A Flux Addition Control Method Based on Ultraviolet Raman Spectroscopy

Abstract

This invention relates to the field of spectral analysis and control technology, and discloses a flux addition control method based on ultraviolet Raman spectroscopy. The method involves continuously scanning the molten material with ultraviolet Raman spectra to acquire real-time spectral data streams; extracting data segments containing characteristic spectral peaks; calculating the dynamic evolution path of spectral features along the time axis based on the peak drift trajectory; identifying the current phase composition mode of the molten material based on the shape characteristics of this path; matching the current mode with a preset phase mode library to obtain the corresponding basic flux requirement; real-time correcting the basic requirement based on the intensity fluctuation period of the spectral data segment to generate a dynamic addition dosage; and controlling the addition mechanism to perform addition according to this dosage. This method achieves precise adaptive control of flux addition by real-time analysis of the dynamic phase evolution path and process state of the molten material, improving the real-time performance and stability of process control.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis control technology, specifically to a flux addition control method based on ultraviolet Raman spectroscopy analysis. Background Technology

[0002] In high-temperature smelting processes such as metallurgy, glass, and ceramics, the control of flux addition directly determines the phase composition and performance indicators of the final product. Current technologies commonly employ ultraviolet Raman spectroscopy as an online monitoring method. This method involves identifying characteristic peaks representing specific crystalline phases in the spectrum. The presence or intensity of these peaks determines the reaction stage of the melt, triggering flux addition or calculating the amount based on empirical formulas.

[0003] This technical approach primarily relies on static analysis of spectral snapshots acquired at discrete time points. However, the physicochemical reactions in high-temperature melt environments are continuous and dynamically evolving, with the formation, transformation, and disappearance of phases often accompanied by complex kinetic processes. Static analysis methods based on single-point or discrete criteria struggle to effectively capture and describe the real-time paths, rates, and intermediate transition states of phase evolution. The system's perception of the process state is fragmented and lagging, failing to reflect the dynamic full picture of phase composition. This leads to discrepancies between flux addition decisions and the actual instantaneous demands of the melt, easily causing problems such as improper timing or inaccurate dosage, affecting process stability and product consistency.

[0004] The core challenge of existing technologies is how to overcome the limitations of static spectral analysis to achieve real-time, continuous tracking and analysis of the dynamic evolution of phases within the melt, and to generate precise, adaptive flux addition strategies based on this dynamic information. The purpose of this invention is to address the problem of insufficient control precision in existing technologies due to their inability to sense and respond to the real-time dynamic evolution path of the melt. Summary of the Invention

[0005] The purpose of this invention is to provide a flux addition control method based on ultraviolet Raman spectroscopy analysis to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a flux addition control method based on ultraviolet Raman spectroscopy analysis, the method comprising:

[0007] The melt was subjected to ultraviolet Raman spectroscopy scanning, and the continuous spectral data stream generated during the scanning process was acquired;

[0008] Extract spectral data segments containing characteristic spectral peaks from the continuous spectral data stream, wherein the characteristic spectral peaks correspond to the vibrational modes of a specific crystalline phase in the melt;

[0009] Based on the spectral peak drift trajectory within the spectral data segment, the dynamic evolution path of the spectral features along the time axis is calculated;

[0010] Based on the shape characteristics of the dynamic evolution path, the current phase composition mode of the melt is identified;

[0011] Match the current phase composition mode with several reference phase composition modes stored in the preset phase composition mode library, and obtain the basic flux requirement associated with the reference phase composition mode that is closest to the current phase composition mode.

[0012] Based on the intensity fluctuation period of the spectral data segment, the basic flux requirement is adjusted to generate a dynamic addition dosage;

[0013] Start the addition actuator and add flux to the melt according to the dynamic addition dosage.

[0014] Preferably, the step of extracting a spectral data segment containing characteristic spectral peaks from the continuous spectral data stream includes:

[0015] Set a sliding analysis window so that the sliding analysis window moves along the time axis of the continuous spectral data stream;

[0016] After each movement of the sliding analysis window, the local kurtosis and skewness values ​​of the spectral data within the window are calculated;

[0017] When the local kurtosis value exceeds a preset kurtosis threshold and the skewness value is lower than a preset skewness threshold, it is determined that the sliding analysis window contains a valid characteristic spectral peak.

[0018] Record the start and end times of the sliding analysis window at this time, and use the start and end times to define the spectral data segment.

[0019] Preferably, calculating the dynamic evolution path of spectral features along the time axis based on the spectral peak drift trajectory within the spectral data segment includes:

[0020] In the spectral data segment, the main peak with the highest intensity is located, and the wavenumber position of the main peak is tracked over time to form a main peak drift trajectory line;

[0021] Meanwhile, within the preset wavenumber neighborhood of the main peak, all secondary peaks are searched, and the intensity change of each secondary peak is tracked to form multiple secondary peak intensity evolution curves.

[0022] The main peak drift trajectory line is synchronously superimposed with the intensity evolution curves of all the secondary peaks to generate a composite dynamic evolution path of spectral features.

[0023] Preferably, identifying the current phase composition mode of the melt based on the shape characteristics of the dynamic evolution path includes:

[0024] Geometric features are extracted from the dynamic evolution path of the spectral features, including the average value of the path curvature change, the number of path inflection points, and the overall slope of the path.

[0025] The extracted geometric features are input into a pre-trained phase pattern discrimination model; the phase pattern discrimination model is trained by analyzing the correlation between the dynamic evolution path of historical spectral features and known phase composition patterns.

[0026] The phase pattern discrimination model outputs a phase composition pattern label that matches the input geometric features with the highest degree of matching, and the phase composition pattern label is defined as the current phase composition pattern.

[0027] Preferably, the matching of the current phase composition mode with several reference phase composition modes stored in a preset phase pattern library includes:

[0028] Calculate the multidimensional Euclidean distance between the geometric feature vector corresponding to the current phase composition mode and the standard geometric feature vector corresponding to each reference phase composition mode in the preset phase pattern library;

[0029] All calculated multidimensional Euclidean distances are sorted, and the reference phase composition mode corresponding to the multidimensional Euclidean distance with the smallest value is selected as the closest reference phase composition mode.

[0030] Preferably, obtaining the basic flux requirement associated with the reference phase configuration mode that is closest to the current phase configuration mode includes:

[0031] In the preset phase model library, a corresponding empirical addition range is preset for each reference phase composition model;

[0032] Query the empirical addition range corresponding to the closest reference phase composition mode, and select the median value from the empirical addition range as the initial basic flux requirement.

[0033] Preferably, the step of adjusting the basic flux requirement based on the intensity fluctuation period of the spectral data segment includes:

[0034] Periodic analysis is performed on the overall intensity sequence of the spectral data segment to detect the fundamental frequency period of intensity fluctuations;

[0035] A correction factor lookup table is established, which defines the correspondence between different fundamental frequency period ranges and dose correction factors;

[0036] Based on the detected fundamental frequency period, the corresponding dose correction coefficient is found in the correction coefficient lookup table;

[0037] The dynamic addition dose is obtained by multiplying the initial flux base requirement by the found dosage correction coefficient.

[0038] Preferably, the training process of the phase pattern discrimination model includes:

[0039] Collect historical spectral data streams of melts under various known process conditions, and perform operations to extract spectral data segments and calculate dynamic evolution paths for each historical spectral data stream to generate a sample set of dynamic evolution paths of historical spectral features;

[0040] Process experts determine the melt state corresponding to each historical spectral feature dynamic evolution path sample and label the phase composition mode.

[0041] The convolutional neural network is trained under supervision using the historical spectral features of the dynamic evolution path samples and their labeled phase composition patterns until the network converges. The trained convolutional neural network is then used as the phase pattern discrimination model.

[0042] Preferably, the activation of the fluxing actuator, which adds flux to the melt according to the dynamic addition dosage, includes:

[0043] The dynamically added dose is converted into a pulse control signal, and the total number of pulses in the pulse control signal is directly proportional to the dynamically added dose.

[0044] The pulse control signal is sent to the stepper motor driver of the added actuator to drive the feeding screw of the added actuator to rotate;

[0045] A metering sensor installed at the end of the feeding screw provides real-time feedback on the cumulative amount of flux added. When the cumulative amount reaches the dynamic addition dose, the pulse control signal is stopped.

[0046] Preferably, after stopping the transmission of the pulse control signal, the method further includes:

[0047] After waiting for the preset process stabilization time, perform another round of ultraviolet Raman spectroscopy scanning on the melt to obtain a new continuous spectral data stream;

[0048] Based on the new continuous spectral data stream, all steps from extracting spectral data segments to generating the dynamic addition dose are repeated to calculate a new round of dynamic addition dose.

[0049] The dynamic dosage added in the new round is compared with the dynamic dosage added in the previous round. If the difference between the two exceeds the set tolerance threshold, the addition execution mechanism is immediately activated to perform a new round of addition; if it does not exceed the threshold, the current state is maintained until the next detection cycle.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] By tracing the real-time drift trajectories of characteristic spectral peaks in a continuous spectral data stream and calculating their dynamic evolution paths over time, the focus of analysis shifts from static peak parameters to dynamic trajectory morphology. The curvature changes and trend directions of this path directly characterize the continuous dynamic processes of phase behavior within the melt, such as crystal growth and phase transformation. This enables the system to identify complex phase composition and transformation modes between initial and final states, rather than simply determining the presence or absence of a particular crystalline phase, thus achieving earlier and more precise monitoring and analysis of the nonlinear evolution of phases.

[0052] Based on the identified current dynamic phase composition mode, it is matched with reference modes in a preset phase mode library to directly map the corresponding basic flux requirement. This transforms the core of the control logic from discrete judgment based on simple thresholds to a correlation mapping based on continuous process modes, making the determination of the basic dosage more closely aligned with actual process dynamics. By analyzing the intensity fluctuation period of the spectral data segment itself, this periodic information reflects the periodic state of mixing, mass transfer, or reaction within the melt, and is used to correct the aforementioned basic requirement in real time. This two-level control mechanism, integrating macroscopic phase evolution modes and microscopic process state feedback, generates a dynamic addition dosage that adapts to transient changes in the melt, ensuring precise synchronization between the flux addition rhythm and dosage and the dynamic process within the melt, reducing control errors caused by fixed parameter control or single-factor feedback lag. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the working principle of the flux addition control method based on ultraviolet Raman spectroscopy analysis described in this invention.

[0054] Figure 2 A flowchart for calculating the dynamic evolution path of spectral features;

[0055] Figure 3 A flowchart for identifying the current phase composition mode of the melt;

[0056] Figure 4 A comparison chart of dynamic flux demand for each cycle under different phase modes;

[0057] Figure 5 Thermograph of the correlation between the control of key parameters for flux addition. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 This invention provides a flux addition control method based on ultraviolet Raman spectroscopy analysis. The method includes: performing ultraviolet Raman spectral scanning on a melt and acquiring a continuous spectral data stream generated during the scanning process; extracting spectral data segments containing characteristic peaks from the continuous spectral data stream, where each characteristic peak corresponds to a vibrational mode of a specific crystalline phase in the melt; calculating the dynamic evolution path of the spectral features along the time axis based on the peak drift trajectory within the spectral data segment; identifying the current phase composition mode of the melt based on the shape characteristics of the dynamic evolution path; matching the current phase composition mode with several reference phase composition modes stored in a preset phase composition mode library to obtain the basic flux requirement associated with the reference phase composition mode closest to the current phase composition mode; adjusting the basic flux requirement based on the intensity fluctuation period of the spectral data segment to generate a dynamic addition dose; and activating the addition actuator to add flux to the melt according to the dynamic addition dose.

[0060] In one embodiment of the present invention, see [reference] Figure 2 A sliding analysis window is set up, moving along the time axis of the continuous spectral data stream. After each movement of the sliding analysis window, the local kurtosis and skewness values ​​of the spectral data within the window are calculated. When the local kurtosis value exceeds a preset kurtosis threshold and the skewness value is lower than a preset skewness threshold, the sliding analysis window is considered to contain a valid characteristic spectral peak. The start and end times of the sliding analysis window are recorded, defining the spectral data segment. Within the spectral data segment, the highest-intensity main peak is located, and its wavenumber position is tracked over time, forming a main peak drift trajectory. Within the preset wavenumber neighborhood of the main peak, all secondary peaks are searched, and the intensity changes of each secondary peak are tracked, forming multiple secondary peak intensity evolution curves. The main peak drift trajectory is synchronously superimposed with all secondary peak intensity evolution curves to generate a composite dynamic evolution path of spectral features.

[0061] In practice, a fixed-width sliding analysis window is set up. This window moves along the time axis of the continuous spectral data stream with a preset step size, and its width is sufficient to cover the time span in which characteristic spectral peaks appear. After each movement of the sliding analysis window, the local kurtosis and skewness values ​​of the spectral data within the window are calculated. The local kurtosis value characterizes the sharpness of the spectral data distribution within the window, while the skewness value characterizes the asymmetry of the distribution. When the calculated local kurtosis value exceeds a preset kurtosis threshold and the calculated skewness value is lower than a preset skewness threshold, the sliding analysis window is determined to contain a valid characteristic spectral peak. The start and end times of the sliding analysis window are recorded at this point, and these times define the spectral data segment.

[0062] In some embodiments, after the spectral data segment is defined, the highest-intensity main peak is located within the spectral data segment, and the wavenumber position of the main peak is tracked over time to form a main peak drift trajectory line. The main peak drift trajectory line reflects the migration of the main peak's center position over time. Simultaneously, a search is performed within a preset wavenumber neighborhood of the main peak. The preset wavenumber neighborhood is determined by extending a fixed width to both sides of the main peak's wavenumber. All secondary peaks within the preset wavenumber neighborhood that meet the intensity threshold are searched, and the intensity change of each secondary peak over time is tracked to form multiple secondary peak intensity evolution curves. The main peak drift trajectory line and all secondary peak intensity evolution curves are synchronously superimposed on the same time coordinate axis to generate a composite dynamic evolution path of spectral features.

[0063] The local kurtosis value within the sliding analysis window can be calculated using a formula based on the fourth-order central moment. In a specific calculation method, for the values ​​contained within the sliding analysis window... For each spectral intensity data point, its local kurtosis value The calculation formula is:

[0064]

[0065] in: This represents the spectral intensity value of the i-th data point within the window. This represents the average of all spectral intensity values ​​within the window. This represents the standard deviation of all spectral intensity values ​​within the window. This indicates summing over all data points.

[0066] Optionally, preset kurtosis and skewness thresholds are set after statistical analysis of data segments in the historical spectral data stream that are known to contain and not contain characteristic spectral peaks. The kurtosis threshold is set to a positive number greater than the kurtosis value of the standard normal distribution, and the skewness threshold is set to a negative number close to zero. After the sliding analysis window has traversed the continuous spectral data stream, it may define multiple spectral data segments, and subsequent processing is performed independently for each spectral data segment.

[0067] In some embodiments, when tracking the wavenumber position of the main peak over time, the center wavenumber of the main peak is determined using the centroid method or Gaussian fitting method for the spectral data at each sampling time, thus obtaining a discrete time-wavenumber point sequence. These points are then connected using linear interpolation to form a continuous main peak drift trajectory. When tracking the intensity evolution curves of secondary peaks, the peak intensity of each secondary peak within its occurrence time range is recorded, and similarly, a continuous curve is formed through interpolation. When superimposing to generate the dynamic evolution path of spectral features, the main peak drift trajectory uses the left ordinate axis to represent wavenumber, the secondary peak intensity evolution curves use the right ordinate axis to represent intensity, and all curves share the same abscissa axis to represent time.

[0068] In one embodiment of the present invention, see [reference] Figure 3 The process involves extracting geometric features from the dynamic evolution path of spectral characteristics. These geometric features include the average value of the path curvature changes, the number of path inflection points, and the overall slope of the path. The extracted geometric features are then input into a pre-trained phase pattern discrimination model. The model outputs a label representing the phase composition pattern that best matches the input geometric features; this label is defined as the current phase composition pattern. The training process of the phase pattern discrimination model includes collecting historical spectral data streams of melts under various known process conditions. For each historical spectral data stream, spectral data segments are extracted and dynamic evolution paths are calculated, generating a sample set of historical spectral feature dynamic evolution paths. Process experts determine the melt state corresponding to each historical spectral feature dynamic evolution path sample and label it with a phase composition pattern label. Using the historical spectral feature dynamic evolution path samples and their labeled phase composition pattern labels, a convolutional neural network is trained under supervision until the network converges. The trained convolutional neural network is then used as the phase pattern discrimination model.

[0069] In practical implementation, geometric features are extracted from the dynamic evolution path of spectral features. These geometric features include the average value of the path curvature change, the number of path inflection points, and the overall path slope. The average value of the path curvature change is obtained by calculating the arithmetic mean of the instantaneous curvature values ​​of a series of sampling points along the evolution path. The number of path inflection points is counted by detecting the number of zero-crossing points of the second derivative of the evolution path. The overall path slope is obtained by linearly fitting the starting and ending points of the evolution path to obtain the absolute value of the slope. The extracted geometric features are input into a pre-trained phase pattern discrimination model. The phase pattern discrimination model outputs the phase composition mode label with the highest matching degree to the input geometric features. The phase composition mode label is defined as the current phase composition mode.

[0070] In some embodiments, the training process of the phase pattern discrimination model includes collecting historical spectral data streams of melts under various known process conditions. For each historical spectral data stream, spectral data segments are extracted and dynamic evolution paths are calculated to generate a sample set of historical spectral feature dynamic evolution paths. Each historical spectral feature dynamic evolution path sample consists of a time series curve and related sampling parameters. Process experts determine the melt state corresponding to each historical spectral feature dynamic evolution path sample and, based on the X-ray diffraction analysis results of the melt and process records, label it with a specific phase composition mode. Using the historical spectral feature dynamic evolution path samples and their labeled phase composition mode tags, a convolutional neural network is trained under supervision until the network converges. The trained convolutional neural network is then used as the phase pattern discrimination model.

[0071] It is understandable that the composition of geometric eigenvectors is fixed, and in a specific calculation method, the average value of the path curvature change is... The calculation formula is:

[0072]

[0073] in: This represents the total number of sampling points uniformly selected along the dynamic evolution path of spectral features. Representative at the The first derivative values ​​obtained by numerical differentiation at each sampling point Representative at the The second derivative values ​​obtained by numerical differentiation at each sampling point This indicates summation over all sampling points.

[0074] Optionally, the convolutional neural network structure of the phase pattern discrimination model includes two convolutional layers and one fully connected classification layer. During training, historical spectral feature dynamic evolution path samples are converted into grayscale images as network input, and phase composition pattern labels are converted into one-hot codes as supervision signals. The network is optimized using the cross-entropy loss function and stochastic gradient descent algorithm. The network is considered converged when the loss function value on the training set no longer decreases for several consecutive training epochs.

[0075] During network training, the network performs nonlinear transformations such as convolution and pooling on the input sample data (usually converted to grayscale image form) layer by layer, ultimately outputting a probability distribution vector at the fully connected layer. Each element in this vector corresponds to the predicted probability of a phase composition pattern label. To evaluate the accuracy of the network's predictions, the predicted probability distribution output by the network needs to be compared with the true phase composition pattern labels. This comparison is achieved through the cross-entropy loss function. The cross-entropy loss function measures the difference between the network's predicted probability distribution and the true label distribution. For a single training sample, its true label is represented as a one-hot encoded vector, where the position corresponding to the correct category is 1, and the other positions are 0; while the network output is a probability vector processed by the softmax function, representing the probability that the network believes the sample belongs to each category. The cross-entropy loss function calculates a scalar loss value based on these two vectors. The smaller the loss value, the closer the network's predicted probability distribution is to the true distribution.

[0076] Stochastic gradient descent (SDD) optimizes network parameters using the calculated loss value. In each iteration, the algorithm first randomly selects a small batch of training samples for forward propagation and calculates the average cross-entropy loss on that batch of data. Then, it calculates the gradient of the loss function with respect to the trainable parameters of each layer in the network using backpropagation. These gradients indicate the direction and rate of change of the loss function with respect to the parameters. SDD updates the network parameters based on the calculated gradient information. Specifically, the update rule is to subtract the product of the learning rate and the gradient value from the current parameter value, thereby adjusting the parameters in the direction of reducing the loss function. The learning rate is a preset hyperparameter that controls the step size of each parameter update. By repeatedly executing this cyclic process of forward propagation, loss calculation, backpropagation, and parameter update, the network parameters are gradually adjusted, causing the average cross-entropy loss on the entire training set to continuously decrease. Ultimately, the network output tends to match the true label, and the model converges.

[0077] In some embodiments, the detection of path inflection points is accomplished by finding the sign change points of the second derivative sequence of the spectral feature dynamic evolution path. In the calculation of the overall path slope, the starting point and the ending point correspond to the feature values ​​at the beginning and end of the time axis of the spectral feature dynamic evolution path, respectively. The trained phase pattern discrimination model performs forward propagation calculation on the input geometric feature vector, generating a probability distribution in the output layer. The category with the highest probability is the output phase composition pattern label.

[0078] In one embodiment of the present invention, matching the current phase composition mode with several reference phase composition modes stored in a preset phase composition mode library includes calculating the multidimensional Euclidean distance between the geometric feature vector corresponding to the current phase composition mode and the standard geometric feature vector corresponding to each reference phase composition mode in the preset phase composition mode library. All calculated multidimensional Euclidean distances are sorted, and the reference phase composition mode corresponding to the smallest multidimensional Euclidean distance is selected as the closest reference phase composition mode. The basic flux requirement associated with the reference phase composition mode closest to the current phase composition mode is obtained, including pre-setting a corresponding empirical addition range for each reference phase composition mode in the preset phase composition mode library. The empirical addition range corresponding to the closest reference phase composition mode is queried, and the median value is selected from the empirical addition range as the initial basic flux requirement.

[0079] Matching the current phase composition mode with several reference phase composition modes stored in the preset phase composition mode library involves calculating the multidimensional Euclidean distance between the geometric feature vector corresponding to the current phase composition mode and the standard geometric feature vector corresponding to each reference phase composition mode in the preset phase composition mode library. The geometric feature vector is composed of three components in sequence: the average value of the path curvature change, the number of path inflection points, and the overall slope of the path. All calculated multidimensional Euclidean distances are sorted in ascending order of distance values, and the reference phase composition mode with the smallest multidimensional Euclidean distance value is selected as the closest reference phase composition mode.

[0080] In some embodiments, obtaining the basic flux requirement associated with the reference phase composition mode closest to the current phase composition mode includes pre-setting an empirical addition range for each reference phase composition mode in a preset phase composition mode library. The empirical addition range is expressed in mass units and associated with a specific melt volume. The empirical addition range corresponding to the closest reference phase composition mode is queried, and the median value is selected from the empirical addition range. The median value is the arithmetic mean of the upper and lower limits of the empirical addition range, and is used as the initial basic flux requirement.

[0081] It is understandable that the calculation of multidimensional Euclidean distance involves the sum of squared differences between corresponding vector components. In a specific calculation method, the geometric eigenvector corresponding to the current phase composition mode is denoted as... The first preset phase model in the library The standard geometric eigenvectors corresponding to the reference phase composition modes are denoted as follows: Then the multidimensional Euclidean distance between the two The calculation formula is:

[0082]

[0083] in: The average value of the path curvature change representing the current phase composition mode. The number of path inflection points representing the current phase composition mode. The overall slope of the path representing the current phase composition mode. , , Representing the first The corresponding geometric characteristic components of each reference phase composition mode.

[0084] Optionally, the preset phase pattern library is stored in the form of a relational database or structured files. Each reference phase pattern record contains a unique pattern identifier, a standard geometric feature vector field, and an empirical addition range field. After calculating all multidimensional Euclidean distances, the pattern identifier associated with the minimum distance value is found by querying the database or traversing the files, thereby locating the corresponding record and reading the empirical addition range data. In some embodiments, the components of the geometric feature vector need to be standardized to eliminate the influence of dimensions. The standardization method is to subtract the mean of all corresponding components of the reference patterns from each component and then divide by the standard deviation. After selecting the median value from the empirical addition range, the initial flux base requirement can be scaled proportionally according to the actual melt volume. The scaling factor is the ratio of the actual melt volume to the standard melt volume associated with the preset phase pattern library.

[0085] In one embodiment of the present invention, the basic flux requirement is adjusted based on the intensity fluctuation period of the spectral data segment. This includes performing periodic analysis on the overall intensity sequence of the spectral data segment to detect the fundamental frequency period of the intensity fluctuation. A correction coefficient lookup table is established, defining the correspondence between different fundamental frequency period ranges and dosage correction coefficients. Based on the detected fundamental frequency period, the corresponding dosage correction coefficient is looked up in the correction coefficient lookup table. The initial basic flux requirement is multiplied by the found dosage correction coefficient to obtain the dynamically added dosage.

[0086] In practice, adjusting the basic flux requirement based on the intensity fluctuation cycle of the spectral data segment involves periodic analysis of the overall intensity sequence of the spectral data segment, detecting the fundamental frequency period of the intensity fluctuations. The overall intensity sequence is obtained by summing the spectral intensity values ​​of all wavenumber channels at each sampling time in the spectral data segment. A correction coefficient lookup table is established, defining the correspondence between different fundamental frequency period ranges and dose correction coefficients. The dose correction coefficient is a dimensionless multiplier factor. Based on the detected fundamental frequency period, the corresponding dose correction coefficient is looked up in the correction coefficient lookup table. The initial basic flux requirement is multiplied by the found dose correction coefficient to obtain the dynamically added dose.

[0087] In some embodiments, the fundamental frequency period for detecting intensity fluctuations is determined using a Fast Fourier Transform (FFT) method. After time-frequency conversion of the overall intensity sequence, the position of the main peak in the power spectrum is identified, and the reciprocal of the frequency corresponding to the main peak is the fundamental frequency period. A correction coefficient lookup table is stored in the control system's memory in tabular form, using the fundamental frequency period as the index key and the dose correction coefficient as the mapping value. The lookup process involves comparing the calculated fundamental frequency period with the fundamental frequency period range of each entry in the correction coefficient lookup table to determine the range to which the fundamental frequency period belongs, and returning the dose correction coefficient value corresponding to that range.

[0088] It can be understood that calculating the dynamic additive amount is a scalar multiplication operation. Let the initial flux base requirement be denoted as... The found dose correction factor is denoted as Then the dosage is added dynamically. The calculation formula is: .in: This represents the initial basic flux requirement obtained from the preset phase pattern library. This represents the dose correction factor obtained from the correction factor lookup table based on the fundamental frequency period. This represents the final dynamically added dose calculated.

[0089] Optionally, the correction coefficient lookup table is constructed based on statistical analysis of historical process data. A correlation regression analysis is performed between the fundamental frequency period of the spectral intensity sequence recorded in historical successful process batches and the finally determined effective flux addition amount to determine the optimization coefficient range corresponding to different period intervals. Refer to Table 1, which shows a simplified correction coefficient lookup table, defining four consecutive fundamental frequency period ranges and their corresponding dosage correction coefficients.

[0090] Table 1: Correspondence between fundamental frequency period and dose correction factor

[0091]

[0092] In some embodiments, periodicity analysis calculates a periodicity significance index while detecting the fundamental frequency period. If this index is below a set threshold, the intensity sequence is determined to have no obvious periodicity, and the default dose correction factor of 1.0 is directly used for calculation. When looking up the correction factor lookup table, if the detected fundamental frequency period happens to fall within the boundary of two defined ranges, the dose correction factor corresponding to the larger range boundary is used. The multiplication of the dose correction factor with the initial flux base requirement is performed in the arithmetic logic unit of the control system.

[0093] See Figure 4This study demonstrates the flux demand distribution for different phase modes (AEs) at four fundamental frequency periods (8s, 25s, 45s, and 70s) during the dynamic dose calculation phase. Based on the correlation rules of "fundamental frequency period - dose correction coefficient" in the project (Table 1), the correction coefficients corresponding to each period are as follows: 8s (β=1.2), 25s (β=1.0), 45s (β=0.8), and 70s (β=0.6). The data in the figure reflects the dynamic difference in demand under the combined effect of "phase mode + period correction": the demand for phase mode D reaches 78.0kg for both 8s and 25s periods, which is the highest among all modes; under the same phase mode, the longer the period (the smaller the correction coefficient), the lower the demand is usually (e.g., in phase mode B, the demand for 8s / 25s period is 65.0kg, which drops to 52.0kg for 45s period, and further to 39.0kg for 70s period); the basic demand for different phase modes differs significantly. For example, the basic demand for phase mode A for 8s / 25s period is 50.0kg, while the basic demand for the corresponding period of mode D is 78.0kg, which reflects the dominant role of phase composition in basic demand.

[0094] In one embodiment of the invention, the addition actuator is activated to add flux to the melt according to a dynamic addition dose. This includes converting the dynamic addition dose into a pulse control signal, where the total number of pulses in the pulse control signal is directly proportional to the dynamic addition dose. The pulse control signal is sent to the stepper motor driver of the addition actuator to drive the feeding screw of the addition actuator to rotate. A metering sensor installed at the end of the feeding screw provides real-time feedback on the cumulative amount of flux added. When the cumulative amount reaches the dynamic addition dose, the transmission of the pulse control signal is stopped. After the transmission of the pulse control signal is stopped, a preset process stabilization time is waited for, and a new round of ultraviolet Raman spectroscopy scanning is performed on the melt to obtain a new continuous spectral data stream. Based on the new continuous spectral data stream, all steps from intercepting the spectral data segment to generating the dynamic addition dose are repeated to calculate the new round of dynamic addition dose. The new round of dynamic addition dose is compared with the previous round of dynamic addition dose. If the difference between the two exceeds a set tolerance threshold, the addition actuator is immediately activated to perform a new round of addition; if it does not exceed the threshold, the current state is maintained until the next detection cycle.

[0095] In practice, the process of adding flux to the molten material according to a dynamic dosage by activating the adding actuator involves converting the dynamic dosage into a pulse control signal. The total number of pulses in the pulse control signal is directly proportional to the dynamic dosage, and the specific conversion depends on a pre-calibrated pulse equivalent parameter. The pulse control signal is sent to the stepper motor driver of the adding actuator, driving the feed screw to rotate. The rotation angle of the feed screw is strictly proportional to the number of received pulses. A metering sensor installed at the end of the feed screw provides real-time feedback on the cumulative amount of flux added. The metering sensor converts physical displacement or pressure signals into electrical signals. When the cumulative amount reaches the dynamic dosage, the control system stops sending pulse control signals to the stepper motor driver.

[0096] In some embodiments, after stopping the transmission of pulse control signals, the control system starts a preset process stabilization timer, which is pre-set based on the thermodynamic properties of the melt. After the preset process stabilization time has elapsed, the control system instructs the spectral analysis equipment to perform another round of ultraviolet Raman spectral scanning on the melt to acquire a new continuous spectral data stream. Based on the new continuous spectral data stream, all steps from truncation of spectral data segments to generation of the dynamic dosage are repeated to calculate the new round of dynamic dosage.

[0097] It is understandable that the conversion from dynamically added dose to pulse control signal involves a linear mapping. The dynamically added dose is denoted as... The preset pulse equivalent is denoted as (i.e., the standard added quality corresponding to each pulse), then the total number of pulses of the pulse control signal. The calculation formula is:

[0098]

[0099] in: This represents the calculated dynamic dosage. This represents a pulse equivalent constant determined through calibration experiments. This represents the floor function, ensuring that the amount added is not less than the computational requirement.

[0100] Optionally, the pulse control signal is generated by the pulse output port of the control system in the form of a square wave electrical signal, and the frequency of the pulse is determined by the maximum response speed of the stepper motor driver. The stepper motor driver receives the pulse control signal and the direction signal, controlling the stepper motor to rotate a specified number of steps, thereby driving the feeding screw to push the flux from the hopper into the conveying pipeline. The metering sensor uses a high-precision rotary encoder or a linear displacement sensor, and its feedback signal is read in real time and compared with... Compare, when satisfied Immediately interrupt pulse output.

[0101] In some embodiments, when comparing the new round of dynamic dosage addition with the previous round of dynamic dosage addition, the absolute value of the difference between the two is calculated. A positive tolerance threshold is set. If the absolute value of the difference exceeds the set tolerance threshold, the control system immediately restarts the addition actuator and executes a new addition process using the new round of dynamic dosage addition as the target value. If the absolute value of the difference does not exceed the set tolerance threshold, the control system maintains the current melt state and does not perform the addition operation until the next preset detection cycle triggers a new round of spectral scanning and analysis.

[0102] See Figure 5 In the flux addition control method based on ultraviolet Raman spectroscopy, the thermogram reveals the linear correlation coefficient distribution among six core parameters: dynamic addition dose, total number of pulses, mean curvature, number of inflection points, matching Euclidean distance, and correction coefficient. Specifically, the correlation coefficient between dynamic addition dose and total number of pulses is 1.00, which perfectly matches the linear mapping mechanism in the implementation logic that "the total number of pulses is directly proportional to the dynamic addition dose." The correlation coefficient between dynamic addition dose and correction coefficient reaches 0.91, reflecting the strong positive regulatory effect of the correction coefficient on the amount of dynamic additive. The correlation coefficient between mean curvature and number of inflection points is -0.73, reflecting a significant negative correlation between the curvature change of the dynamic evolution path of spectral features and the number of inflection points, which corresponds to the inherent correlation characteristics of geometric features in phase mode discrimination. The correlation coefficient between matching Euclidean distance and dynamic addition dose is -0.53, reflecting the inverse influence of phase mode matching results on the amount of additive. In terms of parameter configuration, the correlation coefficient of the heat map is calculated based on the statistical analysis of multiple rounds of process data, and its value can serve as a quantitative basis for subsequent parameter optimization and control model adjustment.

[0103] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for controlling flux addition based on ultraviolet Raman spectroscopy analysis, characterized in that, include: The melt was subjected to ultraviolet Raman spectroscopy scanning, and the continuous spectral data stream generated during the scanning process was acquired; Extract spectral data segments containing characteristic spectral peaks from the continuous spectral data stream; Based on the spectral peak drift trajectory within the spectral data segment, the dynamic evolution path of the spectral features along the time axis is calculated, including: In the spectral data segment, the main peak with the highest intensity is located, and the wavenumber position of the main peak is tracked over time to form a main peak drift trajectory line; Meanwhile, within the preset wavenumber neighborhood of the main peak, all secondary peaks are searched, and the intensity change of each secondary peak is tracked to form multiple secondary peak intensity evolution curves. The main peak drift trajectory line is synchronously superimposed with the intensity evolution curves of all the secondary peaks to generate a composite dynamic evolution path of spectral features. Based on the shape characteristics of the dynamic evolution path, the current phase composition mode of the melt is identified, including: Geometric features are extracted from the dynamic evolution path of the spectral features, including the average value of the path curvature change, the number of path inflection points, and the overall slope of the path. The extracted geometric features are input into a pre-trained phase pattern discrimination model; the phase pattern discrimination model is trained by analyzing the correlation between the dynamic evolution path of historical spectral features and known phase composition patterns. The phase pattern discrimination model outputs a phase composition pattern label that matches the input geometric features with the highest degree of matching, and the phase composition pattern label is defined as the current phase composition pattern; Match the current phase composition mode with several reference phase composition modes stored in the preset phase composition mode library, and obtain the initial flux basic requirement associated with the reference phase composition mode that is closest to the current phase composition mode. Based on the intensity fluctuation period of the spectral data segment, the basic flux requirement is adjusted to generate a dynamic addition dosage, including: Periodic analysis is performed on the overall intensity sequence of the spectral data segment to detect the fundamental frequency period of intensity fluctuations; A correction factor lookup table is established, which defines the correspondence between different fundamental frequency period ranges and dose correction factors; Based on the detected fundamental frequency period, the corresponding dose correction coefficient is found in the correction coefficient lookup table; The dynamic addition dosage is obtained by multiplying the initial flux base requirement by the found dosage correction coefficient. The fundamental frequency period for detecting intensity fluctuations is determined by using the Fast Fourier Transform (FFT) method. After performing time-frequency conversion on the overall intensity sequence, the position of the main peak in the power spectrum is identified, and the reciprocal of the frequency corresponding to the main peak is the fundamental frequency period. Start the addition actuator and add flux to the melt according to the dynamic addition dosage.

2. The flux addition control method based on ultraviolet Raman spectroscopy analysis according to claim 1, characterized in that, Extracting spectral data segments containing characteristic spectral peaks from the continuous spectral data stream includes: Set a sliding analysis window so that the sliding analysis window moves along the time axis of the continuous spectral data stream; After each movement of the sliding analysis window, the local kurtosis and skewness values ​​of the spectral data within the window are calculated; When the local kurtosis value exceeds a preset kurtosis threshold and the skewness value is lower than a preset skewness threshold, it is determined that the sliding analysis window contains a valid characteristic spectral peak. Record the start and end times of the sliding analysis window at this time, and use the start and end times to define the spectral data segment.

3. The flux addition control method based on ultraviolet Raman spectroscopy analysis according to claim 2, characterized in that, The matching of the current phase composition mode with several reference phase composition modes stored in a preset phase mode library includes: Calculate the multidimensional Euclidean distance between the geometric feature vector corresponding to the current phase composition mode and the standard geometric feature vector corresponding to each reference phase composition mode in the preset phase pattern library; All calculated multidimensional Euclidean distances are sorted, and the reference phase composition mode corresponding to the multidimensional Euclidean distance with the smallest value is selected as the closest reference phase composition mode.

4. The flux addition control method based on ultraviolet Raman spectroscopy analysis according to claim 3, characterized in that, The acquisition of the initial flux base requirement associated with the reference phase configuration closest to the current phase configuration includes: In the preset phase model library, a corresponding empirical addition range is preset for each reference phase composition model; Query the empirical addition range corresponding to the closest reference phase composition mode, and select the median value from the empirical addition range as the initial flux basic requirement.

5. The flux addition control method based on ultraviolet Raman spectroscopy analysis according to claim 4, characterized in that, The training process of the phase pattern discrimination model includes: Collect historical spectral data streams of melts under various known process conditions, and perform operations to extract spectral data segments and calculate dynamic evolution paths for each historical spectral data stream to generate a sample set of dynamic evolution paths of historical spectral features; Process experts determine the melt state corresponding to each historical spectral feature dynamic evolution path sample and label the phase composition mode. The convolutional neural network is trained under supervision using the historical spectral features of the dynamic evolution path samples and their labeled phase composition patterns until the network converges. The trained convolutional neural network is then used as the phase pattern discrimination model.

6. The flux addition control method based on ultraviolet Raman spectroscopy analysis according to claim 5, characterized in that, The activation and addition mechanism adds flux to the melt according to the dynamic addition dosage, including: The dynamically added dose is converted into a pulse control signal, and the total number of pulses in the pulse control signal is directly proportional to the dynamically added dose. The pulse control signal is sent to the stepper motor driver of the added actuator to drive the feeding screw of the added actuator to rotate; A metering sensor installed at the end of the feeding screw provides real-time feedback on the cumulative amount of flux added. When the cumulative amount reaches the dynamic addition dose, the pulse control signal is stopped.

7. The flux addition control method based on ultraviolet Raman spectroscopy analysis according to claim 6, characterized in that, After ceasing the transmission of the pulse control signal, the method further includes: After waiting for the preset process stabilization time, perform another round of ultraviolet Raman spectroscopy scanning on the melt to obtain a new continuous spectral data stream; Based on the new continuous spectral data stream, all steps from extracting spectral data segments to generating the dynamic addition dose are repeated to calculate a new round of dynamic addition dose. The dynamic dosage added in the new round is compared with the dynamic dosage added in the previous round. If the difference between the two exceeds the set tolerance threshold, the addition execution mechanism is immediately activated to perform a new round of addition; if it does not exceed the threshold, the current state is maintained until the next detection cycle.

Images