A smelting furnace system fault early warning method based on big data

Abstract

The present invention relates to the field of big data technology and discloses a big data-based method for smelting furnace system fault warning. The method comprises: collecting multi-source sensor time series data during smelting furnace operation; extracting the temperature signal gradient feature of the smelting temperature signal and the signal mean feature of the electrode current signal; decomposing the crucible vibration signal into preset frequency bands and calculating the product of the standard deviation of the energy value of each frequency band and the energy ratio of the adjacent frequency bands to generate a frequency domain energy gradient feature vector; generating a multidimensional feature matrix based on the signal mean feature, the temperature signal gradient feature, and the frequency domain energy gradient feature vector; inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output the smelting furnace's fault boundary; and when the fault boundary reaches a preset trigger condition, triggering a warning instruction corresponding to the smelting furnace's graded warning signal, using the warning instruction to issue a fault warning for the smelting furnace. The present invention can improve the accuracy of smelting furnace system fault warnings.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a smelting furnace system fault early warning method and system based on big data. Background Art

[0002] Melting furnaces are core equipment in fields such as metallurgy and materials processing. The stability of their operating status directly impacts product quality and production safety. In actual production, melting furnace failures (such as electrode overload, crucible breakage, and temperature runaway) are often caused by a combination of factors, necessitating early warning through monitoring of key operating parameters.

[0003] Existing methods typically use a single sensor or multiple sensors to independently collect data, lacking a timestamp synchronization mechanism. For example, the acquisition frequencies of temperature and vibration sensors are out of sync, resulting in mismatched temperature and vibration data at the same point in time. Subsequent feature analysis can generate errors due to inconsistent data baselines, making it difficult to accurately correlate multi-parameter changes with faults. This results in low accuracy in predicting smelting furnace system faults. Summary of the Invention

[0004] The present invention provides a smelting furnace system fault early warning method and system based on big data, the main purpose of which is to solve the problem of low accuracy in smelting furnace system fault early warning.

[0005] To achieve the above objectives, the present invention provides a smelting furnace system fault early warning method based on big data, comprising:

[0006] Collecting multi-source sensor time series data during the operation of the smelting furnace, wherein the multi-source sensor time series data includes a smelting temperature signal, a crucible vibration signal, and an electrode current signal;

[0007] extracting a temperature signal gradient feature of the smelting temperature signal, and extracting a signal mean feature of the electrode current signal;

[0008] The crucible vibration signal is decomposed into preset frequency bands through a preset wavelet packet decomposition, and the product of the standard deviation of the energy value of each frequency band and the energy ratio of the adjacent frequency bands is calculated to generate a frequency domain energy gradient feature vector; a multidimensional feature matrix is ​​generated according to the signal mean feature, the temperature signal gradient feature and the frequency domain energy gradient feature vector; the multidimensional feature matrix is ​​input into a pre-trained lightweight gradient boosting tree model to output the fault boundary of the melting furnace; when the fault boundary reaches a preset trigger condition, the warning instruction corresponding to the graded warning signal of the melting furnace is triggered, and the warning instruction is used to issue a fault warning to the melting furnace.

[0009] Optionally, collecting multi-source sensor time series data during operation of the smelting furnace includes:

[0010] The melting temperature signal in the multi-source sensor time series data is obtained synchronously by using a preset dual-wavelength infrared temperature measurement and an embedded thermocouple, wherein the dual-wavelength infrared temperature measurement points are arranged at equal intervals based on the crucible axis of the melting furnace;

[0011] The crucible vibration signal in the multi-source sensor time series data is collected by a three-axis acceleration sensor fixed to the bearing base of the smelting furnace.

[0012] Optionally, extracting the temperature signal gradient feature of the smelting temperature signal includes:

[0013] generating a temperature signal sequence of a smelting furnace according to the dual-wavelength infrared temperature measurement points;

[0014] Calculating the first-order difference absolute value of the smelting temperature signal in a continuous time window through the temperature signal sequence;

[0015] Perform sliding average filtering on the first-order difference absolute value to generate the temperature signal gradient feature.

[0016] Optionally, decomposing the crucible vibration signal into preset frequency bands by using a preset wavelet packet decomposition includes:

[0017] Construct decomposition filter banks using preset wavelet basis functions;

[0018] determining the number of decomposition layers of the crucible vibration signal according to the rated vibration spectrum characteristics of the smelting furnace;

[0019] The crucible vibration signal is subjected to wavelet packet decomposition according to the number of decomposition layers by means of the analysis filter bank to obtain preset frequency bands of equal width.

[0020] Optionally, calculating the product of the standard deviation of the energy value of each frequency band and the energy ratio of adjacent frequency bands to generate the frequency domain energy gradient feature vector includes:

[0021] Updating the energy ratio of the adjacent frequency bands according to a preset exponential weight;

[0022] When the updated energy ratio of the adjacent frequency bands exceeds the preset transition threshold, a logarithmic operation is performed on the updated energy ratio of the adjacent frequency bands to obtain a target energy ratio of the adjacent frequency bands;

[0023] When the updated energy ratio of the adjacent frequency bands does not exceed the preset transition threshold, the updated energy ratio of the adjacent frequency bands is used as the target energy ratio of the adjacent frequency bands;

[0024] The product of the standard value of the energy of each frequency band and the target energy ratio of the adjacent frequency bands is calculated to generate a frequency domain energy gradient feature vector.

[0025] Optionally, generating a multidimensional feature matrix according to the signal mean feature, the temperature signal gradient feature, and the frequency domain energy gradient feature vector includes:

[0026] Mapping the frequency domain energy gradient eigenvector to a preset vibration dimension;

[0027] Mapping the signal mean feature to a preset electrical dimension;

[0028] Mapping the temperature signal gradient characteristics to a preset thermodynamic dimension;

[0029] A multi-dimensional feature matrix is ​​generated according to a combination of the vibration dimension, the electrical dimension, and the thermodynamic dimension feature dimensions.

[0030] Optionally, before inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output the fault boundary of the smelting furnace, the method further includes:

[0031] Collecting historical multi-source sensor time series data and extracting fault labels corresponding to the historical multi-source sensor time series data;

[0032] generating a historical multi-dimensional feature matrix based on the historical multi-source sensor time series data;

[0033] Using the historical multidimensional feature matrix and the fault label as historical fault samples;

[0034] The preset gradient boosting decision tree is iteratively trained using the historical fault samples to obtain a lightweight gradient boosting tree model.

[0035] Optionally, taking the historical multidimensional feature matrix and the fault label as historical fault samples includes:

[0036] Extract the energy attenuation waveform characteristics corresponding to the draft tube in the smelting furnace;

[0037] Calculating the sliding standard deviation of the energy values ​​corresponding to the energy decay waveform characteristics within the continuous time window;

[0038] When the sliding standard deviation exceeds a preset sliding threshold, marking the energy decay waveform feature as an additional feature, and adding the additional feature to the historical multidimensional feature matrix;

[0039] Using the historical multidimensional feature matrix after adding the additional features as input sample data, and using the fault label as output sample data;

[0040] A historical fault sample is generated according to the input sample data and the output sample data.

[0041] Optionally, inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output the fault boundary of the smelting furnace includes:

[0042] Inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output a failure probability value of the smelting furnace;

[0043] Identifying the real-time operating time and the rated operating cycle of the smelting furnace, and calculating a dynamic time adjustment factor of the smelting furnace based on the real-time operating time and the rated operating cycle;

[0044] Determining a failure boundary value of the smelting furnace according to the failure probability value and the dynamic duration adjustment factor;

[0045] The fault boundary of the smelting furnace is determined according to the fault boundary value.

[0046] Optionally, the warning instruction corresponding to the graded warning signal triggering the smelting furnace includes:

[0047] comparing the fault boundary with a preset multi-level threshold interval, wherein the multi-level threshold interval includes at least two ordered threshold boundaries; triggering a first-level warning signal of a hierarchical warning signal of the smelting furnace when the fault boundary is within the preset first-level threshold interval, and determining a parameter adjustment instruction based on the first-level warning signal;

[0048] When the fault boundary is within a preset second-level threshold range, a second-level warning signal in the graded warning signal of the smelting furnace is triggered, and a safety coordinated stop instruction is determined according to the second-level warning signal;

[0049] The parameter adjustment instruction and the safety coordinated stop instruction are determined as early warning instructions.

[0050] The embodiment of the present invention uses a synchronous clock module to achieve timestamp alignment of the melting temperature signal, crucible vibration signal, and electrode current signal, ensuring that data collected by different sensors at the same time point can be directly correlated and analyzed; by extracting the gradient characteristics of the temperature signal and the mean characteristics of the electrode current signal, and combining wavelet packet decomposition to extract the frequency domain energy gradient feature vector of the crucible vibration signal, a multi-dimensional characterization of the thermal state, electrical state, and structural vibration state of the melting furnace is achieved; a pre-trained lightweight gradient boosting tree model is used to improve data processing efficiency while ensuring recognition accuracy by reducing the number of decision trees and simplifying node splitting rules; the warning level is determined based on the fault boundary and the dynamic duration adjustment factor, and targeted intervention instructions are matched. Therefore, the melting furnace system fault warning method based on big data proposed in the present invention can solve the problem of low accuracy when performing melting furnace system fault warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1A schematic flow chart of a big data-based smelting furnace system fault warning method provided in one embodiment of the present invention.

[0052] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0053] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] The embodiment of the present application provides a smelting furnace system fault early warning method based on big data. The execution subject of the smelting furnace system fault early warning method based on big data includes but is not limited to at least one of the electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiment of the present application. In other words, the smelting furnace system fault early warning method based on big data can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0055] Reference Figure 1 FIG. 1 is a flow chart of a method for early warning of a smelting furnace system failure based on big data according to an embodiment of the present invention. In this embodiment, the method for early warning of a smelting furnace system failure based on big data includes:

[0056] S1. Collecting multi-source sensor time series data during the operation of a smelting furnace, wherein the multi-source sensor time series data includes a smelting temperature signal, a crucible vibration signal, and an electrode current signal.

[0057] In an embodiment of the present invention, the multi-source sensor time series data is a comprehensive data set collected by multiple types of sensors at continuous time points, which can reflect the operating status of the smelting furnace; the smelting temperature signal is a signal used to characterize the temperature of the smelting material inside the smelting furnace, and can reflect the thermal state of the smelting process; the crucible vibration signal is a signal used to characterize the vibration state of the crucible during the smelting process, and can reflect the structural stability of the crucible and the movement state of the material; the electrode current signal is a signal used to characterize the current size of the smelting furnace electrode when it is working, and can reflect the energy consumption and working state of the electrode.

[0058] In the embodiment of the present invention, the collecting of multi-source sensor time series data during operation of the smelting furnace includes:

[0059] The melting temperature signal in the multi-source sensor time series data is obtained synchronously by using a preset dual-wavelength infrared temperature measurement and an embedded thermocouple, wherein the dual-wavelength infrared temperature measurement points are arranged at equal intervals based on the crucible axis of the melting furnace;

[0060] The crucible vibration signal in the multi-source sensor time series data is collected by a three-axis acceleration sensor fixed to the bearing base of the smelting furnace.

[0061] Specifically, equipment is deployed to collect melting temperature signals, using dual-wavelength infrared temperature measurement and embedded thermocouples for simultaneous acquisition. In practice, dual-wavelength infrared temperature measurement points are evenly spaced along the crucible axis of the melting furnace. For example, on a one-meter-long crucible, five temperature measurement points are set every 20 centimeters. Furthermore, embedded thermocouples are embedded into the inner wall of the melting furnace, corresponding to the temperature measurement points. The thermocouples' temperature measurement ends overlap with the infrared temperature measurement points' monitoring areas, ensuring that both capture temperature data from the same area. Secondly, sensors are deployed to collect crucible vibration signals. A triaxial accelerometer is mounted on the crucible's bearing base, positioned in the middle to minimize interference from the mounting position on the vibration signal. The sensor's X, Y, and Z axes correspond to the crucible's radial, axial, and vertical directions, respectively, to comprehensively capture vibration information. Thirdly, sensors are deployed to collect electrode current signals. These sensors are connected in series to the electrode's power supply circuit, for example, in the cable between the electrode and the power control cabinet, to ensure accurate acquisition of current changes during electrode operation. Finally, the timestamp alignment of multi-source sensor time series data is achieved through the synchronous clock module. The synchronous clock module sends a unified time signal to each sensor, so that each sensor marks the same timestamp when collecting data. For example, a timestamp is generated every 10 milliseconds to ensure that the data collected by different sensors at the same time point can accurately correspond.

[0062] Specifically, based on the melting temperature signal, crucible vibration signal, and electrode current signal, the problem of state analysis error caused by one-sided single sensor data and asynchronous multi-source data in the existing technology is solved. Through multi-source synchronous acquisition, complete and consistent basic data is provided for subsequent feature extraction.

[0063] Furthermore, the collected multi-source sensor time series data is the original input for extracting the temperature signal gradient features and the electrode current signal mean features. Without synchronization of the multi-source sensor time series data, feature extraction will produce deviations due to inconsistent data benchmarks.

[0064] S2. Extracting the temperature signal gradient feature of the smelting temperature signal, and extracting the signal mean feature of the electrode current signal.

[0065] In the embodiment of the present invention, the temperature signal gradient feature is the amount of change in the melting temperature signal per unit time, which can reflect the rate of change of the melting temperature and can be used to determine whether the temperature change is stable.

[0066] In an embodiment of the present invention, extracting the temperature signal gradient feature of the smelting temperature signal includes:

[0067] generating a temperature signal sequence of a smelting furnace according to the dual-wavelength infrared temperature measurement points;

[0068] Calculating the first-order difference absolute value of the smelting temperature signal in a continuous time window through the temperature signal sequence;

[0069] Perform sliding average filtering on the first-order difference absolute value to generate the temperature signal gradient feature.

[0070] In detail, for extracting the temperature signal gradient characteristics of the smelting temperature signal, first, a temperature signal sequence of the smelting furnace is generated based on the synchronously collected dual-wavelength infrared temperature measurement and thermocouple data, and then the temperature signal data collected by the infrared temperature measurement and the thermocouple at the same time stamp are fused. For example, a weighted average method is used, with a weight of 0.6 for infrared data and a weight of 0.4 for thermocouple data, to obtain the temperature value at each time point, and then arrange them in chronological order to form a continuous temperature signal sequence, such as one data point every 10 milliseconds. Continuous collection for 10 minutes can obtain a sequence consisting of 60,000 data points.

[0071] Specifically, the absolute value of the first-order difference within a continuous time window is calculated. The time window size is selected as 5 data points (i.e., 50 milliseconds). The temperature difference between two adjacent time points within the time window is calculated, and then the absolute value is taken to obtain a sequence of first-order difference absolute values ​​of the smelting temperature signal within the continuous time window. For example, if the temperature signal sequence is [1200°C, 1202°C, 1205°C, 1204°C, 1206°C], the absolute value of the first-order difference is [2°C, 3°C, 1°C, 2°C]. A sliding average filter is performed on the sequence of first-order difference absolute values. The sliding window size is set to 3 data points. The arithmetic average of the absolute values ​​of the first-order differences within the time window is taken to obtain the temperature signal gradient feature. For example, the absolute value of the first-order difference after sliding average is [(2+3+1) / 3=2°C / 50ms, (3+1+2) / 3=2°C / 50ms], i.e., 0.04°C / ms and 0.04°C / ms.

[0072] In the embodiment of the present invention, the signal mean feature is the average value of the electrode current signal within a certain time range, which can reflect the overall level of the electrode current within the time period and can be used to determine whether the electrode working state is stable.

[0073] In detail, a preset time window is set, for example, the window size is 1 minute and the sliding step is 30 seconds, that is, a 1-minute current signal segment is intercepted every 30 seconds, and the arithmetic mean of the electrode current signal in each time window is calculated. For example, the electrode current signal in a certain time window is [500A, 502A, 498A, ..., 501A] (a total of 600 data points, 10 milliseconds / point). These current values ​​are added and divided by the number of data points to obtain the current average value of the window, which is the signal mean feature corresponding to the window.

[0074] Specifically, by extracting gradient and mean features, the problem in the existing technology that only using the original temperature and current signals is difficult to reflect their changing trends is solved. It can more intuitively reflect the temperature change rate and the overall current level, providing a more effective feature basis for subsequent fault analysis.

[0075] Furthermore, the extracted temperature signal gradient features and signal mean features will serve as the basic data for generating a multidimensional feature matrix. The accuracy of feature extraction directly affects the quality of the subsequent matrix.

[0076] S3. Decompose the crucible vibration signal into preset frequency bands by using a preset wavelet packet decomposition, and calculate the product of the standard deviation of the energy value of each frequency band and the energy ratio of the adjacent frequency bands to generate a frequency domain energy gradient feature vector.

[0077] In the embodiments of the present invention, wavelet packet decomposition is a signal processing method that can perform detailed analysis of signals in both the time and frequency domains. It decomposes the original signal into multiple independent frequency bands, facilitating the extraction of signal characteristics in different frequency bands. The preset frequency band is a signal band with a specific frequency range obtained after wavelet packet decomposition. This frequency range is determined based on the characteristics of the crucible vibration signal.

[0078] In an embodiment of the present invention, decomposing the crucible vibration signal into preset frequency bands by using a preset wavelet packet decomposition includes:

[0079] Construct decomposition filter banks using preset wavelet basis functions;

[0080] determining the number of decomposition layers of the crucible vibration signal according to the rated vibration spectrum characteristics of the smelting furnace;

[0081] The crucible vibration signal is subjected to wavelet packet decomposition according to the number of decomposition layers by means of the analysis filter bank to obtain preset frequency bands of equal width.

[0082] Specifically, the crucible vibration signal is decomposed into preset frequency bands through a preset wavelet packet decomposition. First, a suitable wavelet basis function is selected to construct a decomposition filter bank, such as the db4 wavelet basis function, which has good time-frequency localization characteristics and is suitable for processing vibration signals. Then, the number of decomposition layers is determined according to the rated vibration spectrum characteristics of the melting furnace. For example, the rated vibration of a certain melting furnace is mainly concentrated in the range of 10Hz-500Hz. In order to ensure that the decomposed frequency band can cover this range and have sufficient frequency resolution, the number of decomposition layers is set to 4, and 16 equal-width frequency bands can be obtained (the width of each frequency band is 500Hz / 16≈31.25Hz). Then, through the constructed decomposition filter bank, the crucible vibration signal is subjected to wavelet packet decomposition according to the determined number of decomposition layers to obtain signals of 16 preset frequency bands, each of which corresponds to a specific frequency range (such as the first frequency band is 10Hz-41.25Hz, the second is 41.25Hz-72.5Hz, and so on).

[0083] In the embodiment of the present invention, the frequency domain energy gradient feature vector is a vector used to comprehensively characterize the energy change characteristics of each preset frequency band, and can reflect the subtle changes of the crucible vibration in the frequency domain.

[0084] In the embodiment of the present invention, calculating the product of the standard deviation of the energy value of each frequency band and the energy ratio of adjacent frequency bands to generate the frequency domain energy gradient feature vector includes:

[0085] Updating the energy ratio of the adjacent frequency bands according to a preset exponential weight;

[0086] When the updated energy ratio of the adjacent frequency bands exceeds the preset transition threshold, a logarithmic operation is performed on the updated energy ratio of the adjacent frequency bands to obtain a target energy ratio of the adjacent frequency bands;

[0087] When the updated energy ratio of the adjacent frequency bands does not exceed the preset transition threshold, the updated energy ratio of the adjacent frequency bands is used as the target energy ratio of the adjacent frequency bands;

[0088] The product of the standard value of the energy of each frequency band and the target energy ratio of the adjacent frequency bands is calculated to generate a frequency domain energy gradient feature vector.

[0089] Specifically, the standard deviation of the energy value of each frequency band is calculated by multiplying the energy ratio of adjacent frequency bands to generate a frequency domain energy gradient feature vector. The energy value of each preset frequency band is then calculated, and the square sum of the signals in each frequency band is performed to obtain the energy value of that frequency band. The standard deviation of the energy value of each frequency band is then calculated. For example, for a frequency band energy value sequence [5, 7, 6, 8, 5], the average value is first calculated as 6.2. The squares of the differences between each energy value and the average value are then averaged to obtain the variance, and the square root is then taken to obtain the standard deviation. Next, the energy ratio of adjacent frequency bands is calculated: that is, the ratio of the energy value of the nth frequency band to the energy value of the n+1th frequency band.

[0090] Specifically, the energy ratios of adjacent frequency bands are updated according to a preset exponential weight. For example, the exponential weight is set to 0.8, and the updated ratio = original ratio × 0.8 + 0.2 × historical average ratio (the historical average ratio is the average ratio of the past five adjacent frequency bands). It is determined whether the updated energy ratios of adjacent frequency bands exceed a preset transition threshold (for example, set to 0.9). If so, a logarithmic operation is performed (taking the natural logarithm as an example) to obtain a target energy ratio of the adjacent frequency bands (ln1.0=0). If not, the ratio is 0.8, and the target energy ratio of the adjacent frequency bands is directly used. Finally, the product of the standard deviation of the energy value of each frequency band and the target energy ratio of the corresponding adjacent frequency band is calculated. The product results of the 16 frequency bands are arranged in frequency band order to generate a frequency domain energy gradient feature vector (for example, [1.2×0.8, 0.9×0.7,..., 1.5×0]), where the preset exponential weight and transition threshold are both system presets, which can be determined through experience or experiments based on the distribution characteristics of the training data and the requirements for recognition accuracy in the actual crucible vibration signal scenario. The preset exponential weight is usually set to 0.1-0.9, which is used to distinguish the accuracy of the energy ratio of adjacent frequency bands; the transition threshold is a critical value determined based on the frequency domain energy characteristic analysis of a large number of historical crucible vibration signals. The optimized threshold is verified through experiments, and different faults are simulated on the actual melting furnace (such as slight loosening of the crucible by human intervention), the vibration signal is collected and the energy ratio of adjacent frequency bands is calculated. The value is compared with the ratio of the initial threshold and the actual fault, and the threshold is adjusted to a value that can accurately distinguish normal fluctuations from fault transitions. Therefore, the energy ratio of adjacent frequency bands mostly fluctuates gently during normal operation, while a sudden jump will occur before a fault occurs (such as the crucible structure begins to loosen). The transition threshold clearly distinguishes the jump from the gentle fluctuation by setting a critical value. The ratio exceeding the threshold is given a more significant feature weight after logarithmic operation, and the ratio not exceeding the threshold is kept at the original value, so that the fault-related energy mutation is more prominent in the feature vector, solving the problem of feature ambiguity caused by not distinguishing the fluctuation type.

[0091] For example, in the processing of a crucible vibration signal, 16 frequency bands are obtained after wavelet packet decomposition, among which the energy value sequence of the third frequency band is [12, 15, 13, 14, 16], and its standard deviation is calculated to be 1.4; the original energy ratio of this frequency band to the fourth frequency band is 15 / 12=1.25, which is 1.1 after exponential weight update, exceeding the transition threshold of 0.9, and is 0.095 after logarithmic operation. The product of the two is 1.4×0.095≈0.133, which is the component of the third frequency band in the eigenvector.

[0092] Furthermore, wavelet packet decomposition and frequency domain energy feature extraction solve the problem in the existing technology that it is difficult to capture subtle changes in crucible vibration only through time domain features. It can effectively reflect the energy change pattern of vibration signals at different frequencies and improve the ability to identify potential crucible faults (such as loose structure and material abnormalities).

[0093] Furthermore, the frequency domain energy gradient feature vector, along with the extracted signal mean feature and temperature signal gradient feature, serves as the input to a multidimensional feature matrix. The quality of these features directly impacts the accuracy of subsequent fault boundary calculations. S4. Generate a multidimensional feature matrix based on the signal mean feature, the temperature signal gradient feature, and the frequency domain energy gradient feature vector.

[0094] In an embodiment of the present invention, the multidimensional feature matrix is ​​a matrix that integrates signal mean features, temperature signal gradient features, and frequency domain energy gradient feature vectors. Its rows represent different time points, and its columns represent different feature dimensions, which can comprehensively reflect the operating characteristics of the smelting furnace in terms of electricity, thermodynamics, and vibration.

[0095] In an embodiment of the present invention, a multidimensional feature matrix is ​​generated according to the signal mean feature, the temperature signal gradient feature, and the frequency domain energy gradient feature vector, including:

[0096] Mapping the frequency domain energy gradient eigenvector to a preset vibration dimension;

[0097] Mapping the signal mean feature to a preset electrical dimension;

[0098] Mapping the temperature signal gradient characteristics to a preset thermodynamic dimension;

[0099] A multi-dimensional feature matrix is ​​generated according to a combination of the vibration dimension, the electrical dimension, and the thermodynamic dimension feature dimensions.

[0100] In detail, the vibration dimension is used to characterize the characteristic dimension corresponding to the frequency domain energy gradient eigenvector; the electrical dimension is used to characterize the characteristic dimension corresponding to the signal mean feature; and the thermodynamic dimension is used to characterize the characteristic dimension corresponding to the temperature signal gradient feature.

[0101] Specifically, the frequency domain energy gradient feature vector is mapped to a preset vibration dimension. For example, if the frequency domain energy gradient feature vector contains 16 components, the vibration dimension is set to 16 dimensions, with each component corresponding to one dimension. Secondly, the signal mean feature is mapped to a preset electrical dimension. Since the signal mean feature is a single value, the electrical dimension is set to 1 dimension. Thirdly, the temperature signal gradient feature is mapped to a preset thermodynamic dimension. Similarly, since the temperature signal gradient feature is a single value, the thermodynamic dimension is set to 1 dimension. Finally, a multidimensional feature matrix is ​​generated based on the feature dimension combination of the vibration dimension, electrical dimension, and thermodynamic dimension. For example, at a certain time point, the 16 components of the vibration dimension are [0.1, 0.2, ..., 0.15], the electrical dimension component is 500 A (mean), and the thermodynamic dimension component is 0.04 °C / ms. Then the matrix row vector corresponding to this time point is [0.1, 0.2, ..., 0.15, 500, 0.04]. The row vectors of consecutive time points are arranged in chronological order to form a multidimensional feature matrix.

[0102] For example, within a 10-minute monitoring period, a feature combination is generated every 30 seconds, and features of 20 time points are generated in total. The features of each time point contain 16 vibration dimension components, 1 electrical dimension component and 1 thermodynamic dimension component. The generated multidimensional feature matrix is ​​a matrix with 20 rows and 18 columns.

[0103] Furthermore, by generating a multi-dimensional feature matrix, the multi-dimensional features are integrated to solve the problem of different types of features being scattered and difficult to comprehensively analyze in the existing technology, providing structured and comprehensive data for subsequent model input, so that the model can comprehensively learn the correlation between the operating status of the smelting furnace and faults.

[0104] Furthermore, the multidimensional feature matrix is ​​a direct input to the lightweight gradient boosting tree model. The matrix's integrity and feature relevance directly impact the accuracy of the fault boundary output by the model. S5: Input the multidimensional feature matrix into the pre-trained lightweight gradient boosting tree model to output the fault boundary of the smelting furnace.

[0105] In an embodiment of the present invention, the lightweight gradient boosting tree model is a simplified and optimized gradient boosting decision tree model. By reducing the number of decision trees and the complexity of nodes, the computing speed is improved while ensuring a certain accuracy, and it is suitable for real-time monitoring scenarios; the fault boundary is a critical characteristic boundary used to divide the normal operation state and the fault state of the smelting furnace. When the operating characteristics of the smelting furnace exceed this boundary, it can be determined that there is a fault risk.

[0106] In an embodiment of the present invention, before inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output the fault boundary of the smelting furnace, the method further includes:

[0107] Collecting historical multi-source sensor time series data and extracting fault labels corresponding to the historical multi-source sensor time series data;

[0108] generating a historical multi-dimensional feature matrix based on the historical multi-source sensor time series data;

[0109] Using the historical multidimensional feature matrix and the fault label as historical fault samples;

[0110] The preset gradient boosting decision tree is iteratively trained using the historical fault samples to obtain a lightweight gradient boosting tree model.

[0111] Specifically, historical multi-source sensor time series data is collected, covering sensor data from both normal operation of the smelting furnace and various fault conditions (such as electrode overload, crucible breakage, and temperature runaway). Furthermore, corresponding fault labels (such as "normal," "electrode overload," and "crucible breakage") are added to this data based on historical operation records and manual annotation. Secondly, a historical multi-dimensional feature matrix is ​​generated based on the historical multi-source sensor time series data. This involves extracting the signal mean feature, temperature signal gradient feature, and frequency domain energy gradient feature vectors from the historical data and combining them into a historical multi-dimensional feature matrix.

[0112] Specifically, historical fault samples refer to the training data of the gradient boosting decision tree, which includes the correspondence between the historical multi-dimensional feature matrix and the fault labels.

[0113] In the embodiment of the present invention, the using the historical multidimensional feature matrix and the fault label as historical fault samples includes:

[0114] Extract the energy attenuation waveform characteristics corresponding to the draft tube in the smelting furnace;

[0115] Calculating the sliding standard deviation of the energy values ​​corresponding to the energy decay waveform characteristics within the continuous time window;

[0116] When the sliding standard deviation exceeds a preset sliding threshold, marking the energy decay waveform feature as an additional feature, and adding the additional feature to the historical multidimensional feature matrix;

[0117] Using the historical multidimensional feature matrix after adding the additional features as input sample data, and using the fault label as output sample data;

[0118] A historical fault sample is generated according to the input sample data and the output sample data.

[0119] Specifically, historical fault samples are processed by extracting the energy attenuation waveform features corresponding to the draft tube in the smelting furnace (collected by additionally deployed acoustic wave sensors); the sliding standard deviation of the energy values ​​corresponding to the energy attenuation waveform features within a continuous time window (such as 5 minutes) is calculated. For example, if the energy values ​​in a certain window are [10, 12, 9, 11, 13], its sliding standard deviation is 1.5. The sliding standard deviation is to first calculate the arithmetic mean of all data points in the window; then calculate the difference between each data point and the arithmetic mean, and square each difference; then calculate the arithmetic mean of these square values ​​to obtain the average of the square differences; finally, perform a square root operation on the average of the square differences, and the result is the standard deviation of the data in the sliding window. When the sliding window moves to a new position at a set step size, the above calculation process is repeated for the data within the new window, sequentially obtaining the standard deviation of each window and forming a sliding standard deviation sequence. The sliding standard deviation dynamically reflects the degree of data dispersion within a continuous time period. By sliding the window, data fluctuations can be tracked in real time. Compared to standard deviation calculations within fixed windows, this method is more adaptable to the temporal characteristics of data and provides a continuous quantitative basis for determining whether abnormal data fluctuations occur. For example, in monitoring the energy decay of a smelting furnace draft tube, changes in the sliding standard deviation can promptly reflect the stability of the energy decay. A sudden increase in the sliding standard deviation indicates a possible draft tube anomaly.

[0120] Specifically, when the sliding standard deviation exceeds a preset sliding threshold (e.g., 2.0), the energy decay waveform feature is marked as an additional feature and added to the historical multidimensional feature matrix. The historical multidimensional feature matrix with the added features is used as input sample data, and the fault label is used as output sample data to generate historical fault samples. Finally, the preset gradient boosting decision tree is iteratively trained using the historical fault samples. For example, the number of decision trees is initially set to 50, and then reduced to 20 through iterative optimization (retaining nodes corresponding to important features), resulting in a lightweight gradient boosting tree model. The preset sliding threshold is system-preset and can be determined through experience or experimentation based on the distribution characteristics of the training data and the recognition accuracy requirements of the actual energy decay waveform features. The preset sliding threshold is typically set to 0.1-0.9 and is used to distinguish whether the energy decay waveform feature should be added to the historical multidimensional feature matrix.

[0121] In an embodiment of the present invention, inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output the fault boundary of the smelting furnace includes:

[0122] Inputting the multidimensional feature matrix into a pre-trained lightweight gradient boosting tree model to output a failure probability value of the smelting furnace;

[0123] Identifying the real-time operating time and the rated operating cycle of the smelting furnace, and calculating a dynamic time adjustment factor of the smelting furnace based on the real-time operating time and the rated operating cycle;

[0124] Determining a failure boundary value of the smelting furnace according to the failure probability value and the dynamic duration adjustment factor;

[0125] The fault boundary of the smelting furnace is determined according to the fault boundary value.

[0126] In detail, the multidimensional feature matrix is ​​input into the pre-trained lightweight gradient boosting tree model to output the fault boundary of the melting furnace. The multidimensional feature matrix is ​​then input into the model, and the model calculates through an internal decision tree combination to output the failure probability value of the melting furnace (ranging from 0 to 1, with the larger the value, the higher the failure risk). Secondly, the real-time operating time and rated operating cycle of the melting furnace are identified. For example, if the real-time operating time is 8 hours and the rated operating cycle is 12 hours, the dynamic time adjustment factor = real-time operating time / rated operating cycle = 8 / 12≈0.67. Thirdly, the fault boundary value is determined based on the fault probability value and the dynamic time adjustment factor. For example, if the fault probability is 0.7, the boundary value = 0.7×0.67≈0.47. Finally, the fault boundary is determined based on the boundary value, that is, when the fault probability corresponding to the real-time feature of the melting furnace exceeds 0.47 after dynamic adjustment. When the fault probability is adjusted, it is judged to be close to the fault boundary. By dynamically adjusting the fault probability and comparing it with the boundary value, the judgment of the fault boundary is more in line with the actual operating status of the equipment. The longer the operating time (the higher the equipment fatigue), the larger the adjustment factor, and the easier it is to trigger the judgment of being close to the fault boundary under the same fault probability, thereby improving the pertinence and accuracy of fault warning.

[0127] For example, the rated operating cycle of a smelting furnace is 12 hours. When it runs in real time for 11 hours (dynamic duration adjustment factor = 0.92), the failure probability value corresponding to the multi-dimensional feature matrix of the input model is 0.6, then the failure boundary value = 0.6×0.92≈0.57. At this time, if the real-time calculated failure probability adjustment value is 0.57 (exceeding 0.47), it is judged to be close to the failure boundary.

[0128] Furthermore, the lightweight model improves computing efficiency while ensuring recognition accuracy, and solves the problems of slow computing speed and difficulty in adapting to real-time monitoring of traditional models. The dynamic adjustment of the fault boundary takes into account the impact of the operating time of the smelting furnace on the fault risk, making fault judgment more in line with actual operating conditions.

[0129] Furthermore, the fault boundary forms the basis for graded warnings. The quality of model training and the accuracy of boundary calculation directly determine the accuracy of the warning. S6. When the fault boundary reaches the preset trigger condition, the warning instruction corresponding to the graded warning signal of the smelting furnace is triggered, and the warning instruction is used to issue a fault warning for the smelting furnace.

[0130] In an embodiment of the present invention, the preset trigger conditions are set based on the problem of "fuzzy fault warning classification and insufficiently targeted intervention measures" existing in the smelting furnace. Specifically, the preset trigger conditions are multi-level threshold intervals divided according to the safety thresholds corresponding to different fault types (such as electrode overload, temperature runaway, and crucible breakage).

[0131] In this embodiment of the present invention, the graded warning signals are classified into different levels based on the severity of the fault boundary, used to distinguish high from low fault risks. The warning instructions are generated based on the graded warning signals and are used to control the smelting furnace to perform corresponding fault processing, thereby enabling timely intervention in the smelting furnace.

[0132] In an embodiment of the present invention, the warning instructions corresponding to the graded warning signals triggering the smelting furnace include:

[0133] comparing the fault boundary with a preset multi-level threshold interval, wherein the multi-level threshold interval includes at least two ordered threshold boundaries; triggering a first-level warning signal of a hierarchical warning signal of the smelting furnace when the fault boundary is within the preset first-level threshold interval, and determining a parameter adjustment instruction based on the first-level warning signal;

[0134] When the fault boundary is within a preset second-level threshold range, a second-level warning signal in the graded warning signal of the smelting furnace is triggered, and a safety coordinated stop instruction is determined according to the second-level warning signal;

[0135] The parameter adjustment instruction and the safety coordinated stop instruction are determined as early warning instructions.

[0136] In detail, a multi-level threshold interval is preset, for example, a two-level threshold interval is set, the first-level threshold interval is a fault boundary value of 0.3-0.6 (corresponding to a mild fault risk. Such as a slight electrode overload, a small temperature fluctuation, and other mild faults), and the second-level threshold interval is a fault boundary value greater than 0.6 (corresponding to a severe fault risk, such as a severe electrode overload, abnormal crucible vibration, and other severe faults). Secondly, the output fault boundary is compared with the preset multi-level threshold interval: when the fault boundary is in the first-level threshold interval, a first-level warning signal is triggered, and a parameter adjustment instruction is determined according to the signal. For example, the electrode current is reduced to 80% of the rated value, and the heating power is adjusted to stabilize the temperature within the target value ±5°C to mitigate the risk of failure. When the fault boundary is in the second-level threshold range, the second-level warning signal is triggered. Based on the signal, the safety coordinated stop instruction is determined and a sequential shutdown operation is executed. For example, the electrode current is gradually reduced to 0, the heating system is shut down, and the cooling device is started to ensure the safe shutdown of the smelting furnace. The threshold range is determined based on the correlation analysis of the fault boundary and the actual fault consequences in 500 sets of historical fault data. For example, historical data shows that when the fault boundary reaches 0.5, 80% of them are minor faults that can be solved by parameter adjustment; when it reaches 0.7, 90% require emergency shutdown.

[0137] For example, when the fault boundary is within the first-level threshold range, a first-level warning instruction is triggered. The technical carrier of this instruction is the control system of the smelting furnace. The specific operations include adjusting the electrode current to 80% of the rated value (such as adjusting it to 800A when the rated current is 1000A), and stabilizing the smelting temperature within the target value ±5°C range (such as controlling it within 1595-1605°C when the target temperature is 1600°C). The adjustment signal is transmitted to the electrode power supply device and heating device through the PLC module of the control system. For example, when the fault margin is 0.4 (in the first-level range), the PLC module sends a command to the power supply unit to "adjust current to 800A" and to the heating unit to "maintain power at 85kW." It then monitors the adjusted data in real time to ensure that the current and temperature remain within the target range within three minutes. When the fault margin reaches the second-level threshold, a second-level warning command is triggered, which also directs the control system to execute a sequenced shutdown: reducing the electrode current from its current value to 300A within 10 seconds; shutting off the heat source input to the heating unit to stop the temperature increase; activating the water cooling system to reduce the crucible temperature at a rate of 5°C / minute; and opening the emergency smoke exhaust system to exhaust the furnace fumes. For example, when the fault margin is 0.7 (in the second-level range), the control system executes the above sequence, providing status feedback to the central control room after each step, ensuring a safe and orderly shutdown process.

[0138] Furthermore, graded warnings and targeted instructions can solve the problem of a single warning method and the inability to perform graded treatment according to the severity of the fault. Timely intervention can be made in the early stages of the fault to prevent the fault from expanding. At the same time, safe shutdown can be ensured in the event of a serious fault to reduce losses.

[0139] Furthermore, the output fault boundary is the basis for triggering the early warning, and the early warning instruction is the final response to the entire fault warning process, realizing closed-loop processing of the melting furnace failure. Fault warning is the process of prompting and handling the potential failure risks of the melting furnace by executing the early warning instruction, aiming to avoid the expansion of the fault or the occurrence of safety accidents.

[0140] The embodiment of the present invention uses a synchronous clock module to achieve timestamp alignment of the melting temperature signal, crucible vibration signal, and electrode current signal, ensuring that data collected by different sensors at the same time point can be directly correlated and analyzed; by extracting the gradient characteristics of the temperature signal and the mean characteristics of the electrode current signal, and combining wavelet packet decomposition to extract the frequency domain energy gradient feature vector of the crucible vibration signal, a multi-dimensional characterization of the thermal state, electrical state, and structural vibration state of the melting furnace is achieved; a pre-trained lightweight gradient boosting tree model is used to improve data processing efficiency while ensuring recognition accuracy by reducing the number of decision trees and simplifying node splitting rules; the warning level is determined based on the fault boundary and the dynamic duration adjustment factor, and targeted intervention instructions are matched. Therefore, the melting furnace system fault warning method based on big data proposed in the present invention can solve the problem of low accuracy when performing melting furnace system fault warning.

[0141] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0142] Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the claims are intended to be embraced therein. Any reference to a figure in a claim should not be construed as limiting the claim to which it relates.

[0143] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results.

[0144] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or systems recited in a system claim may also be implemented by a single unit or system through software or hardware. Terms such as "first" and "second" are used to indicate names and do not imply any particular order.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smelting furnace system fault early warning method based on big data, characterized in that: The method comprises: Collect multi-source sensor time series data during the operation of the smelting furnace, where the multi-source sensor time series data includes the smelting temperature signal, the crucible vibration signal, and the electrode current signal. The time stamps of the multi-source sensor time series data are aligned through the synchronization clock module. The synchronization clock module sends a unified time signal to each sensor, so that each sensor is marked with the same time stamp when collecting data, ensuring that the data collected by different sensors at the same time point can accurately correspond. extracting a temperature signal gradient feature of the smelting temperature signal, and extracting a signal mean feature of the electrode current signal; Decomposing the crucible vibration signal into preset frequency bands by using a preset wavelet packet decomposition, and calculating the product of the standard deviation of the energy value of each frequency band and the energy ratio of the adjacent frequency bands to generate a frequency domain energy gradient feature vector; generating a multidimensional feature matrix based on the signal mean feature, the temperature signal gradient feature, and the frequency domain energy gradient feature vector; The multidimensional feature matrix is ​​input into a pre-trained lightweight gradient boosting tree model to output a failure probability value of the smelting furnace; the real-time operating time and the rated operating cycle of the smelting furnace are identified, and the dynamic time adjustment factor of the smelting furnace is calculated according to the real-time operating time and the rated operating cycle; the failure boundary value of the smelting furnace is determined according to the failure probability value and the dynamic time adjustment factor; the failure boundary of the smelting furnace is determined according to the failure boundary value; when the fault boundary reaches a preset trigger condition, the warning instruction corresponding to the graded warning signal of the smelting furnace is triggered, and the warning instruction is used to perform a fault warning on the smelting furnace.

2. The method for early warning of smelting furnace system failure based on big data according to claim 1, characterized in that: The multi-source sensor time series data collected during the operation of the smelting furnace includes: The melting temperature signal in the multi-source sensor time series data is obtained synchronously by using a preset dual-wavelength infrared temperature measurement and an embedded thermocouple, wherein the dual-wavelength infrared temperature measurement points are arranged at equal intervals based on the crucible axis of the melting furnace; The crucible vibration signal in the multi-source sensor time series data is collected by a three-axis acceleration sensor fixed to the bearing base of the smelting furnace.

3. The method for early warning of smelting furnace system failure based on big data according to claim 2, characterized in that: The step of extracting the temperature signal gradient feature of the smelting temperature signal includes: generating a temperature signal sequence of a smelting furnace according to the dual-wavelength infrared temperature measurement points; Calculating the first-order difference absolute value of the smelting temperature signal in a continuous time window through the temperature signal sequence; Perform sliding average filtering on the first-order difference absolute value to generate the temperature signal gradient feature.

4. The method for early warning of smelting furnace system failure based on big data according to claim 1, characterized in that: Decomposing the crucible vibration signal into preset frequency bands by using a preset wavelet packet decomposition includes: Construct decomposition filter banks using preset wavelet basis functions; determining the number of decomposition layers of the crucible vibration signal according to the rated vibration spectrum characteristics of the smelting furnace; The crucible vibration signal is subjected to wavelet packet decomposition according to the number of decomposition layers by means of the analysis filter bank to obtain preset frequency bands of equal width.

5. The method for early warning of smelting furnace system failure based on big data according to claim 1, characterized in that: The step of calculating the product of the standard deviation of the energy value of each frequency band and the energy ratio of adjacent frequency bands to generate a frequency domain energy gradient feature vector includes: Updating the energy ratio of the adjacent frequency bands according to a preset exponential weight; When the updated energy ratio of the adjacent frequency bands exceeds the preset transition threshold, a logarithmic operation is performed on the updated energy ratio of the adjacent frequency bands to obtain a target energy ratio of the adjacent frequency bands; When the updated energy ratio of the adjacent frequency bands does not exceed the preset transition threshold, the updated energy ratio of the adjacent frequency bands is used as the target energy ratio of the adjacent frequency bands; The product of the standard value of the energy of each frequency band and the target energy ratio of the adjacent frequency bands is calculated to generate a frequency domain energy gradient feature vector.

6. The method for early warning of smelting furnace system failure based on big data according to claim 1, characterized in that: Generating a multidimensional feature matrix according to the signal mean feature, the temperature signal gradient feature, and the frequency domain energy gradient feature vector includes: Mapping the frequency domain energy gradient eigenvector to a preset vibration dimension; Mapping the signal mean feature to a preset electrical dimension; Mapping the temperature signal gradient characteristics to a preset thermodynamic dimension; A multi-dimensional feature matrix is ​​generated according to a combination of the vibration dimension, the electrical dimension, and the thermodynamic dimension feature dimensions.

7. The method for early warning of smelting furnace system failure based on big data according to claim 1, characterized in that: Before inputting the multidimensional feature matrix into the pre-trained lightweight gradient boosting tree model to output the fault boundary of the smelting furnace, the method further includes: Collecting historical multi-source sensor time series data and extracting fault labels corresponding to the historical multi-source sensor time series data; generating a historical multi-dimensional feature matrix based on the historical multi-source sensor time series data; Using the historical multidimensional feature matrix and the fault label as historical fault samples; The preset gradient boosting decision tree is iteratively trained using the historical fault samples to obtain a lightweight gradient boosting tree model.

8. The method for early warning of smelting furnace system failure based on big data according to claim 7, characterized in that: The using the historical multidimensional feature matrix and the fault label as historical fault samples includes: Extract the energy attenuation waveform characteristics corresponding to the draft tube in the smelting furnace; Calculating the sliding standard deviation of the energy values ​​corresponding to the energy decay waveform characteristics within the continuous time window; When the sliding standard deviation exceeds a preset sliding threshold, marking the energy decay waveform feature as an additional feature, and adding the additional feature to the historical multidimensional feature matrix; Using the historical multidimensional feature matrix after adding the additional features as input sample data, and using the fault label as output sample data; A historical fault sample is generated according to the input sample data and the output sample data.

9. The method for early warning of smelting furnace system failure based on big data according to claim 1, characterized in that: The warning instructions corresponding to the graded warning signals triggering the smelting furnace include: comparing the fault boundary with a preset multi-level threshold interval, wherein the multi-level threshold interval includes at least two ordered threshold boundaries; triggering a first-level warning signal of a hierarchical warning signal of the smelting furnace when the fault boundary is within the preset first-level threshold interval, and determining a parameter adjustment instruction based on the first-level warning signal; When the fault boundary is within a preset second-level threshold range, a second-level warning signal in the graded warning signal of the smelting furnace is triggered, and a safety coordinated stop instruction is determined according to the second-level warning signal; The parameter adjustment instruction and the safety coordinated stop instruction are determined as early warning instructions.

Images