Sensorless molten steel underwater slagging nozzle closing prediction method

Abstract

The invention belongs to the technical field of steel and iron manufacturing, and particularly relates to a sensorless method for predicting closing of a molten steel underwater slagging nozzle. A sensorless molten steel slag discharging nozzle closing prediction method comprises the following steps that slag discharging is monitored in real time, various indirect data of a sensorless environment in the steel production process are collected, and a data sequence of the change of the weight of molten steel along with time is obtained through the indirect data; monitoring the slag discharging process in real time by using a linear prediction method of a sensorless environment to obtain linear correlation between indirect data and a slag discharging state, and predicting the stopping time of slag discharging and water gap closing equipment; and by predicting the stopping time of slag discharging and water gap closing equipment, if an alarm condition is triggered, a water gap is closed. According to the method, the time for closing the water gap during slag tapping can be quickly and accurately predicted, the use of a traditional slag tapping sensor is avoided, the equipment cost and the maintenance complexity are reduced, and the requirements of industrial production for low-delay and high-precision prediction are met.

Description

Technical Field

[0001] The present invention belongs to the technical field of steel manufacturing, and particularly relates to a method for predicting the closing of a submerged slag nozzle of molten steel without sensors. Background Art

[0002] In the process of steel manufacturing, the monitoring and control of molten steel slag are crucial for ensuring the quality of steel products and production efficiency. Slag flowing out with molten steel not only affects the purity of steel products but may also cause damage to production equipment. Therefore, accurately predicting and controlling the closing time of the slag nozzle is an important link in steel production. In the refining process of steel production, accurate detection of molten steel slag and timely closing of the nozzle are crucial for the quality control of molten steel. The defects and deficiencies of the existing technologies are mainly reflected in the following aspects:

[0003] 1. Sensor-related problems

[0004] High cost and easy to be damaged: The existing methods for predicting the closing of the submerged slag nozzle mainly rely on various sensors, such as electromagnetic sensors, optical sensors, vibration sensors, etc. These sensors are expensive, and their procurement costs, installation and commissioning costs, and subsequent maintenance costs are all very high. Moreover, in the harsh steel production environment, sensors are easily damaged by factors such as high temperature, high humidity, strong electromagnetic fields, and dust, and need to be replaced frequently, further increasing the production cost.

[0005] Complex installation and maintenance: The installation of sensors requires precise operation by professional technicians, and the continuous casting equipment needs to be modified to adapt to the sensors. During the installation process, various factors such as the position, angle, and protection of the sensors need to be considered. Slight carelessness may affect the measurement accuracy. In addition, sensors need to be calibrated, cleaned, and maintained regularly during use, which requires a large amount of manpower and time, increasing the difficulty of production management.

[0006] Data quality affected by environmental interference: The steel production site has a complex environment, and the data collected by sensors is easily interfered with. For example, high temperature may cause thermal drift of the sensors, resulting in deviation of the measured values; strong electromagnetic fields may interfere with the signal transmission of electromagnetic sensors, causing data loss or errors; dust may block the detection components of the sensors, affecting their sensitivity and accuracy. These factors lead to unstable data quality collected by sensors, thereby affecting the reliability of the prediction results.

[0007] 2. Limitations in data processing and methods

[0008] Difficulty in multi-sensor data fusion: The data formats, ranges, accuracies, etc. of the data collected by different types of sensors are all different. Fusing this data for use in prediction methods is a complex process. It is necessary to design complex data fusion algorithms to coordinate the relationships between different sensor data, determine the weights and priorities of the data, which increases the difficulty and computational amount of data processing and is prone to introducing errors.

[0009] Trade-off between method complexity and accuracy: Prediction methods based on sensor data are usually complex to adapt to the input of various sensor data and complex production conditions. However, complex methods often have the risk of overfitting and may not be able to accurately generalize to different working conditions in actual production, resulting in unstable prediction accuracy. At the same time, complex methods have high computational complexity and require high-performance computing equipment support, increasing system costs and running times, and it is difficult to meet the real-time requirements of industrial production. Summary of the Invention

[0010] In view of the above problems, the object of the present invention is to provide a sensorless prediction method for the closing of the slag notch of molten steel. According to the real-time collected molten steel weight and time data, by determining the linear relationship between the change trend of the molten steel weight and time, the time for closing the slag notch to close the nozzle is predicted quickly and accurately. This method avoids the use of traditional slag sensors, reduces equipment costs and maintenance complexity, and at the same time uses a simple and efficient linear method to meet the requirements of industrial production for low-latency and high-precision prediction.

[0011] The technical solution of the present invention is as follows: A sensorless prediction method for the closing of the slag notch of molten steel includes the following steps:

[0012] S1: Real-time monitoring of slag discharge, collecting a variety of indirect data in a sensorless environment during the steel production process. The indirect data includes heat number, ladle weight, weight when the ladle reaches a specific position, weight when the ladle leaves a specific position, weight of the empty ladle, net weight of molten steel, tundish temperature, maximum value of the ladle temperature, minimum value of the ladle temperature, thickness of the ladle slag, start time of tundish pouring, end time of tundish pouring. Through the indirect data, a data sequence of the change of molten steel weight over time is obtained;

[0013] S2: Using a linear prediction method in a sensorless environment to monitor the slag discharge process in real time, obtaining the linear correlation between the indirect data and the slag discharge state, reconstructing the signal characteristics of the slag discharge process, realizing real-time tracking of the slag discharge situation, and then predicting the stop time of the equipment for closing the slag notch;

[0014] S3: By predicting the stop time of the equipment for closing the slag notch, it is judged whether the alarm condition is triggered. If it is triggered, the nozzle is closed.

[0015] In step S2, a linear prediction method for a sensorless environment is used to monitor the slagging process in real time, obtaining a linear correlation between the indirect data and the slagging state. The specific process is as follows:

[0016] S21: Data decomposition;

[0017] For the data of the molten steel weight changing with time, the moving average kernel method is adopted to decompose the data sequence into two key components: the trend component and the remainder component. The trend component is used to accurately capture the overall change trend of the molten steel weight over a long time span, reflecting the long-term trend of the molten steel weight change. The remainder component covers the relatively short-term fluctuation information and the existing seasonal change characteristics, and is used to reveal the change law and periodic fluctuation of the data in a short time period. Specifically:

[0018] Let the original data sequence be where y t represents the molten steel weight at time t. Through the moving average kernel method, it is decomposed into the following two parts:

[0019] y t = T t + R t (1)

[0020] In the formula, T t represents the trend component, which is used to describe the long-term change trend, and R t represents the remainder component, which is used to describe the short-term fluctuation and seasonal change. The trend component T t is obtained through a smoothing operation. The specific calculation method is as follows:

[0021]

[0022] where: k is a parameter of the window width, representing the time span involved in the smoothing process, 2k + 1 is the length of the smoothing window, that is, the width of the moving average kernel, y t+i represents the value of the original data at time t + i. When t + i exceeds the sequence range, boundary processing methods such as the mirror method or interpolation method can be adopted. The meaning of this formula is to calculate the mean value of the original data using a sliding window, thereby eliminating short-term fluctuations;

[0023] The remainder component R t is obtained by subtracting the trend component from the original data:

[0024] R t = y t - T t (3)

[0025] The remainder component reflects the part of the data that is not captured by the trend component, including short-term fluctuations and possible periodic characteristics;

[0026] S22: Linear layer application

[0027] After the decomposition of the data sequence is completed, an independent single-layer linear layer is equipped for the decomposed trend component and remainder component respectively. For the trend component, its corresponding linear layer learns a set of weight parameters through deep learning and transforms the trend information into a feature representation form that makes a key contribution to the prediction of the stopping time of the slag-dropping and nozzle-closing device. For the remainder component, its corresponding linear layer also extracts the feature information closely related to the device stopping time contained therein through linear transformation operations;

[0028] S23: Add the features obtained after being processed by the two linear layers to obtain the final prediction result.

[0029] In step S22, for the trend component and the remainder component, their corresponding linear layers learn a set of weight parameters through deep learning and transform the trend information and the remainder information into feature representation forms that make key contributions to the prediction of the stopping time of the slag-dropping and nozzle-closing device, specifically as follows:

[0030] Let the time series data be t = [t1, t2,... t n , and the corresponding molten steel weight change data be ω = [ω1, ω2,... ω n . The stopping time of the slag-dropping and nozzle-closing device predicted by the linear prediction method is calculated through the following linear combination:

[0031]

[0032] In the formula, α, β, γ, and δ are a set of weight parameters learned through deep learning. α controls the influence degree of the product term of molten steel weight and time on the prediction result, reflecting the factor of the interaction between weight change and time; β mainly reflects the contribution of the molten steel weight change alone to the prediction result; γ focuses on the influence of the time factor alone; δ, as a bias term, is used to adjust the overall prediction value to adapt to different working conditions and data distribution characteristics.

[0033] To optimize the weight parameters of α, β, γ, and δ, the mean square error MSE is used as the loss function, and the calculation formula of MSE is as follows:

[0034]

[0035] In the formula, m is the number of training samples, is the predicted device stopping time of the jth training sample, and T i is the actual device stopping time of the jth training sample;

[0036] Given the actual slag stopper nozzle closing time as T, by minimizing the MSE, continuously adjust the parameters α, β, γ, and δ to make the predicted value as close as possible to the true equipment stop time T.

[0037] The network structure of deep learning in step S22 includes an input layer, a hidden layer, and an output layer, where:

[0038] The input layer has two nodes, which receive molten steel weight change data and time data respectively. These data can be obtained through high-precision weighing sensors installed on the ladle support structure and synchronous time acquisition devices. The accuracy of the weighing sensor can reach ±0.1%, which can accurately measure the tiny changes in the molten steel weight. The accuracy of the time acquisition device is at the millisecond level to ensure the accuracy and synchronization of the time data. The input layer directly passes these two data to the hidden layer;

[0039] The hidden layer, that is, the linear transformation layer, consists of a group of linear neurons and performs linear transformation operations. For the input molten steel weight data ω and time data t, the hidden layer performs a linear transformation through the weight matrix W and the bias vector b. The calculation formula for the output vector h is:

[0040]

[0041] In the formula, W is a 2×p weight matrix, p is the number of neurons in the hidden layer, b is a p-dimensional bias vector. The role of the hidden layer is to preliminarily integrate and extract features from the input molten steel weight and time data, mine the potential information in the data through linear transformation, and prepare for outputting an accurate predicted value of the equipment stop time;

[0042] The output layer receives the output h of the hidden layer and calculates the final prediction result through a linear weight vector ω0 and a bias term b0 The calculation formula is:

[0043]

[0044] In the formula, w0 is a p-dimensional weight vector. The output layer performs the final linear combination of the feature information extracted by the hidden layer and outputs the predicted value of the slag stopper nozzle closing equipment stop time.

[0045] For the parameter initialization of the network structure of deep learning in step S22, among the weight parameters, including the weight matrix W of the hidden layer and the weight vector w0 of the output layer, a random initialization method is adopted, and the initial weight values are randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.01. The bias terms, including the bias vector b of the hidden layer and the bias term b0 of the output layer, are initialized to 0.

[0046] In the parameter optimization of the deep learning network structure in step S22, an optimization algorithm based on gradient descent is used to update the parameters. By calculating the gradient of the loss function with respect to each parameter and according to the direction and magnitude of the gradient, the parameters are updated at a certain learning rate. The update formula for the weight parameters is as follows:

[0047]

[0048] In the formula, α is the learning rate. At the initial stage of training, a relatively large learning rate of 0.01 is selected. Then, as the training progresses, a learning rate decay strategy is adopted to gradually reduce the learning rate. Every 10 training cycles, the learning rate decays to 0.9 times the original value. At the same time, to prevent overfitting, a regularization technique such as L2 regularization is used, that is, the sum of the squares of the weight parameters is added to the loss function to constrain the weights. After adding L2 regularization, the loss function becomes:

[0049]

[0050] In the formula, λ is the regularization parameter, which is used to control the strength of regularization.

[0051] The technical effects of the present invention are as follows: 1. According to the molten steel weight and time data collected in real time, the present invention quickly and accurately predicts the time to close the nozzle for slagging by determining the linear relationship between the change trend of the molten steel weight and time, and uses a simple and efficient linear method to meet the requirements of industrial production for low-latency and high-precision prediction; 2. The present invention does not rely on the slagging nozzle closing time monitoring method in the prior art, which highly depends on various slag sensors. It only relies on the molten steel weight change data, time data and other easily obtainable relevant data, greatly simplifies the hardware structure, reduces the hardware cost, saves a large amount of capital investment for enterprises, and at the same time reduces the downtime caused by hardware failures and improves production efficiency; 3. Compared with the slag sensor, which is easily interfered with and damaged in a harsh industrial environment, such as high temperature, strong electromagnetic field, dust and other factors that will affect the performance and measurement accuracy of the sensor, resulting in inaccurate or lost monitoring data, thereby affecting the prediction accuracy of the slagging nozzle closing time and even possibly causing production accidents. The present invention does not rely on sensors, avoiding the direct impact of these environmental factors on the monitoring system, and significantly enhancing the stability and reliability of the system.

[0052] The following will be further described with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flowchart of a method for predicting the closing of a molten steel slagging nozzle without a sensor according to the present invention.

[0054] Figure 2It is the linear prediction flow chart of the sensorless environment of the present invention. Detailed implementation mode

[0055] Embodiment 1

[0056] As Figure 1 、 Figure 2 shown, a sensorless prediction method for closing the submerged nozzle of molten steel underwater includes the following steps:

[0057] S1: Real-time monitoring of slag entrainment, collecting a variety of indirect data in the sensorless environment during the steel production process. The indirect data includes heat number, ladle weight, weight when the ladle arrives at a specific position, weight when the ladle leaves a specific position, weight of the empty ladle, net weight of molten steel, ladle temperature, maximum value of ladle temperature, minimum value of ladle temperature, thickness of ladle slag, start time of pouring from the tundish, end time of pouring from the tundish. Through the indirect data, a data sequence of the change of molten steel weight over time is obtained;

[0058] S2: Using the linear prediction method in the sensorless environment to monitor the slag entrainment process in real time, obtaining the linear correlation between the indirect data and the slag entrainment state, reconstructing the signal characteristics of the slag entrainment process, realizing real-time tracking of the slag entrainment situation, and further predicting the stop time of the submerged nozzle closing device;

[0059] S3: By predicting the stop time of the submerged nozzle closing device, it is judged whether the alarm condition is triggered. If it is triggered, the nozzle is closed.

[0060] In the step S2, using the linear prediction method in the sensorless environment to monitor the slag entrainment process in real time and obtaining the linear correlation between the indirect data and the slag entrainment state, the specific process is as follows:

[0061] S21: Data decomposition;

[0062] For the data sequence of the change of molten steel weight over time, using the moving average kernel method, the data sequence is decomposed into two key components, namely the trend component and the remainder component. The trend component is used to accurately capture the overall change trend of the molten steel weight over a long time span, reflecting the long-term trend of the change of molten steel weight. The remainder component covers the relatively short-term fluctuation information and the existing seasonal change characteristics, and is used to reveal the change law and periodic fluctuation situation of the data in a short time period. Specifically:

[0063] Let the original data sequence be where y t represents the molten steel weight at time t. Through the moving average kernel method, it is decomposed into the following two parts:

[0064] y t =T t +R t (1)

[0065] In the formula, T t represents the trend component, which is used to describe the long-term change trend, and R t represents the remainder component, which is used to describe the short-term fluctuations and seasonal variations. The trend component T t is obtained through a smoothing operation, and the specific calculation method is as follows:

[0066]

[0067] where: k is a parameter of the window width, representing the time span involved in the smoothing process, 2k + 1 is the length of the smoothing window, that is, the width of the moving average kernel, and y t+i represents the value of the original data at time t + i. When t + i exceeds the sequence range, boundary processing methods such as the mirror method or interpolation method can be used. The significance of this formula is to calculate the mean value of the original data using a sliding window, thereby eliminating short-term fluctuations;

[0068] The remainder component R t is obtained by subtracting the trend component from the original data:

[0069] R t = y t - T t (3)

[0070] The remainder component reflects the part of the data that is not captured by the trend component, including short-term fluctuations and possible periodic characteristics;

[0071] S22: Linear layer application

[0072] After the data sequence decomposition is completed, an independent single-layer linear layer is assigned to the decomposed trend component and remainder component respectively. For the trend component, its corresponding linear layer obtains a set of weight parameters through deep learning and transforms the trend information into a feature representation form that makes a key contribution to the prediction of the stopping time of the slag-dropping and nozzle-closing device. For the remainder component, its corresponding linear layer also extracts the feature information closely related to the device stopping time contained therein through a linear transformation operation; through this separate processing method, the valuable information contained in different components of the data can be fully mined, avoiding information confusion and loss;

[0073] S23: Add the features obtained after being processed by the two linear layers to obtain the final prediction result.

[0074] The present invention performs an addition operation on the features obtained after processing through two linear layers to obtain the final prediction result. This processing method ingeniously integrates the information provided by the trend component and the remainder component, enabling the prediction result to comprehensively consider the long-term trend and short-term fluctuation characteristics of the data. By fully leveraging the respective advantages of these two components, it is possible to more accurately predict the stopping time of the slag-dropping and nozzle-closing device, providing a reliable basis for relevant decisions in the industrial production process. In the actual application scenario, it is possible to timely and accurately predict the closing time of the slag-dropping nozzle based on the real-time collected data of the molten steel weight change, helping the operator to make preparations in advance, optimize the production process, and improve production efficiency and product quality.

[0075] In step S22, for the trend component and the remainder component, their corresponding linear layers learn a set of weight parameters through deep learning, and convert the trend information and the remainder information into a feature representation form that makes a key contribution to the prediction of the stopping time of the slag-dropping and nozzle-closing device. Specifically:

[0076] Let the time series data be t = [t1, t2,... t n , and the corresponding molten steel weight change data be ω = [ω1, ω2,... ω n . The stopping time of the slag-dropping and nozzle-closing device predicted by the linear prediction method is calculated through the following linear combination:

[0077]

[0078] In the formula, α, β, γ, and δ are a set of weight parameters learned through deep learning. α controls the influence degree of the product term of the molten steel weight and time on the prediction result, reflecting the interaction factor between the weight change and time; β mainly reflects the contribution of the molten steel weight change alone to the prediction result; γ focuses on the influence of the time factor alone; δ, as a bias term, is used to adjust the overall prediction value to adapt to different working conditions and data distribution characteristics.

[0079] To optimize the weight parameters α, β, γ, and δ, the mean square error MSE is used as the loss function. The calculation formula of MSE is as follows:

[0080]

[0081] In the formula, m is the number of training samples, is the predicted device stopping time of the jth training sample, and T i is the actual device stopping time of the jth training sample;

[0082] Given the actual slag-dropping and nozzle-closing time as T, by minimizing the MSE, the parameters α, β, γ, and δ are continuously adjusted to make the predicted value As close as possible to the actual equipment stop time T.

[0083] In the step S22, the network structure of deep learning includes an input layer, a hidden layer, and an output layer, where:

[0084] The input layer has two nodes, which respectively receive the molten steel weight change data and time data. These data can be obtained through high-precision weighing sensors installed on the ladle support structure and synchronous time acquisition devices. The accuracy of the weighing sensor can reach ±0.1%, which can accurately measure the tiny changes in the molten steel weight. The accuracy of the time acquisition device is in milliseconds, ensuring the accuracy and synchronization of the time data. The input layer directly transmits these two data to the hidden layer;

[0085] The hidden layer, that is, the linear transformation layer, consists of a group of linear neurons and performs a linear transformation operation. For the input molten steel weight data ω and time data t, the hidden layer performs a linear transformation through the weight matrix W and the bias vector b. The calculation formula for the output vector h is:

[0086]

[0087] In the formula, W is a 2×p weight matrix, p is the number of neurons in the hidden layer, b is a p-dimensional bias vector. The role of the hidden layer is to initially integrate and extract features from the input molten steel weight and time data, and mine the potential information in the data through linear transformation to prepare for outputting an accurate predicted value of the equipment stop time;

[0088] Due to the adoption of linear transformation in the present invention, the calculation speed is fast and can meet the requirements of low latency;

[0089] The output layer receives the output h of the hidden layer and calculates the final prediction result through a linear weight vector ω0 and a bias term b0 The calculation formula is:

[0090]

[0091] In the formula, w0 is a p-dimensional weight vector. The output layer performs the final linear combination of the feature information extracted by the hidden layer and outputs the predicted value of the tapping and tundish nozzle closing equipment stop time.

[0092] For the parameter initialization of the network structure of deep learning in the step S22, among the weight parameters, including the weight matrix W of the hidden layer and the weight vector w0 of the output layer, a random initialization method is adopted, and the initial weight values are randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.01. The bias terms, including the bias vector b of the hidden layer and the bias term b0 of the output layer, are initialized to 0.

[0093] The initialization method of the present invention provides diverse starting points for the training of the method, enabling the method to gradually adjust the weights and biases according to the characteristics and distribution of the molten steel weight and time data during the subsequent training process, so as to learn the optimal linear relationship between the data and the equipment stop time.

[0094] For the parameter optimization of the deep learning network structure in step S22, an optimization algorithm based on gradient descent is used to update the parameters. By calculating the gradient of the loss function with respect to each parameter and according to the direction and magnitude of the gradient, the parameters are updated at a certain learning rate. The update formula for the weight parameters is as follows:

[0095]

[0096] In the formula, α is the learning rate. At the initial stage of training, a relatively large learning rate of 0.01 is selected. Then, as the training progresses, a learning rate decay strategy is adopted to gradually reduce the learning rate. Every 10 training epochs, the learning rate decays to 0.9 times the original. At the same time, to prevent overfitting, regularization techniques such as L2 regularization are used, that is, the sum of the squares of the weight parameters is added to the loss function to constrain the weights, so that the method has better generalization ability. After adding L2 regularization, the loss function becomes:

[0097]

[0098] In the formula, λ is the regularization parameter, which is used to control the strength of regularization.

[0099] By introducing the regularization parameter λ, the present invention can avoid the overfitting problem while ensuring the fitting ability of the method, and improve the performance and stability of the method in practical applications.

[0100] Based on the real-time collected molten steel weight and time data, the present invention quickly and accurately predicts the time to close the nozzle for slagging by determining the linear relationship between the change trend of the molten steel weight and time. Using a simple and efficient linear method, it meets the requirements of industrial production for low-latency and high-precision prediction. The present invention can effectively achieve low-latency and high-precision prediction of the equipment stop time for closing the nozzle for slagging only relying on the molten steel weight change data and time data, providing reliable technical support for the sensorless slagging monitoring scheme.

[0101] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A sensorless method for predicting the closing of a molten steel slag nozzle, characterized in that: The following steps are involved: S1: Real-time monitoring of slag discharge, collecting various indirect data in the sensorless environment during steel production, including furnace number, ladle weight, weight of ladle when it arrives at a specific position, weight of ladle when it leaves a specific position, weight of ladle when empty, net weight of molten steel, ladle temperature, maximum ladle temperature, minimum ladle temperature, ladle slag thickness, ladle pouring start time, ladle pouring end time. Through the indirect data, a data sequence of molten steel weight changes over time is obtained; S2: Use the linear prediction method in a sensorless environment to monitor the slag-laying process in real time, obtain the linear correlation between the indirect data and the slag-laying state, reconstruct the signal characteristics of the slag-laying process, realize real-time tracking of the slag-laying situation, and then predict the stop time of the slag-laying nozzle closing equipment; S3: By predicting the stop time of the slag discharge and water nozzle closing equipment, it is determined whether the alarm condition is triggered. If triggered, the water nozzle is closed.

2. According to claim 1, a sensorless method for predicting the closing of a molten steel slag nozzle is characterized in that: In step S2, the slag-laying process is monitored in real time using a linear prediction method in a sensorless environment to obtain a linear correlation between indirect data and the slag-laying state. The specific process is as follows: S21: data decomposition; For the data of changes in molten steel weight over time, the moving average kernel method is used to decompose the data series into two key components: trend component and remainder component. The trend component is used to accurately capture the overall change trend of molten steel weight over a longer time span, reflecting the long-term trend of molten steel weight changes. The remainder component covers relatively short-term fluctuation information and seasonal change characteristics, and is used to reveal the change law and periodic fluctuation of data in a shorter time period, specifically: Assume the original data sequence is where y t It represents the weight of molten steel at time t, and is decomposed into the following two parts by the moving average kernel method: y t =T t +R t (1) Where, T t Represents the trend component, which is used to describe the long-term trend. t Represents the remainder component, which is used to describe short-term fluctuations and seasonal changes, and the trend component T t It is obtained through smoothing operation. The specific calculation method is: Where: k is the parameter of the window width, which represents the time span involved in the smoothing process, 2k+1 is the length of the smoothing window, that is, the width of the moving average kernel, and y t+i It represents the value at time t+i in the original data. When t+i exceeds the sequence range, boundary processing methods such as mirroring or interpolation can be used. The significance of this formula is to use a sliding window to calculate the average of the original data, thereby eliminating short-term fluctuations. Remainder component R t Subtracting the trend component from the original data gives: R t =y t -T t (3) The residual component reflects the part of the data that is not captured by the trend component, including short-term fluctuations and possible cyclical characteristics; S22: Linear layer application After the data sequence is decomposed, an independent linear layer is provided for the trend component and the remainder component obtained by the decomposition. For the trend component, the corresponding linear layer transforms the trend information into a feature representation that has a key contribution to the prediction of the stop time of the slag closing nozzle equipment through deep learning of a set of weight parameters. For the remainder component, the corresponding linear layer also extracts the feature information that is closely related to the equipment stop time through linear transformation operations. S23: Add the features obtained after processing by the two linear layers to obtain the final prediction result.

3. According to claim 2, a sensorless method for predicting the closing of a molten steel slag nozzle is characterized in that: In step S22, for the trend component and the remainder component, the corresponding linear layer converts the trend information and the remainder information into feature representation forms that have a key contribution to the stop time prediction of the slag closing nozzle equipment through deep learning of a set of weight parameters, specifically: Assume that the time series data is t=[t1, t2, ...t n ], the corresponding molten steel weight change data is ω=[ω1,ω2,...ω n ], the stopping time of the slag closing nozzle equipment predicted by the linear prediction method Computed as the following linear combination: Where α, β, γ and δ are a set of weight parameters through deep learning. α controls the influence of the product of molten steel weight and time on the prediction result, reflecting the interaction between weight change and time; β mainly reflects the contribution of molten steel weight change to the prediction result; γ focuses on the influence of time factor alone; δ is used as a bias term to adjust the overall prediction value to adapt to different working conditions and data distribution characteristics.

4. According to claim 3, a sensorless method for predicting the closing of a molten steel slag nozzle is characterized in that: In order to optimize the α, β, γ and δ weight parameters, the mean square error MSE is used as the loss function. The MSE calculation formula is as follows: In the formula, m is the number of training samples, is the predicted equipment downtime of the jth training sample, T i is the actual equipment stop time of the jth training sample; Given the actual closing time of the slag discharge nozzle as T, by minimizing the MSE, the parameters α, β, γ and δ are continuously adjusted to make the predicted value As close as possible to the actual device stopping time T.

5. According to claim 2, a sensorless method for predicting the closing of a molten steel slag nozzle is characterized in that: The network structure of the deep learning in step S22 includes an input layer, a hidden layer and an output layer, wherein: The input layer has two nodes, which receive the molten steel weight change data and time data respectively. These data can be obtained through the high-precision weighing sensor installed on the ladle support structure and the synchronous time acquisition device. The weighing sensor has an accuracy of up to ±0.1%, which can accurately measure the slight changes in the weight of the molten steel. The accuracy of the time acquisition device is at the millisecond level, ensuring the accuracy and synchronization of the time data. The input layer directly passes these two data to the hidden layer; The hidden layer, i.e., the linear transformation layer, is composed of a group of linear neurons and performs linear transformation operations. For the input molten steel weight data ω and time data t, the hidden layer performs linear transformation through the weight matrix W and the bias vector b. The calculation formula of the output vector h is: Where W is a 2×p weight matrix, p is the number of neurons in the hidden layer, and b is a p-dimensional bias vector. The hidden layer is used to perform preliminary integration and feature extraction on the input molten steel weight and time data, and to mine the potential information in the data through linear transformation, so as to prepare for outputting accurate equipment stop time prediction values. The output layer receives the output h of the hidden layer and calculates the final prediction result through a linear weight vector ω0 and a bias term b0 The calculation formula is: In the formula, w0 is a p-dimensional weight vector. The output layer performs the final linear combination on the feature information extracted by the hidden layer and outputs the predicted value of the stop time of the slag closing nozzle equipment.

6. A sensorless method for predicting closure of molten steel slag nozzle according to claim 5, characterized in that: In the step S22, the parameters of the deep learning network structure are initialized, wherein the weight parameters, including the weight matrix W of the hidden layer and the weight vector w0 of the output layer, are randomly initialized by randomly sampling from a normal distribution with a mean of 0 and a standard deviation of 0.01 to generate initial weight values, and the bias terms, including the bias vector b of the hidden layer and the bias term b0 of the output layer, are initialized to 0.

7. A sensorless method for predicting closure of molten steel slag nozzle according to claim 5, characterized in that: The parameter optimization of the deep learning network structure in step S22 adopts an optimization algorithm based on gradient descent to update the parameters. The gradient of the loss function for each parameter is calculated, and the parameters are updated at a certain learning rate according to the direction and size of the gradient. The update formula of the weight parameter is as follows: In the formula, α is the learning rate. At the beginning of training, a relatively large learning rate of 0.01 is selected. Then, as the training progresses, the learning rate decay strategy is adopted to gradually reduce the learning rate. After every 10 training cycles, the learning rate decays to 0.9 times the original value. At the same time, in order to prevent overfitting, regularization techniques such as L2 regularization are used, that is, the square sum of weight parameters is added to the loss function to constrain the weights. The loss function after adding L2 regularization becomes: Where λ is the regularization parameter, which is used to control the strength of regularization.

Images