A real-time acquisition system for glass melting process parameters

Abstract

The application provides a glass melting process parameter real-time acquisition system, comprising: dividing sensor process parameter upstream and downstream parameter groups according to physical cause and effect chain; extracting parameter short-term fluctuation and relative long-term baseline drift characteristics, obtaining process stability characteristic vector by means of attention network fusion, and setting a reference sampling period accordingly; using upstream parameter stability characteristics to feed forward adjust downstream sampling period and error code decision threshold through a time delay model; constructing a target parameter local prediction model through meta-learning according to the stability characteristics; when the parameter measured value exceeds the prediction confidence interval, switching the target and strongly correlated parameters to a high-frequency calibration period; under high-frequency calibration, if the time series residual energy of the measured value and the predicted value is continuously higher than the threshold value, and after excluding real working condition mutations through synchronous change verification of strongly correlated parameters, it is determined that there is an error code, the measured output is isolated and the predicted value is used instead; after the residual energy is lower than the recovery threshold, the isolation is removed and the reference sampling period is restored.

Description

Technical Field

[0001] This application belongs to the field of real-time data acquisition, and in particular relates to a real-time data acquisition system for glass melting process parameters. Background Technology

[0002] Real-time acquisition of process parameters such as liquid level, temperature, pressure, and gas flow rate during glass melting furnace operation is crucial for ensuring stable operation of the automatic control system, improving the quality of finished glass products, and reducing energy consumption. However, the internal environment of a glass melting furnace is complex and harsh. Sensors are exposed to high temperatures, dust, and strong electromagnetic interference for extended periods, making them prone to performance degradation, measurement drift, or sudden errors. Traditional parameter acquisition systems typically use a fixed sampling frequency, which can lead to wasted computing resources and data storage space when operating conditions are stable. Furthermore, it is difficult to increase the sampling density in time to capture critical transient changes during periods of drastic fluctuations or sudden sensor anomalies. In addition, there is a significant time delay between upstream and downstream glass melting process parameters. Traditional acquisition methods struggle to fully utilize the strong correlations between parameters for feedforward adjustment and collaborative monitoring, resulting in a lag in the system's response to changes in operating conditions. Existing mechanisms for detecting and handling process parameter anomalies often fail to effectively distinguish between normal short-term process fluctuations, long-term equipment baseline drift, and genuine sensor errors, easily leading to false alarms or missed alarms. Once a bit error is confirmed, existing data acquisition systems mostly handle it by either removing outliers or maintaining the old values ​​from the previous moment, lacking a flexible fault-tolerance mechanism based on prediction results. When harsh or distorted data enters the downstream closed-loop control system, it may cause abrupt changes in control signals and violent oscillations in actuators, disrupting the thermal balance of the glass melting process, leading to a decrease in product yield and even inducing production safety risks. Furthermore, existing technologies lack a smooth transition mechanism during the recovery phase after anomaly resolution, making it difficult to assess whether the sensor has returned to stability based on the actual residual state. Therefore, the industry urgently needs a real-time parameter acquisition method that can integrate the process causal chain, intelligently adjust sampling strategies, and possess reliable bit error detection, flexible isolation, and smooth recovery capabilities to improve the safety, continuity, and data reliability of the glass melting process. Summary of the Invention

[0003] To enhance the anti-interference capability and data reliability of the process parameter acquisition system, according to a first aspect of the present invention, a real-time acquisition system for glass melting process parameters is provided, the system comprising the following modules:

[0004] The fusion module is used to divide the process parameters collected by the sensor into parameter groups with upstream and downstream relationships according to the preset physical process causal chain; calculate the short-term fluctuation characteristics of each parameter and the drift characteristics between the current collected value and the long-term process baseline; fuse the short-term fluctuation characteristics and drift characteristics through the attention mechanism network to generate a process stability feature vector, and set the benchmark sampling period according to the process stability feature vector. The switching module is used to adjust the baseline sampling period and error decision threshold of the downstream parameter group by using the process stability feature vector of the upstream parameter group and processing it with a time delay model; based on the process stability feature vector, a meta-learning framework is used to generate a local prediction model for the target parameter; when a single sampling value of the target parameter exceeds the prediction confidence interval of the local prediction model, the sampling period of the target parameter and the strongly correlated parameter is switched to a high-frequency verification period. The acquisition module is used to determine that a bit error has occurred and isolate the control output of the actual acquired value during the high-frequency verification period if the time series residual energy value between the actual acquired value and the theoretical predicted value of the target parameter continuously exceeds the bit error decision threshold, and after the synchronous change verification of the strongly correlated parameters excludes the sudden change of the actual working condition, the module outputs the theoretical predicted value during the isolation period. When the time series residual energy value between the subsequent actual acquired value and the theoretical predicted value of the target parameter is lower than the recovery threshold during the isolation period, the module releases the isolation and restores the system to the reference sampling period.

[0005] According to a second aspect of the present invention, a method for real-time acquisition of glass melting process parameters is provided, the method comprising: Based on the preset physical process causal chain, the process parameters collected by the sensor are divided into parameter groups with upstream and downstream relationships; the short-term fluctuation characteristics of each parameter are calculated, and the drift characteristics between the current collected value and the long-term process baseline are calculated; the short-term fluctuation characteristics and drift characteristics are fused through the attention mechanism network to generate a process stability feature vector, and the benchmark sampling period is set according to the process stability feature vector. Using the process stability feature vector of the upstream parameter group, after processing by the time delay model, the baseline sampling period and error decision threshold of the downstream parameter group are adjusted by feedforward; based on the process stability feature vector, a local prediction model for the target parameter is generated using a meta-learning framework; when the single sampling value of the target parameter exceeds the prediction confidence interval of the local prediction model, the sampling period of the target parameter and the strongly correlated parameter is switched to a high-frequency verification period. During the high-frequency verification period, if the time series residual energy value between the actual collected value and the theoretical predicted value of the target parameter continuously exceeds the bit error decision threshold, and after the synchronous change verification of the strongly correlated parameter eliminates the sudden change in the actual working condition, a bit error is determined to have occurred and the control output of the actual collected value is isolated. During the isolation period, the theoretical predicted value is output. When the time series residual energy value between the subsequent actual collected value and the theoretical predicted value of the target parameter is lower than the recovery threshold during the isolation period, the isolation is lifted and the system is restored to the reference sampling period.

[0006] This invention groups process parameters based on the physical causal chain in the glass melting process and integrates short-term fluctuation characteristics and long-term drift characteristics using an attention mechanism to assess process stability and reasonably determine the baseline sampling period. By using a time delay model to feed forward upstream parameter characteristics to the sampling period and error judgment threshold of downstream parameters, the system's response time to changes in process status is improved. A local prediction model for the target parameter is constructed using a meta-learning framework, switching to a high-frequency verification state when the parameter acquisition value exceeds the prediction confidence interval to enhance the ability to identify sudden anomalies. Furthermore, error judgment is performed based on the time-series residual energy value, and synchronous changes of strongly correlated parameters are used to verify and eliminate sudden changes in actual operating conditions. After confirming the error, the abnormal actual acquisition value is isolated, and the theoretical prediction value is used to replace the output, thereby preventing erroneous data from interfering with the production control process. When the residual energy meets the recovery threshold condition, the system smoothly releases the isolation and restores the baseline sampling period, thereby improving the anti-interference capability, fault tolerance, and data reliability of the glass melting process parameter acquisition system. Attached Figure Description

[0007] Figure 1 The graph shows the number of iterations for fine-tuning versus the prediction error. Figure 2 This is a graph showing the relationship between residual energy and error decision threshold. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0009] This application discloses a real-time acquisition system for glass melting process parameters, the system comprising: The fusion module is used to divide the process parameters collected by the sensor into parameter groups with upstream and downstream relationships according to the preset physical process causal chain; calculate the short-term fluctuation characteristics of each parameter and the drift characteristics between the current collected value and the long-term process baseline; fuse the short-term fluctuation characteristics and drift characteristics through the attention mechanism network to generate a process stability feature vector, and set the benchmark sampling period according to the process stability feature vector.

[0010] A directed acyclic graph (DAG) model is constructed based on the physical process of glass melting. The kiln head feed rate and fuel flow rate are used as starting nodes, the melting zone temperature and refining zone temperature as intermediate nodes, and the liquid level in the feed channel as the ending node. The process parameters from each sensor are sorted and grouped according to the flow of process information, resulting in upstream and downstream parameter groups. When intermediate process identifiers need to be retained, the parameter group located between the starting and ending nodes is marked as an intermediate transition parameter group. In the feedforward calculation, this intermediate transition parameter group is treated as a downstream parameter group relative to the preceding parameter group and as an upstream parameter group relative to the following parameter group. Real-time time series data for each parameter group are stored separately. A first sliding time window and a second sliding time window are set for each process parameter. First-order difference processing is performed on the data within the first sliding time window to obtain a discrete difference sequence. The sum of squares and square root operations are then performed on the discrete difference sequence to extract short-term fluctuation characteristics. Simultaneously, the arithmetic mean of historical collected values ​​is calculated within the second sliding time window as the long-term process baseline. The current collected value is subtracted from the long-term process baseline, and the absolute value is taken to obtain the drift characteristics.

[0011] A fusion network is constructed by first performing dimensionless processing on short-term fluctuation characteristics and drift characteristics according to the engineering range, historical standard deviation, or preset reference scale corresponding to each process parameter, resulting in dimensionless short-term fluctuation characteristics and dimensionless drift characteristics. These dimensionless short-term fluctuation characteristics and dimensionless drift characteristics are then input into the linear mapping layer of the attention mechanism network. The corresponding attention weights are obtained through the dot product operation of the query vector and the key vector, followed by normalization. The dimensionless short-term fluctuation characteristics and dimensionless drift characteristics are then multiplied by their respective attention weights. Subsequently, the weighted short-term fluctuation characteristics and drift characteristics are concatenated along their dimensional direction to output a two-dimensional process stability feature vector. The L2 norm of this process stability feature vector is calculated as its modulus, which characterizes the intensity of process fluctuations. The modulus is converted by a ratio according to a preset reference modulus to obtain a dimensionless fluctuation intensity value. The target value is obtained by dividing the reference time constant after the direction angle correction by the largest of the dimensionless fluctuation intensity value and the preset lower limit value. The target value is written into the sampling clock configuration register and set as the reference sampling period of the current parameter.

[0012] In some implementations, the process parameters collected by the sensors are divided into parameter groups with upstream and downstream relationships according to a preset physical process causal chain, including: Extract the spatial coordinates of each sensor in the three-dimensional solid model of the glass furnace, as well as the material flow vector indicating the glass melt transport route; Calculate the spatial straight-line distance between each sensor and identify the physical medium phase of each sensor. Sensors in the same physical medium phase and whose spatial straight-line distance is less than a preset distance threshold are grouped into the same control unit. The dataset collected by the control unit constitutes a parameter group. Along the material flow direction vector, calculate the projected position of each control unit on the transport route. Based on the order of the projected positions, the parameter group corresponding to the control unit that is relatively close to the starting source is divided into the upstream parameter group, and the parameter group corresponding to the control unit that is relatively close to the ending position is divided into the downstream parameter group.

[0013] Extract the 3D spatial coordinates from the 3D model of the kiln, as well as the material flow vector along the length of the kiln, such as setting it as a positive vector along the longitudinal axis. The physical medium phases identified by the sensors include the gas phase (the upper combustion space of the kiln), the liquid phase (molten glass), and the solid phase (the refractory material of the kiln walls and bottom). The preset distance threshold is preferably in the range of 0.5 to 2.0 meters, with an example value of 1.2 meters. Calculate the spatial distance between two sensors; for example, if temperature sensor A and level sensor B are both in the liquid phase and the spatial distance is 0.8 meters, they are classified as belonging to the same control unit. Calculate the 3D coordinates of the geometric center of each control unit and perform a dot product operation with the flow vector to obtain the projected position on the transport path. Sort the units according to their projection values ​​from smallest to largest; for example, classify units with a projection value of 5 meters as the upstream parameter group in the melting section, and units with a projection value of 15 meters as the downstream parameter group in the refining or forming section, thereby establishing the physical topology and data causal chain. When there is a control unit whose projection position is between the upstream parameter group and the downstream parameter group, it is used as an intermediate transition parameter group and participates in the step-by-step feedforward adjustment from upstream to downstream according to the order between adjacent control units.

[0014] In some implementations, the calculation of the short-term fluctuation characteristics of each parameter includes: The length of the first sliding time window used to calculate short-term fluctuation characteristics is set to a preset time span; Within the first sliding time window, time series data of the target parameter are acquired according to a preset sampling frequency; The first-order difference calculation is performed on the collected values ​​at adjacent sampling times in the time series data to obtain a discrete difference sequence; Calculate the sum of squares of each element in the discrete difference sequence, and use the square root of the sum of squares as the short-term fluctuation characteristic of the parameter.

[0015] The preset time span of the first sliding time window is preferably set to 30 to 300 seconds, with an example value of 60 seconds. The preset sampling frequency is preferably set to 1Hz to 10Hz, with an example value of 1Hz. During operation, 60 consecutive data points are extracted from the time window to form time series data. First-order difference calculations are performed on the values ​​at adjacent time points, i.e., the value at the later time point is subtracted from the value at the previous time point, generating a discrete difference sequence of length 59. For example, for the kiln top temperature parameter, if the consecutive sampling values ​​are 1500.1℃, 1500.3℃, and 1499.8℃, the resulting difference values ​​are 0.2℃ and -0.5℃, respectively. The sum of squares is performed on all elements of this discrete difference sequence to obtain the difference energy value. If the sum of squares is 0.29, its square root is taken to obtain 0.538. 0.538 is used as the short-term fluctuation characteristic representing the short-term high-frequency oscillation intensity of this parameter, thereby filtering out the interference of baseline offset on the short-term fluctuation assessment in the form of the L2 norm.

[0016] In some implementations, calculating the drift characteristics of the current acquired value relative to the long-term process baseline includes: The length of the second sliding time window used to calculate the drift characteristics is set to be greater than the preset long span of the first sliding time window; Calculate the arithmetic mean of all historical collected values ​​of the parameter within the second sliding time window, and use the arithmetic mean as the long-term process baseline; Calculate the difference between the current acquired value and the long-term process baseline, and use the absolute value of the difference as the drift characteristic of the parameter.

[0017] The preset duration of the second sliding time window is preferably set to 30 to 120 minutes, with an example duration of 60 minutes (3600 seconds). At a sampling frequency of 1 Hz, 3600 consecutive historical data points are buffered within the window.

[0018] A long-term process baseline is established by calculating the arithmetic mean of these 3600 data points. For example, if the average glass melt level over the past hour is 1250.5 mm, then 1250.5 mm is determined as the long-term process baseline at the current moment. The latest single-data point is extracted, for example, 1251.8 mm, and the difference between it and the long-term process baseline is calculated to be 1.3 mm. The absolute value of this difference is taken as 1.3, and this value is used as the drift characteristic. This calculation process filters out local transient spikes without using filters, and extracts the macroscopic system-level deviation caused by changes in the firing process or the feed formula.

[0019] In some implementations, the process stability feature vector is generated by fusing short-term fluctuation features and drift features through an attention mechanism network, including: The dimensionless short-term fluctuation features and the drift features are input into the attention mechanism network, and the attention weights of the two are calculated by performing a dot product operation between the query vector in the network and the corresponding key vectors and then normalizing the result. The attention weights are used to calculate the short-term fluctuation features and drift features respectively. The weighted short-term fluctuation features and drift features are concatenated into a two-dimensional feature vector, and the two-dimensional feature vector is output as the process stability feature vector.

[0020] The attention mechanism network adopts a self-attention architecture based on a multilayer perceptron. The structure includes parallel query linear mapping layers and key linear mapping layers, a feature dot product calculation layer, a soft maximum normalization layer, and a terminal feature concatenation layer. The network input consists of short-term fluctuation features and drift features extracted in the preceding steps, and the network output is a two-dimensional process stability feature vector representing the current combined operating condition of the system. The short-term fluctuation features and drift features, as input, are respectively input to the corresponding linear mapping layers, and a query vector and a key vector of the same dimension are generated through projection using a linear transformation matrix, for example, uniformly set as an eight-dimensional feature vector.

[0021] The dot product similarity between the query vector and each key vector is calculated, and then scaled to between 0 and 1 using a soft-maximum normalization layer to obtain attention weights that sum to 1. The calculation formula is as follows: ,in Attention weights and This represents the dot product similarity value. In the example operating condition, if it is determined to be in the feeding disturbance stage, the network will adaptively output a short-term fluctuation feature weight of 0.75 and a drift feature weight of 0.25. By multiplying the dimensionless feature value by its corresponding weight, such as obtaining 0.325, the weighted calculation is completed. The weighted scalar values ​​are concatenated in the dimensional direction through the terminal feature concatenation layer to generate a two-dimensional dimensionless feature vector composed of 0.4035 and 0.325. This vector is defined and output as the process stability feature vector representing the current composite operating condition of the system.

[0022] In some implementations, setting the reference sampling period based on the process stability feature vector includes: Two components of the process stability feature vector are extracted, and the square root of the sum of the squares of the two components is calculated as the modulus of the feature vector. The modulus is used to characterize the intensity of process fluctuations. The direction angle of the feature vector is calculated by taking the two components as input and using the two-parameter arctangent function. The dimensionless fluctuation intensity value is obtained by performing a ratio conversion on the preset reference modulus value. Based on the preset reference time constant, the reference sampling period is calculated by combining the dimensionless fluctuation intensity value and the direction angle. The calculation formula is: the reference sampling period is equal to the reference time constant corrected based on the direction angle divided by the largest of the dimensionless fluctuation intensity value and the preset lower limit value, so that the larger the dimensionless fluctuation intensity value, the shorter the reference sampling period.

[0023] Extracting the components of the two-dimensional process stability feature vector and As mentioned above with 0.4035 and 0.325, calculating the square root of the sum of squares yields a modulus M of approximately 0.518. This can be achieved using the arctangent function. The azimuth angle is calculated, yielding an example angle of approximately 38.8°. The preset reference time constant is preferably set to 10 to 60 seconds, for example, 20 seconds. The reference time constant is compensated using the azimuth angle correction function. The azimuth angle correction coefficient f = 1 + 0.5 × sin(2θ) is calculated, and the reference time constant is multiplied by the azimuth angle correction coefficient to obtain the compensated corrected time constant, such as 29.8 seconds. The modulus value is divided by a preset standard stable reference modulus value, for example, set to 0.1, to achieve dimensionless processing, resulting in a dimensionless value of 5.18. The reference sampling period is 5.75 seconds, rounded to the nearest 6 seconds. When the kiln stabilizes, the modulus value approaches zero. A preset lower limit is used to avoid division by zero, and a preset upper limit is used to limit the sampling period to no more than 30 seconds, establishing a closed-loop balance between communication load control and transient disturbance monitoring.

[0024] The switching module is used to adjust the baseline sampling period and error decision threshold of the downstream parameter group by using the process stability feature vector of the upstream parameter group and processing it with a time delay model; based on the process stability feature vector, a local prediction model for the target parameter is generated using a meta-learning framework; when a single sampling value of the target parameter exceeds the prediction confidence interval of the local prediction model, the sampling period of the target parameter and the strongly correlated parameter is switched to a high-frequency verification period.

[0025] The process stability feature vector of the upstream parameter set is extracted, and the maximum value of the cross-correlation coefficient between the upstream and downstream time series, along with the corresponding lag time, is calculated. A time delay model is constructed using this lag time as the core parameter. The expected response time of the downstream parameters is determined based on the lag time. Before this expected response time, the upstream process stability feature vector is input into a multilayer perceptron network, and a time adjustment factor and a threshold adjustment factor are output in advance. The time adjustment factor is multiplied by the original downstream baseline sampling period, and the threshold adjustment factor is multiplied by a default base threshold coefficient with the same dimensions as the time series residual energy value to obtain the feedforward corrected base threshold coefficient. Subsequent acquisition modules use this feedforward corrected base threshold coefficient as the basis for calculating the error decision threshold scaling, completing the feedforward adjustment and updating it in the controller memory.

[0026] An adaptive model initialization mechanism based on meta-learning initializes the network weights of the Long Short-Term Memory (LSTM) network under the constraint of the current process stability feature vector. Using a large amount of historical normal process parameter data as the meta-training set, inner-loop gradient descent updates and outer-loop parameter optimization are performed, enabling the model to quickly adapt to the specific distribution of the current target parameter and generate a local prediction model for the target parameter. The local prediction model is then used for forward propagation inference for the next time step, outputting the theoretical prediction value and prediction variance. Based on the normal distribution assumption, the theoretical prediction value plus or minus three times the prediction standard deviation is calculated as the prediction confidence interval. The prediction variance is generated from the variance output header of the local prediction model or estimated based on the variance of recent historical prediction residuals. When the actual target parameter value transmitted back by the sensor in a single acquisition exceeds the upper limit of the confidence interval or falls below the lower limit, i.e., exceeds the prediction confidence interval, the directed acyclic graph is queried to obtain strongly correlated parameters with edges connected to the target parameter. The hardware sampling clock period of the target parameter and strongly correlated parameter is switched to a preset high-frequency verification period by using control commands. The preset high-frequency verification period is less than the corresponding reference sampling period, for example, 0.5 seconds, thereby switching to the high-frequency verification period state.

[0027] In some implementations, the generation of a local prediction model for the target parameters based on the process stability feature vector using a meta-learning framework includes: Using the current process stability feature vector as the search keyword, calculate the Euclidean distance between the feature vector and each historical feature vector in the preset historical operating condition model library; The model weight parameters corresponding to the historical feature vectors with the smallest Euclidean distance and less than a preset similarity threshold are selected as the initial weights of the local prediction model. Obtain a preset number of historical data points of the target parameter before the current time, and use the historical data points to perform multiple rounds of gradient descent iterative fine-tuning on the initial weights to obtain the trained local prediction model.

[0028] The system has a pre-set historical operating condition model library, which is trained using historical operating condition samples. It also pre-stores prior weight parameters of a long short-term memory network (LSTM) corresponding to 50 to 100 typical glass melting conditions, as well as the central feature vectors corresponding to each typical condition. The LTM network structure includes interconnected forgetting gates, input gates, cell state update units, and output gates. The network input is time-series data of historical process parameters during kiln operation, and the network output is the theoretical prediction of the future trend of the target parameter. Each gate structure uses an activation function to control the flow of feature information along the time step. For example, the calculation formula for the input gate is... , where I is the input gate vector and X is the input feature at the current time step, while H is the hidden state vector at the previous time step.

[0029] The preset similarity threshold is preferably set to 0.15. By calculating the Euclidean distances between the current two-dimensional feature vector (e.g., 0.4035 and 0.325 as mentioned above) and each vector in the database, the nearest neighbor historical vector is matched, and its distance is 0.08 (less than 0.15). At this point, the neural network weight matrix and bias parameters bound to this matched vector are retrieved as the initial weights for the local prediction model of the Long Short-Term Memory network. 600 time-series data points collected within 5 to 15 minutes (e.g., 600 seconds) prior to the current moment are retrieved to form a supervised learning set and fed into the network input layer. A learning rate such as 0.001 is set, and 3 to 5 rounds of gradient descent fine-tuning iterations are performed. This mechanism can complete the initialization of the local prediction model within a 1 to 2 second online computation cycle, ensuring that its network output fits and converges to the real-time boundary conditions of the current thermodynamic state of the molten glass. The iteration fine-tuning number-prediction error curve is shown below. Figure 1 As shown.

[0030] The acquisition module is used to determine that a bit error has occurred and isolate the control output of the actual acquired value during the high-frequency verification period if the time series residual energy value between the actual acquired value and the theoretical predicted value of the target parameter continuously exceeds the bit error decision threshold, and after the synchronous change verification of the strongly correlated parameters excludes the sudden change of the actual working condition, the module outputs the theoretical predicted value during the isolation period. When the time series residual energy value between the subsequent actual acquired value and the theoretical predicted value of the target parameter is lower than the recovery threshold during the isolation period, the module releases the isolation and restores the system to the reference sampling period.

[0031] After the equipment enters the high-frequency verification cycle, it sequentially reads the actual collected values ​​of the target parameters and simultaneously calls the local prediction model to synchronously generate theoretical prediction values ​​for multiple consecutive time steps. It synchronously reads the actual collected values ​​of strongly correlated parameters within the corresponding verification evaluation window to determine whether the target parameter residual exceeding the limit is accompanied by synchronous changes in operating conditions that conform to the physical causal chain. For each high-frequency time step, it calculates the difference between the actual collected value and the theoretical prediction value, squares the differences for multiple consecutive time steps within the verification evaluation window, and calculates the cumulative sum to obtain the time series residual energy value. Before determining the bit error, it calculates the direction of change, amplitude of change, or residual energy of the strongly correlated parameters within the same delay compensation window and performs consistency verification with the change trend of the target parameters. If the strongly correlated parameters meet the preset operating condition change consistency condition, the current anomaly is identified as a real operating condition change, the actual collected values ​​of the target parameters are not isolated, and high-frequency verification continues until the operating conditions stabilize.

[0032] When the residual energy value of the time series exceeds the aforementioned bit error decision threshold for three consecutive verification evaluation windows, and the strongly correlated parameter does not meet the preset operating condition change consistency condition, a bit error is determined by setting the Boolean fault flag to true. The control output of the actual acquired value channel is immediately blocked, and the control feedback output channel is switched to the theoretical prediction value channel. During this isolation period, the theoretical prediction value calculated by the local prediction model is written into the digital-to-analog converter register via the internal data bus, serving as a replacement output for the control feedback signal to maintain the stability of the glass melting furnace level and temperature system. Simultaneously, the time series residual energy value of the latest actual acquired value and theoretical prediction value of the target parameter during the isolation period is continuously calculated. The recent actual data change trend and theoretical change trend sequence are time-aligned, and the recovery evaluation residual energy is obtained based on the sum of squares of the corresponding point differences after alignment. The recovery threshold is determined based on the mean and standard deviation of the residual energy under historical normal operating conditions, or by multiplying the bit error decision threshold by a recovery coefficient less than 1, ensuring that the recovery threshold and the time series residual energy value have the same sum of squares dimension.

[0033] When the residual energy value of the time series drops below the recovery threshold, the controller clears the Boolean fault flag and pulls the digital output enable pin high to release the isolation. By modifying the division factor in the system timer configuration register, the data sampling period is restored to the aforementioned reference sampling period. For strongly correlated parameters that enter the high-frequency verification period synchronously with the target parameter but are not identified as errors, their sampling period is synchronously restored to the corresponding reference sampling period after the target parameter is released from isolation or after a preset number of verification evaluation windows have not exceeded their limits.

[0034] In some implementations, if the time series residual energy value between the actual collected value and the theoretical predicted value of the target parameter continuously exceeds the error decision threshold, and after verifying the synchronous change of strongly correlated parameters to exclude sudden changes in the actual operating conditions, then a bit error is determined to have occurred, and the control output of the actual collected value is isolated, including: The verification evaluation window of the high-frequency verification period is set to a preset time length, and the actual collected value and theoretical predicted value of the target parameter are obtained according to the high-frequency sampling period within the verification evaluation window. The difference between the actual collected value and the theoretical predicted value is calculated point by point to obtain the residual sequence, and the sum of squares of each residual term in the residual sequence is calculated as the time series residual energy value of the current window. The modulus of the process stability feature vector is extracted, converted into a dimensionless fluctuation intensity value according to a preset reference modulus value, and then squared. Combined with a preset basic threshold coefficient, a stretched error decision threshold is calculated, so that the larger the dimensionless fluctuation intensity value, the larger the error decision threshold. Obtain the synchronous change results of strongly correlated parameters within the verification and evaluation window, and determine whether they meet the preset working condition change consistency condition. If, within multiple consecutive verification and evaluation windows, the residual energy value of the time series is greater than the error decision threshold, and the synchronous change result does not meet the preset working condition change consistency condition, then it is determined that the target parameter has an error and the control output of the actual collected value is isolated.

[0035] The preset time length of the verification and evaluation window is preferably set to 5 to 15 seconds, with an example of 10 seconds. Within this range, dense readings are performed with a high-frequency sampling period of preferably 0.5 seconds, ensuring that at least 20 sampling comparison points are generated within a single window. The 20 sets of actual collected values ​​are subtracted point by point from the theoretical output of the prediction model to obtain the residual sequence, and the sum of the squares of all residual terms is calculated. If the sensor encounters electromagnetic interference and data jumps, the time series residual energy value of the current window will increase to, for example, 150.0. At the same time, the magnitude of the extracted process stability feature vector is 0.0268, which is converted to a dimensionless fluctuation intensity value of 0.268 according to the preset reference magnitude of 0.1. Combined with the feedforward correction base threshold coefficient, which is preset to 100.0 and has the same dimension as the time series residual energy value, the scaling error decision threshold of 107.18 is calculated by multiplying the base threshold coefficient by 1 and the sum of the squares of the dimensionless fluctuation intensity value.

[0036] Within the current verification and evaluation window, the synchronous change results of the strongly correlated parameters are further read. If the direction and magnitude of the change of the strongly correlated parameters meet the preset working condition change consistency condition, it is determined to be a real working condition disturbance and high-frequency tracking continues; if the strongly correlated parameters do not meet the consistency condition, and the residual energy value of the target parameter continues to exceed the limit, the error judgment process is entered. The window is judged to be out of limit after detecting 150.0 greater than 107.18. The continuous judgment condition is preferably set to 3 windows. Once the condition of exceeding the limit 3 times consecutively is met, and the synchronous change results of the strongly correlated parameters do not meet the preset working condition change consistency condition, the control output of the actual acquired value channel is shielded in the control logic of the programmable logic controller, and the control feedback output channel is switched to the theoretical prediction value channel. The smooth theoretical prediction value is used to replace the execution feedback control to prevent oscillation of the physical actuator of the gas regulating valve due to transient errors. The residual energy and error judgment threshold curves are shown below. Figure 2 As shown.

[0037] This application discloses a method for real-time acquisition of glass melting process parameters, the method comprising: Based on the preset physical process causal chain, the process parameters collected by the sensor are divided into parameter groups with upstream and downstream relationships; the short-term fluctuation characteristics of each parameter are calculated, and the drift characteristics between the current collected value and the long-term process baseline are calculated; the short-term fluctuation characteristics and drift characteristics are fused through the attention mechanism network to generate a process stability feature vector, and the benchmark sampling period is set according to the process stability feature vector. Using the process stability feature vector of the upstream parameter group, after processing by the time delay model, the baseline sampling period and error decision threshold of the downstream parameter group are adjusted by feedforward; based on the process stability feature vector, a local prediction model for the target parameter is generated using a meta-learning framework; when the single sampling value of the target parameter exceeds the prediction confidence interval of the local prediction model, the sampling period of the target parameter and the strongly correlated parameter is switched to a high-frequency verification period. During the high-frequency verification period, if the time series residual energy value between the actual collected value and the theoretical predicted value of the target parameter continuously exceeds the bit error decision threshold, and after the synchronous change verification of the strongly correlated parameter eliminates the sudden change in the actual working condition, a bit error is determined to have occurred and the control output of the actual collected value is isolated. During the isolation period, the theoretical predicted value is output. When the time series residual energy value between the subsequent actual collected value and the theoretical predicted value of the target parameter is lower than the recovery threshold during the isolation period, the isolation is lifted and the system is restored to the reference sampling period.

[0038] In some implementations, the process parameters collected by the sensors are divided into parameter groups with upstream and downstream relationships according to a preset physical process causal chain, including: Extract the spatial coordinates of each sensor in the three-dimensional solid model of the glass furnace, as well as the material flow vector indicating the glass melt transport route; Calculate the spatial straight-line distance between each sensor and identify the physical medium phase of each sensor. Sensors in the same physical medium phase and whose spatial straight-line distance is less than a preset distance threshold are grouped into the same control unit. The dataset collected by the control unit constitutes a parameter group. Along the material flow direction vector, calculate the projected position of each control unit on the transport route. Based on the order of the projected positions, the parameter group corresponding to the control unit that is relatively close to the starting source is divided into the upstream parameter group, and the parameter group corresponding to the control unit that is relatively close to the ending position is divided into the downstream parameter group.

[0039] In some implementations, the calculation of the short-term fluctuation characteristics of each parameter includes: The length of the first sliding time window used to calculate short-term fluctuation characteristics is set to a preset time span; Within the first sliding time window, time series data of the target parameter are acquired according to a preset sampling frequency; The first-order difference calculation is performed on the collected values ​​at adjacent sampling times in the time series data to obtain a discrete difference sequence; Calculate the sum of squares of each element in the discrete difference sequence, and use the square root of the sum of squares as the short-term fluctuation characteristic of the parameter.

[0040] In some implementations, calculating the drift characteristics of the current acquired value relative to the long-term process baseline includes: The length of the second sliding time window used to calculate the drift characteristics is set to be greater than the preset long span of the first sliding time window; Calculate the arithmetic mean of all historical collected values ​​of the parameter within the second sliding time window, and use the arithmetic mean as the long-term process baseline; Calculate the difference between the current acquired value and the long-term process baseline, and use the absolute value of the difference as the drift characteristic of the parameter.

[0041] In some implementations, the process stability feature vector is generated by fusing short-term fluctuation features and drift features through an attention mechanism network, including: The dimensionless short-term fluctuation features and the drift features are input into the attention mechanism network, and the attention weights of the two are calculated by performing a dot product operation between the query vector in the network and the corresponding key vectors and then normalizing the result. The attention weights are used to calculate the short-term fluctuation features and drift features respectively. The weighted short-term fluctuation features and drift features are concatenated into a two-dimensional feature vector, and the two-dimensional feature vector is output as the process stability feature vector.

[0042] In some implementations, setting the reference sampling period based on the process stability feature vector includes: Two components of the process stability feature vector are extracted, and the square root of the sum of the squares of the two components is calculated as the modulus of the feature vector. The modulus is used to characterize the intensity of process fluctuations. The direction angle of the feature vector is calculated by taking the two components as input and using the two-parameter arctangent function. The dimensionless fluctuation intensity value is obtained by performing a ratio conversion on the preset reference modulus value. Based on the preset reference time constant, the reference sampling period is calculated by combining the dimensionless fluctuation intensity value and the direction angle. The calculation formula is: the reference sampling period is equal to the reference time constant corrected based on the direction angle divided by the largest of the dimensionless fluctuation intensity value and the preset lower limit value, so that the larger the dimensionless fluctuation intensity value, the shorter the reference sampling period.

[0043] In some implementations, if the time series residual energy value between the actual collected value and the theoretical predicted value of the target parameter continuously exceeds the error decision threshold, and after verifying the synchronous change of strongly correlated parameters to exclude sudden changes in the actual operating conditions, then a bit error is determined to have occurred, and the control output of the actual collected value is isolated, including: The verification evaluation window of the high-frequency verification period is set to a preset time length, and the actual collected value and theoretical predicted value of the target parameter are obtained according to the high-frequency sampling period within the verification evaluation window. The difference between the actual collected value and the theoretical predicted value is calculated point by point to obtain the residual sequence, and the sum of squares of each residual term in the residual sequence is calculated as the time series residual energy value of the current window. The modulus of the process stability feature vector is extracted, converted into a dimensionless fluctuation intensity value according to a preset reference modulus value, and then squared. Combined with a preset basic threshold coefficient, a stretched error decision threshold is calculated, so that the larger the dimensionless fluctuation intensity value, the larger the error decision threshold. Obtain the synchronous change results of strongly correlated parameters within the verification and evaluation window, and determine whether they meet the preset working condition change consistency condition. If, within multiple consecutive verification and evaluation windows, the residual energy value of the time series is greater than the error decision threshold, and the synchronous change result does not meet the preset working condition change consistency condition, then it is determined that the target parameter has an error and the control output of the actual collected value is isolated.

[0044] In some implementations, the generation of a local prediction model for the target parameters based on the process stability feature vector using a meta-learning framework includes: Using the current process stability feature vector as the search keyword, calculate the Euclidean distance between the feature vector and each historical feature vector in the preset historical operating condition model library; The model weight parameters corresponding to the historical feature vectors with the smallest Euclidean distance and less than a preset similarity threshold are selected as the initial weights of the local prediction model. Obtain a preset number of historical data points of the target parameter before the current time, and use the historical data points to perform multiple rounds of gradient descent iterative fine-tuning on the initial weights to obtain the trained local prediction model.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A real-time acquisition system for glass melting process parameters, characterized in that, include: The fusion module is used to divide the process parameters collected by the sensors into parameter groups with upstream and downstream relationships according to the preset physical process causal chain. Calculate the short-term fluctuation characteristics of each parameter, and calculate the drift characteristics between the current collected value and the long-term process baseline; The process stability feature vector is generated by fusing short-term fluctuation features and drift features through an attention mechanism network, and a benchmark sampling period is set according to the process stability feature vector. The switching module is used to adjust the baseline sampling period and error decision threshold of the downstream parameter group by using the process stability feature vector of the upstream parameter group and processing it with a time delay model; based on the process stability feature vector, a meta-learning framework is used to generate a local prediction model for the target parameter; when the single sampling value of the target parameter exceeds the prediction confidence interval of the local prediction model, the sampling period of the target parameter and the strongly correlated parameter is switched to a high-frequency verification period. The acquisition module is used to determine that a bit error has occurred and isolate the control output of the actual acquired value within the high-frequency verification period if the time series residual energy value between the actual acquired value and the theoretical predicted value of the target parameter continuously exceeds the bit error decision threshold, and after the synchronous change verification of the strongly correlated parameter excludes the sudden change of the actual working condition, the module outputs the theoretical predicted value during the isolation period. When the time series residual energy value between the subsequent actual collected value and the theoretical predicted value of the target parameter during the isolation period is lower than the recovery threshold, the isolation is lifted and the system is restored to the baseline sampling period.

2. The system according to claim 1, characterized in that, The process parameters collected by the sensors are divided into parameter groups with upstream and downstream relationships according to a preset physical process causal chain, including: Extract the spatial coordinates of each sensor in the three-dimensional solid model of the glass furnace, as well as the material flow vector indicating the glass melt transport route; Calculate the spatial straight-line distance between each sensor and identify the physical medium phase of each sensor. Sensors in the same physical medium phase and whose spatial straight-line distance is less than a preset distance threshold are grouped into the same control unit. The dataset collected by the control unit constitutes a parameter group. Along the material flow direction vector, calculate the projected position of each control unit on the transport route. Based on the order of the projected positions, the parameter group corresponding to the control unit that is relatively close to the starting source is divided into the upstream parameter group, and the parameter group corresponding to the control unit that is relatively close to the ending position is divided into the downstream parameter group.

3. The system according to claim 1, characterized in that, The calculation of the short-term fluctuation characteristics of each parameter includes: The length of the first sliding time window used to calculate short-term fluctuation characteristics is set to a preset time span; Within the first sliding time window, time series data of the target parameter are acquired according to a preset sampling frequency; The first-order difference calculation is performed on the collected values ​​at adjacent sampling times in the time series data to obtain a discrete difference sequence; Calculate the sum of squares of each element in the discrete difference sequence, and use the square root of the sum of squares as the short-term fluctuation characteristic of the parameter.

4. The system according to claim 3, characterized in that, The calculation of the drift characteristics between the current acquired value and the long-term process baseline includes: The length of the second sliding time window used to calculate the drift characteristics is set to be greater than the preset long span of the first sliding time window; Calculate the arithmetic mean of all historical collected values ​​of the parameter within the second sliding time window, and use the arithmetic mean as the long-term process baseline; Calculate the difference between the current acquired value and the long-term process baseline, and use the absolute value of the difference as the drift characteristic of the parameter.

5. The system according to claim 1, characterized in that, The process stability feature vector is generated by fusing short-term fluctuation features and drift features through an attention mechanism network, including: The dimensionless short-term fluctuation features and the drift features are input into the attention mechanism network, and the attention weights of the two are calculated by performing a dot product operation between the query vector in the network and the corresponding key vectors and then normalizing the result. The attention weights are used to calculate the short-term fluctuation features and drift features respectively. The weighted short-term fluctuation features and drift features are concatenated into a two-dimensional feature vector, and the two-dimensional feature vector is output as the process stability feature vector.

6. The system according to claim 5, characterized in that, The step of setting the benchmark sampling period based on the process stability feature vector includes: Two components of the process stability feature vector are extracted, and the square root of the sum of the squares of the two components is calculated as the modulus of the feature vector. The modulus is used to characterize the intensity of process fluctuations. The two components are used as inputs, and the direction angle of the feature vector is calculated using a two-parameter arctangent function. The dimensionless fluctuation intensity value is obtained by performing a ratio conversion on the preset reference modulus value. Based on the preset reference time constant, the reference sampling period is calculated by combining the dimensionless fluctuation intensity value and the direction angle. The calculation formula is: the reference sampling period is equal to the reference time constant corrected based on the direction angle divided by the largest of the dimensionless fluctuation intensity value and the preset lower limit value, so that the larger the dimensionless fluctuation intensity value, the shorter the reference sampling period.

7. The system according to claim 6, characterized in that, If the time series residual energy value between the actual collected value and the theoretical predicted value of the target parameter continuously exceeds the error judgment threshold, and after verifying the synchronous change of strongly correlated parameters to exclude sudden changes in the actual operating conditions, an error is determined to have occurred, and the control output of the actual collected value is isolated, including: The verification evaluation window of the high-frequency verification period is set to a preset time length, and the actual collected value and theoretical predicted value of the target parameter are obtained according to the high-frequency sampling period within the verification evaluation window. The difference between the actual collected value and the theoretical predicted value is calculated point by point to obtain the residual sequence, and the sum of squares of each residual term in the residual sequence is calculated as the time series residual energy value of the current window. The modulus of the process stability feature vector is extracted, converted into a dimensionless fluctuation intensity value according to a preset reference modulus value, and then squared. Combined with a preset basic threshold coefficient, a stretched error decision threshold is calculated, so that the larger the dimensionless fluctuation intensity value, the larger the error decision threshold. Obtain the synchronous change results of strongly correlated parameters within the verification and evaluation window, and determine whether they meet the preset working condition change consistency condition. If, within multiple consecutive verification and evaluation windows, the residual energy value of the time series is greater than the error decision threshold, and the synchronous change result does not meet the preset working condition change consistency condition, then it is determined that the target parameter has an error and the control output of the actual collected value is isolated.

8. The system according to claim 1, characterized in that, The method of generating a local prediction model for the target parameters based on the process stability feature vector using a meta-learning framework includes: Using the current process stability feature vector as the search keyword, calculate the Euclidean distance between the feature vector and each historical feature vector in the preset historical operating condition model library; The model weight parameters corresponding to the historical feature vectors with the smallest Euclidean distance and less than a preset similarity threshold are selected as the initial weights of the local prediction model. Obtain a preset number of historical data points of the target parameter before the current time, and use the historical data points to perform multiple rounds of gradient descent iterative fine-tuning on the initial weights to obtain the trained local prediction model.

9. A method for real-time acquisition of glass melting process parameters, characterized in that, include: Based on the preset physical process causal chain, the process parameters collected by the sensor are divided into parameter groups with upstream and downstream relationships; Calculate the short-term fluctuation characteristics of each parameter, and calculate the drift characteristics between the current collected value and the long-term process baseline; The process stability feature vector is generated by fusing short-term fluctuation features and drift features through an attention mechanism network, and a benchmark sampling period is set according to the process stability feature vector. Using the process stability feature vector of the upstream parameter group, after processing by the time delay model, the baseline sampling period and error decision threshold of the downstream parameter group are adjusted by feedforward; based on the process stability feature vector, a local prediction model for the target parameter is generated using a meta-learning framework; when the single sampling value of the target parameter exceeds the prediction confidence interval of the local prediction model, the sampling period of the target parameter and the strongly correlated parameter is switched to a high-frequency verification period. During the high-frequency verification period, if the time series residual energy value between the actual collected value and the theoretical predicted value of the target parameter continuously exceeds the bit error decision threshold, and after the synchronous change verification of the strongly correlated parameter eliminates the sudden change of the actual working condition, a bit error is determined to have occurred and the control output of the actual collected value is isolated, while the theoretical predicted value is output during the isolation period. When the time series residual energy value between the subsequent actual collected value and the theoretical predicted value of the target parameter during the isolation period is lower than the recovery threshold, the isolation is lifted and the system is restored to the baseline sampling period.

10. The method according to claim 9, characterized in that, The process parameters collected by the sensors are divided into parameter groups with upstream and downstream relationships according to a preset physical process causal chain, including: Extract the spatial coordinates of each sensor in the three-dimensional solid model of the glass furnace, as well as the material flow vector indicating the glass melt transport route; Calculate the spatial straight-line distance between each sensor and identify the physical medium phase of each sensor. Sensors in the same physical medium phase and whose spatial straight-line distance is less than a preset distance threshold are grouped into the same control unit. The dataset collected by the control unit constitutes a parameter group. Along the material flow direction vector, calculate the projected position of each control unit on the transport route. Based on the order of the projected positions, the parameter group corresponding to the control unit that is relatively close to the starting source is divided into the upstream parameter group, and the parameter group corresponding to the control unit that is relatively close to the ending position is divided into the downstream parameter group.

Images