increase, It will get smaller and smaller;

[0068] Define sample confidence level:

[0069] ;

[0070] in, Controlling the intensity of entropy penalty For the predicted sample type, The maximum probability predicted by the model directly reflects the confidence in the most likely class, where entropy... The calculation formula is:

[0071]

[0072] in, It is a penalty term based on normalized entropy, entropy Maximum value Normalizing to the [0, 1] interval, when entropy is high (uncertainty is high), this penalty term approaches 0, thus reducing the overall confidence level. , Indicates "known" In the case of prediction results, the type is The probability of;

[0073] Step 4: In the abnormal operating condition identification stage, it is constructed as a hypothesis testing problem:

[0074] (1) Event The sample belongs to the category ;

[0075] (2) Event The sample does not belong to any known category;

[0076] Using sample features Compared with the predicted category prototype mean Similarity between As a test statistic, it is compared with the detection threshold ( ) compare, if The high score indicates that the characteristics of the sample are very close to the prototype of its class, so there is no reason for this invention to reject it. ,if The low value indicates that the sample is far from the center of all known classes, and this invention tends to reject it. ,accept The detection threshold for the above hypothesis test is based on the historical distribution of similarity. Dynamic setting, i.e. , Indicates the first The monitoring threshold for sample types is adjusted by... It can flexibly control the sensitivity to abnormal operating conditions and achieve robust anomaly detection;

[0077] Step 5: Based on the above mechanism for anomaly detection of the current state, in order to achieve early warning of abnormal operating conditions, a Transformer-based time series prediction model is adopted to perform multi-step prediction of the future time series data of the blast furnace operation process to obtain the future state trajectory.

[0078] Traditional early warning mechanisms with fixed time windows are ill-suited to the dynamic characteristics of blast furnaces, often leading to delayed responses or false alarms. To address this, this invention proposes a strategy for adjusting the early warning window length based on the time scale of the blast furnace system, grounded in control theory. As a complex nonlinear system, the dynamic behavior of a blast furnace is influenced by multiple coupled mechanisms, including heating processes (furnace temperature, heat flow) and material-gas processes (pressure, charge drop), exhibiting significant multi-time-scale characteristics. However, under small perturbations near steady state, the main inertial component of the system can be effectively approximated by a second-order underdamped system, thus simplifying the model while preserving key dynamic characteristics and providing a scientific basis for setting the early warning time scale.

[0079] For a second-order underdamped system, its step response is described by the following equation:

[0080] ;

[0081] in, This is the steady-state value. The system's natural frequency. Damping ratio ( ), Given the damped oscillation frequency, the dominant time constant of the system can be defined as:

[0082] ;

[0083] The physical meaning of this time constant is the time required for the system envelope to decay to a 36.8% deviation from its steady-state value, characterizing the overall response speed of the system. Approaching 1 (critical damping), the system degenerates into a fast convergent process without oscillations. When the temperature is small, the system exhibits oscillating characteristics, which is consistent with the temperature and pressure fluctuations observed in actual blast furnace operation. The observation of actual blast furnace fluctuations is consistent with the characteristics of small... The second-order system oscillation characteristics are consistent, indicating that during the dynamic process of the blast furnace, there is significant energy exchange between energy storage elements (heat capacity, gas capacity), while dissipation (damping) is relatively small.

[0084] For the blast furnace operation process, the optimal natural frequency is to obtain ( ) and damping ratio ( The method is based on recursive subspace system identification using operational data, combined with adaptive filtering techniques. This method does not require active interference with production. It collects multivariate input-output data (such as air volume, pulverized coal injection rate, furnace temperature, and pressure) in real time during normal blast furnace operation. It then uses a recursive subspace identification algorithm to extract the system's state-space model online and parses the dominant feature values ​​from it, thereby directly calculating the system's state space model. and Then, adaptive Kalman filtering is used to smooth the parameters and evaluate their confidence level, ensuring that the estimation results are robust and reliable, and can dynamically track the time-varying characteristics of the system, providing accurate and adaptive parameter basis for real-time adjustment of the early warning window. As a system time constant, it can serve as a unified indicator for quantifying the speed of system response.

[0085] Industrial process data is collected periodically by sensors; it is discrete, not continuous. Let the time interval between each sampling be... Then the number of warning steps Defined as:

[0086] ;

[0087] in, The rounding up sign means that the value within the parentheses is rounded up to the smallest integer not less than it. This mapping ensures that the warning window covers the main part of the system's dynamic process, synchronizing the warning with the physical process in the time dimension. Because This represents the system's response speed and the length of the warning window. and" "quite;

[0088] Step 6: To achieve the transformation from feature prediction to operational condition early warning, multi-step probabilistic prediction is used to quantify the uncertainty of future states. The future states are modeled as follows:

[0089]

[0090] in, For warning steps, Indicates known The prediction result is The probability of this, under the Gaussian assumption:

[0091]

[0092] in, This is the deterministic part, provided by the trained prediction model, representing the "mean" of the predictions. It is the random component, representing the first... The uncertainty or error in the step prediction is modeled as a matrix with a mean of 0 and a covariance matrix of... Gaussian noise, It also follows a Gaussian distribution with a mean of . Covariance is ,Right now .

[0093] As a further improvement, at the decision-making level, a loss function matrix is ​​defined. To characterize the cost of different misjudgments, let the Bayesian probability be:

[0094] ;

[0095] This probability describes the decision to be made. The expected cost, defining the loss function matrix. To characterize the cost of different misjudgments.

[0096] As a further improvement, this probability describes the decision to be made. Expected cost, optimal decision This is obtained by minimizing this probability.

[0097] After the above steps are performed to train and optimize the model, the blast furnace abnormal operating condition early warning method based on conditional manifold alignment of the present invention can obtain a well-trained blast furnace abnormal operating condition early warning model if its performance reaches the preset threshold condition. This model can output specific operating condition types and achieve effective early warning.

[0098] Compared with the prior art, the beneficial effects of the present invention are:

[0099] (1) The latent space representation learning model based on temporal neighborhood comparison proposed in this invention enhances the ability of the representation learning model to capture the local temporal consistency of process data. It designs a latent space training mechanism based on future state prediction task and causal prediction coding loss function, which enables the encoder to learn a predictable data feature evolution pattern and realizes the joint representation of the stability and dynamic characteristics of the blast furnace operation process.

[0100] (2) The class conditional manifold alignment model proposed in this invention designs a loss function based on the maximum mean difference to accurately achieve the alignment of intra-class manifold structures. Under the premise of ensuring inter-class separability, it effectively reduces the distribution difference between the target domain and the source domain, thereby solving the knowledge transfer problem caused by data distribution drift and operating condition differences between different blast furnaces.

[0101] (3) The prototype dynamic update mechanism based on confidence adjustment designed in this invention proposes a confidence-weighted adaptive update strategy, studies the early warning window length calculation strategy based on system time constant, realizes the synchronization of abnormal working condition early warning with the physical process of blast furnace in the time dimension, and improves the flexibility and accuracy of abnormal working condition early warning.

[0102] (4) This invention performs latent space temporal neighborhood comparison modeling on the blast furnace operation process, performs class conditional manifold alignment on the target domain and source domain datasets, and accurately classifies the operating conditions based on dynamically updated prototypes, thereby achieving accurate early warning of real blast furnace abnormal operating conditions, providing a new technical path for early warning of blast furnace abnormal operating conditions. Attached Figure Description

[0103] Figure 1 This is a flowchart of the method according to an embodiment of the present invention;

[0104] Figure 2 This describes the dataset distribution of the source and target domains in an embodiment of the present invention.

[0105] Figure 3 This is the confusion matrix of the experimental results of the embodiments of the present invention. Detailed Implementation

[0106] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0107] A method for early warning of abnormal operating conditions in blast furnaces based on conditional manifold alignment, characterized by the following steps:

[0108] (1) Latent space representation learning model based on temporal neighborhood comparison;

[0109] (2) A domain-adaptive class-conditional manifold alignment model;

[0110] (3) Based on the confidence-adjusted prototype dynamic update mechanism, the latent space representation learning based on temporal neighborhood in step (1) specifically includes the following steps:

[0111] Step 1: Establish the mathematical expression of the physical continuity assumption, assuming that the blast furnace operation process satisfies local stationarity within a short time window:

[0112] ;

[0113] in, This indicates the time required for a significant change to occur in the system. Represents time series observations, Is with Small correction terms of the same order, Indicates that in the known In the case of Conditional expectation, Indicates time interval, Represents any time , This indicates that the time interval is less than the time required for a significant change to occur in the system. This means that when the time interval is less than the time required for a significant change in the system, for any given moment... All satisfied ;

[0114] For time Observations Its temporal neighborhood is defined as:

[0115] ;

[0116] in, Represents the neighborhood radius (in time steps). Represents the values ​​in the temporal neighborhood. and Indicates a time index. Represents the observed value The neighborhood radius is The set of observations.

[0117] To construct a latent space model that can characterize the evolution pattern of blast furnace time-series data, this invention designs a latent space representation learning model based on the fundamental principles of contrastive learning. Simultaneously, it utilizes temporal neighborhoods to construct physically real positive and negative sample pairs. The designed temporal neighborhood contrastive loss function is as follows:

[0118] ;

[0119] in, This represents the set of all possible time steps from the entire time series of the training data. That is , Represents L2 normalized cosine similarity. Indicates the anchor point. These represent the original data points. via encoder The resulting low-dimensional representation vector Indicates the batch size during training. This indicates that negative samples come from all other samples within the same batch, and a batch contains a total of One negative sample, one positive sample. Represents the set of time steps All moments Then for all those belonging to of Calculate mathematical expectation. These represent the original data points. via encoder The resulting low-dimensional representation vector;

[0120] Step 2: Use the system's intrinsic dynamics as a guiding principle for learning good representations:

[0121] ;

[0122] in, This represents the deterministic drift term, which is the inherent and decisive dynamic law of the system. This represents the random diffusion term, which is the random disturbance or noise experienced by the system. Indicates the intensity of noise. This represents white noise with a mean of zero.

[0123] Step 3: The trajectory of the blast furnace during operation in the latent space should be predictable, just like its trajectory in the original physical space. Dataset, defining the autoregressive prediction task:

[0124] ;

[0125] in, It is an autoregressive prediction model, specifically an LSTM model, with the following parameters: , To predict the step size, The predicted value of the feature. for Feature values ​​at and before time point;

[0126] Causal prediction coding loss Its definition is:

[0127] ;

[0128] in, This indicates that the gradient has stopped, meaning that to prevent representation collapse, during forward propagation... The output value is It is exactly the same as without it; during backpropagation, when calculating the gradient, It is treated as a constant, its gradient is forced to zero, and it is not propagated back to its internal variables. Therefore, it will not be transmitted back to the source. encoder (i.e.) );

[0129] Gradient flow direction predictor Predictor It still receives the full gradient from the loss, and it is trained better and better to make its output... Approaching a "fixed" target At the same time, gradients do not flow to the encoder; the encoder will not adjust its future output to accommodate gradients. It is easier to predict and optimize, and it no longer has an incentive to collapse into a constant, because "keeping the target fixed" no longer provides gradient rewards for the encoder;

[0130] Step 4: The encoder generates an information-rich, non-collapsed representation space, while the predictor learns the dynamics of the system within this established space. The two loss functions work together to create a high-quality latent space that can distinguish different states and reflect the laws of state evolution.

[0131] The joint loss function is defined as follows:

[0132] ;

[0133] in, , is the regularization term. , and The weights of the three losses are respectively. The method for preparing the domain migration based on class-conditional manifold alignment in (2) includes the following steps:

[0134] Step 1: We now possess a large amount of labeled data from the source domain (i.e., ) and a small amount of labeled data in the target domain (i.e. The core objective of domain adaptation is to enable the model to perform well in another target domain. To reduce the target domain error, the optimization direction is clearly to minimize the distribution difference between the target domain and the source domain while ensuring model performance.

[0135] The class-conditional alignment loss is designed based on the maximum mean difference, and is applied to each class. ( (where the total number of categories is 1), calculate the distribution distance between samples of that class in the target domain and the source domain respectively:

[0136] ;

[0137] in, This indicates that the actual label in the source domain is The sample distribution Indicates that the predicted category in the target domain is Regarding the sample distribution, in actual training, this invention uses the model's current prediction results to assign labels to the target domain samples. These labels change dynamically during the training process. This represents the maximum mean difference, used to calculate the distribution distance between various samples in the target domain and the source domain;

[0138] Step 2: The class-conditional alignment loss is combined by weighted summation of the contributions from each class.

[0139] ;

[0140] Among them, weight Based on the number of samples in each category To mitigate the impact of class imbalance on alignment, the method for preparing abnormal operating condition early warning based on prototype dynamic updates in step (3) includes the following steps:

[0141] The design incorporates a prototype dynamic update mechanism based on confidence level adjustment. This mechanism adaptively adjusts the prototype update rate using confidence level weighting, ensuring that updates are only performed when the confidence level is high. Simultaneously, it calculates the warning window length using the system time constant, synchronizing the warning with the physical processes of the blast furnace in the time dimension. This improves the accuracy and flexibility of abnormal operating condition warnings. The specific steps include:

[0142] Step 1: Based on the prototype support set For each work condition category, a probabilistic prototype is established, treating the prototype as a sufficient statistic for the category. The prototype of a class is typically defined as the mean vector of all samples in the feature space of that class, under the Gaussian assumption.

[0143] ;

[0144] ;

[0145] ;

[0146] in, Indicates the first Class prototype sample, for The Middle Latent space characteristics of a sample Class conditional probability, representing the probability given a known class. Under these conditions, the observed features The probability, The mean represents the category. The "central position" of the sample set in the feature space. This represents the covariance matrix, which describes the categories. The distribution of the sample set, i.e., the "size" and "shape" of the category, where "·" represents "T", indicating the transpose of the matrix;

[0147] Step 2: Classification decision based on Bayesian criteria:

[0148] ;

[0149] in, Let be the posterior probability, representing the probability given a feature vector. Under the condition that the sample belongs to category The goal of classification is to find the category that maximizes the probability of a given probability. Let be the prior probability, representing the category. The probability of occurrence in the entire sample set, compared with specific observations The fact that it is irrelevant reflects the prevalence of this category. Marginal probability represents the observed feature. The overall probability is independent of the category;

[0150] Step 3: To address noise interference in industrial data, a confidence-weighted exponential moving average strategy was designed for prototype updates.

[0151] ;

[0152] in, and They are respectively Time and The mean of the class prototype at time step. For the first One in Latent space characteristics at time step, adaptive learning rate Defined as:

[0153] ;

[0154] It's the base learning rate, which controls the overall update speed; it's a hyperparameter. The confidence level of the sample is determined based on the current sample. Adjust its influence based on the predictive certainty. The time decay factor is a constant less than 1, which increases with time steps. Increase, It will get smaller and smaller;

[0155] Define sample confidence level:

[0156] ;

[0157] in, Controlling the intensity of entropy penalty For the predicted sample type, The maximum probability predicted by the model directly reflects the confidence in the most likely class, where entropy... The calculation formula is:

[0158]

[0159] in, It is a penalty term based on normalized entropy, entropy Maximum value Normalizing to the [0, 1] interval, when the entropy is high, this penalty term approaches 0, thus reducing the overall confidence level. , Indicates "known" In the case of prediction results, the type is The probability of;

[0160] Step 4: In the abnormal operating condition identification stage, it is constructed as a hypothesis testing problem:

[0161] (1) Event The sample belongs to the category ;

[0162] (2) Event The sample does not belong to any known category;

[0163] Using sample features Compared with the predicted category prototype mean Similarity between As a test statistic, it is compared with the detection threshold ( ) compare, if The high score indicates that the characteristics of the sample are very close to the prototype of its class, so there is no reason for this invention to reject it. ,if The low value indicates that the sample is far from the center of all known classes, and this invention tends to reject it. ,accept The detection threshold for the above hypothesis test is based on the historical distribution of similarity. Dynamic setting, i.e. , Indicates the first The monitoring threshold for sample types is adjusted by... It can flexibly control the sensitivity to abnormal operating conditions and achieve robust anomaly detection;

[0164] Step 5: Based on the above mechanism for anomaly detection of the current state, in order to achieve early warning of abnormal operating conditions, a Transformer-based time series prediction model is adopted to perform multi-step prediction of the future time series data of the blast furnace operation process to obtain the future state trajectory.

[0165] Under small perturbations near steady state, the main inertial component of the system can be effectively approximated by a second-order underdamped system. This simplifies the model while preserving key dynamic characteristics, providing a scientific basis for setting the time scale for early warning.

[0166] For a second-order underdamped system, its step response is described by the following equation:

[0167] ;

[0168] in, This is the steady-state value. The system's natural frequency. Damping ratio ( ), Given the damped oscillation frequency, the dominant time constant of the system can be defined as:

[0169] ;

[0170] The physical meaning of this time constant is the time required for the system envelope to decay to a 36.8% deviation from its steady-state value, characterizing the overall response speed of the system. Approaching 1 (critical damping), the system degenerates into a fast convergent process without oscillations. When the value is small, the system exhibits oscillatory characteristics;

[0171] For the blast furnace operation process, the optimal natural frequency (i.e. ) and damping ratio (i.e. The method is based on recursive subspace system identification using operational data, combined with adaptive filtering techniques. This method requires no active interference with production. It collects multivariate input and output data from the blast furnace in real time during normal operation, uses the recursive subspace identification algorithm to extract the system's state-space model online, and parses the dominant feature values ​​from it, thereby directly calculating... and Then, adaptive Kalman filtering is used to smooth the parameters and evaluate their confidence level, ensuring that the estimation results are robust and reliable, and can dynamically track the time-varying characteristics of the system, providing accurate and adaptive parameter basis for real-time adjustment of the early warning window. As a system time constant, it can serve as a unified indicator for quantifying the speed of system response.

[0172] Industrial process data is collected periodically by sensors; it is discrete, not continuous. Let the time interval between each sampling be... Then the number of warning steps Defined as:

[0173] ;

[0174] in, The rounding up sign means rounding the value within the parentheses to the smallest integer not less than it. (Also refers to the warning window length.) and quite;

[0175] Step 6: To achieve the transformation from feature prediction to operational condition early warning, multi-step probabilistic prediction is used to quantify the uncertainty of future states. The future states are modeled as follows:

[0176]

[0177] in, For warning steps, Indicates known The prediction result is The probability of this, under the Gaussian assumption:

[0178]

[0179] in, This is the deterministic part, provided by the trained prediction model, representing the "mean" of the predictions. It is the random component, representing the first... The uncertainty or error in the step prediction is modeled as a matrix with a mean of 0 and a covariance matrix of... Gaussian noise, It also follows a Gaussian distribution with a mean of . Covariance is ,Right now At the decision-making level, the loss function matrix is ​​defined. To characterize the cost of different misjudgments, let the Bayesian probability be:

[0180] ;

[0181] This probability describes the decision to be made. The expected cost, and the loss function matrix are defined. To characterize the cost of different misjudgments, the optimal decision is... ,

[0182] This embodiment is based on the above method, taking a domestic steel plant's 2650 as an example. The blast furnace process is first analyzed. First, blast furnace process datasets from both the source and target domains are acquired and preprocessed, including cleaning, alignment, and standardization. Then, a temporal neighborhood dataset and positive / negative sample pairs are constructed in the source domain. A Time-Nearest Neighborhood (TNC) loss function, an autoregressive prediction task, and a CPC loss function are designed to fully exploit the temporal semantic features of the source domain. Next, the optimization direction for the target domain is determined, and a loss function based on class-conditional alignment is designed to achieve conditional alignment of cross-domain features. Subsequently, a prototype support set for the target domain is constructed, and probabilistic models are established for different operating conditions. A dynamic update strategy for the prototype is designed to support the model's temporal prediction and operating condition early warning tasks in the target domain. Following this, the overall model is jointly trained and its parameters are optimized. Finally, based on the performance validation results, it is determined whether the threshold conditions are met. If they are met, the optimal model parameters are output, and early warning results for abnormal operating conditions at multiple future time steps are provided.

[0183] In this embodiment, some results are shown in [the original text]. Figure 2 and Figure 3 As shown, by Figure 2 As can be seen, after dimensionality reduction through t-distributed random neighborhood embedding, the five types of operating data in the source domain (blast furnace 1) and the target domain (blast furnace 2) show obvious distribution differences, indicating that there is a significant feature shift between the two domains, which verifies the necessity of cross-domain modeling. Figure 3 The confusion matrix obtained from the experiment is presented. Numbers 0, 1, 2, 3, and 4 represent five operating conditions: "Normal," "Suspended Material," "Pipeline," "Slipping Material," and "Collapsing Material," respectively. The latter four are all abnormal operating conditions. Figure 3 As can be seen, the overall early warning accuracy of the model reaches 94.68%. Except for the fourth type of working condition, which has a slightly lower accuracy due to the influence of sample complexity, the recognition accuracy of the other types of working conditions all exceed 99%, which fully demonstrates the effectiveness of the proposed method in cross-domain working condition recognition and early warning tasks.

[0184] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A blast furnace abnormal condition early warning method based on quasi-conditional manifold alignment, characterized in that, Includes the following steps: (1) Latent space representation learning model based on temporal neighborhood comparison; (2) A domain-adaptive class-conditional manifold alignment model; (3) Prototype dynamic update mechanism based on confidence adjustment.

2. The blast furnace abnormal condition early warning method based on quasi-condition manifold alignment according to claim 1, characterized in that, The latent space representation learning based on temporal neighborhood in step (1) specifically includes the following steps: Step 1: Establish the mathematical expression of the physical continuity assumption, assuming that the blast furnace operation process satisfies local stationarity within a short time window: ； wherein, denotes the time required for a significant change in the system to occur, denotes the time series observation, is a small correction term of the same order as , denotes the conditional expectation of given the knowledge of , denotes the time interval, denotes that for any time , denotes that the time interval is less than the time required for a significant change in the system to occur, denotes that for any time , is satisfied when the time interval is less than the time required for a significant change in the system to occur; For an observation value at time instant , its temporal neighborhood is defined as: ； wherein, denotes the neighborhood radius (in time steps), denotes the values in the time neighborhood, and denotes the time index, denotes the observation of the observation set with neighborhood radius . To construct a latent space model that can characterize the evolution pattern of blast furnace time-series data, this invention designs a latent space representation learning model based on the fundamental principles of contrastive learning. Simultaneously, it utilizes temporal neighborhoods to construct physically real positive and negative sample pairs. The designed temporal neighborhood contrastive loss function is as follows: ； wherein, denotes the set of all possible time steps from the time series of the whole training data, i.e. , denotes the L2 normalized cosine similarity, denotes the anchor point, denotes the original data points through an encoder to get the low-dimensional representation vectors, denotes the batch size during training, denotes that the negative samples come from all other samples within the same batch, a batch contains negative samples and 1 positive sample, denotes the mathematical expectation of for all time points in the set , , denotes the original data points, through an encoder to get the low-dimensional representation vectors;​ Step 2: Utilize the system's intrinsic dynamics as a guiding principle for learning good representations: ； wherein, represents a deterministic drift term, which is the intrinsic, deterministic dynamics of the system, represents a stochastic diffusion term, which is the random disturbance or noise to which the system is subjected, represents the intensity of the noise, represents a white noise, whose mean is zero; Step 3: The trajectory of the blast furnace operation in the latent space should be as predictable as the trajectory in the original physical space, for which a dataset, define an autoregressive prediction task: ； wherein, is an autoregressive prediction model, which is an LSTM model, and the parameters of the LSTM model are , is a prediction step, is a predicted value of the feature, is the feature values at the time instant and before. Causal prediction coding loss defined as: ； wherein, represents a stop gradient, i.e. to prevent collapse, when forward-propagating, the output value of is exactly the same as without it, when back-propagating, when calculating the gradient, is treated as a constant, its gradient is forced to be zero, and it does not back-propagate to the variables inside it and, in turn, to the encoder that produced (i.e. ); Gradient flow predictor , the predictor Still receiving the full gradient from the loss, it gets trained better and better to make its output approach the "fixed” target At the same time, the gradient does not flow to the encoder, which does not optimize to make its future outputs easier to predict, it has no more incentive to collapse into a constant, because there is no gradient reward for the encoder to fix the target; Step 4: The encoder generates an information-rich, non-collapsed representation space, while the predictor learns the dynamics of the system within this established space. The two loss functions work together to create a high-quality latent space that can distinguish different states and reflect the laws of state evolution. The joint loss function is defined as follows: ； in, , is the regularization term. , and These are the weights of the three losses.

3. The method for early warning of abnormal operating conditions of a blast furnace based on conditional manifold alignment according to claim 2, characterized in that, The method for preparing domain migration based on class-conditional manifold alignment in (2) includes the following steps: Step 1: We now possess a large amount of labeled data from the source domain (i.e., ) and a small amount of labeled data in the target domain (i.e. The core objective of domain adaptation is to enable the model to perform well in another target domain. To reduce the target domain error, the optimization direction is clearly to minimize the distribution difference between the target domain and the source domain while ensuring model performance. The class-conditional alignment loss is designed based on the maximum mean difference, and is applied to each class. ( (where the total number of categories is 1), calculate the distribution distance between samples of that class in the target domain and the source domain respectively: ； in, This indicates that the actual label in the source domain is The sample distribution Indicates that the predicted category in the target domain is Regarding the sample distribution, in actual training, this invention uses the model's current prediction results to assign labels to the target domain samples. These labels change dynamically during the training process. This represents the maximum mean difference, used to calculate the distribution distance between various samples in the target domain and the source domain; Step 2: The class-conditional alignment loss is combined by weighted summation of the contributions from each class. ； Among them, weight Based on the number of samples in each category Settings are configured to mitigate the impact of category imbalance on alignment.

4. The method for early warning of abnormal operating conditions of blast furnaces based on conditional manifold alignment according to claim 3, characterized in that, The method for preparing abnormal operating condition early warning based on prototype dynamic update in step (3) includes the following steps: The design incorporates a prototype dynamic update mechanism based on confidence level adjustment. This mechanism adaptively adjusts the prototype update rate using confidence level weighting, ensuring that updates are only performed when the confidence level is high. Simultaneously, it calculates the warning window length using the system time constant, synchronizing the warning with the physical processes of the blast furnace in the time dimension. This improves the accuracy and flexibility of abnormal operating condition warnings. The specific steps include: Step 1: Based on the prototype support set For each work condition category, a probabilistic prototype is established, treating the prototype as a sufficient statistic for the category. The prototype of a class is typically defined as the mean vector of all samples in the feature space of that class, under the Gaussian assumption. ； ； ； in, Indicates the first Class prototype sample, for The Middle Latent space characteristics of a sample Class conditional probability, representing the probability given a known class. Under these conditions, the observed features The probability, The mean represents the category. The "central position" of the sample set in the feature space. This represents the covariance matrix, which describes the categories. The distribution of the sample set, i.e., the "size" and "shape" of the category, where "·" represents "T", indicating the transpose of the matrix; Step 2: Classification decision based on Bayesian criteria: ； in, Let be the posterior probability, representing the probability given a feature vector. Under the condition that the sample belongs to category The goal of classification is to find the category that maximizes the probability of a given probability. Let be the prior probability, representing the category. The probability of occurrence in the entire sample set, compared with specific observations The fact that it is irrelevant reflects the prevalence of this category. Marginal probability represents the observed feature. The overall probability is independent of the category; Step 3: To address noise interference in industrial data, a confidence-weighted exponential moving average strategy was designed for prototype updates. ； in, and They are respectively Time and The mean of the class prototype at time step. For the first One in Latent space characteristics at time step, adaptive learning rate Defined as: ； It's the base learning rate, which controls the overall update speed; it's a hyperparameter. The confidence level of the sample is determined based on the current sample. Adjust its influence based on the predictive certainty. The time decay factor is a constant less than 1, which increases with time steps. Increase, It will get smaller and smaller; Define sample confidence level: ； in, Controlling the intensity of entropy penalty For the predicted sample type, The maximum probability predicted by the model directly reflects the confidence in the most likely class, where entropy... The calculation formula is: ； in, It is a penalty term based on normalized entropy, entropy Maximum value Normalizing to the [0,1] interval, when the entropy is high, this penalty term approaches 0, thus reducing the overall confidence level. , Indicates known In the case of prediction results, the type is The probability of; Step 4: In the abnormal operating condition identification stage, it is constructed as a hypothesis testing problem: (1) Event The sample belongs to the category ; (2) Event The sample does not belong to any known category; Using sample features Compared with the predicted category prototype mean Similarity between As a test statistic, it is compared with the detection threshold ( ) compare, if The high score indicates that the characteristics of the sample are very close to the prototype of its class, so there is no reason for this invention to reject it. ,if The low value indicates that the sample is far from the center of all known classes, and this invention tends to reject it. ,accept The detection threshold for the above hypothesis test is based on the historical distribution of similarity. Dynamic setting, i.e. , Indicates the first The monitoring threshold for sample types is adjusted by... It can flexibly control the sensitivity to abnormal operating conditions and achieve robust anomaly detection. Step 5: Based on the above mechanism for anomaly detection of the current state, in order to achieve early warning of abnormal operating conditions, a Transformer-based time series prediction model is adopted to perform multi-step prediction of the future time series data of the blast furnace operation process to obtain the future state trajectory. Under small perturbations near steady state, the main inertial component of the system can be effectively approximated by a second-order underdamped system. This simplifies the model while preserving key dynamic characteristics, providing a scientific basis for setting the time scale for early warning. For a second-order underdamped system, its step response is described by the following equation: ； in, This is the steady-state value. The system's natural frequency. Damping ratio ( ), Given the damped oscillation frequency, the dominant time constant of the system can be defined as: ； The physical meaning of this time constant is the time required for the system envelope to decay to a 36.8% deviation from its steady-state value, characterizing the overall response speed of the system. Approaching 1 (critical damping), the system degenerates into a fast convergent process without oscillations. When the value is small, the system exhibits oscillatory characteristics; For the blast furnace operation process, the optimal natural frequency (i.e. ) and damping ratio (i.e. The method is based on recursive subspace system identification using operational data, combined with adaptive filtering techniques. This method requires no active interference with production. It collects multivariate input and output data from the blast furnace in real time during normal operation, uses the recursive subspace identification algorithm to extract the system's state-space model online, and parses the dominant feature values ​​from it, thereby directly calculating... and Then, adaptive Kalman filtering is used to smooth the parameters and evaluate their confidence level, ensuring that the estimation results are robust and reliable, and can dynamically track the time-varying characteristics of the system, providing accurate and adaptive parameter basis for real-time adjustment of the early warning window. As a system time constant, it can serve as a unified indicator for quantifying the speed of system response. Industrial process data is collected periodically by sensors; it is discrete, not continuous. Let the time interval between each sampling be... Then the number of warning steps Defined as: ； in, The rounding up sign means rounding the value within the parentheses to the smallest integer not less than it. (Also refers to the warning window length.) and quite; Step 6: To achieve the transformation from feature prediction to operational condition early warning, multi-step probabilistic prediction is used to quantify the uncertainty of future states. The future states are modeled as follows: ； in, For warning steps, Indicates known The prediction result is The probability, Under Gaussian assumptions: ； in, This is the deterministic part, provided by the trained prediction model, representing the "mean" of the predictions. It is the random component, representing the first... The uncertainty or error in the step prediction is modeled as a matrix with a mean of 0 and a covariance matrix of... Gaussian noise, It also follows a Gaussian distribution with a mean of . Covariance is ,Right now .

5. The method for early warning of abnormal operating conditions of a blast furnace based on conditional manifold alignment according to claim 4, characterized in that, At the decision-making level, the loss function matrix is ​​defined. To characterize the cost of different misjudgments, let the Bayesian probability be: ； This probability describes the decision to be made. The expected cost, defining the loss function matrix. To characterize the cost of different misjudgments.

6. The method for early warning of abnormal operating conditions of a blast furnace based on conditional manifold alignment according to claim 5, characterized in that, The optimal decision is .

Images