Machine learning based real-time analysis method for sintering furnace atmosphere uniformity

Abstract

This invention relates to the field of sintering furnace atmosphere monitoring technology, specifically to a real-time analysis method for sintering furnace atmosphere uniformity based on machine learning. The method includes: acquiring various monitoring data from each monitoring point within the sintering furnace during the current sintering process, and using one type of monitoring data as target data; obtaining a local sequence of the target data at a monitoring point at a given time, and obtaining the correlation characteristic value of the target data at that monitoring point at that time; acquiring the uniformity anomaly factor of the target data at a monitoring point at a given time; segmenting the uniformity anomaly factor of the target data at each time point of a monitoring point and obtaining target segments; acquiring the anomaly characterization value and time characteristic value of each target segment; then acquiring the warning index of the current target segment corresponding to the target data at a monitoring point; and issuing a warning based on the warning index of the current target segment corresponding to various monitoring data at each monitoring point. This application can effectively monitor the sintering furnace atmosphere.

Description

Technical Field

[0001] This invention relates to the field of sintering furnace atmosphere monitoring technology, specifically to a real-time analysis method for sintering furnace atmosphere uniformity based on machine learning. Background Technology

[0002] Sintering furnaces, as crucial heat treatment equipment in powder metallurgy, electronic ceramics, and advanced materials manufacturing processes, significantly influence the sintering quality, microstructure, and final properties of materials through their internal atmosphere. During sintering, the furnace atmosphere typically consists of various gases such as nitrogen, hydrogen, carbon monoxide, and carbon dioxide. The distribution of these different gas components directly affects the reduction, oxidation, and diffusion reactions of the materials. Therefore, maintaining the stability and uniformity of the furnace atmosphere is essential for ensuring consistent sintered product quality, reducing defects, and improving production efficiency. However, in actual production, due to the complex structure and large furnace space of sintering furnaces, coupled with the influence of gas flow patterns, temperature variations, and other factors, significant differences in gas concentration, temperature, and pressure often exist in different areas of the furnace, leading to uneven atmosphere distribution. Insufficient atmosphere uniformity can easily cause abnormal reaction conditions in localized areas, resulting in problems such as under-sintering, over-burning, or fluctuations in material properties. Therefore, it is necessary to monitor and analyze the sintering furnace atmosphere in real time to improve the stability and controllability of the sintering process.

[0003] Existing sintering furnace atmosphere monitoring technologies largely rely on single-point sensor data or simple numerical judgments. However, during actual sintering furnace operation, parameters such as gas concentration, temperature, and pressure are often affected by various factors, including gas flow patterns and changes in operating conditions, exhibiting significant time-varying and nonlinear fluctuation characteristics. Furthermore, parameters at different locations are coupled with each other due to gas flow within the furnace. Under these circumstances, the aforementioned monitoring methods may struggle to accurately depict the dynamic changes in atmosphere parameters, leading to inaccurate analysis results on the uniformity of the sintering furnace atmosphere. Consequently, this affects the accuracy of intelligent assessment and effective early warning of the furnace atmosphere. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a real-time analysis method for the atmosphere uniformity of sintering furnaces based on machine learning. The specific technical solution adopted is as follows:

[0005] One embodiment of the present invention provides a real-time analysis method for the atmosphere uniformity of a sintering furnace based on machine learning, the method comprising:

[0006] Acquire various monitoring data from each monitoring point inside the sintering furnace during the current sintering process, and use one type of monitoring data as the target data; perform frequency domain analysis on the target data of a monitoring point during different sintering processes in history to obtain the window of the target data at each time point;

[0007] The target data within a window of a monitoring point at a given time are used to form a local sequence of the target data at that monitoring point at that time. The local sequence of the target data at a monitoring point at a given time is then used to perform correlation analysis with the target data at other monitoring points to obtain the correlation characteristic value of the target data at that monitoring point at that time.

[0008] The uniformity anomaly factor of the target data at a given time is obtained based on the difference in correlation characteristic values ​​between the target data at a given time and the difference in the second-order difference sequence of the local sequence; the uniformity anomaly factor of the target data at each time of a given time is segmented and the target segment is obtained;

[0009] The anomaly characterization value of a target segment is obtained based on the uniformity anomaly factor and the duration of the target segment; the time distance between the initial time of a target segment and the initial time of the corresponding sintering process is used as the time feature value of the target segment; based on the differences in time feature values, the target data of a monitoring point is clustered into target segments corresponding to different sintering processes, and the anomaly characterization value of the current target segment corresponding to the target data of the monitoring point is combined to obtain the early warning index of the current target segment; an early warning is issued based on the early warning index of the current target segment corresponding to various monitoring data of each monitoring point.

[0010] Preferably, frequency domain analysis is performed on the target data of a monitoring point during different sintering processes in history to obtain the window of the target data at each time point, including:

[0011] The target data at each moment of a monitoring point during a historical sintering process are compiled into a target data sequence for that monitoring point. Similarly, target data sequences for the same monitoring point during other historical sintering processes are obtained. The target data sequence for that monitoring point during a historical sintering process is processed using a Short-Time Fourier Transform (SFT) to obtain the frequencies and their corresponding amplitudes within a given SFT window. The amplitude of each frequency within the SFT window is multiplied to obtain the multiplication result. The sum of these multiplication results is then compared with the sum of the amplitudes of each frequency within the SFT window. The reciprocal is used to obtain the period length value at that moment in the target data sequence of the monitoring point during the sintering process. Similarly, the period length values ​​at each moment in the target data sequence of the monitoring point during the sintering process can be obtained. The median of the period length values ​​at a moment in the target data sequence of the monitoring point during different historical sintering processes is obtained as the reference period length at that moment in the target data sequence of the monitoring point. Similarly, the reference period lengths at other moments in the target data sequence of the monitoring point are obtained. Thus, the reference period lengths at each moment in the target data sequence of the monitoring point are arranged into a reference period length sequence according to the time sequence. The window of the target data at the monitoring point at a moment is a window with the reference period length corresponding to that moment in the reference period length sequence as its size.

[0012] Preferably, the correlation characteristic value of the target data at a monitoring point at a given time is obtained by performing a correlation analysis on the local sequence of target data at a monitoring point and target data at other monitoring points, including:

[0013] A local sequence of target data at a monitoring point at a given time is denoted as the sequence to be analyzed. Target data at each time point from one of the other monitoring points that are in the same position as the target data in the sequence to be analyzed are selected to form a comparison sequence corresponding to that monitoring point. The average grey relational coefficient between the sequence to be analyzed and the target data in the comparison sequence corresponding to another monitoring point is obtained using grey relational analysis, and this average grey relational coefficient is denoted as the average grey relational coefficient. The mean of the average grey relational coefficients corresponding to the comparison sequences of the other monitoring points is then obtained and used as the correlation characteristic value of the target data at that monitoring point at that time.

[0014] Preferably, the uniformity anomaly factor of the target data at a given time is obtained based on the difference in correlation characteristic values ​​between a monitoring point and target data at other monitoring points at a given time, and the difference in the second-order difference sequence of the local sequence, including:

[0015] The sum of the absolute values ​​of the second-order differences in the second-order difference sequence of the local sequence of the target data at a monitoring point at a certain time is denoted as the second-order difference sum corresponding to the target data at that monitoring point at that time. The difference between the second-order difference sum corresponding to the target data at that monitoring point at that time and the second-order difference sum corresponding to the target data at that monitoring point at that time is mapped using an exponential function with the natural constant as the base, to obtain the first mapping value corresponding to the target data at that monitoring point at that time. The ratio of the correlation feature value of the target data at that monitoring point at that time to the correlation feature value of the target data at that monitoring point at that time is obtained and denoted as the comparative correlation corresponding to the target data at that monitoring point at that time. The first mapping value corresponding to the target data at that monitoring point at that time is used as a weight to calculate the weighted average of the comparative correlations corresponding to the target data at that monitoring point at that time, thus obtaining the uniformity anomaly factor of the target data at that monitoring point at that time.

[0016] Preferably, the uniformity anomaly factor of the target data at each time point of a monitoring point is segmented and the target segment is obtained, including:

[0017] The uniformity anomaly factors of the target data at each time point of a monitoring point are used to form a uniformity anomaly factor sequence of the target data at that monitoring point; the uniformity anomaly factor sequence of the target data at that monitoring point is traversed, and the uniformity anomaly factors that are greater than the first preset value and are continuous are used to form different target segments.

[0018] Preferably, obtaining the anomaly characterization value of a target segment based on the uniformity anomaly factor in the target segment and the duration of the target segment includes:

[0019] Linear fitting is performed on the uniformity anomaly factor in a target segment to obtain the fitting slope; the product of the fitting slope corresponding to the target segment and the mean of the uniformity anomaly factor in the target segment is normalized to obtain the first normalized value; the duration of the target segment is normalized to obtain the second normalized value; the first normalized value and the second normalized value are weighted and summed to obtain the anomaly characterization value of the target segment.

[0020] Preferably, based on the differences in time characteristic values, the target data of a monitoring point is clustered into target segments corresponding to different sintering processes, and the early warning index of the current target segment is obtained by combining the abnormal characterization value of the current target segment corresponding to the target data of that monitoring point, including:

[0021] The target data of a monitoring point is divided into target segments for different sintering processes to form a target segment set. The normalized value of the absolute difference between the time feature values ​​of two target segments in the target segment set is calculated to obtain the clustering metric distance between the two target segments. The target segments in the target segment set are clustered using the clustering metric distance between each pair of target segments to obtain the clusters corresponding to the target data of the monitoring point. The current target segment corresponding to the target data in the current sintering process of the monitoring point is obtained. The cluster to which the current target segment belongs is determined. The difference between the abnormal characterization value of the current target segment and the abnormal characterization value of a target segment in the cluster to which the current target segment belongs is mapped using an exponential function with the natural constant as the base to obtain the second mapping value corresponding to the target segment in the cluster to which the current target segment belongs. The second mapping values ​​corresponding to each target segment in the cluster to which the current target segment belongs are averaged to obtain the early warning index of the current target segment.

[0022] Preferably, early warnings are issued based on the early warning indicators for the current target segment corresponding to various monitoring data at each monitoring point, including:

[0023] For the target data of a monitoring point in the current sintering process, the warning threshold of the target data of the current monitoring point in the current sintering process is obtained by combining the mean and standard deviation of the warning indicators of each target segment of the target data of the monitoring point in the historical sintering process with the three-times standard deviation method. Similarly, the warning thresholds of various monitoring data of each monitoring point in the current sintering process can be obtained. If the warning indicator of the current target segment corresponding to a certain type of monitoring data of a monitoring point in the current sintering process is greater than the warning threshold of that type of monitoring data of that monitoring point in the current sintering process, a warning is issued.

[0024] The embodiments of the present invention have at least the following beneficial effects: Firstly, this application acquires various monitoring data from each monitoring point within the sintering furnace during the current sintering process, and uses one type of monitoring data as the target data. Then, it performs frequency domain analysis on the target data from a monitoring point during different historical sintering processes to obtain the window size for each moment of the target data at that monitoring point. This allows for adaptive determination of the window size, thereby more accurately extracting the correlation features between the local time-series data of each monitoring point, i.e., obtaining more accurate correlation feature values ​​for the target data at each moment of each monitoring point. Next, based on the difference in correlation feature values ​​between the target data of one monitoring point and other monitoring points at a given moment, and the difference in the second-order difference sequence of the local sequence, it obtains the uniformity anomaly factor of the target data at that monitoring point at that moment, used to characterize the deviation between the changing trends of monitoring data at different monitoring points and the overall trend. Subsequently, it analyzes the target data at each moment of a monitoring point... The data uniformity anomaly factor is segmented and target segments are obtained. Anomaly characterization values ​​are constructed by combining the anomaly duration and change trend. Then, based on the difference in time characteristic values, the target data of a monitoring point is clustered into target segments corresponding to different sintering processes. The warning index of the current target segment is obtained by combining the anomaly characterization value of the current target segment corresponding to the target data of the monitoring point. Warnings are issued based on the warning index of the current target segment corresponding to various monitoring data of each monitoring point. Compared with traditional monitoring methods based on single-point data or simple threshold judgment, this application can comprehensively consider the time-series change periodic characteristics of various monitoring data in the sintering furnace, the dynamic correlation between multiple monitoring points, and the gas change law under different sintering stages. This significantly improves the accuracy and stability of sintering furnace atmosphere uniformity monitoring and anomaly warning, and can effectively reduce misjudgments caused by factors such as short-term airflow disturbances, thereby improving the quality control level of the sintering process. Attached Figure Description

[0025] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0026] Figure 1 This application provides a flowchart of a real-time analysis method for the atmosphere uniformity of a sintering furnace based on machine learning, as an embodiment of the present application. Detailed Implementation

[0027] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a machine learning-based real-time analysis method for sintering furnace atmosphere uniformity proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0029] The following description, in conjunction with the accompanying drawings, details a specific scheme for a real-time analysis method for the atmosphere uniformity of a sintering furnace based on machine learning, provided by the present invention.

[0030] In this embodiment, the main application scenario of the present invention is as follows: This application provides a real-time analysis method for the atmosphere uniformity of a sintering furnace based on machine learning. By comprehensively modeling and correlating the multi-source sensor data of the sintering furnace, the method can achieve real-time and accurate assessment and early warning of anomalies in the atmosphere uniformity of the furnace.

[0031] Please see Figure 1 The diagram illustrates a method flowchart for real-time analysis of sintering furnace atmosphere uniformity based on machine learning, according to an embodiment of the present invention. The method includes the following steps:

[0032] Step S1: Obtain various monitoring data of each monitoring point in the furnace of the sintering furnace during the current sintering process, and take one monitoring data as the target data; perform frequency domain analysis on the target data of a monitoring point in different sintering processes in history to obtain the window of the target data of the monitoring point at each time.

[0033] This study focuses on more accurate monitoring and analysis of the atmosphere uniformity in the sintering furnace. Therefore, appropriate sensors were first deployed at key locations within the furnace to collect real-time data on the furnace atmosphere. These key locations included the furnace center, inlet and outlet, top and bottom of the furnace, etc., each designated as a monitoring point. Gas concentration sensors, temperature sensors, and pressure sensors were deployed at each monitoring point to collect data on various gas concentrations (e.g., carbon monoxide and carbon dioxide), temperatures, and pressures at different times. Each gas concentration and temperature data was recorded as a different type of monitoring data, thus obtaining various monitoring data from each monitoring point within the furnace during the current sintering process. Subsequently, to integrate historical data for analysis, monitoring data from each monitoring point within the furnace during each complete historical sintering process also needed to be obtained. Simultaneously, the acquired monitoring data underwent standardization to prevent the dimensions from affecting subsequent analysis. This resulted in a multi-monitoring time-series monitoring dataset. The specific data types and sensor configurations can be selected based on the actual application scenario and monitoring requirements, and are not specifically limited here. For ease of explanation later, one type of monitoring data will be used as the target data for subsequent explanation. The analysis will be conducted with one monitoring point and one type of monitoring data as a unit of analysis. The analysis process for other types of monitoring data at other monitoring points is the same.

[0034] A sintering furnace is an industrial device used for high-temperature heating of powder materials or green bodies. It is widely used in powder metallurgy, ceramics, and electronic materials manufacturing. Its main function is to promote diffusion bonding between material particles under specific temperature and atmosphere conditions, thereby forming products with a certain strength and density. During the sintering process, atmospheric parameters such as gas concentration, temperature, and pressure inside the furnace directly affect the sintering reaction process and the final product quality. Therefore, it is necessary to maintain a relatively uniform spatial distribution of the atmosphere within the furnace to avoid inconsistent sintering states in localized areas. For this reason, real-time monitoring and analysis of the atmosphere uniformity in the sintering furnace are required.

[0035] However, in actual operation, the furnace atmosphere parameters are often affected by a variety of factors such as gas flow state, gas supply strategy and operating conditions, exhibiting obvious time-varying and nonlinear fluctuation characteristics. Existing monitoring methods mostly rely on single-point sensor data or simple threshold judgment, which makes it difficult to fully reflect the dynamic correlation between different monitoring points in the time dimension, thus affecting the accurate assessment of the uniformity of the sintering furnace atmosphere and the effective early warning of abnormal conditions.

[0036] To achieve more accurate monitoring and analysis of the atmosphere uniformity within the sintering furnace, thereby addressing the shortcomings of traditional methods for monitoring and analyzing the atmosphere uniformity of sintering furnaces.

[0037] This application analyzes the temporal variation characteristics of gas concentration, temperature, pressure, and other data collected by sensors at various monitoring points by establishing windows, and comprehensively evaluates the atmosphere distribution state inside the furnace by combining the consistency and correlation of changes between different monitoring points, thereby achieving a more accurate determination of the atmosphere uniformity inside the sintering furnace.

[0038] During actual operation, the gas variation rhythm in different sintering stages of a sintering furnace may exhibit different cycles over time due to factors such as gas supply regulation, material reaction rate, and gas flow state within the furnace. Therefore, when using windowing for time-series data analysis, the window size for different time periods needs to be matched with the actual sintering process. This ensures that the selected local data can better reflect the main characteristics of gas changes at that stage, thereby achieving a more accurate analysis of the uniformity of gas parameters within the sintering furnace.

[0039] Therefore, frequency domain analysis is performed on the target data of a monitoring point during different sintering processes in history to obtain the window of the target data at each moment of the monitoring point.

[0040] Specifically, the target data at each moment of a monitoring point during a historical sintering process are used to form a target data sequence for that monitoring point. Similarly, target data sequences for the same monitoring point during other historical sintering processes are obtained. The target data sequence for that monitoring point during a historical sintering process is processed using a Short-Time Fourier Transform (STFT) to obtain the frequencies and their corresponding amplitudes within a STFT window at a given moment. A frequency within the STFT window at that moment is multiplied by its corresponding amplitude to obtain the multiplication result. The sum of these multiplication results is then compared with the sum of the corresponding amplitudes within the STFT window at that moment, and the reciprocal is taken to obtain the period length of that moment in the target data sequence for that monitoring point during the sintering process. Similarly, the period length value at each moment in the target data sequence of the monitoring point during the sintering process can be obtained. The median of the period length values ​​at a moment in the target data sequence of the monitoring point during different historical sintering processes can be obtained as the reference period length at that moment in the target data sequence of the monitoring point. Similarly, the reference period lengths at other moments in the target data sequence of the monitoring point can be obtained. Thus, the reference period lengths at each moment in the target data sequence of the monitoring point are arranged into a reference period length sequence according to the time sequence. The window of the target data at the monitoring point at a moment is a window with the reference period length corresponding to that moment in the reference period length sequence as its size.

[0041] The specific model for calculating the period length at a given moment is as follows:

[0042] ,

[0043] in, This represents the period length at time t in the target data sequence of a monitoring point during a historical sintering process; and These represent the maximum and minimum frequencies within the short-time Fourier transform window corresponding to time t in the target data sequence when the target data sequence at this monitoring point in the sintering process is processed using short-time Fourier transform. and Let represent a frequency and its amplitude within the short-time Fourier transform window at time t in the target data sequence, respectively. This is the result of the multiplication corresponding to this frequency; This represents the reciprocal of the frequency weighting value within the short-time Fourier transform window at time t in the target data sequence of the monitoring point during the sintering process. It also represents the length of the main change cycle of the time series data at that time, which is the cycle length value.

[0044] It should be noted that this application selects 20 sintering processes from historical data (the implementer can choose the number of sintering processes according to the actual situation). That is, for a monitoring point, the target data sequence corresponding to 20 historical sintering processes can be obtained. In other words, 20 target data sequences were historically obtained for this monitoring point. Then, for this monitoring point and the target data, the period length value of each moment in these 20 target data sequences can be obtained. For the same moment, there are 20 period length values. Taking the median of them, the reference period length of this moment can be obtained. Similarly, the reference period length of each moment in the target data sequence corresponding to the target data at this monitoring point can be obtained, forming a reference period length sequence. The same applies to other types of data at other monitoring points. The reference period length sequence corresponding to each type of monitoring data at each monitoring point can be obtained. The reference period length of a moment is the time length or size of the window at that moment. In addition, for a window at a moment, that moment is located at the center of the window.

[0045] Through the above operations, when acquiring the current sintering furnace operating data during actual monitoring, the appropriate window length can be determined based on the periodic reference sequence corresponding to the current moment, so that the window can cover a major periodic interval of the current gas change. This allows for more accurate extraction of the local correlation between the data at each monitoring point when performing local data similarity or correlation analysis, thereby improving the accuracy and stability of the assessment of the atmosphere uniformity in the sintering furnace.

[0046] Step S2: The target data within a window of a monitoring point at a certain time are used to form a local sequence of the target data of that monitoring point at that time; the local sequence of the target data of a monitoring point at a certain time is used to perform correlation analysis with the target data of other monitoring points to obtain the correlation feature value of the target data of that monitoring point at that time.

[0047] Based on the above analysis, a reference period length sequence in the time dimension can be obtained for any type of monitoring data time series based on historical sample data. The main purpose of this monitoring is to maintain the relative uniformity of the furnace atmosphere in spatial distribution, so as to avoid the problem of inconsistent sintering state in local areas.

[0048] Therefore, further, by using the windows of the target data at each monitoring point at each time moment, the local analysis window corresponding to the current time moment can be determined. That is, the target data within the window of the target data at a monitoring point at a given time moment is used to form the local sequence of the target data at that monitoring point at that time moment. Subsequently, based on the grey relational analysis method, the correlation characteristic value of the local sequence of the target data at any monitoring point at that time moment relative to the other monitoring points is calculated.

[0049] Therefore, by performing correlation analysis on the local sequence of target data at a monitoring point at a certain time with target data from other monitoring points, the correlation characteristic value of the target data at that monitoring point at that time can be obtained.

[0050] Specifically, a local sequence of target data at a monitoring point at a given time is denoted as the sequence to be analyzed. Target data at each time point from one of the other monitoring points that are in the same position as the target data in the sequence to be analyzed are selected to form a comparison sequence corresponding to that monitoring point. The average grey relational coefficient between the sequence to be analyzed and the target data in the comparison sequence corresponding to another monitoring point is obtained using grey relational analysis, and this average grey relational coefficient is denoted as the average grey relational coefficient. The mean of the average grey relational coefficients corresponding to the comparison sequences of the other monitoring points is then obtained and used as the correlation characteristic value of the target data at that monitoring point at that time.

[0051] The specific calculation model for correlation feature values ​​is as follows:

[0052] ,

[0053] in, The correlation feature value of the target data at time t of the a-th monitoring point is represented, which is the correlation between the local sequence (the sequence to be analyzed) corresponding to the target data at time t of the a-th monitoring point and the comparison sequence corresponding to other monitoring points; m represents the number of monitoring points, and m-1 is the number of other monitoring points besides the a-th monitoring point. The average grey correlation coefficient (AMC) represents the average grey correlation coefficient between the local sequence corresponding to the target data at time t of the target data at monitoring point a and the corresponding target data in the comparison sequence of the b-th monitoring point. It can also be called the degree of correlation between the local target data at time t of the target data at monitoring point a and the local target data at the same position in the b-th monitoring point. In other words, it represents the local correlation value between the monitoring time series and the time series of the other monitoring points.

[0054] Similarly, the correlation characteristics of various monitoring data at various monitoring points at various times can be obtained.

[0055] Step S3: Obtain the uniformity anomaly factor of the target data at a given time based on the difference in correlation feature value between a monitoring point and the target data at another monitoring point at a given time and the difference in the second-order difference sequence of the local sequence; segment the uniformity anomaly factor of the target data at each time of a monitoring point and obtain the target segment.

[0056] Through the above analysis, the correlation characteristics of various monitoring data at each monitoring point at each time point were obtained. When the distribution of gas, temperature and pressure in the furnace is relatively uniform, the change characteristics of various monitoring data at each monitoring point in time series often have a high degree of consistency, and the corresponding correlation characteristics are also relatively high. However, when the correlation of the monitoring data at a certain monitoring point is low relative to the other monitoring points, it indicates that the change trend of the monitoring point at that location deviates from the overall trend, which may indicate that the atmosphere uniformity in that area is poor.

[0057] Therefore, based on the correlation feature value of the target data at any monitoring point at a certain time, the uniformity anomaly factor value is constructed; the uniformity anomaly factor of the target data at that monitoring point at that time is obtained according to the difference between the correlation feature value of the target data at a certain monitoring point and that of the target data at other monitoring points at a certain time and the difference between the second-order difference sequence of the local sequence.

[0058] Specifically, the sum of the absolute values ​​of the second-order differences in the second-order difference sequence of the local sequence of the target data at a monitoring point at a given time is denoted as the second-order difference sum corresponding to the target data at that monitoring point at that time. A first mapping value corresponding to the target data at that time is obtained by mapping the difference between the second-order difference sum corresponding to the target data at that monitoring point and the second-order difference sum corresponding to the target data at that time at another monitoring point using an exponential function with the natural constant as the base. The ratio of the correlation feature value of the target data at that time at another monitoring point to the correlation feature value of the target data at that time at the first-order difference sequence is obtained, and denoted as the comparative correlation corresponding to the target data at that time at the first-order difference sequence at another monitoring point. Finally, the uniformity anomaly factor of the target data at that time at the first-order difference sequence at each monitoring point is obtained by weighting the comparative correlations of the target data at each monitoring point at that time.

[0059] The specific calculation model for the uniformity anomaly factor of target data at a monitoring point at a given time is as follows:

[0060] ,

[0061] in, The uniformity anomaly factor of the target data at time t at the a-th monitoring point; and These represent the correlation characteristic values ​​of the target data at time t for the a-th monitoring point and the b-th monitoring point among other monitoring points. This indicates the comparative correlation of the target data at time t for the b-th monitoring point among other monitoring points; and Let represent the sum of the absolute values ​​of the second-order differences in the second-order difference sequence of the target data at time t at monitoring point a, and the sum of the absolute values ​​of the second-order differences in the second-order difference sequence of the target data at time t at monitoring point b, respectively. These two sums can represent the stability of the temporal variation of the monitoring data represented by the target data; e represents the natural constant. The first mapping value, representing the target data at time t of the b-th monitoring point among other monitoring points, is used as the weight for the comparative correlation of the target data at time t of the b-th monitoring point among other monitoring points. This weight is used to adjust the impact of the stability of data at different monitoring points on the anomaly judgment result. Then, the first mapping value corresponding to the target data at time t of each of the other monitoring points is used as a weight to calculate the weighted average of the comparative correlation of the target data at time t of each of the other monitoring points. The larger the value, the greater the deviation of the monitoring point from the overall target data trend, and the higher the degree of abnormality in its atmosphere uniformity.

[0062] If the distribution of various monitoring data at a certain monitoring point is uneven, it may be due to insufficient local reaction, poor gas flow, or gas stagnation, causing the trend of various monitoring data at that monitoring point to deviate from that of other monitoring points. Therefore, when conducting comparative analysis, if the correlation characteristic value of a certain monitoring point is relatively low compared to other monitoring points, and its local fluctuation characteristics are relatively stable, it indicates that the state of various monitoring data at that location has deviated from the overall trend for a long time, and its atmospheric uniformity may be poor.

[0063] Through the above calculations, the uniformity anomaly factor of the target data at each monitoring point at each time can be obtained, thus forming a uniformity anomaly factor value sequence in the time dimension. Since different sintering stages may be affected by factors such as the gas flow state in the furnace, the material reaction rate, and the gas supply strategy, the manifestation and duration of gas uniformity anomalies in the time dimension may differ at different stages.

[0064] To further obtain the non-uniformity characteristics of any type of monitoring data throughout the sintering process, the uniformity anomaly factor of the target data at each time point of a monitoring point is segmented and the target segment is obtained.

[0065] Specifically, the uniformity anomaly factors of the target data at each time point of a monitoring point are used to form a uniformity anomaly factor sequence of the target data at that monitoring point; the uniformity anomaly factor sequence of the target data at that monitoring point is traversed, and the uniformity anomaly factors that are greater than a first preset value and are continuous are used to form different target segments.

[0066] The first preset value is 1. When the uniformity anomaly factor in the uniformity anomaly factor sequence is greater than 1, it indicates a relative anomaly at that moment; if it is less than or equal to 1, it indicates relative normality. There may be a certain degree of atmospheric non-uniformity anomaly in the target segment compared to other monitoring points. When the distribution of the monitored objects (gas concentration, temperature, or pressure) corresponding to the target data in the furnace is relatively uniform, the change characteristics of the target data at each monitoring point often have high consistency, and their corresponding correlation characteristic values ​​are also relatively high. However, if this point is abnormal, its correlation with other monitoring points is weak, its distribution is small, and the calculated result will be greater than the first preset value, which may indicate an anomaly.

[0067] Similarly, the uniformity anomaly factors of various monitoring data at each monitoring point at each time can be segmented and the corresponding target segments can be obtained.

[0068] Step S4: Obtain the anomaly characterization value of a target segment based on the uniformity anomaly factor and the duration of the target segment; use the time distance between the initial time of a target segment and the initial time of the sintering process corresponding to the target segment as the time feature value of the target segment; cluster the target data of a monitoring point in different sintering processes according to the differences in the time feature values, and obtain the early warning index of the current target segment by combining the anomaly characterization value of the current target segment corresponding to the target data of the monitoring point; issue an early warning based on the early warning index of the current target segment corresponding to various monitoring data of each monitoring point.

[0069] For a target segment, an anomaly characterization value is constructed based on its changing trend and temporal duration characteristics. Specifically, a linear fit is performed on the uniformity anomaly factor in a target segment to obtain the fitting slope; the product of the fitting slope corresponding to the target segment and the mean of the uniformity anomaly factor in the target segment is normalized to obtain a first normalized value; the duration of the target segment is normalized to obtain a second normalized value; and the first and second normalized values ​​are weighted and summed to obtain the anomaly characterization value of the target segment.

[0070] The specific calculation model for the anomaly representation value of the target segment is as follows:

[0071] ,

[0072] In the formula, This represents the anomaly representation value of the Lth target segment; α represents the weight parameter, which is 0.6 (empirical value); norm() represents the linear normalization function; This represents the mean value of the uniformity anomaly factor for the Lth target segment. The larger this value, the greater the anomaly indication. The slope of the linear fit obtained from the uniformity anomaly factor value sequence of the Lth target segment is represented by the slope of the fit. The larger the value, the more abnormal the segment is. This is the first normalized value; This represents the duration of the Lth target segment (i.e., the amount of data contained or the duration of the duration). The larger the value, the longer the abnormal state lasts in the time dimension. Therefore, it is appropriate to increase the weight of this indicator to reduce the impact of random fluctuations such as short-term gas disturbances.

[0073] Furthermore, the above calculations can yield anomaly characteristics for any target segment. Considering the differences in gas flow characteristics and reaction processes within the furnace at different sintering stages, the anomaly patterns may also differ. Subsequent cluster analysis is required.

[0074] First, for a target segment, the time distance between the initial time of the target segment and the initial time of the sintering process corresponding to the target segment is used as the time feature value of the target segment, and the time feature value of any target segment can be obtained.

[0075] Furthermore, the target data of a monitoring point is divided into target segments corresponding to different sintering processes to form a target segment set of the target data of that monitoring point; the normalized value of the absolute value of the difference between the time feature values ​​of two target segments in the target segment set is calculated to obtain the clustering metric distance between the two target segments; the target segments in the target segment set are clustered using the clustering metric distance between each pair of target segments to obtain the various clusters corresponding to the target data of that monitoring point; the current target segment corresponding to the target data in the current sintering process of that monitoring point is obtained; the cluster to which the current target segment belongs is determined, and the difference between the abnormal characterization value of the current target segment and the abnormal characterization value of a target segment in the cluster to which the current target segment belongs is mapped using an exponential function with the natural constant as the base to obtain the second mapping value corresponding to the target segment in the cluster to which the current target segment belongs; the second mapping values ​​corresponding to each target segment in the cluster to which the current target segment belongs are averaged to obtain the early warning index of the current target segment.

[0076] It should be noted that the method for obtaining the current target segment is the same as the method described above. If a current target segment exists, the uniformity anomaly factor corresponding to the target data at the current moment within that segment must be greater than the first preset value. Conversely, a current target segment may not exist; in this case, the uniformity anomaly factor corresponding to the target data at the current moment is less than or equal to the first preset value, indicating that no anomaly exists. Furthermore, when determining the cluster to which the current target segment belongs, it is necessary to calculate the clustering metric distance between the current target segment and the centers of each cluster. The cluster with the smallest clustering metric distance is the cluster to which the current target segment belongs. The clustering method used is the HDBSCAN clustering algorithm (an upgraded version of the DBSCAN clustering algorithm, which can perform data clustering analysis using the algorithm's default values ​​without preset parameters, and will not be elaborated upon here). This algorithm performs clustering analysis on the target segments corresponding to the current monitoring data and selected historical samples, thereby identifying typical patterns of atmosphere anomalies at different sintering stages and further assisting in determining whether the current abnormal state belongs to normal process fluctuations or a potential abnormal state.

[0077] After clustering, the overall distribution characteristics of the cluster to which the current target segment belongs are combined with the comparison and analysis of the degree of abnormality of the current target segment with the other target segments in the cluster. This allows for the construction of an early warning index for the current target segment, which is used to determine whether the abnormality belongs to the regular fluctuation under the normal sintering stage or may be an abnormality of non-uniform atmosphere.

[0078] The specific calculation model for the clustering metric distance between two target segments is as follows:

[0079] ,

[0080] In the formula, This represents the clustering metric distance between target segment j and target segment L. , These represent the time distances between the initial time of the i-th target segment and the initial time of their respective sintering processes, i.e., their respective time characteristic values. The absolute value of the difference between the time characteristic values ​​of the i-th target segment and the L-th target segment indicates that the two target segments are in the same reaction stage.

[0081] The calculation model for the early warning indicators of the current target segment is as follows:

[0082] ,

[0083] in, This represents the early warning indicator for the current target segment corresponding to the target data at a monitoring point. and These represent the anomaly representation value of the current target segment and the anomaly representation value of the s-th target segment in the cluster to which the current target segment belongs, respectively. This indicates the number of target segments in the cluster described by the current target segment. This is the second mapping value corresponding to the s-th target segment in the cluster to which the current target segment belongs. This indicates the relative degree of anomalousness of the current target segment's anomalous characterization value compared to the other anomalous characterization values ​​in the same cluster. The larger this value, the more likely the atmospheric inhomogeneity at this time is not a regular reaction unique to the sintering furnace, but may be a true manifestation of inhomogeneity.

[0084] Similarly, corresponding early warning indicators can be obtained for the current target segment corresponding to various monitoring data at each monitoring point. Using the above method, based on the characteristics of historical sintering processes, the abnormal state of the current sintering furnace atmosphere can be more reliably identified, thereby effectively improving the accuracy and stability of sintering furnace atmosphere uniformity monitoring and early warning analysis.

[0085] Through the above analysis, the early warning indicators corresponding to the current target segments of various monitoring data at each monitoring point during the current sintering process can be obtained. Furthermore, for the early warning indicator of the current target segment of a certain type of monitoring data at a monitoring point during the current sintering process, combined with the early warning indicators of each target segment of that type of monitoring data at that monitoring point during the current sintering process and the distribution characteristics of the early warning indicators of each target segment of that type of monitoring data at that monitoring point during historical sintering processes, an abnormal situation can be judged by setting a reasonable threshold. The threshold can be determined using statistical methods such as fixed empirical thresholds, box plot outlier judgment methods, or the three-standard-deviation method, thereby issuing an early warning. In this embodiment, the three-standard-deviation method is used as the method for determining the early warning threshold.

[0086] Warnings are issued based on the warning indicators for the current target segment corresponding to various monitoring data at each monitoring point. Specifically, for the target data at a monitoring point in the current sintering process, the warning threshold for the target data at that monitoring point in the current sintering process is obtained by combining the mean and standard deviation of the warning indicators for each target segment of the target data at that monitoring point in historical sintering processes using the three-standard-deviation method. Similarly, the warning thresholds for various monitoring data at each monitoring point in the current sintering process can be obtained. If, in the current sintering process, at least one monitoring point has a warning indicator for the current target segment corresponding to a certain type of monitoring data that is greater than the warning threshold for that type of monitoring data at that monitoring point in the current sintering process, a warning is issued. It should be noted that the three-standard-deviation method for determining thresholds is a well-known technique and will not be elaborated upon here.

[0087] When issuing an early warning, the system indicates the location of the warning and the type of monitoring data, thereby enabling timely identification and early warning of uneven atmosphere within the furnace.

[0088] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0089] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time analysis method for the atmosphere uniformity of a sintering furnace based on machine learning, characterized in that, The method includes: The temperature, concentrations of various gases, and pressures at various monitoring points within the sintering furnace during the current sintering process are acquired as various monitoring data, and one type of monitoring data is selected as the target data. Frequency domain analysis is performed on the target data of a monitoring point during different sintering processes in history to obtain the window of the target data at each moment of that monitoring point. The target data within a window of a monitoring point at a given time are used to form a local sequence of the target data at that monitoring point at that time. The local sequence of the target data at a monitoring point at a given time is then used to perform correlation analysis with the target data at other monitoring points to obtain the correlation characteristic value of the target data at that monitoring point at that time. The uniformity anomaly factor of the target data at a given time is obtained based on the difference in correlation characteristic values ​​between the target data at a given time and the difference in the second-order difference sequence of the local sequence; the uniformity anomaly factor of the target data at each time of a given time is segmented and the target segment is obtained; The anomaly characterization value of a target segment is obtained based on the uniformity anomaly factor and the duration of the target segment; the time distance between the initial time of a target segment and the initial time of the corresponding sintering process is used as the time feature value of the target segment; based on the differences in time feature values, the target data of a monitoring point is clustered into target segments corresponding to different sintering processes, and the anomaly characterization value of the current target segment corresponding to the target data of the monitoring point is combined to obtain the early warning index of the current target segment; an early warning is issued based on the early warning index of the current target segment corresponding to various monitoring data of each monitoring point.

2. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The method of obtaining the window for each moment of the target data at a monitoring point by performing frequency domain analysis on the target data at a monitoring point during different sintering processes in history includes: The target data at each moment of a monitoring point during a historical sintering process are compiled into a target data sequence for that monitoring point. Similarly, target data sequences for the same monitoring point during other historical sintering processes are obtained. The target data sequence for that monitoring point during a historical sintering process is processed using a Short-Time Fourier Transform (SFT) to obtain the frequencies and their corresponding amplitudes within a given SFT window. The amplitude of each frequency within the SFT window is multiplied to obtain the multiplication result. The sum of these multiplication results is then compared with the sum of the amplitudes of each frequency within the SFT window. The reciprocal is used to obtain the period length value at that moment in the target data sequence of the monitoring point during the sintering process. Similarly, the period length values ​​at each moment in the target data sequence of the monitoring point during the sintering process can be obtained. The median of the period length values ​​at a moment in the target data sequence of the monitoring point during different historical sintering processes is obtained as the reference period length at that moment in the target data sequence of the monitoring point. Similarly, the reference period lengths at other moments in the target data sequence of the monitoring point are obtained. Thus, the reference period lengths at each moment in the target data sequence of the monitoring point are arranged into a reference period length sequence according to the time sequence. The window of the target data at the monitoring point at a moment is a window with the reference period length corresponding to that moment in the reference period length sequence as its size.

3. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The method involves performing correlation analysis on the local sequence of target data at a monitoring point at a certain time with target data from other monitoring points to obtain the correlation feature value of the target data at that monitoring point at that time. A local sequence of target data at a monitoring point at a given time is denoted as the sequence to be analyzed. Target data at each time point from one of the other monitoring points that are in the same position as the target data in the sequence to be analyzed are selected to form a comparison sequence corresponding to that monitoring point. The average grey relational coefficient between the sequence to be analyzed and the target data in the comparison sequence corresponding to another monitoring point is obtained using grey relational analysis, and this average grey relational coefficient is denoted as the average grey relational coefficient. The mean of the average grey relational coefficients corresponding to the comparison sequences of the other monitoring points is then obtained and used as the correlation characteristic value of the target data at that monitoring point at that time.

4. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The step of obtaining the uniformity anomaly factor of the target data at a given time point based on the difference in correlation characteristic values ​​between a monitoring point and target data at other monitoring points at a given time point and the difference in the second-order difference sequence of the local sequence includes: The sum of the absolute values ​​of the second-order differences in the second-order difference sequence of the local sequence of the target data at a monitoring point at a certain time is denoted as the second-order difference sum corresponding to the target data at that monitoring point at that time. The difference between the second-order difference sum corresponding to the target data at that monitoring point at that time and the second-order difference sum corresponding to the target data at that monitoring point at that time is mapped using an exponential function with the natural constant as the base, to obtain the first mapping value corresponding to the target data at that monitoring point at that time. The ratio of the correlation feature value of the target data at that monitoring point at that time to the correlation feature value of the target data at that monitoring point at that time is obtained and denoted as the comparative correlation corresponding to the target data at that monitoring point at that time. The first mapping value corresponding to the target data at that monitoring point at that time is used as a weight to calculate the weighted average of the comparative correlations corresponding to the target data at that monitoring point at that time, thus obtaining the uniformity anomaly factor of the target data at that monitoring point at that time.

5. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The step of segmenting the uniformity anomaly factor of the target data at each time point of a monitoring point and obtaining the target segment includes: The uniformity anomaly factors of the target data at each time point of a monitoring point are used to form a uniformity anomaly factor sequence of the target data at that monitoring point; the uniformity anomaly factor sequence of the target data at that monitoring point is traversed, and the uniformity anomaly factors that are greater than the first preset value and are continuous are used to form different target segments.

6. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The step of obtaining the anomaly characterization value of a target segment based on the uniformity anomaly factor and the duration of the target segment includes: Linear fitting is performed on the uniformity anomaly factor in a target segment to obtain the fitting slope; the product of the fitting slope corresponding to the target segment and the mean of the uniformity anomaly factor in the target segment is normalized to obtain the first normalized value; the duration of the target segment is normalized to obtain the second normalized value; the first normalized value and the second normalized value are weighted and summed to obtain the anomaly characterization value of the target segment.

7. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The method of clustering target data at a monitoring point across different sintering processes based on differences in time characteristic values ​​and combining this with the anomaly representation value of the current target segment corresponding to the target data at that monitoring point to obtain the early warning index for the current target segment includes: The target data of a monitoring point is divided into target segments for different sintering processes to form a target segment set. The normalized value of the absolute difference between the time feature values ​​of two target segments in the target segment set is calculated to obtain the clustering metric distance between the two target segments. The target segments in the target segment set are clustered using the clustering metric distance between each pair of target segments to obtain the clusters corresponding to the target data of the monitoring point. The current target segment corresponding to the target data in the current sintering process of the monitoring point is obtained. The cluster to which the current target segment belongs is determined. The difference between the abnormal characterization value of the current target segment and the abnormal characterization value of a target segment in the cluster to which the current target segment belongs is mapped using an exponential function with the natural constant as the base to obtain the second mapping value corresponding to the target segment in the cluster to which the current target segment belongs. The second mapping values ​​corresponding to each target segment in the cluster to which the current target segment belongs are averaged to obtain the early warning index of the current target segment.

8. The real-time analysis method for sintering furnace atmosphere uniformity based on machine learning according to claim 1, characterized in that, The method of issuing early warnings based on the early warning indicators of the current target segment corresponding to various monitoring data at each monitoring point includes: For the target data of a monitoring point in the current sintering process, the warning threshold of the target data of the current monitoring point in the current sintering process is obtained by combining the mean and standard deviation of the warning indicators of each target segment of the target data of the monitoring point in the historical sintering process with the three-times standard deviation method. Similarly, the warning thresholds of various monitoring data of each monitoring point in the current sintering process can be obtained. If the warning indicator of the current target segment corresponding to a certain type of monitoring data of a monitoring point in the current sintering process is greater than the warning threshold of that type of monitoring data of that monitoring point in the current sintering process, a warning is issued.

Images