Audio-based Monitoring Method and System for Slag Splashing State

Abstract

The present invention relates to an audio-based monitoring method and system for the slag splashing state. After the slag splashing in the converter starts, real-time data required for model calculation during the slag splashing process is collected; the collected real-time data is analyzed and processed; a slag starting threshold is set, and the temporal characteristics of the sound intensity during slag starting are observed and learned to establish an automatic detection model for the slag starting point to determine the slag starting point; a slag splashing warning point detection model is established to determine the warning point; a slag splashing end point detection model is established to determine the slag splashing end point, realizing the automatic prediction of the slag splashing end point. The present invention monitors the audio signal during the slag splashing process of the converter in real time, extracts the audio temporal characteristics, and combines machine learning algorithms and end point prediction technologies to realize the detection of the slag starting point and the accurate prediction of the slag splashing end point. This method improves the production efficiency of the slag splashing process in the converter, reduces the cost of manual intervention, avoids the negative impact on the converter life due to too long slag splashing time, and provides strong technical support for the continuous progress of the steelmaking industry.

Description

Technical Field

[0001] The present invention relates to the technical field of converter steelmaking, and particularly to an audio-based monitoring method and system for slag splashing state. Background Art

[0002] Slag splashing for converter lining protection is a major advancement in converter lining protection technology. The application of converter slag splashing technology can not only effectively improve the protection effect of the furnace lining, extend the service life of the converter, but also significantly reduce the consumption of refractory materials, which is of great significance for improving the economy and reliability of steelmaking production.

[0003] At present, steel enterprises at home and abroad have adopted a series of measures and means such as static control models to optimize slag splashing operations. However, these methods cannot achieve real-time monitoring of the entire slag splashing process and automatic prediction of the end point of converter slag splashing. Traditionally, relying on the personal experience of operators increases the cost of manual intervention and labor intensity. From the statistical data of the end point time of converter slag splashing in steel plants, the proportion of furnace batches with too long or too short slag splashing time due to improper operation is larger. On the one hand, too long slag splashing time will increase the erosion of the inner wall of the converter and shorten the service life of the converter. To extend the service life of the converter, a large amount of brick patching is required to maintain the furnace condition, wasting a lot of manpower and reducing production efficiency. On the other hand, too short slag splashing time may lead to poor slag quality and affect the final quality of the product. According to on-site experience and industry knowledge, the slag starting time during the slag splashing process has a crucial impact on the slag splashing duration, and the change in sound can reflect the slag state in real time.

[0004] Therefore, how to achieve automatic prediction of the end point of converter slag splashing has become a key challenge in slag splashing technology for converter lining protection. Summary of the Invention

[0005] To solve at least one of the above technical problems, the present invention provides an audio-based monitoring method for slag splashing state, including:

[0006] Collecting historical data during the slag splashing process, including audio data, lance height data, and slag adjusting agent data; and annotating the slag starting point, warning point, and slag splashing end point for the historical data;

[0007] Establishing a slag starting point detection model according to the slag starting point and the historical data before it;

[0008] Establishing a slag splashing warning point detection model according to the historical data from the slag starting point to the warning point;

[0009] Establishing a slag splashing end point detection model according to the historical data from the warning point to the slag splashing end point;

[0010] Collect the real-time data of the slag splashing process, including audio data, lance height data, and slag modifier data. First, input the data into the slag starting point detection model to predict the slag starting point; then, input the real-time data after the slag starting point into the slag splashing warning point detection model to predict the warning point; then, input the real-time data after the warning point into the slag splashing end point detection model to predict the slag splashing end point.

[0011] Further, collect the historical data of the slag splashing process, including: generating an audio intensity change curve reflecting the slag splashing process based on the audio data, lance height data, and slag modifier data; specifically:

[0012] Extract the time-domain and frequency-domain features of the audio data, and screen the audio channels that can effectively describe the slag splashing process from multiple frequency bands;

[0013] According to the selected audio channel values and combined with the data sampling rate, calculate the average value of the audio intensity data per second as the instantaneous audio intensity value at the current moment ;

[0014] Calculate the lance position compensation value and the slag modifier compensation value respectively according to the lance height data and the slag modifier data to compensate the instantaneous audio intensity value and obtain the compensated instantaneous audio intensity value , so as to generate an audio intensity change curve.

[0015] Further, the lance position compensation value , is calculated by the formula ), where, is the lance position compensation coefficient, is the instantaneous lance position height, and the initial lance position value ; the slag modifier compensation value , is calculated by the formula , where, is the slag modifier compensation coefficient, Δm(t) is the instantaneous change amount of the slag modifier, is the real-time correction coefficient, Collect the actual addition amount during tapping; the compensated instantaneous audio intensity value , is calculated by the formula .

[0016] Further, after calculating that the compensated instantaneous audio intensity value is , it also includes:

[0017] S141: Set a moving window to calculate the sliding average value of the instantaneous audio intensity value ; ;

[0018] S142: Set the exponentially weighted moving average weighting factor ewmaAlpha to assign a higher weight to the instantaneous audio intensity value at the current moment;

[0019] S143: Calculate the final instantaneous sound intensity value based on the instantaneous sound intensity value, the sliding average value, and the exponentially weighted moving average weighting factor , and then generate a sound intensity change curve

[0020] Furthermore, the final instantaneous sound intensity value , is calculated through the formula for calculation

[0021] Furthermore, the steps for establishing a slag point detection model include

[0022] According to the historical data at and before the slag starting point, statistically analyze the slag starting parameter thresholds for different furnace batches, including any one or more of the following: lance height change, sound intensity magnitude and mutation, sound intensity curve trend slope range, movement of frequency peak, amplitude change of the sound waveform in the time domain, and slag starting duration

[0023] Construct a slag point detection model with the slag starting parameter thresholds and historical data as inputs and the slag starting point as the output, for learning the first sound intensity time series feature

[0024] Through multi-furnace batch learning with historical data, fine-tune the parameter settings of the slag point detection model to obtain a trained slag point detection model for automatically predicting the slag starting point

[0025] Furthermore, the steps for establishing a slag splashing warning point detection model include

[0026] Statistically analyze the instantaneous sound intensity values from the slag starting point to the warning point for different furnace batches, and construct an instantaneous sound intensity time series data set

[0027] Construct a supervised deep learning time series prediction model, input the instantaneous sound intensity time series data set into the time series prediction model to learn the second audio time series feature, and obtain a trained time series prediction model as the slag splashing warning point detection model for automatically predicting the warning point

[0028] Furthermore, the steps for establishing a slag splashing end point detection model include

[0029] According to the historical data from the warning point to the slag splashing end point, statistically analyze the end point parameter thresholds for different furnace batches, including any one or more of the following: slag splashing end point sound intensity magnitude and increase / decrease speed, sound intensity curve trend slope range, frequency distribution and movement of peak value, amplitude change and energy distribution of the sound waveform in the time domain, and slag splashing duration

[0030] Construct a slag splashing end point detection model with the end point parameter thresholds and historical data as inputs and the slag splashing end point as the output, for learning the third sound intensity time series feature

[0031] Divide historical data into different slag splashing modes, conduct learning for different slag splashing modes, fine-tune the parameter settings of the slag splashing endpoint inspection model, obtain the trained slag splashing endpoint detection model, and automatically predict the slag splashing endpoint.

[0032] On the other hand, the present invention also provides an audio-based slag splashing state monitoring system, including: an audio data acquisition device, a multi-band audio analyzer, a PLC data acquisition device, and an audio slag melting industrial control computer, which are used to implement any of the above-mentioned slag splashing state monitoring methods.

[0033] Furthermore, the audio data acquisition device includes an audio sensor, which is arranged outside the converter fireproof wall and is used to collect the audio signal of nitrogen impacting the slag during slag splashing.

[0034] The present invention provides an audio-based slag splashing state monitoring method and system to achieve the full-process monitoring and endpoint prediction of the slag splashing process, so as to solve the technical problems that in the process of slag splashing for furnace lining protection, it overly relies on the experience and intuition of operators, resulting in too long or too short slag splashing time, affecting the converter life and production efficiency, poor slag splashing effect, and high manual intervention cost. Description of the Drawings

[0035] Figure 1 It is a schematic flow chart of an embodiment of an audio-based slag splashing state monitoring method of the present invention;

[0036] Figure 2 It is a schematic structural diagram of an embodiment of a prediction model of an audio-based slag splashing state monitoring method of the present invention;

[0037] Figure 3 It is a schematic structural diagram of an embodiment of an audio-based slag splashing state monitoring system of the present invention. Detailed Embodiments

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0039] It should be noted that if there are directional indications involved in the embodiments of the present invention, such as up, down, left, right, front, back..., then such directional indications are only used to explain the relative positional relationship, movement conditions, etc. between components in a certain specific posture. If the specific posture changes, the directional indications will also change accordingly. In addition, if there are descriptions such as "first, second", "S1, S2", "step one, step two", etc. involved in the embodiments of the present invention, then such descriptions are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features or indicating the execution order of the method, etc. Those skilled in the art can understand that all those that do not violate the inventive concept of the invention should be included in the protection scope of the present invention.

[0040] As Figure 1 shown in -2, the present invention provides an audio-based monitoring method for the slag splashing state, including:

[0041] S1: Collect historical data during the slag splashing process, including audio data, lance height data, and slag adjusting agent data; and label the slag starting point , warning point and slag splashing end point ;

[0042] Specifically, for step S1, as Figure 3 shown, an audio-based monitoring system for the slag splashing state is given, which may include: an audio data acquisition device A, a multi-band audio analyzer B, a PLC data acquisition device C, and an audio slag melting industrial control computer D.

[0043] The audio data acquisition device A may include an audio sensor, a sound collection tube, etc.; the intelligent audio sensor is installed outside the converter fire wall, and the sound collection tube in the intelligent purging protection box collects the audio signal during slag splashing, such as the audio signal emitted by nitrogen impacting the slag; the multi-band audio analyzer B is responsible for preprocessing the collected original audio signal, converting the signal into channel data of multiple frequency bands, reducing noise interference, and thus extracting useful slag splashing information. More specifically, the audio data may include audio intensity values, and the optional unit sampling rate is 8192. The entire audio signal of nitrogen impacting the slag during slag splashing is collected through the audio acquisition device, and the original audio intensity values after frequency division are obtained in real time by using the multi-band audio analyzer. The central frequencies where different audio intensity values are located include: 107HZ, 130HZ, 155HZ, 190HZ, 225HZ, 267HZ.

[0044] The PLC data acquisition device C, optionally including a liquid level gauge, a weight sensor, etc., is responsible for accurately collecting and transmitting the lance height data and slag adjusting agent data in the converter, and can ensure the real-time monitoring and accurate recording of these key process parameters, providing data support for the control and optimization of the slag splashing process. More specifically, the lance height data can be the lance position height, that is, the height difference between the oxygen lance and the furnace bottom; the slag adjusting agent data can optionally include the type of slag adjusting agent, the addition time and weight of each slag adjusting agent; for example, the types of slag adjusting agents include: limestone, dolomite, iron oxide scale, lime, return ore, iron powder pellets.

[0045] The audio slag splashing industrial control computer D is the core of the entire system, responsible for real-time monitoring and analyzing data such as audio, implementing this monitoring method, displaying the detected slag starting point, and automatically predicting the warning point and the slag splashing end point according to the analysis results. Through the intelligent processing and analysis of the industrial control computer, operators can more intuitively understand the state of the slag splashing process, and thus make more accurate operation decisions.

[0046] More specifically, for historical data, it is optional to observe and analyze at least 100 furnaces of historical data covering all shifts, and based on the change characteristics of data such as audio and lance height at each stage, combined with the experience of on-site operators, flame images, etc., mark the actual slag starting time as the slag starting point; mark the time when the lance position reaches the lowest point at the end of slag splashing, that is, the slag splashing warning start time , as the warning point; mark the time corresponding to the slag splashing end point , as the slag splashing end point; after marking, the historical data is divided into three parts.

[0047] Preferably, step S1 further includes: generating an audio intensity change curve reflecting the slag splashing process according to the audio data, lance height data and slag adjusting agent data;

[0048] Specifically, for the subsequent steps S2 - S4, to better learn the audio time series characteristics, step S1 can also optionally perform one or more processing processes such as frequency division data analysis, instantaneous audio intensity noise reduction processing, curve dynamic compensation processing, and audio intensity curve weighting processing on the acquired original audio data according to the audio data, lance height data and slag adjusting agent data, so as to more effectively and accurately describe the audio change.

[0049] Preferably, it includes:

[0050] S11: Extract the time-domain and frequency-domain features of the audio data, and screen out the audio channels that can effectively describe the slag splashing process from multiple frequency bands. Specifically, the frequency division data analysis function needs to observe the time-domain and frequency-domain features of the sound intensity data in the slag splashing stage, and screen out the audio channels that can effectively and accurately describe the slag splashing process from multiple frequency bands. The data of these channels will be used as the basis for drawing the sound intensity curve. Specifically, when the slag starts to splash, it is affected by the slag temperature and the final slag composition, and the sound is relatively soft. As the high-pressure nitrogen cools the slag, the lance position drops, the slag polymerizes, and the thickness decreases, the sound in the furnace will gradually increase to the maximum.

[0051] S12: According to the selected audio channel values and combined with the data sampling rate, calculate the average value of the sound intensity data per second as the instantaneous sound intensity value at the current moment. Specifically, this step: the instantaneous sound intensity noise reduction processing function is to reduce the influence of background noise on the signal and enhance the stability of the signal.

[0052] S13: Calculate the lance position compensation value and the slag conditioner compensation value respectively according to the lance height data and the slag conditioner data to compensate the instantaneous sound intensity value , and the compensated instantaneous sound intensity value is to generate the sound intensity change curve.

[0053] Specifically, the curve dynamic compensation function needs to calculate the lance position compensation value and the slag conditioner compensation value in real time according to the lance position change and the time and weight of the slag conditioner addition during slag splashing. Optionally, on the one hand, the real-time lance position compensation value is calculated by the formula ), where is the lance position compensation coefficient, is the instantaneous lance position height, and the initial lance position is adjusted according to the three sub-mode setting strategies of the hearth bottom height (low hearth bottom, normal hearth bottom, and high hearth bottom). On the other hand, the real-time slag conditioner compensation value is calculated by the formula , where is the slag conditioner compensation coefficient, Δm(t) is the instantaneous change amount of the slag conditioner. For the newly added slag conditioner modification function during the tapping process, the actual addition amount during tapping is collected, and the corresponding real-time correction coefficient is set. This correction coefficient is related to the type of slag conditioner, the specific furnace situation, the environment, etc. According to the above formula, the compensated instantaneous sound intensity value is .

[0054] More preferably, after the compensated instantaneous sound intensity value obtained in step S13 is , it further includes:

[0055] S141: Set a moving window to calculate the instantaneous sound intensity value of the sliding average value ;

[0056] S142: Set the exponentially weighted moving average weighting factor ewmaAlpha to give the instantaneous sound intensity value at the current moment a higher weight;

[0057] S143: Calculate the final instantaneous sound intensity value based on the instantaneous sound intensity value, the sliding average value, and the exponentially weighted moving average weighting factor , and then generate a sound intensity change curve.

[0058] Specifically, to further utilize the sound intensity time series data and extract its short-term and long-term trend characteristics, which helps to smooth the random fluctuations of the slag splashing sound intensity in the short term and helps to identify potential changes in the sound intensity trend, meeting the requirements of the trend curve analysis for audio-guided slag splashing furnace lining protection. Optionally, first, set a moving window with a size within 5 to 10 and calculate the sliding average value of the instantaneous sound intensity ; Second, set the exponentially weighted moving average weighting factor ewmaAlpha in the range of 0.8 to 0.9 to give a higher weight to the sound intensity data at the current moment; finally, according to the formula , calculate the final instantaneous sound intensity value for describing the slag splashing audio trend curve.

[0059] S2: Establish a slagging point detection model based on the slagging point and historical data before it;

[0060] Specifically, optionally, based on the slagging point and historical data before it, learn the sound intensity time series characteristics during this period through a neural network method to establish a slagging point detection model; the specific model structure can adopt any method of neural network, such as a convolutional neural network, etc.

[0061] Preferably, step S2 includes:

[0062] S21: According to the slagging point and historical data before it, statistically analyze the slagging parameter thresholds of the slagging points of different furnace campaigns, including but not limited to: any one or more of the lance height change, sound intensity magnitude and mutation, sound intensity curve trend slope range, movement of the frequency peak, amplitude change of the sound waveform in the time domain, and slagging duration;

[0063] S22: Construct a slagging point detection model with the slagging parameter thresholds and historical data as inputs and the slagging point as the output to learn the first sound intensity time series characteristics;

[0064] S23: Through multi-furnace campaign learning with historical data, fine-tune the parameter settings of the slagging point detection model to obtain a trained slagging point detection model for automatically predicting the slagging point.

[0065] Preferably, step S2 further includes screening the historical data according to the slagging parameter threshold, for example, eliminating abnormal furnace batches including overoxidized slag, and then performing multi-furnace learning.

[0066] S3: Establish a slag splashing warning point detection model based on the historical data from the slagging point to the warning point;

[0067] Specifically, optionally, according to the historical data from the slagging point to the warning point, the temporal characteristics of the sound intensity in this period can also be learned through a neural network method to establish a slag splashing warning point detection model; the specific model structure can adopt any method of neural network.

[0068] Preferably, step S3 includes:

[0069] S31: Statistically analyze the instantaneous sound intensity values from the slagging point to the warning point of different furnace batches to construct an instantaneous sound intensity time series data set;

[0070] S32: Construct a supervised deep learning time series prediction model, input the instantaneous sound intensity time series data set into the time series prediction model, learn the second audio time series characteristics, and obtain the trained time series prediction model as the slag splashing warning point detection model to automatically predict the warning point;

[0071] Specifically, optionally, observe and analyze at least 100 historical furnace batches covering all shifts, and mark the time when the lance position reaches the lowest range at the end of slag splashing , as the start time of slag splashing warning. For these furnace batches, the slagging point to the warning point the instantaneous sound intensity values during this period are used as the time series data set for training the slag splashing warning point detection model, and a supervised deep learning time series prediction method is designed to automatically learn the corresponding audio time series characteristics , and extract the audio time series characteristics at the end of slag splashing to dynamically detect the start signal of slag splashing warning.

[0072] S4: Establish a slag splashing end point detection model based on the historical data from the warning point to the slag splashing end point;

[0073] Specifically, similar to steps S2 and S3, optionally, according to the historical data from the warning point to the slag splashing end point, the temporal characteristics of the sound intensity in this period can be learned through a neural network method to establish a slag splashing end point detection model; the specific model structure can adopt any method of neural network, such as a convolutional neural network, etc.

[0074] Preferably, step S4 includes:

[0075] S41: According to the historical data from the warning point to the slag splashing end point, count the end point parameter thresholds for different furnace heats, including but not limited to any one or more of the following: the magnitude and increase / decrease speed of the sound intensity at the slag splashing end point, the slope range of the sound intensity curve trend, the frequency distribution and the movement of the peak value, the amplitude change and energy distribution of the sound waveform in the time domain, and the slag splashing duration.

[0076] S42: Construct a slag splashing end point detection model with the end point parameter thresholds and historical data as inputs and the slag splashing end point as the output, for learning the third sound intensity time series features.

[0077] S43: Divide the historical data into different slag splashing modes, conduct learning for different slag splashing modes, fine-tune the parameter settings of the slag splashing end point inspection model, and obtain the trained slag splashing end point detection model to automatically predict the slag splashing end point.

[0078] Specifically, collect the instantaneous sound intensity values in real time from the start to the current moment, and extract the sound intensity features under different slag splashing modes (lower furnace bottom, rising furnace bottom, protecting the molten pool). Then, combine the on-site operation experience and the audio change characteristics at the end of slag splashing to set the end point thresholds. The end point thresholds that need to be fine-tuned for different converters. Finally, predict the time corresponding to the slag splashing end point based on the satisfaction of the sound intensity time series features and the end point thresholds. 。

[0079] S5: Collect the real-time data during the slag splashing process, including audio data, lance height data, and slag adjusting agent data. First, input it into the slag starting point detection model to predict the slag starting point; then input the real-time data after the slag starting point into the slag splashing warning point detection model to predict the warning point; then input the real-time data after the warning point into the slag splashing end point detection model to predict the slag splashing end point.

[0080] Specifically, monitor in real time, collect the factual data during the slag splashing process. First, input it into the previously trained slag starting point detection model, combine the slag starting threshold, and after predicting the slag starting point, input the subsequent data into the previously trained slag splashing warning point detection model. After predicting the warning point, then input the subsequent data into the previously trained slag splashing end point detection model, combine the end point threshold, to predict the slag splashing end point.

[0081] In this embodiment, a method for monitoring the slag splashing state based on audio of the present invention is given. The converter monitoring is divided into three stages: the slag starting stage, the early warning stage, and the end point stage. It mainly includes a slag starting point detection model, a slag splashing early warning point detection model, and a slag splashing end point detection model. It also involves multiple key process parameters such as the original audio corresponding to the slag splashing time, the lance height, limestone, dolomite, mill scale, lime, return fines, and the addition time and weight of iron powder pellets. 1. The slag starting point is a key turning point in the slag splashing process. Its accurate detection is crucial for ensuring the accuracy of the slag splashing end point prediction. Because before the slag starting point, there is more noise and interference in the audio data, and it is difficult to extract the audio time series features. After the slag starting point, the quality of the audio time series features is significantly improved and becomes clearer and more useful. Setting the slag starting point detection model before this helps to obtain more accurate audio time series features, reduce data noise interference, and improve the quality and accuracy of feature extraction. The optimization of these features helps to improve the prediction accuracy and robustness of the entire slag splashing end point prediction model, and further provides more accurate operation guidance for operators, thereby optimizing the slag splashing process and improving production efficiency. In the preferred solution, through the precise detection of the slag starting point and the analysis of the audio time series features, the slag splashing sound intensity change curve can more accurately reflect the state of the slag splashing process, quickly respond to the changes in the slag splashing process, improve the stability and production efficiency of the slag splashing process, reduce manual observation and judgment, and further improve the end point prediction accuracy. 2. The early warning point is the time point when the lance position reaches the lowest range at the end of the slag splashing. After that, the lance position no longer changes significantly. Before the end point prediction, adding an early warning point and constructing a slag splashing early warning detection model helps to reduce the end point misjudgment during the lance lowering process and enhance the robustness of the model. 3. The slag splashing end point is the slag splashing termination point. By analyzing the audio time series features after the early warning point and establishing a slag splashing end point detection model to determine the slag splashing end point, the hit rate of the slag splashing end point prediction can be improved, unnecessary impact and damage to the furnace lining can be avoided, thereby extending the service life of the furnace lining, reducing labor costs and operation risks.

[0082] In summary, the present invention provides a method and device for automatically predicting the slagging start and slag splashing end points based on audio time-series features, realizing the whole-process monitoring and end-point prediction of the slag splashing process, so as to solve the technical problems that in the process of slag splashing for converter lining protection, excessive reliance on the experience and intuition of operators leads to too long or too short slag splashing time, affecting the converter life and production efficiency, poor slag splashing effect, and high manual intervention cost. It should be noted that the first key of the present invention is to divide the slag splashing process into three stages. In the first stage, the slagging start point is determined to bypass the audio data with large previous noise interference and difficult feature extraction; in the second stage, the early warning point is determined to give an early warning and reduce false positives; in the third stage, the slag splashing end point is determined to determine the end of slag splashing, avoiding unnecessary impact and damage to the furnace lining, thereby prolonging the service life of the furnace lining, reducing labor costs and operation risks; the second key lies in how to process the audio data to further facilitate the subsequent extraction of audio time-series features; as for the specific model structure, any other learning algorithms such as random forest, support vector machine, and neural network can be used to train the model, and deep learning technologies such as convolutional neural network and recurrent neural network are used to process the time-series data to extract the complex non-linear relationships between various features and learn the first, second, and third audio data features.

[0083] In a preferred embodiment, in step S1, several preferred embodiments for generating the sound intensity change curve are given, which can facilitate the subsequent steps S2-S4 and better learn the audio time-series features; in step S2, according to the actual situation of slagging, there are some parameter thresholds that meet the slagging start point, and a slagging start point detection model for learning the first audio time-series feature in combination with the slagging threshold is proposed, which is more in line with the actual situation; in step S3, since the indicators in this time period do not change much, only the audio time-series features need to be considered. Therefore, to simplify the model structure, only the instantaneous sound intensity values in this time period are statistically analyzed, and a time-series prediction model is trained to learn the second audio time-series feature, obtaining a slag splashing early warning point detection model; in step S4, according to the actual situation of the slag splashing end point, there is a parameter threshold that meets the end point, and a slag splashing end point detection model for learning the third audio time-series feature in combination with the slagging end point is proposed. More importantly, since the features are significantly different for different slag splashing modes, the historical data for different modes are trained separately to specifically extract the time-series features of different modes, and the prediction results are more accurate, which can greatly avoid situations such as incorrect end point prediction, further improve the efficiency and effect of the slag splashing operation, reduce material consumption, shorten the smelting cycle, and achieve the purpose of cost reduction and efficiency increase.

[0084] On the other hand, the present invention also provides an audio-based slag splashing state monitoring system, as Figure 3 shown, including: an audio data acquisition device A, a multi-band audio analyzer B, a PLC data acquisition device C, and an audio slag melting industrial control computer D, which are used to implement the above-mentioned audio-based slag splashing state monitoring method.

[0085] The above-mentioned slag splashing state monitoring system is created based on the above-mentioned slag splashing state monitoring method, and its technical functions and beneficial effects will not be elaborated here. The technical features of the above-mentioned embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0086] The above-mentioned embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for monitoring slag splashing status based on audio, characterized in that: include: Collect historical data of the slag splashing process, including audio data, gun height data and slag conditioning agent data; And mark the slag starting point, warning point and slag splashing end point of the historical data; According to the slag starting point and previous historical data, a slag starting point detection model is established; According to the historical data from the slag starting point to the early warning point, a slag splashing early warning point detection model is established; According to the historical data from the early warning point to the slag splashing endpoint, a slag splashing endpoint detection model is established; Collect real-time data of the slag splashing process, including audio data, gun height data and slag adjusting agent data, first input the slag starting point detection model to predict the slag starting point; then input the real-time data after the slag starting point into the slag splashing warning point detection model to predict the warning point; Then the real-time data after the warning point is input into the slag splashing endpoint detection model to predict the slag splashing endpoint; Collecting historical data of the slag splashing process also includes: generating a sound intensity change curve reflecting the slag splashing process based on audio data, gun height data and slag adjusting agent data; specifically: Extract the time domain and frequency domain features of audio data, and select the audio channels that effectively describe the slag splashing process from multiple frequency bands; According to the selected audio channel value, combined with the data sampling rate, the average value of the sound intensity data per second is calculated as the instantaneous sound intensity value at the current moment. ; According to the gun height data and slag adjusting agent data, the gun position compensation value and slag adjusting agent compensation value are calculated to compensate for the instantaneous sound intensity value, and the compensated instantaneous sound intensity value is obtained. , to generate a sound intensity variation curve; Gun position compensation value , through the formula Calculate, where is the gun position compensation coefficient, is the instantaneous gun height, the initial value of the gun position ; Slag adjustment agent compensation value , through the formula Calculate, where is the slag adjustment agent compensation coefficient, is the instantaneous change of slag adjusting agent, is the real-time correction coefficient, Collect the actual amount of steel added during tapping; the instantaneous sound intensity value after compensation , through the formula calculate.

2. The slag splashing state monitoring method according to claim 1, characterized in that: The instantaneous sound intensity after compensation is calculated as After that, it also includes: S141: Set the moving window to calculate the instantaneous sound intensity value The sliding average ; S142: Set the exponentially weighted moving average weighting factor ewmaAlpha to assign the instantaneous sound intensity value at the current moment Higher weight; S143: Calculate the final instantaneous sound intensity value according to the instantaneous sound intensity value, the sliding average value and the exponentially weighted moving average weighting factor , and regenerate the sound intensity change curve.

3. The slag splashing state monitoring method according to claim 2, characterized in that: Final instantaneous sound intensity , through the formula calculate.

4. The slag splashing state monitoring method according to claim 1, characterized in that: The steps to establish a slag point detection model include: According to the slag starting point and previous historical data, the slag starting parameter thresholds of different furnace slag starting points are counted, including any one or more of: change in gun height, sound intensity and sudden change, trend slope range of sound intensity curve, movement of frequency peak, amplitude change of sound waveform in time domain, and slag starting duration; A slag point detection model is constructed with slag parameter threshold and historical data as input and slag point as output, which is used to learn the first sound intensity time series characteristics; Through multi-furnace learning based on historical data, the parameter settings of the slag point detection model are fine-tuned to obtain the trained slag point detection model, and the slag point is automatically predicted.

5. The slag splashing state monitoring method according to claim 1, characterized in that: The steps of establishing a slag splash warning point detection model include: The instantaneous sound intensity values ​​from the slag starting point to the warning point of different furnaces are counted to construct a time series data set of instantaneous sound intensity; A supervised deep learning time series prediction model is constructed, the instantaneous sound intensity time series dataset is input into the time series prediction model, the second audio time series features are learned, and the trained time series prediction model is obtained as a slag splash warning point detection model to automatically predict the warning point.

6. The slag splashing state monitoring method according to any one of claims 1 to 5, characterized in that: The steps of establishing the slag splash endpoint detection model include: According to the historical data from the early warning point to the slag splashing endpoint, the endpoint parameter thresholds of different furnaces are counted, including: the intensity and increase and decrease speed of the slag splashing endpoint sound, the trend slope range of the sound intensity curve, the frequency distribution and the movement of the peak value, the amplitude change and energy distribution of the sound waveform in the time domain, and any one or more of the slag splashing duration; A slag splashing endpoint detection model is constructed with endpoint parameter threshold and historical data as input and slag splashing endpoint as output, which is used to learn the third tone intensity time series characteristics; The historical data is divided into different slag splashing modes, different slag splashing modes are learned, and the parameter settings of the slag splashing endpoint inspection model are fine-tuned to obtain the trained slag splashing endpoint detection model to automatically predict the slag splashing endpoint.

7. An audio-based slag splashing status monitoring system, characterized in that: include: An audio data acquisition device, a multi-band audio analyzer, a PLC data acquisition device and an audio slag-removing industrial computer are used to implement the slag splashing state monitoring method described in any one of claims 1-6.

8. The slag splashing state monitoring system according to claim 7, characterized in that: The audio data acquisition device comprises an audio sensor which is arranged on the outer side of the converter fire retaining wall and is used for collecting the audio signal of nitrogen impacting the slag during slag splashing.

Images