Vanadium-nitrogen alloy finished product preparation double push plate kiln fault early warning control system

Abstract

The present application belongs to the technical field of intelligent control, and discloses a kind of vanadium-nitrogen alloy finished product preparation double push plate kiln fault early warning control system;Including:Data acquisition module: collect original perception data, original perception data include mechanical data and process data;Feature extraction module: construct the benchmark model of equipment, and generate quality disturbance fingerprint in combination with original perception data;Analysis and early warning module: quality disturbance fingerprint is used as the input of associated prediction model, and quality index prediction value is obtained, whether quality early warning signal is triggered according to quality index prediction value;Composite control module: according to the quality early warning signal output by analysis and early warning module, feedback suppression instruction is sent to hydraulic servo system, and process data adjustment amount is obtained according to quality expected deviation;The present application greatly improves the vanadium-nitrogen alloy product pass rate and production process stability, reduces equipment failure and production loss.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically, to a fault early warning and control system for a double-pusher kiln used in the preparation of vanadium-nitrogen alloy products. Background Technology

[0002] In the double-pusher kiln system for preparing vanadium-nitrogen alloy products, the servo valve of the pusher hydraulic system is the core actuator, and its dynamic accuracy directly determines the stability of the sintering process. However, due to internal hydraulic forces, oil micro-cavitation, or interference from electrical control signals, the servo valve experiences tiny self-excited oscillations with frequencies far exceeding the conventional monitoring range. These oscillations have extremely small amplitudes in the time domain and cannot trigger conventional alarm thresholds based on displacement or pressure. However, their high-frequency energy is precisely transmitted to the vanadium-nitrogen alloy green billet undergoing solid-state reaction and nitriding processes through the hydraulic pipeline and mechanical pusher plate in the form of stress waves. This causes microstructural cracks to form in the vanadium-nitrogen alloy green billet during the solid-state reaction. The cracked vanadium-nitrogen alloy green billet will shatter in the cooling section due to thermal stress concentration. The shattered powder will drift to the heating element, causing a short circuit, and will also cause the downstream atmosphere purity to become uncontrolled. Such continuous micro-impacts disrupt the contact and diffusion dynamics between powder particles, inducing irreversible grain boundary microcracks at the microscale of the product. This leads to irregular brittle fracture of the sintered product during subsequent cooling or processing, resulting in the implicit scrapping of the entire batch. Furthermore, conventional monitoring systems cannot detect these faults, which restricts the consistency and yield of high-end vanadium-nitrogen alloy products.

[0003] Research and practical application of the above methods and existing technologies have revealed that they have at least the following shortcomings:

[0004] The servo valve of the pusher plate hydraulic system generates high-frequency self-excited oscillation due to trace cavitation in the oil. This oscillation is transmitted to the product during the sintering process through the pusher plate, causing microstructural cracks to form in the solid-phase reaction. The cracked product will pulverize due to thermal stress concentration in the cooling section. The pulverized powder will drift to the heating element and cause a short circuit. At the same time, it will also cause the purity of the downstream atmosphere to become uncontrolled, leading to a series of production and quality problems.

[0005] In view of this, the present invention proposes a fault early warning and control system for a double-pusher kiln for the preparation of vanadium-nitrogen alloy products to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a fault early warning and control system for a double-pusher kiln in the preparation of vanadium-nitrogen alloy products, comprising:

[0007] Data acquisition module: Collects raw sensing data, including mechanical data and process data; mechanical data includes servo valve vibration signals. and hydraulic pressure signal Process data includes micro-differential pressure sequences. Temperature sequence Voltage sequences associated with resistivity ;

[0008] Feature extraction module: Constructs a baseline model of the device and generates a quality perturbation fingerprint by combining the raw sensing data;

[0009] Analysis and early warning module: The quality disturbance fingerprint is used as the input of the correlation prediction model to obtain the predicted value of the quality index, and the quality early warning signal is triggered based on the predicted value of the quality index.

[0010] Composite control module: Based on the quality warning signal output by the analysis and early warning module, it sends a feedback suppression command to the hydraulic servo system, analyzes the expected quality deviation, and obtains the adjustment amount of process data.

[0011] Furthermore, methods for constructing benchmark models include:

[0012] Collect raw sensing data within a preset time period and preprocess it to obtain baseline sensing data;

[0013] The model order is obtained by analyzing the baseline sensing data using the subspace identification method.

[0014] A state-space model is established based on the model order, and the prediction error is calculated based on the actual output and the linear prediction output of the state-space model.

[0015] Reconstructed samples are obtained by reconstructing the baseline sensing data and linear prediction output based on the sliding window length;

[0016] The reconstructed samples are used as input to the LSTM model to obtain the reconstructed predicted values;

[0017] Based on the actual output, reconstructed predicted value, and prediction error, a mean squared error loss function is constructed. The LSTM model is optimized with the goal of minimizing the loss function, and the optimized LSTM model is used as the baseline model.

[0018] Furthermore, methods for obtaining the model order include:

[0019] Construct the Hankel matrix for the baseline sensing data. Each column block, consisting of all variable data from the first p time steps, forms the past window matrix, labeled as the upper block; subsequent... Each column block, composed of all variable data at each time point, forms the future window matrix and is labeled as the lower block; the upper and lower blocks are stacked to form a Hankel matrix;

[0020] Singular value decomposition of the Hankel matrix yields the left singular vector matrix, the singular value matrix, and the right singular vector matrix.

[0021] Normalize all singular values ​​in the singular value matrix to obtain standardized singular values. Obtain the index of the first standardized singular value that is not greater than the singular threshold. Then, subtract 1 from the index value to obtain the model order.

[0022] Furthermore, methods for obtaining the length of the sliding window include:

[0023] Calculate the cross-correlation function between any two variables in the baseline sensing data, obtain the maximum dominant delay time corresponding to the maximum value of the cross-correlation function, obtain the optimal lag order set based on the maximum dominant delay time, and set the sliding window length by combining the maximum dominant delay time and the optimal lag order set.

[0024] Furthermore, methods for obtaining the optimal set of lag orders include:

[0025] A set of candidate lag orders is set based on the maximum dominant delay time, where the candidate lag orders are integers ranging from 0 to twice the maximum dominant delay time;

[0026] For each group of variables and Iterate through the set of candidate lag orders, construct an autoregressive model for each candidate lag order, calculate the AIC value corresponding to each candidate lag order, and select the candidate lag order with the smallest AIC value as the variable pair. and The optimal lag order is obtained by organizing the optimal lag orders of all variable pairs to obtain the set of optimal lag orders.

[0027] Furthermore, methods for generating quality perturbation fingerprints by combining raw perceptual data include:

[0028] The raw sensing data is used as input to the baseline model to obtain the ideal prediction value; the residual between each raw sensing data and the corresponding ideal prediction value is calculated to obtain the variable residual; the variable residual includes mechanical residual and process residual; the process residual is time offset calibrated based on the average transmission delay time; where the average transmission delay time is the average of the dominant delay times of the process data.

[0029] The variable-wise residuals are denoised to obtain the calibrated and denoised residuals; each calibrated and denoised residual is subjected to CWT transformation to obtain the time-frequency distribution matrix; the time-frequency distribution matrix is ​​amplitude-scaled to obtain the amplitude matrix corresponding to the original sensing data in sequence.

[0030] right and The corresponding amplitude matrix is ​​integrated to calculate the periodic band energy of the push plate.

[0031] according to and The corresponding amplitude matrix is ​​used to calculate the mechanical disturbance coherence coefficient;

[0032] According to process data , and The peak value of high-frequency noise amplitude is obtained by filtering the corresponding amplitude matrix;

[0033] The standard deviation of the residuals is calculated based on the residuals corresponding to the process data.

[0034] according to , and The corresponding amplitude matrix and the frequency band energy of the push plate period are used to calculate the time-frequency coordination coefficient.

[0035] The quality disturbance fingerprint is obtained by normalizing the frequency band energy of the push plate cycle, the coherence coefficient of mechanical disturbance, the peak value of high frequency noise amplitude, the standard deviation of residual and the time-frequency coherence coefficient and then splicing them together.

[0036] Furthermore, methods for determining whether to trigger a quality early warning signal based on the predicted value of the quality indicator include:

[0037] The difference between the predicted value of the quality indicator and the allowable threshold of the quality indicator is calculated to obtain the expected deviation of quality. When the expected deviation of quality is not lower than the preset deviation threshold, a quality warning signal is triggered.

[0038] Furthermore, methods for sending feedback suppression commands to the hydraulic servo system include:

[0039] When a quality warning signal is triggered, a PI control command is constructed based on the energy of the pusher period frequency band. The PI control command is adjusted according to the range of the command amplitude to obtain a PI adjustment command, which is then sent to the servo controller.

[0040] Furthermore, methods for obtaining process data adjustment amounts include:

[0041] A linear correlation model between expected quality deviation and adjustment amount of process data is constructed based on historical data;

[0042] The adjustment amount of process data output by the linear correlation model is constrained according to the process constraints to obtain the adjustment amount of process data.

[0043] Furthermore, the micro-pressure difference sequence was obtained. The methods include:

[0044] Micro-differential pressure sensors are installed at the pusher plate inlet, sintering center, and pusher plate outlet. The raw voltage values ​​collected by the micro-differential pressure sensors are used to calculate the micro-differential pressure calibration coefficient based on the actual voltage and theoretical voltage. The actual micro-differential pressure is then converted by combining the voltage range of the acquisition card, the sensor range, and the micro-differential pressure calibration coefficient. The data from the three monitoring points are then stitched together to obtain the micro-differential pressure sequence. .

[0045] The technical effects and advantages of the fault early warning control system for a double-pusher kiln in the preparation of vanadium-nitrogen alloy products of the present invention are as follows:

[0046] This invention deploys multiple types of sensors at key nodes in the flutter propagation chain to comprehensively collect and accurately process raw mechanical and process sensing data. Based on high-quality, flutter-free operating data, a benchmark model integrating subspace identification and LSTM is constructed to form an ideal quality state reference. Subsequently, through residual analysis and time-frequency transformation, cross-physical domain quality disturbance fingerprints are extracted, transforming invisible micro-flutter into quantifiable features. A correlation prediction model is then used to predict quality risks and trigger early warnings. Finally, a composite control logic combining PI feedback suppression and process data compensation is employed to suppress high-frequency micro-flutter in servo valves at the source and compensate for process deviations. This not only precisely solves a series of problems caused by untriggered alarm flutter in servo valves due to trace oil cavitation, such as micro-cracks in the solid phase reaction of the product, cooling section crushing, short circuits in heating elements, and loss of atmosphere purity, but also achieves a shift from passive fault response to proactive quality prediction and precise intervention at the source. This significantly improves the yield rate and production process stability of vanadium-nitrogen alloy products, reduces equipment failures and production losses, and provides an effective reference for micro-disturbance control in similar high-precision sintering production. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the fault early warning control system for a double-pusher kiln used in the preparation of vanadium-nitrogen alloy products according to the present invention.

[0048] Figure 2 This is a schematic diagram of the method for constructing a benchmark model according to the present invention;

[0049] Figure 3 This is a schematic diagram of the method for obtaining the model order according to the present invention;

[0050] Figure 4 This is a schematic diagram of the method for generating quality perturbation fingerprints by combining raw sensing data according to the present invention. Detailed Implementation

[0051] 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.

[0052] Example 1:

[0053] Please see Figure 1 As shown, this embodiment provides a fault early warning and control system for a double-pusher kiln used in the preparation of vanadium-nitrogen alloy products, including:

[0054] Data acquisition module: Collects raw sensing data, including mechanical data and process data; mechanical data includes servo valve vibration signals. and hydraulic pressure signal Process data includes micro-differential pressure sequences. Temperature sequence Voltage sequences associated with resistivity ;

[0055] By deploying sensors along the flutter propagation chain, raw sensing data is collected. The flutter propagation chain includes the servo valve body and adjacent hydraulic lines of the double-pusher kiln, the inlet and return oil lines of the servo valve, and the interior of the double-pusher kiln sintering chamber. Acceleration sensors are deployed on the servo valve body and adjacent hydraulic lines to collect servo valve vibration signals. Dynamic pressure sensors are deployed on the inlet and return oil lines of the servo valve to collect hydraulic pressure signals. Process data sensors, including micro-differential pressure sensors, thermocouples, and non-contact eddy current sensors, are deployed inside the double-pusher kiln sintering chamber. The micro-differential pressure sensors are installed in the atmosphere flow path inside the kiln to monitor changes in atmosphere flow resistance in different areas of the double-pusher kiln. Thermocouples are arranged in the core sintering area of ​​the product to capture minute temperature fluctuations. Non-contact eddy current sensors are aligned with the vanadium-nitrogen alloy billet during the sintering process and indirectly sense changes in the surface resistivity of the billet through electromagnetic induction.

[0056] Obtaining micro-pressure differential sequences The methods include:

[0057] Micro-differential pressure sensors are installed at the pusher plate inlet, sintering center, and pusher plate outlet. The raw voltage values ​​collected by the micro-differential pressure sensors are used to calculate the micro-differential pressure calibration coefficient based on the actual voltage and theoretical voltage. The actual micro-differential pressure is then converted by combining the voltage range of the acquisition card, the sensor range, and the micro-differential pressure calibration coefficient. The data from the three monitoring points are then stitched together to obtain the micro-differential pressure sequence. Such as the micro-pressure differential calibration coefficient. ,in, Number of calibration points; For the first The original voltage values ​​at each calibration point; For the first The theoretical voltage value at each calibration point , For the first Standard voltage values ​​at each calibration point This is the minimum range of the differential pressure sensor. This represents the maximum range of the differential pressure sensor. This is the minimum voltage value for the data acquisition card. This represents the maximum voltage of the data acquisition card; the actual micro-voltage difference. ,in, for The original voltage value at that moment; This refers to the full-scale range of the differential pressure sensor.

[0058] Obtain resistivity-related voltage sequences The methods include:

[0059] The sensor's open-circuit voltage without a billet and its real-time voltage with a billet are collected using an eddy current sensor. Based on the open-circuit voltage and real-time voltage, and combined with a correlation calibration coefficient, the resistivity-related voltage is calculated. (Example: Resistivity-related voltage) ,in, The correlation calibration coefficients are obtained by fitting historical data. This is the real-time voltage when a blank is present. This is the open-circuit voltage of the sensor when there is no blank.

[0060] The data acquisition module deploys corresponding sensors at key nodes in the flutter propagation chain, such as the servo valve body and adjacent hydraulic pipelines, the servo valve inlet and return oil lines, and the sintering chamber of the double-push plate kiln. It accurately collects mechanical data such as servo valve vibration signals and hydraulic pressure signals, as well as process data such as micro-pressure difference sequences, temperature sequences, and resistivity-related voltage sequences. This multi-dimensional raw sensing data comprehensively captures the high-frequency micro-flutter caused by trace oil cavitation in the servo valve that does not trigger a displacement alarm, and the impact of this flutter propagating to the sintering chamber on the microstructure of the product's solid-phase reaction stage, the flow resistance of the kiln atmosphere, and slight temperature fluctuations. This provides core data support for subsequent analysis of flutter propagation patterns, diagnosis of its impact on product quality, establishment of early warning models, and implementation of compensation control. Furthermore, it captures flutter signals at the source and tracks their interference with the product and process, laying a data foundation for preventing microstructural cracks in the product, avoiding subsequent cooling section pulverization, heating element short circuits, and uncontrolled atmosphere purity.

[0061] Feature extraction module: Constructs a baseline model of the device and generates a quality perturbation fingerprint by combining the raw sensing data;

[0062] Reference Figure 2 Methods for constructing benchmark models include:

[0063] Select a preset time period, namely the period when the equipment is free from vibration, production is continuous, and product quality meets the standards, collect raw sensing data and preprocess it to obtain baseline sensing data;

[0064] The model order is obtained by processing the baseline sensing data through subspace identification.

[0065] Reference Figure 3 Methods for obtaining the model order include:

[0066] Construct the Hankel matrix of the benchmark sensing data, from the previous Each column block, composed of all variable data at each time point, forms the past window matrix, i.e., the corresponding... The number of rows in each block is marked as the previous block; subsequent blocks... Each column block, composed of all variable data at each time point, forms the future window matrix, i.e., the corresponding... Each prediction step is labeled as the next block; the upper and lower blocks are stacked to form a Hankel matrix;

[0067] Perform singular value decomposition on the Hankel matrix to obtain the left singular vector matrix, the singular value matrix, and the right singular vector matrix; for example... ,in, This is the Hankel matrix; It is a left singular vector matrix; It is a singular value matrix; It is the transpose of the right singular vector matrix;

[0068] The left singular vector matrix, singular value matrix, and right singular vector matrix are all divided according to the first n singular values ​​and the remaining singular values ​​to obtain the left singular vector label matrix, singular value label matrix, and right singular vector label matrix, as well as the left singular vector residual matrix, singular value residual matrix, and right singular vector residual matrix.

[0069] The left singular vector label matrix is ​​segmented to obtain the upper block. and the next block The observability matrix is ​​obtained by combining the singular value matrix; the controllability matrix is ​​obtained by calculating the Hankel matrix, singular value matrix, and right singular vector matrix; specifically, through... and Calculate the observability matrix The observability matrix is ​​a power-law combination of the prediction step size of the system output matrix and the state transition matrix. Observability matrix The former piece; Label the singular value matrix; by Calculate and obtain the controllability matrix The controllability matrix is ​​a power-law combination of the prediction step size of the system input matrix and the state transition matrix; where, Hankel matrix The former piece; The transpose of the matrix is ​​used to label the right singular vectors;

[0070] Normalize all singular values ​​in the singular value matrix to obtain standardized singular values. Obtain the index value corresponding to the first standardized singular value that is not greater than the singular threshold. Obtain the value corresponding to the index value minus 1 as the model order.

[0071] Calculate whether the rank of the observability matrix and the rank of the controllability matrix are equal to the model order. If yes, output the model order; otherwise, adjust the number of rows in the adjustment block. and prediction step size Until the rank of the observability matrix and the rank of the controllability matrix equal the model order.

[0072] A state-space model is established based on the model order, and the prediction error is calculated based on the actual output and the linear prediction output of the state-space model.

[0073] Based on the sliding window length, the baseline sensing data and linear prediction output are reconstructed to obtain reconstructed samples. The reconstructed samples include input samples and output samples. The input samples consist of all baseline sensing data from time k-L+1 to time k, and the output samples consist of all linear prediction outputs at time k-L+1. k is time; L is the sliding window length.

[0074] Methods for obtaining the length of a sliding window include:

[0075] Calculate any two variables in the baseline sensing data and The cross-correlation function between the two variables is used to obtain the maximum dominant delay time corresponding to the maximum value of the cross-correlation function. Based on the maximum dominant delay time, the optimal lag order set is obtained. The sliding window length is set in combination with the maximum dominant delay time and the optimal lag order set, such as setting the sliding window length to three times the maximum dominant delay time. Specifically, and Cross-correlation function between ,in, for Time of the first One variable; for Time of the first One variable, For time delay; For expectation operator; when When taking the maximum value, the corresponding That is right The maximum dominant delay time.

[0076] Methods for obtaining the optimal set of lag orders include:

[0077] A set of candidate lag orders is set based on the maximum dominant delay time, where the candidate lag orders are integers ranging from 0 to twice the maximum dominant delay time;

[0078] For each group of variables and Iterate through the set of candidate lag orders, construct an autoregressive model for each candidate lag order, calculate the AIC value corresponding to each candidate lag order, and select the candidate lag order with the smallest AIC value as the variable pair. and The optimal lag order is obtained by organizing the optimal lag orders of all variable pairs. For example, in an autoregressive model... ,in, For constant terms; The coefficient of the lagged term; For residuals; For lagged terms; candidate lag order Corresponding AIC value ,in, The number of model parameters; This is the maximum likelihood estimate of the model.

[0079] The reconstructed samples are used as input to the LSTM model to obtain the reconstructed predicted values;

[0080] Based on the actual output, reconstructed predicted value, and prediction error, a mean squared error loss function is constructed. The LSTM model is optimized with the goal of minimizing the loss function, and the optimized LSTM model is used as the baseline model.

[0081] The above steps select data from periods of equipment operation without chatter, with continuous production and quality meeting standards as benchmark sensing data. Using the ideal quality state as the core of modeling, the data is first processed using a subspace identification method. A Hankel matrix is ​​constructed to integrate past and future window data of variables. Singular value decomposition is performed to extract observability and controllability matrices, accurately capturing the dynamic coupling relationship between mechanical and process variables. Singular value normalization and rank verification are combined to determine the model order that can characterize the true dynamic characteristics of the system. Then, based on the maximum dominant delay time and optimal lag order obtained from cross-correlation analysis, the sliding window length is determined to ensure the model covers the complete delay process of chatter propagation from the mechanical domain to the process domain. Finally, the reconstructed samples are input into the LSTM model, and a loss function optimization model is constructed by combining the actual output and prediction error, forming a benchmark model that can accurately characterize the dynamic response law of the system under chatter-free conditions. This benchmark model is used to identify high-frequency micro-channels caused by trace amounts of oil cavitation in servo valves. Small chatter provides an ideal quality state reference standard. The dynamic coupling relationship between mechanical and process variables captured by subspace identification can accurately reflect the normal correlation law on the chatter propagation path. The time-series prediction capability of LSTM and the sliding window design that adapts to the propagation delay enable the model to accurately output the predicted values ​​of mechanical and process data under ideal working conditions. When the servo valve experiences small chatter, the deviation between the actual collected data and the predicted value of the benchmark model can intuitively quantify the degree of disturbance of variables by chatter. In particular, it can capture the slight fluctuation of process data caused by chatter after it is transmitted to the sintering chamber. This provides the core normal and abnormal comparison basis for subsequent extraction of quality disturbance fingerprints through residual analysis, generation of hydraulic servo system feedback suppression commands and process data adjustment. This enables early identification and precise intervention of the risk of micro-cracks in products caused by small chatter, preventing a series of problems such as product crushing in the cooling section, short circuit of heating elements and downstream atmosphere runaway from the root.

[0082] Reference Figure 4 Methods for generating quality perturbation fingerprints by combining raw sensing data include:

[0083] The raw sensor data is used as input to the baseline model to obtain the ideal prediction value;

[0084] Calculate the residual between each original sensing data and the corresponding ideal prediction value to obtain the variable-wise residual; the variable-wise residual includes mechanical residual and process residual;

[0085] Time offset calibration of process residuals is performed based on the average propagation delay time; where the average propagation delay time is the average of the dominant delay times of the process data.

[0086] Wavelet threshold denoising method is used to denoise mechanical residuals to obtain denoised mechanical residuals; Kalman filtering denoising method is used to denoise process residuals to obtain denoised process residuals; the residuals after calibration and denoising are summarized.

[0087] For each calibration-denoised residual, perform CWT transformation to obtain the time-frequency distribution matrix. Then, perform amplitude scaling on the time-frequency distribution matrix to obtain the amplitude matrices corresponding to the original sensing data. Specifically, the... The time-frequency distribution matrix of the residual after calibration and denoising. ,in, For time shift parameters, , The translation interval for the time translation parameter. This represents the upper limit of the translation interval. This is the lower limit of the translation interval. For a specific moment; The complex conjugate of the mother wavelet function; For frequency The corresponding time scale parameters; For the first The residual after calibration and denoising; amplitude matrix ,in, For the real part of a complex number; It represents the imaginary part of a complex number;

[0088] right and Integrating the corresponding amplitude matrix, the energy of the pusher period band is calculated; for example... ,in, The energy of the push plate period frequency band; For the residual frequency band, This represents the upper limit of the residual frequency band. This represents the lower limit of the residual frequency band. for The corresponding magnitude matrix; for The corresponding magnitude matrix;

[0089] according to and The corresponding amplitude matrix is ​​used to calculate the mechanical disturbance coherence coefficient; for example, the mechanical disturbance coherence coefficient. ;

[0090] According to process data , and The peak amplitude of high-frequency noise is obtained by filtering the corresponding amplitude matrix; for example... ,in, This represents the peak value of the high-frequency noise. High-frequency threshold; for The corresponding magnitude matrix; for The corresponding magnitude matrix; for The corresponding magnitude matrix;

[0091] The standard deviation of the residuals is calculated based on the residuals corresponding to the process data; for example, the standard deviation of the residuals. ,in, For the first Each type of process data, namely: , and The corresponding residual standard deviation; The first output of the baseline model Each type of process data, namely: , and The corresponding residual standard deviation;

[0092] according to , and The corresponding amplitude matrix and the frequency band energy of the push plate period are used to calculate the time-frequency coordination coefficient; such as the time-frequency coordination coefficient. ;

[0093] The quality disturbance fingerprint is obtained by normalizing the frequency band energy of the push plate cycle, the coherence coefficient of mechanical disturbance, the peak value of high frequency noise amplitude, the standard deviation of residual and the time-frequency coherence coefficient and then splicing them together.

[0094] The method for generating a quality disturbance fingerprint involves inputting raw sensing data into a benchmark model to obtain ideal predicted values, calculating and distinguishing between mechanical and process-related variable residuals, calibrating the time offset of process-related residuals using the average propagation delay time, and then optimizing the two types of residuals using wavelet thresholding and Kalman filtering respectively. Subsequently, the calibrated and denoised residuals are subjected to CWT and amplitude normalization processing. From the amplitude matrix, the push plate period band energy, mechanical disturbance coherence coefficient, high-frequency noise amplitude peak value, residual standard deviation, and time-frequency coherence coefficient are extracted and normalized to form a unique quality disturbance fingerprint characterizing the flutter disturbance mode. This method overcomes the limitation that high-frequency micro-flutter in servo valves is difficult to identify due to its small amplitude and failure to trigger an alarm. By using residual analysis, it amplifies the small offsets in mechanical and process-related data caused by flutter. By leveraging time-frequency analysis, the propagation characteristics of chatter from the hydraulic system through the pusher plate to the sintering chamber across physical domains were accurately captured. In particular, by observing the high-frequency fluctuations, statistical distribution changes, and cross-domain synergy coefficients of process residuals, the microstructural changes caused by chatter during the solid-state reaction stage of the product can be indirectly perceived. The generated quality disturbance fingerprint transforms the invisible micro-chatter into a quantifiable feature vector, providing a core identification basis for establishing correlation prediction models and determining whether chatter is about to cause micro-cracks in the product. This lays the foundation for accurately generating feedback suppression commands for the hydraulic servo system and process data compensation adjustments, thus blocking the negative impact of chatter on product quality from the disturbance identification level and avoiding a series of problems such as product crushing in the cooling section, short circuits in heating elements, and uncontrolled purity of downstream atmosphere.

[0095] Analysis and early warning module: The quality disturbance fingerprint is used as the input of the correlation prediction model to obtain the predicted value of the quality index, and the quality early warning signal is triggered based on the predicted value of the quality index.

[0096] Methods for determining whether to trigger a quality warning signal based on predicted quality indicator values ​​include:

[0097] The difference between the predicted value of the quality indicator and the allowable threshold of the quality indicator is calculated to obtain the expected deviation of quality. When the expected deviation of quality is not lower than the preset deviation threshold, a quality warning signal is triggered. The preset deviation threshold is obtained based on historical data statistics.

[0098] The analysis and early warning module inputs the quality disturbance fingerprint, which characterizes the cross-domain disturbance mode of high-frequency micro-chatter in servo valves, into the prediction model to accurately obtain the predicted values ​​of key product quality indicators. Then, by calculating the difference between the predicted value and the allowable threshold of the quality indicator, the expected quality deviation is obtained. When the expected quality deviation is not lower than the preset deviation threshold, a quality early warning signal is triggered, directly linking the chatter disturbance to the risk of product quality defects. This breaks through the limitation of traditional methods that can only detect equipment vibration but cannot predict the impact on quality. It can identify in advance the risk that the micro-chatter caused by trace cavitation of oil in the servo valve will cause micro-structural cracks in the solid-phase reaction stage of the product. This avoids the product with cracks from being crushed due to thermal stress concentration after entering the cooling section, thereby preventing short circuits in heating elements and loss of downstream atmosphere purity caused by the dispersion of crushed powder. It realizes the transformation from passively responding to faults to actively predicting quality risks, buying time for subsequent precise suppression and compensation control, and blocking the transmission chain of a series of production and quality problems from the early warning level.

[0099] Composite control module: Based on the quality warning signal output by the analysis and early warning module, it sends a feedback suppression command to the hydraulic servo system, analyzes the expected quality deviation, and obtains the adjustment amount of process data.

[0100] Methods for sending feedback suppression commands to a hydraulic servo system include:

[0101] When a quality warning signal is triggered, a PI control command is constructed based on the energy of the pusher cycle frequency band. The PI control command is adjusted according to the command amplitude range to obtain a PI adjustment command, which is then sent to the servo controller. The command amplitude range is obtained based on historical data statistics to avoid overloading the servo system due to excessively strong commands. When the PI control command exceeds the command amplitude range, the value of the PI control command is set to the value corresponding to the command amplitude range closest to the current value. For example, if the PI control command exceeds the upper limit of the command amplitude range, the value of the PI control command is set to the upper limit of the command amplitude range; if the PI control command is lower than the lower limit of the command amplitude range, the value of the PI control command is set to the lower limit of the command amplitude range.

[0102] Methods for obtaining process data adjustment amounts include:

[0103] A linear correlation model between expected quality deviation and adjustment amount of process data is constructed based on historical data;

[0104] The adjustment amount of process data output by the linear correlation model is constrained according to the process constraints to obtain the adjustment amount of process data.

[0105] After the quality warning signal is triggered by the analysis and warning module, the composite control module, on the one hand, constructs a PI control command based on the frequency band energy of the pusher plate period and adjusts the amplitude range of the command by combining historical data statistics. The adjusted PI control command is then sent to the servo controller to suppress the high-frequency self-excited oscillation of the servo valve of the pusher plate hydraulic system caused by minor oil cavitation at the source, reducing the transmission of chatter to the product through the pusher plate. On the other hand, it constructs a linear correlation model between the expected quality deviation and the adjustment amount of process data based on historical data. Combined with process constraints, it constrains the adjustment amount output by the model to obtain the process data adjustment amount adapted to actual production. This compensates for the negative impact of chatter on the solid-phase reaction stage of the product, avoids the formation of microstructural cracks in the product, and prevents the product with cracks from shattering due to thermal stress concentration in the cooling section, as well as preventing short circuits of heating elements and loss of purity of downstream atmosphere caused by the dispersion of shattered powder. Through the composite control logic of source chatter suppression and process quality compensation, a closed-loop intervention is formed to block a series of production and quality problems from the execution level.

[0106] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0107] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault early warning and control system for a double-pusher kiln in the preparation of vanadium-nitrogen alloy products, characterized in that, include: Data acquisition module: Collects raw sensing data, including mechanical data and process data; Mechanical data includes servo valve vibration signals. and hydraulic pressure signal Process data includes micro-differential pressure sequences. Temperature sequence Voltage sequences associated with resistivity ; Feature extraction module: Building the baseline model of the device: Collect raw sensing data within a preset time period and preprocess it to obtain baseline sensing data; The model order is obtained by analyzing the baseline sensing data using the subspace identification method. A state-space model is established based on the model order, and the prediction error is calculated based on the actual output and the linear prediction output of the state-space model. Reconstructed samples are obtained by reconstructing the baseline sensing data and linear prediction output based on the sliding window length; The reconstructed samples are used as input to the LSTM model to obtain the reconstructed predicted values; The mean squared error loss function is constructed based on the actual output, reconstructed prediction value and prediction error. The LSTM model is optimized with the goal of minimizing the loss function to obtain the optimized LSTM model. The optimized LSTM model is then used as the baseline model. Generate quality perturbation fingerprints by combining raw sensor data: The raw sensor data is used as input to the baseline model to obtain the ideal prediction value; Calculate the residual between each raw sensed data and the corresponding ideal predicted value to obtain the variable-wise residual; The variable-wise residuals are denoised to obtain the calibrated and denoised residuals; CWT transformation is performed on each calibration and denoising residual to obtain the time-frequency distribution matrix. The time-frequency distribution matrix is ​​then amplitude-scaled to obtain the amplitude matrix corresponding to the original sensing data. right and The corresponding amplitude matrix is ​​integrated to calculate the periodic band energy of the push plate. according to and The corresponding amplitude matrix is ​​used to calculate the mechanical disturbance coherence coefficient; According to process data , and The peak value of high-frequency noise amplitude is obtained by filtering the corresponding amplitude matrix; The standard deviation of the residuals is calculated based on the residuals corresponding to the process data. according to , and The corresponding amplitude matrix and the frequency band energy of the push plate period are used to calculate the time-frequency coordination coefficient. The quality disturbance fingerprint is obtained by normalizing the frequency band energy of the push plate cycle, the coherence coefficient of mechanical disturbance, the peak value of high frequency noise amplitude, the standard deviation of residual and the time-frequency coordination coefficient and then splicing them together. Analysis and Early Warning Module: This module uses the quality disturbance fingerprint as input to the correlation prediction model to obtain predicted values ​​for quality indicators. Based on these predicted values, it determines whether to trigger a quality early warning signal. The difference between the predicted value of the quality indicator and the allowable threshold of the quality indicator is calculated to obtain the expected deviation of quality. When the expected deviation of quality is not lower than the preset deviation threshold, a quality warning signal is triggered. Composite control module: Based on the quality warning signal output by the analysis and early warning module, it sends a feedback suppression command to the hydraulic servo system, analyzes the expected quality deviation, and obtains the adjustment amount of process data.

2. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 1, characterized in that, Methods for obtaining the model order include: Construct the Hankel matrix for the baseline sensing data. Each column block, consisting of all variable data from the first p time steps, forms the past window matrix, labeled as the upper block; subsequent... Each column block, composed of all variable data at each time point, forms the future window matrix and is labeled as the lower block; the upper and lower blocks are stacked to form a Hankel matrix; Singular value decomposition of the Hankel matrix yields the left singular vector matrix, the singular value matrix, and the right singular vector matrix. Normalize all singular values ​​in the singular value matrix to obtain standardized singular values. Obtain the index of the first standardized singular value that is not greater than the singular threshold. Then, subtract 1 from the index value to obtain the model order.

3. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 1, characterized in that, Methods for obtaining the length of a sliding window include: Calculate the cross-correlation function between any two variables in the baseline sensing data, obtain the maximum dominant delay time corresponding to the maximum value of the cross-correlation function, obtain the optimal lag order set based on the maximum dominant delay time, and set the sliding window length by combining the maximum dominant delay time and the optimal lag order set.

4. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 3, characterized in that, Methods for obtaining the optimal set of lag orders include: A set of candidate lag orders is set based on the maximum dominant delay time, where the candidate lag orders are integers ranging from 0 to twice the maximum dominant delay time; For each group of variables and Iterate through the set of candidate lag orders, construct an autoregressive model for each candidate lag order, calculate the AIC value corresponding to each candidate lag order, and select the candidate lag order with the smallest AIC value as the variable pair. and The optimal lag order is obtained by organizing the optimal lag orders of all variable pairs to obtain the set of optimal lag orders.

5. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 1, characterized in that, The variable-wise residuals include mechanical residuals and process residuals; the process residuals are time-offset calibrated based on the average propagation delay time; wherein, the average propagation delay time is the average of the dominant delay times of the process data.

6. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 1, characterized in that, Methods for sending feedback suppression commands to a hydraulic servo system include: When a quality warning signal is triggered, a PI control command is constructed based on the energy of the pusher period frequency band. The PI control command is adjusted according to the range of the command amplitude to obtain a PI adjustment command, which is then sent to the servo controller.

7. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 1, characterized in that, Methods for obtaining process data adjustment amounts include: A linear correlation model between expected quality deviation and adjustment amount of process data is constructed based on historical data; The adjustment amount of process data output by the linear correlation model is constrained according to the process constraints to obtain the adjustment amount of process data.

8. The fault early warning control system for a double-pusher kiln for preparing vanadium-nitrogen alloy products according to claim 1, characterized in that, Obtaining micro-pressure differential sequences The methods include: Micro-differential pressure sensors are installed at the pusher plate inlet, sintering center, and pusher plate outlet. The raw voltage values ​​collected by the micro-differential pressure sensors are used to calculate the micro-differential pressure calibration coefficient based on the actual voltage and theoretical voltage. The actual micro-differential pressure is then converted by combining the voltage range of the acquisition card, the sensor range, and the micro-differential pressure calibration coefficient. The data from the three monitoring points are then stitched together to obtain the micro-differential pressure sequence. .

Images