Intelligent identification method for tin smelting furnace condition of top-blown furnace

Abstract

The embodiment of the invention provides an intelligent identification method for the tin smelting furnace condition of a top-blown furnace, and belongs to the technical field of data processing, and the method specifically comprises the steps: 1, synchronously collecting multi-source heterogeneous data in the tin smelting process of the top-blown furnace, and carrying out the preprocessing of the multi-source heterogeneous data; step 2, based on the preprocessed data, extracting process features of different time scales by adopting a multi-stage sliding time window, and extracting time delay features corresponding to the process features by using a time sequence convolutional network; step 3, performing time alignment on the multi-source data after the time delay feature extraction, and inputting the multi-source data into a multi-mode Transform fusion module to obtain fusion features; and step 4, inputting the fusion features into a classifier to obtain an identification result of furnace condition categories, constructing a multi-stage early warning mechanism according to the identification result and time-varying feature trend analysis, and sending out an early warning signal before key parameter abnormity occurs. Through the scheme of the invention, the identification efficiency, accuracy and adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for intelligent identification of the furnace condition of a top-blown tin smelting furnace. Background Technology

[0002] Currently, tin metal, as an important strategic resource, is widely used in new energy, electronics, aerospace, and other fields. Especially with the rapid development of lithium batteries, photovoltaics, and semiconductor technology, the market demand for high-purity tin is constantly increasing. This places higher demands on tin smelting processes, particularly top-blown furnace tin refining technology. Top-blown furnace tin refining involves injecting oxygen-enriched air and fuel into the molten pool to complete the smelting process at high temperatures, involving a strong redox dynamic equilibrium. The success of this process directly affects the quality of the final product and production costs. In actual production, top-blown furnace tin refining presents complex multi-physics coupling problems. Complex interactions exist between the gas flow, melt, and solid particles in the molten pool, leading to high sensitivity of key parameters such as the oxygen-to-material ratio and lance height. Slight fluctuations can cause changes in slag tin content and a decrease in metal recovery. Simultaneously, multiple chemical reactions within the furnace (such as SnS oxidation, SnO2 reduction, and Fe3O4 formation) occur at different temperatures. The equilibrium constants of these reactions change significantly with temperature, making the reaction process complex and difficult to control. Since the reactions inside the molten pool cannot be directly observed, existing detection methods can only obtain partial data through offline analysis, which has significant time lag and limitations. Existing furnace condition identification methods rely on manual experience and periodic sampling analysis, failing to monitor furnace condition changes in real time, and the monitoring frequency is low. The lag in detection results and subjective judgment make it difficult to implement precise control and optimization in the production process. Abnormal furnace conditions (such as excessively high or low oxygen potential, excessively high or low furnace temperature) can seriously affect product quality and production efficiency, and may even cause equipment damage. Therefore, establishing a real-time and accurate abnormal furnace condition prediction system to identify potential risks in advance and take preventative measures is particularly important. Existing solutions, such as mechanistic modeling methods and traditional machine learning methods, have certain limitations. Mechanistic models struggle to accurately adapt to complex multi-physics coupling and time-delay effects, while traditional machine learning methods also face challenges in feature engineering, time-delay modeling, multimodal data fusion, and real-time performance. In summary, existing smelting process control methods suffer from problems such as time lag, subjectivity, and poor model adaptability.

[0003] It is evident that there is an urgent need for an intelligent identification method for the furnace condition of top-blown tin smelting furnaces, which can identify changes in furnace condition during the top-blown tin smelting process in real time and accurately, and can effectively provide early warnings. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide an intelligent identification method for the condition of a top-blown tin smelting furnace, which at least partially solves the problems of poor identification accuracy and adaptability in the prior art.

[0005] This invention provides a method for intelligent identification of the furnace condition of a top-blown tin smelting furnace, including:

[0006] Step 1: Synchronously collect multi-source heterogeneous data during the top-blown furnace tin smelting process and preprocess it. The multi-source heterogeneous data includes state parameter sequences, slag composition data and slag analysis composition data. The preprocessing includes outlier removal, missing value compensation and data standardization.

[0007] Step 2: Based on the preprocessed data, a multi-level sliding time window is used to extract process features at different time scales, and a temporal convolutional network is used to extract the time delay features corresponding to the process features. The temporal convolutional network adopts a dilated causal convolutional structure, and its dilation rate is set according to an exponential law to capture the dynamic delay patterns between parameters from the minute level to the hour level.

[0008] Step 3: Time alignment is performed on the multi-source data after time-delay feature extraction, and the data is input into the multi-modal Transformer fusion module to obtain fused features. The multi-modal Transformer fusion module adaptively learns the correlation weights between different modal data through a cross-modal attention mechanism to achieve deep fusion of multi-source heterogeneous data.

[0009] Step 4: Input the fused features into the classifier to obtain the identification results of the furnace condition category, and construct a multi-level early warning mechanism based on the identification results and time-varying feature trend analysis to issue early warning signals before key parameters become abnormal.

[0010] According to a specific implementation of the present invention, the acquisition frequency of the state parameter sequence is 15-30 seconds, including the rising flue temperature and flue gas component concentration; the slag component data is acquired by a laser detector, with a sampling interval of 30 minutes; the slag analysis component data is obtained by sampling and testing twice in each furnace cycle.

[0011] According to a specific implementation of an embodiment of the present invention, the preprocessing step includes:

[0012] The 3σ principle is used to identify and remove outlier data points;

[0013] For missing data points, linear interpolation is used for compensation;

[0014] The slag composition data were normalized to the [0,1] interval using Min-Max.

[0015] The flue gas composition data and flue temperature data were standardized using Z-score and converted into a distribution with a mean of 0 and a standard deviation of 1.

[0016] According to a specific implementation of an embodiment of the present invention, the multi-level sliding time window includes a first window for capturing rapid response parameter fluctuations, a second window for capturing the mid-term effects of raw material melting, and a third window for capturing the long-term cumulative effects of process conditions.

[0017] According to a specific implementation of an embodiment of the present invention, the temporal convolutional network TCN has an 8-layer dilated causal convolutional structure, with dilation rates d set sequentially to 1, 2, 4, 8, 16, 32, 64, 128, kernel width K=5, and LeakyReLU is used as the activation function with a negative slope coefficient α=0.1.

[0018] According to a specific implementation of an embodiment of the present invention, step 3 specifically includes:

[0019] Step 3.1: Use the dynamic time warping algorithm to perform time alignment on the multi-source data after time delay feature extraction;

[0020] Step 3.2: Input the time-aligned data into a multimodal Transformer fusion module with an 8-head attention mechanism and a 256-dimensional hidden layer. Perform feature fusion using a weighted summation method through a cross-modal attention mechanism to obtain the fused features:

[0021]

[0022] in, Indicates the number of modes. For the weights of the i-th mode in adaptive learning, This represents the original high-level feature representation of the i-th modality.

[0023] According to a specific implementation of an embodiment of the present invention, before step 4, the method further includes:

[0024] The classifier is trained using the training and validation sets, and weighted cross-entropy is used as the loss function.

[0025]

[0026] in, The loss weights represent the values ​​of outlier samples. This represents the standard cross-entropy classification loss;

[0027] The learning rate is adjusted using a cosine annealing strategy, and the learning rate update formula is as follows:

[0028] ,

[0029] in, , These represent the initial minimum learning rate and the initial maximum learning rate, respectively. This indicates the number of training cycles that have been executed so far. This indicates the maximum number of training cycles set.

[0030] According to a specific implementation of an embodiment of the present invention, the furnace condition categories include normal furnace condition, insufficient reducing atmosphere and high furnace temperature, insufficient reducing atmosphere and low furnace temperature, excessively strong reducing atmosphere and high furnace temperature, and excessively strong reducing atmosphere and low furnace temperature; the warning signal is issued in advance when the temperature or oxygen potential parameters show an abnormal trend, but have not yet reached the abnormal threshold.

[0031] The intelligent identification scheme for top-blown furnace tin smelting conditions in this embodiment of the invention includes: Step 1, synchronously collecting multi-source heterogeneous data during the top-blown furnace tin smelting process and preprocessing it, wherein the multi-source heterogeneous data includes state parameter sequences, slag composition data, and slag analysis composition data, and the preprocessing includes outlier removal, missing value compensation, and data standardization; Step 2, based on the preprocessed data, using a multi-level sliding time window to extract process features at different time scales, and using a temporal convolutional network to extract the time delay features corresponding to the process features, wherein the temporal convolutional network adopts a dilated causal convolution structure, and its dilation rate is set according to an exponential law. Step 3: Time alignment is performed on the multi-source data after time delay feature extraction, and the data is input into the multi-modal Transformer fusion module to obtain fusion features. The multi-modal Transformer fusion module adaptively learns the correlation weights between different modal data through a cross-modal attention mechanism to achieve deep fusion of multi-source heterogeneous data. Step 4: The fusion features are input into a classifier to obtain the furnace condition category identification result. Based on the identification result and time-varying feature trend analysis, a multi-level early warning mechanism is constructed to issue early warning signals before key parameters become abnormal.

[0032] The beneficial effects of the embodiments of the present invention are as follows:

[0033] 1. Accurate Time Delay Effect Modeling: By replacing the traditional LSTM model with a Temporal Convolutional Network (TCN), and utilizing a dilated causal convolutional structure, the model effectively captures time delay effects at different time scales, particularly suitable for the variable time delay problem in the top-blown tin refining process. This innovation improves the accuracy of time delay modeling and solves the problem that traditional methods cannot accurately handle dynamic time delay effects.

[0034] 2. Multimodal Data Fusion: An innovative cross-modal attention mechanism is introduced, utilizing a multimodal Transformer architecture to achieve deep fusion of different data sources (such as flue gas temperature, slag composition, and flue gas composition), automatically learning the correlation weights between various modal data. This fusion significantly improves the accuracy and reliability of furnace condition identification, especially in the processing of multi-source heterogeneous data.

[0035] 3. Improved Real-Time Performance: By designing a lightweight model and employing depthwise separable convolution and model pruning techniques, the number of model parameters was successfully compressed, improving the model's inference speed and meeting the high real-time requirements of industrial environments. This innovation solves the performance bottleneck of complex algorithms in real-time applications.

[0036] 4. Early Warning Mechanism: A multi-level early warning mechanism is established, which, through time-varying characteristic analysis and trend prediction, can issue warnings some time before temperature anomalies occur. This provides ample time for process adjustments, significantly improving production safety and reducing the potential risks brought about by abnormal operating conditions.

[0037] 5. Adaptive Data Sampling and Alignment: By developing an adaptive sampling and alignment mechanism, the problem of fusing multi-source heterogeneous data is solved, ensuring high recognition accuracy even in the event of missing data. Furthermore, physical constraint feature enhancement and dynamic attention weight allocation techniques further enhance the system's robustness and practicality. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 A flowchart illustrating an intelligent identification method for the condition of a top-blown tin smelting furnace provided in an embodiment of the present invention;

[0040] Figure 2 This is a flowchart illustrating the algorithm framework of an intelligent identification method for the condition of a top-blown tin smelting furnace provided in an embodiment of the present invention.

[0041] Figure 3This is a 3D visualization diagram of the field of view of a dilated causal convolution kernel provided in an embodiment of the present invention. (a) shows that when the dilation rate is 1, the convolution kernel is tightly connected on the input sequence, with a receptive field of 5, used to capture short-range, continuous temporal dependencies. (b) shows that when the dilation rate increases to 2, the convolution kernel structure becomes sparse, and the receptive field expands to 9, enabling the capture of longer-period temporal patterns without changing the number of parameters. (c) shows that when the dilation rate further increases to 4, the convolution kernel becomes extremely sparse, and the receptive field expands dramatically to 17, achieving coverage of extremely long-range, macroscopic evolutionary trends in the input sequence.

[0042] Figure 4 This is a flowchart of a multimodal data fusion layer provided in an embodiment of the present invention. Detailed Implementation

[0043] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0044] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0045] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0046] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0047] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0048] This invention provides an intelligent identification method for the furnace condition of a top-blown tin smelting furnace, which can be applied to the tin smelting process in smelting scenarios.

[0049] See Figure 1 This is a flowchart illustrating a method for intelligent identification of the furnace condition in a top-blown tin smelting furnace according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method mainly includes the following steps:

[0050] Step 1: Synchronously collect multi-source heterogeneous data during the top-blown furnace tin smelting process and preprocess it. The multi-source heterogeneous data includes state parameter sequences, slag composition data and slag analysis composition data. The preprocessing includes outlier removal, missing value compensation and data standardization.

[0051] In practical implementation, the system employs a multi-source heterogeneous data synchronous acquisition scheme during the data acquisition phase. The data primarily originates from real-time records of the top-blown furnace tin smelting process collected by tin smelting units. After investigation, analysis is mainly based on three factors: state parameters (including flue gas composition and rising flue temperature), slag composition obtained by the laser detector, and slag analysis results. State parameters are obtained through real-time sensor detection, while slag composition and analysis results require operators to obtain slag from the furnace using a sampling rod, resulting in varying frequencies. Operators sample every 30 minutes, allowing the laser detector to obtain slag composition data once, while slag composition analysis is only performed twice per furnace cycle.

[0052] The data preprocessing workflow includes rigorous quality control steps. The system uses the 3σ principle to automatically identify and remove outlier data points. This principle is based on the statistical properties of the normal (Gaussian) distribution. For a dataset that follows a normal distribution, approximately 99.73% of the data will fall within three standard deviations of the mean. Therefore, any data point deviating from the mean by more than 3 standard deviations has an extremely low probability of occurrence (less than 0.3%), and can be considered a low-probability event. This is usually caused by abnormal reasons such as measurement errors, transmission errors, or severe operational disturbances, and should therefore be removed to ensure data quality. Secondly, the system uses a linear interpolation method to compensate for missing values. This method assumes that the data value changes at a constant rate (linearly) between two known data points. When a sensor loses individual data points due to a momentary malfunction, the system will use the nearest valid data before and after that point, through... Reasonable estimates were calculated to fill the gaps and maintain the continuity of the time series. Data from different sources were standardized: slag composition was normalized to the [0,1] range using Min-Max normalization to eliminate the order-of-magnitude differences between different components, making all components fall within a similar numerical range, thereby accelerating model convergence and improving model stability and accuracy; flue gas data and flue gas temperature were standardized using Z-score normalization, which transforms the original data into a standard normal distribution (or close to a standard normal distribution) with a mean of 0 and a standard deviation of 1. These processing methods ensured the comparability of multi-source data and the stability of model training.

[0053] Step 2: Based on the preprocessed data, a multi-level sliding time window is used to extract process features at different time scales, and a temporal convolutional network is used to extract the time delay features corresponding to the process features. The temporal convolutional network adopts a dilated causal convolutional structure, and its dilation rate is set according to an exponential law to capture the dynamic delay patterns between parameters from the minute level to the hour level.

[0054] In practice, the feature extraction stage employs multi-scale time-varying feature processing technology. The system sets three sliding windows: 5-15 minutes, 20-40 minutes, and 50-90 minutes, to capture feature changes at different time scales, such as rapid response parameters, raw material melting effects, and cumulative process influences. Time-delay feature extraction uses an 8-layer dilated causal convolutional network, with its core calculation formula being... The dilation rate d is set to increase exponentially from 1, 2, 4, 8, 16, 32, 64, 128, the kernel width K=5, and the LeakyReLU activation function is used. This effectively captures the dynamic delay patterns between parameters. Its computational field of view changes with the expansion rate as follows: Figure 3 As shown. The expression for the LeakyReLU activation function is: ,in In this system, the value is set to 0.1. By introducing a very small negative slope (α=0.1), LeakyReLU ensures that there is also a small gradient in the negative region, thereby alleviating the neuron death problem, enhancing the training stability and expressiveness of the model, and introducing a nonlinear transformation into the neural network, enabling it to fit complex nonlinear relationships.

[0055] Step 3: Time alignment is performed on the multi-source data after time-delay feature extraction, and the data is input into the multi-modal Transformer fusion module to obtain fused features. The multi-modal Transformer fusion module adaptively learns the correlation weights between different modal data through a cross-modal attention mechanism to achieve deep fusion of multi-source heterogeneous data.

[0056] In specific implementation, such as Figure 3 As shown, the system first employs a dynamic time warping algorithm to achieve time alignment. This algorithm measures the similarity between two time series of different lengths or with nonlinear time offsets. It aligns the sequences by finding an optimal warping path, thereby eliminating distortion on the time axis. The calculation method involves constructing an N × M distance matrix D given two sequences P (a state parameter sequence of length N) and G (a flue gas sequence of length M), where D(i, j) represents a point... and The distance (e.g., Euclidean distance). Its core is calculating the cumulative distance matrix γ:

[0057]

[0058] Ultimately, the distance between the two sequences is γ(N, M), and the optimal normalized path is found through backtracking. The Transformer module is configured with an 8-head attention mechanism and a 256-dimensional hidden layer, and its calculation formula is as follows: By using a learnable position encoding matrix and temperature parameter settings, a deep correlation is established between state parameters. Cross-modal fusion employs a weighted approach. ,in Represents two modes, These are modality weights for adaptive learning. This design preserves the characteristics of each modality's data while also uncovering potential cross-modal correlations.

[0059] The model training process emphasizes the specific needs of industrial scenarios. The training dataset must contain at least 60% abnormal furnace condition samples, cover more than five operating conditions, and the data collection period must fully record at least 400 furnace runs. The optimization algorithm employs a cosine annealing strategy, and the learning rate update formula is as follows: The initial learning rate The learning rate decreases slowly, which is better than a step-like decrease, allowing the model to converge to local optima more finely in the later stages of training. The system employs a batch size of 32 and a weighted cross-entropy loss function. The loss function uses weighted cross-entropy: The weight of outlier samples is set to 1.5, which amplifies the gradient signal caused by model errors on outliers during backpropagation. This forces the model to focus more on correctly classifying outliers, thus significantly improving recall and sensitivity for key operating conditions. An early stopping mechanism is implemented during training, automatically terminating training when the evidence set loss does not decrease for 10 consecutive rounds.

[0060] Step 4: Input the fused features into the classifier to obtain the identification results of the furnace condition category, and construct a multi-level early warning mechanism based on the identification results and time-varying feature trend analysis to issue early warning signals before key parameters become abnormal.

[0061] In practical implementation, after obtaining the fused features and the trained classifier, the fused features can be input into the trained classifier for furnace condition category identification. Simultaneously, a multi-level early warning mechanism can be constructed, achieving early identification of abnormal operating conditions through time-varying feature analysis and trend prediction. The system can issue warnings some time before temperature anomalies occur, providing a sufficient time window for process adjustments and significantly improving production safety. This early warning mechanism, based on deep analysis of multi-parameter coupling relationships, can identify early abnormal features that are difficult to detect with the human eye.

[0062] The intelligent furnace condition identification method for top-blown tin smelting furnaces provided in this embodiment replaces the traditional LSTM model with a Temporal Convolutional Network (TCN). Utilizing a dilated causal convolutional structure, it effectively captures time-delay effects at different time scales, particularly adapting to the variable time-delay problem in top-blown tin smelting processes. This innovation improves the accuracy of time-delay modeling and solves the problem of traditional methods failing to accurately handle dynamic time-delay effects. It innovatively introduces a cross-modal attention mechanism, using a multimodal Transformer architecture to achieve deep fusion of different data sources (such as flue gas temperature, slag composition, and flue gas composition), automatically learning the correlation weights between different modalities. This fusion significantly improves the accuracy and reliability of furnace condition identification, especially performing excellently in processing multi-source heterogeneous data. Through lightweight model design, employing deep separable convolution and model pruning techniques, it successfully compresses the number of model parameters, improving the model's inference speed and meeting the high real-time requirements of industrial environments. This innovation solves the performance bottleneck of complex algorithms in real-time applications. A multi-level early warning mechanism is constructed, which, through time-varying feature analysis and trend prediction, can issue early warnings some time before temperature anomalies occur. This provides ample time for process adjustments, significantly improves production safety, and reduces potential risks from abnormal operating conditions;

[0063] By developing an adaptive sampling alignment mechanism, the problem of fusing multi-source heterogeneous data was solved, ensuring high recognition accuracy even in the event of missing data. Furthermore, physical constraint feature enhancement and dynamic attention weight allocation techniques further enhanced the system's robustness and practicality.

[0064] The method of the present invention will be further described below with reference to a specific embodiment. To verify the effectiveness of the present invention, we conducted rigorous experimental verification based on actual production data of a top-blown furnace in a tin smelting enterprise. The experiment aims to demonstrate the accuracy, real-time performance, and reliability of the proposed solution in the task of identifying the furnace condition of a top-blown furnace for tin smelting.

[0065] 1. Explanation of Experimental Data

[0066] The experimental data came from the company's historical database of top-blown furnace production processes from January 2023 to June 2024. The dataset fully records the production processes of more than 250 heats, covering normal furnace conditions and four typical abnormal furnace conditions (insufficient reducing atmosphere and high furnace temperature, insufficient reducing atmosphere and low furnace temperature, excessively strong reducing atmosphere and high furnace temperature, and excessively strong reducing atmosphere and low furnace temperature).

[0067] The data used in the experiment were consistent with those described in Part 5, and mainly included three categories: (1) State parameter sequence (high frequency): the sampling frequency was 15-30 seconds, including the rising flue temperature and flue gas composition (concentration of CO, CO2, O2 and SO2 in the flue gas); (2) Slag composition (low frequency): samples were taken every 30 minutes, and the tin content of the slag was obtained by laser detector; (3) Slag analysis composition (ultra-low frequency): slag samples were taken only twice for analysis in each furnace period to obtain the slag composition analysis table.

[0068] The labeling frequency of the data is consistent with the sampling frequency of the slag composition. Experienced engineers conduct a comprehensive analysis based on the slag morphology, flue gas change trends, and slag tin content data for the corresponding time period, and label the furnace condition category (normal or one of the four abnormal categories mentioned above) within that time window as the true label for model training and evaluation.

[0069] 2. Data Preprocessing Flow

[0070] The preprocessing procedure is executed exactly as described in Part 5, with the specific parameters and steps as follows:

[0071] The first step is outlier removal, which strictly applies the 3σ principle to process each sensor data stream. The mean (μ) and standard deviation (σ) of each parameter sequence are calculated, and data points falling outside the range of (μ-3σ, μ+3σ) are automatically identified and removed.

[0072] The second step is missing value compensation. For missing data points caused by momentary sensor malfunctions, linear interpolation is used to fill them in. The interpolation formula is: .in, It is the nearest valid data value before and after the missing point. That is the corresponding timestamp.

[0073] The third step is data standardization. For slag composition data (including laser analyzer data and laboratory data), Min-Max normalization is used to scale the data to the [0, 1] interval. For flue gas composition data and rising flue temperature data, Z-score standardization is used to convert the data into a distribution with a mean of 0 and a standard deviation of 1.

[0074] 3. Feature Extraction and Model Training

[0075] First, in the feature extraction stage, multi-scale temporal feature processing technology is employed. Specifically, three sliding time windows of 5 minutes, 30 minutes, and 60 minutes are set to capture the fluctuations of rapidly responding operating parameters, the mid-term effects of raw material melting, and the long-term cumulative effects of process conditions, respectively. Based on this, time-delay features are extracted using an 8-layer dilated causal convolutional network (TCN). The network's dilation rate (d) is set exponentially to [1, 2, 4, 8, 16, 32, 64], the kernel width (K) is fixed at 5, and a LeakyReLU activation function (negative slope coefficient α = 0.1) is used. This structure can effectively model the complex dynamic delay patterns between process parameters, ranging from minutes to hours.

[0076] Subsequently, multimodal data fusion is performed. Since the state parameter sequence and the flue gas composition sequence are acquired at different frequencies, a Dynamic Time Warping (DTW) algorithm is first applied to align them in time to eliminate time axis distortion. The aligned multimodal data is then fed into the fusion layer for processing. The core of this fusion layer is a Transformer module configured with an 8-head self-attention mechanism and a hidden layer dimension of 256. Attention calculation follows the formula: This is to uncover deep correlations between different modalities of data. The final cross-modal fusion is achieved using a weighted summation method, i.e. Where M=2 represents two modes (state parameters and slag composition), and α is the modal weights for adaptive learning during model training. This design can both preserve the characteristics of each mode and explore their potential correlations.

[0077] Regarding model training details, the dataset was strictly divided into training (70%), validation (15%), and test (15%) sets according to each furnace cycle, ensuring complete data independence between different sets and avoiding information leakage. The Adam optimizer was used for optimization, employing a cosine annealing learning rate scheduling strategy (initial learning rate...). Minimum learning rate The model aims for stable convergence. A weighted cross-entropy loss function is used. To address the class imbalance problem caused by the scarcity of outlier samples in industrial data, the loss weight for outlier samples is set to 1.5 to improve the model's sensitivity in identifying key operating conditions. The training batch size is set to 32, the maximum training epoch is 100, and an early stopping mechanism is implemented (training automatically terminates when the validation set loss does not decrease for 10 consecutive epochs) to prevent overfitting and ensure the model with optimal generalization ability is obtained.

[0078] 4. Experimental Results Analysis and Visualization

[0079] On the independent test set, the proposed model achieved the following performance metrics: (1) Overall accuracy: 90.46%; (2) Recall rate for identifying abnormal furnace conditions: 89.33%.

[0080] The above experimental results show that the intelligent identification method for top-blown tin smelting furnace conditions based on TCN-Transformer proposed in this invention has detailed and feasible solutions in terms of data processing, feature extraction, model training and practical application. It can efficiently, accurately and in real time complete the task of furnace condition identification and early warning, which is significantly better than the existing technical solutions. It provides solid data support and experimental evidence for the industrial application of this technology.

[0081] Compared with the prior art, the method of the present invention has the following significant technical advantages:

[0082] 1. Significantly improves the timeliness and early warning capability of furnace condition identification.

[0083] The method of this invention significantly improves inference speed on edge computing platforms through lightweight model design (depthiable separable convolution and model pruning), meeting the real-time monitoring requirements with millisecond-level response. Compared to the detection lag in manual sampling analysis in existing technologies and the real-time inference that is difficult to achieve with traditional machine learning models, this solution can realize real-time monitoring of furnace conditions.

[0084] The multi-level early warning mechanism constructed by the method of the present invention can identify early signs before the occurrence of abnormalities in key parameters such as temperature through time-varying feature analysis and trend prediction. It provides operators with sufficient time window (e.g., several minutes to tens of minutes in advance) to make process adjustments, transforming passive alarms into proactive early warnings. This fundamentally changes the "post-event remediation" control mode of existing technologies, effectively avoiding production accidents and improving production safety.

[0085] 2. Significantly improves the accuracy and reliability of furnace condition identification.

[0086] Existing technologies struggle to accurately model the dynamic time delay effects and multi-physics coupling relationships in the top-blown furnace tin refining process. The core innovation of this invention lies in employing a Temporal Convolutional Network (TCN) to explicitly model multi-scale time delay relationships. The dilated causal convolutional structure of the TCN, with its exponentially increasing expansion rate (1, 2, 4, ..., 128), effectively captures different time delay effects ranging from minutes to hours, accurately characterizing the differentiated delay patterns of parameter influence on furnace conditions, thus solving the technical problem of inaccurate time delay modeling in existing models.

[0087] To address the issue of insufficient fusion of multi-source heterogeneous data, this invention introduces a cross-modal attention mechanism and a multimodal Transformer architecture. This architecture adaptively learns the correlation weights between different modalities of data (such as state parameters, flue gas composition, and slag data), achieving deep information fusion and establishing a complete causal chain of "process operation - physical field change - furnace condition response," significantly improving the model's recognition accuracy and robustness under complex operating conditions. Compared to traditional methods that simply stitch together data or rely on subjective feature engineering, this method can automatically uncover deep, nonlinear correlations.

[0088] 3. Enhance the system's adaptability and robustness to complex operating conditions.

[0089] The method of this invention effectively unifies data streams of different frequencies through an adaptive sampling alignment mechanism (such as a dynamic time warping algorithm), solving the problem of time scale mismatch in multi-source heterogeneous data. The system can maintain a high recognition accuracy even when some data is missing.

[0090] The weighted cross-entropy loss function (with anomaly class samples weighted at 1.5) and extensive data covering multiple operating conditions employed during model training compel the model to focus more on identifying abnormal furnace conditions. This significantly improves the recall and sensitivity for key abnormal conditions (such as excessively high / low oxygen potential combined with abnormal furnace temperature), reducing the risk of missed detections. Compared to existing technologies that rely excessively on operator experience, this method provides more consistent and objective criteria.

[0091] 4. To provide technical support for achieving refined control and resource conservation.

[0092] By providing real-time and accurate furnace condition identification and early warning, the method of this invention can guide operators to adjust key parameters such as the real-time flow rate of reducing coal and combustion coal in a timely manner, helping to stabilize the oxygen potential and temperature in the furnace within the optimal range. This is expected to bring the following benefits:

[0093] (1) Improve metal recovery rate (direct recovery rate): by reducing abnormal situations such as excessive tin content in slag or loss of metal mist; (2) Reduce energy consumption: by optimizing combustion and control processes to avoid unnecessary energy waste; (3) Improve product quality stability: ensure more stable crude tin composition to provide a guarantee for the production of high-purity tin (≥99.95%); (4) Extend furnace life: by preventing damage to equipment caused by severe abnormal furnace conditions.

[0094] The method of this invention systematically solves the core problems of existing technologies, such as inaccurate time delay modeling, insufficient multimodal fusion, poor real-time performance, and delayed early warning. It demonstrates significant technological progress in terms of the timeliness, accuracy, reliability, and adaptability of furnace condition identification, and provides a practical and efficient technical solution for the intelligent upgrading and refined control of the top-blown furnace tin smelting process.

[0095] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent identification of the condition of a top-blown tin smelting furnace, characterized in that, include: Step 1: Synchronously collect multi-source heterogeneous data during the top-blown furnace tin smelting process and preprocess it. The multi-source heterogeneous data includes state parameter sequences, slag composition data and slag analysis composition data. The preprocessing includes outlier removal, missing value compensation and data standardization. Step 2: Based on the preprocessed data, a multi-level sliding time window is used to extract process features at different time scales, and a temporal convolutional network is used to extract the time delay features corresponding to the process features. The temporal convolutional network adopts a dilated causal convolutional structure, and its dilation rate is set according to an exponential law to capture the dynamic delay patterns between parameters from the minute level to the hour level. Step 3: Time alignment is performed on the multi-source data after time-delay feature extraction, and the data is input into the multi-modal Transformer fusion module to obtain fused features. The multi-modal Transformer fusion module adaptively learns the correlation weights between different modal data through a cross-modal attention mechanism to achieve deep fusion of multi-source heterogeneous data. Step 4: Input the fused features into the classifier to obtain the furnace condition category identification result, and construct a multi-level early warning mechanism based on the identification result and time-varying feature trend analysis to issue early warning signals before key parameters become abnormal.

2. The method according to claim 1, characterized in that, The state parameter sequence is collected at a frequency of 15-30 seconds, including the rising flue temperature and flue gas component concentration; the slag component data is acquired by a laser detector with a sampling interval of 30 minutes; the slag analysis component data is obtained by sampling and testing twice during each furnace cycle.

3. The method according to claim 2, characterized in that, The preprocessing steps include: The 3σ principle is used to identify and remove outlier data points; For missing data points, linear interpolation is used for compensation; The slag composition data were normalized to the [0,1] interval using Min-Max. The flue gas composition data and flue temperature data were standardized using Z-score and converted into a distribution with a mean of 0 and a standard deviation of 1.

4. The method according to claim 1, characterized in that, The multi-level sliding time window includes a first window for capturing rapid response parameter fluctuations, a second window for capturing the mid-term effects of raw material melting, and a third window for capturing the long-term cumulative effects of process conditions.

5. The method according to claim 4, characterized in that, The Temporal Convolutional Network (TCN) has an 8-layer dilated causal convolutional structure with dilation rates d set sequentially to 1, 2, 4, 8, 16, 32, 64, and 128. The kernel width K is 5, and LeakyReLU is used as the activation function with a negative slope coefficient α = 0.

1.

6. The method according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Use the dynamic time warping algorithm to perform time alignment on the multi-source data after time delay feature extraction; Step 3.2: Input the time-aligned data into a multimodal Transformer fusion module with an 8-head attention mechanism and a 256-dimensional hidden layer. Perform feature fusion using a weighted summation method through a cross-modal attention mechanism to obtain the fused features: ； in, Indicates the number of modes. For the weights of the i-th mode in adaptive learning, This represents the original high-level feature representation of the i-th modality.

7. The method according to claim 1, characterized in that, Before step 4, the method further includes: The classifier is trained using the training and validation sets, and weighted cross-entropy is used as the loss function. ； in, The loss weights represent the values ​​of outlier samples. This represents the standard cross-entropy classification loss; The learning rate is adjusted using a cosine annealing strategy, and the learning rate update formula is as follows: ； in, , These represent the initial minimum learning rate and the initial maximum learning rate, respectively. This indicates the number of training cycles that have been executed so far. This indicates the maximum number of training cycles set.

8. The method according to claim 1, characterized in that, The furnace condition categories include normal furnace condition, insufficient reducing atmosphere and high furnace temperature, insufficient reducing atmosphere and low furnace temperature, excessively strong reducing atmosphere and high furnace temperature, and excessively strong reducing atmosphere and low furnace temperature; the warning signal is issued in advance when the temperature or oxygen potential parameters show an abnormal trend, but have not yet reached the abnormal threshold.

Images