Cupman network-based hybrid hopper level control system and method

Abstract

The application discloses a mixed tank level control system and method based on a Kupman network, and belongs to the technical field of tank level control. The method specifically comprises the following steps: collecting key process data of a mixed tank, and constructing an original state vector of tank level control; inputting the original state vector into a pre-trained Kupman network; mapping a nonlinear dynamic process of the mixed tank level to a high-dimensional observable space; establishing a linear prediction model; combining a tank level balance to generate a tank level prediction trajectory; based on the tank level prediction trajectory, using a predictive compound PID strategy to solve a comprehensive feeding amount adjustment amount, and issuing the comprehensive feeding amount adjustment amount to a feeding execution unit to maintain the tank level within a target range; after executing the comprehensive feeding amount adjustment amount, calculating a prediction deviation, and iteratively correcting the Kupman network and the tank level balance. The application improves the tank level control accuracy and continuity, reduces the risk of full tank or lack of material, and enhances the self-adaptive ability to working condition switching and sudden disturbance.

Description

Technical Field

[0001] This invention belongs to the field of material level control technology for troughs, specifically a mixing trough level control system and method based on Koopman networks. Background Technology

[0002] In industries such as metallurgy, mineral processing, and building materials, mixing tanks are crucial for continuous material supply and homogenization, and their material level stability directly impacts the continuity of downstream sintering or conveying processes and product quality. However, the dynamic process of material level in mixing tanks typically exhibits nonlinear, multi-disturbance, and strongly coupled characteristics: upstream belt conveying and feeding at the collection point are intermittent and fluctuating; in-tank stirring and auxiliary gate operation introduce non-uniform flow; and the discharge process suffers from lag effects due to the sintering machine speed and roller rotation speed.

[0003] Existing technologies typically rely on manual adjustment of the total feed volume or closed-loop control using empirical rules, with some employing neural networks or model predictive control for level prediction and regulation. While these methods can improve control accuracy to some extent, they have certain limitations: first, their ability to model complex nonlinear coupling relationships is limited; second, the accuracy of multi-step predictions is easily affected by error accumulation; and third, they lack adaptive adjustment capabilities to changes in operating conditions and sudden disturbances. Therefore, in high-disturbance, multi-condition scenarios, existing control methods still struggle to achieve continuous and stable level control. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a mixing tank level control system and method based on Koopman networks. The Koopman network is effectively applied to the mixing tank level control. At the same time, by combining multi-source data such as feed rate, equipment parameters, and auxiliary gate actions, an adaptive and iteratively updated prediction and control closed loop is established to achieve stable level control under multiple disturbances, nonlinearity, and operating condition switching conditions.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for controlling the level of a mixing tank based on a Koopman network includes:

[0007] Collect key process data of the mixing tank and construct the original state vector for material level control. The key process data includes feeding amount, material amount at each mixing point, roller speed, auxiliary door opening, sintering machine speed and current material level in the tank.

[0008] The original state vector is input into a pre-trained Koopman network to map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a material level prediction trajectory by combining the material level balance. The material level balance represents the coupled influence of the total amount of material fed, equipment parameters and auxiliary gate opening on the material level increase and decrease trend. The linear prediction model satisfies the one-step prediction error, multi-step rolling prediction error and high-dimensional linear evolution consistency constraints.

[0009] Based on the predicted material level trajectory, a predictive composite PID strategy is used to solve for the comprehensive material feeding adjustment amount, and then the adjustment amount is sent to the feeding execution unit to maintain the material level within the target range. The predictive composite PID strategy introduces the predicted material level change trend as a feedforward correction into the PID control process.

[0010] After adjusting the overall feed rate, the actual material level and key process data are collected in real time, the prediction deviation is calculated, and the Koopman network and material level balance are iteratively corrected online based on the prediction deviation.

[0011] Specifically, the process of collecting key process data from the mixing tank and constructing the original state vector for material level control includes:

[0012] The feeding amount of each feeding source in the mixing tank is collected in real time, and the data of each feeding source is synchronously arranged according to the timestamp to form a feeding status sub-vector;

[0013] Collect the material level of each mixing point in the mixing tank, and associate and mark it according to the material flow direction and mixing sequence to construct a sub-vector of material state at the mixing point;

[0014] Obtain the motion parameters of the trough and auxiliary doors, including the rotational speed of the rollers and the opening degree of each auxiliary door, and map them into equipment action state sub-vectors according to their spatial position and motion sequence;

[0015] Get the current machine speed of the sintering machine and the current material level in the trough, and align them with the collected material status at the feeding and mixing points in time to generate a sub-vector that associates the material level history with the instantaneous status.

[0016] The generated sub-vectors are combined and normalized according to the working condition code and time series to construct the original state vector representing the current comprehensive operating state of the mixing tank.

[0017] Specifically, the original state vector is input into a pre-trained Koopman network to map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a predicted material level trajectory by combining the material level balance, including:

[0018] The constructed original state vector is input into a pre-trained Koopman network to generate an upgraded state representation in a high-dimensional observable space, and to distinguish and label the nonlinear features of different working conditions.

[0019] A linear prediction model is established in the up-dimensional state representation. The future material level evolution sequence is generated by linear combination and time recursion of each up-dimensional state representation, and the prediction confidence interval information of each time step is recorded.

[0020] By combining the feeding amount, roller speed, auxiliary door opening and sintering machine speed with the dimension-upgrading status, and using the material level balance to perform coupled analysis on the material level increase and decrease trend, a material level balance adjustment item is generated.

[0021] The future material level evolution sequence is fused with the material level balance adjustment term to form a material level prediction trajectory, wherein the material level prediction value at each time step simultaneously includes nonlinear dynamic response and coupled balance constraint information.

[0022] The predicted material level trajectory is subjected to rolling verification to ensure that it meets the one-step prediction error constraint, the multi-step rolling prediction error constraint, and the linear evolution consistency constraint.

[0023] Specifically, the process combines the feeding rate, roller speed, auxiliary door opening, and sintering machine speed with the dimensional state representation, and uses material level balance to perform coupled analysis on the material level increase / decrease trend to generate material level balance adjustment items, including:

[0024] The upgraded state representation is time-aligned and operating condition-matched with the current data collected, such as the feeding amount, roller speed, auxiliary door opening, and sintering machine speed, to generate a combined state vector. The combined state vector is used to represent the current comprehensive operating conditions of the trough.

[0025] Material level balance analysis is performed on the combined state vector. By mapping the contribution of each feeding point, equipment operation and auxiliary door opening change to the material level increase / decrease trend space, a material level coupling relationship matrix is ​​formed, and the weight distribution of different influencing factors is marked.

[0026] Based on the material level coupling relationship matrix, the contributions of each influencing factor are aggregated and calculated to generate material level balance adjustment terms, where each adjustment term corresponds to a prediction correction vector of the material level change trend in future time steps.

[0027] Specifically, the future material level evolution sequence is integrated with the material level balance adjustment term to form a material level prediction trajectory, including:

[0028] The future material level evolution sequence is unfolded according to time steps and matched with the material level balance adjustment item in time to form a time-aligned prediction combination unit;

[0029] For each time step within the prediction combination unit, a weighted fusion is performed according to the coupling relationship between the dimensionality-upgraded state weights and the material level balance constraints to generate a fused single-step prediction value vector.

[0030] The single-step prediction vectors fused from each time step are connected sequentially to form a continuous material level prediction trajectory. The material level prediction value at each time step simultaneously reflects the nonlinear dynamic evolution and the coupling balance constraint information of each control parameter on the material level.

[0031] The generated material level prediction trajectory is rolled over and verified. By checking the evolution consistency and coupling constraint compliance between consecutive time steps, the predicted material level values ​​in the material level prediction trajectory are adjusted to generate the final material level prediction trajectory.

[0032] Specifically, the step of using a predictive composite PID strategy based on the predicted material level trajectory to obtain the comprehensive feeding rate adjustment, and then sending it to the feeding execution unit to maintain the material level within the target range, includes:

[0033] Based on the predicted material level trajectory, the predicted material level value, the direction of material level change, and the trend information of the material level and the boundary of the target range are extracted in the order of prediction time to generate the predicted trend input.

[0034] The predicted trend input is compared with the preset target range to determine the material level deviation type in the current control cycle. The material level deviation type includes material level rise deviation, material level fall deviation, full material approach deviation, and material shortage approach deviation.

[0035] A feedforward correction amount is generated based on the material level deviation type, and the feedforward correction amount is introduced into the PID control process. It is then fused with the feedback control amount formed based on the current actual material level deviation to obtain the predicted composite PID control amount.

[0036] The predicted composite PID adjustment is subjected to process boundary processing to generate a comprehensive feed rate adjustment. The process boundary processing includes upper and lower limits of comprehensive feed rate, feed rate change range constraint, adjustment direction consistency verification, and adjustment cycle limit matching the sintering machine speed change state.

[0037] The comprehensive feeding quantity adjustment is sent to the feeding execution unit, and after the issuance, the predicted trend input, feedforward correction, feedback adjustment and material level feedback data within the corresponding control cycle are recorded.

[0038] Specifically, a feedforward correction amount is generated based on the material level deviation type, and this feedforward correction amount is introduced into the PID control process. It is then fused with the feedback control amount based on the current actual material level deviation to obtain a predicted composite PID control amount, including:

[0039] Based on the material level deviation type, the predicted material level change trend is decomposed by factor, and the trend components corresponding to different material level deviation types are labeled as upward trend component, downward trend component, full material approach component and short material approach component, respectively.

[0040] For each trend component, a feedforward correction is generated. The predictive effect of each trend component is time-matched with the corresponding control period to form a comprehensive feedforward correction for this control period.

[0041] The comprehensive feedforward correction amount and the feedback adjustment amount calculated based on the current actual material level deviation are linearly or weightedly fused to form a fused adjustment vector. The fusion process takes into account the predicted trend weight and deviation priority, while taking into account both nonlinear dynamic response and instantaneous deviation control.

[0042] The fusion adjustment vector is defined as the predicted composite PID adjustment amount, used as the comprehensive feed rate adjustment input for the next control cycle, and as reference data for subsequent online correction and control parameter updates.

[0043] Specifically, the predicted composite PID adjustment is processed by process boundary conditions to generate a comprehensive feed rate adjustment, including:

[0044] The predicted composite PID adjustment is compared with the preset upper and lower limits of the feed amount. Boundary constraints are applied to the adjustment amount that exceeds the upper and lower limits to form the initial constraint adjustment value.

[0045] The change range of the initial constraint adjustment value is compared with that of the feeding amount in the previous control cycle, and the adjustment amount that exceeds the allowable change range is compressed.

[0046] Check the consistency between the adjustment amount after amplitude compression and the feeding direction, correct any possible conflict in the adjustment direction, and match the adjustment cycle according to the changes in the sintering machine speed to generate an adjustment plan that conforms to the process rhythm.

[0047] The adjustment plan is defined as the comprehensive feeding quantity adjustment amount and is issued to the feeding execution unit.

[0048] Specifically, after adjusting the overall material feeding rate, the actual material level and key process data are collected in real time, the prediction deviation is calculated, and the Koopman network and material level balance are iteratively corrected online based on the prediction deviation, including:

[0049] After adjusting the overall feeding amount, the actual material level and feeding amount in the mixing tank, the rotation speed of the roller, the opening degree of the auxiliary door, and the speed of the sintering machine are collected in real time, and time stamps and working condition codes are performed in the order of the control cycle.

[0050] The real-time collected data is compared with the material level prediction trajectory to calculate the prediction deviation at each time step, and a prediction deviation vector is generated based on the deviation magnitude and trend.

[0051] Based on the predicted deviation vector, the upgraded state mapping parameters in the Koopman network and the coupling weights in the level balance are iteratively corrected online.

[0052] The modified Koopman network is used as the prediction input for the next control cycle, while the deviation data, adjustment parameters and corresponding process data of the current control cycle are recorded for rolling optimization and adaptive reconfiguration of operating conditions.

[0053] A mixing tank level control system based on Koopman network is used to implement the mixing tank level control method based on Koopman network, including: a vector construction module, a level prediction module, a level control module, and an iterative correction module.

[0054] The vector construction module is used to collect key process data of the mixing tank and construct the original state vector of the material level control. The key process data includes the feeding amount, the amount of material at each mixing point, the rotation speed of the roller, the opening degree of the auxiliary door, the speed of the sintering machine, and the current material level of the tank.

[0055] The material level prediction module is used to input the original state vector into a pre-trained Koopman network, map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a material level prediction trajectory in combination with the material level balance. The material level balance represents the coupled influence of the total amount of material fed, equipment parameters and auxiliary door opening on the material level increase or decrease trend. The linear prediction model satisfies one-step prediction error, multi-step rolling prediction error and linear evolution consistency constraints.

[0056] The material level control module is used to solve the comprehensive feeding quantity adjustment based on the predicted material level trajectory using a predictive composite PID strategy, and then send it to the feeding execution unit to maintain the material level within the target range. The predictive composite PID strategy introduces the predicted material level change trend as a feedforward correction into the PID adjustment process.

[0057] The iterative correction module is used to collect actual material level and key process data in real time after the comprehensive feeding amount adjustment is executed, calculate the prediction deviation, and perform online iterative correction of the Koopman network and material level balance based on the prediction deviation.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] This invention proposes a mixing tank level control system and method based on a Koopman network. By constructing a state vector and inputting it into a pre-trained Koopman network for high-dimensional mapping, a predicted level trajectory is generated by combining level balance. The comprehensive feed rate adjustment is calculated using a predictive composite PID strategy for regulation and control. This method can achieve stable maintenance of the mixing tank level under multiple operating conditions, multiple disturbances, and nonlinear coupling conditions, improving the accuracy and continuity of level control, reducing the risk of full or insufficient material, and enhancing the adaptive capability to operating condition switching and sudden disturbances, ensuring the smooth operation of downstream sintering or conveying processes and overall production efficiency. Attached Figure Description

[0060] Figure 1 Flowchart of the mixing tank level control method based on Koopman network provided by the present invention;

[0061] Figure 2 A schematic diagram of the Koopman network provided by this invention;

[0062] Figure 3 This is a schematic diagram of the predictive composite PID control process provided by the present invention;

[0063] Figure 4 The diagram shows the architecture of the mixing tank level control system based on the Koopman network provided by this invention. Detailed Implementation

[0064] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0066] It should be noted that, unless there is a conflict, the various features in the embodiments of this application can be combined with each other, all of which are within the protection scope of this application. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. In addition, the terms "first," "second," and "third" used in this application do not limit the data or execution order, but only distinguish identical or similar items with essentially the same function and effect.

[0067] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0068] Example 1:

[0069] Please see Figure 1The present invention provides an embodiment of a method comprising the following specific steps:

[0070] Step S1: Collect key process data of the mixing tank and construct the original state vector of material level control. The key process data includes feeding amount, material amount at each mixing point, roller speed, auxiliary door opening, sintering machine speed and current material level of the tank.

[0071] The specific steps of step S1 are as follows:

[0072] Step S101: Collect the feeding amount of each feeding source in the mixing tank in real time, and arrange the data of each feeding source synchronously according to the timestamp to form a feeding status sub-vector.

[0073] In this embodiment, to meet the feeding control requirements of the mixing tank, the feeding amount of each feeding source is collected in real time by a data acquisition device. The collected data from each feeding source is aligned and synchronized according to timestamps to ensure the temporal consistency of the data from each feeding source within the same control cycle. The specific operation includes sampling, filtering, and data verification of the sensor output signals of each feeding source. The valid data is stored in the control recording unit according to the sampling time order. The data is identified according to the location and conveying order of the feeding source. The instantaneous feeding flow rate, cumulative feeding weight, and corresponding equipment operation status identifiers of each feeding source within the same time window are arranged in order to form a one-dimensional feeding status sub-vector with the same dimension as the number of feeding sources.

[0074] Step S102: Collect the material quantity of each mixing point in the mixing tank, and associate and mark it according to the material flow direction and mixing sequence to construct a sub-vector of the material state of the mixing point.

[0075] In this embodiment, material level detection devices installed at key mixing points in the mixing tank are used to collect the material level at each mixing point in real time. The collected data is then time- and space-associated and marked according to the material flow direction and mixing sequence to ensure the continuity and logical order of the material state at each mixing point. The specific operation includes sampling the weight or height sensor signal at each mixing point, cleaning the data and removing anomalies, and identifying and encoding the collected values ​​with the corresponding mixing sequence to construct a one-dimensional mixing point material state sub-vector with the same dimension as the total number of mixing points.

[0076] Step S103: Obtain the motion parameters of the trough and auxiliary doors, including the rotation speed of the rollers and the opening degree of each auxiliary door, and map them into equipment action state sub-vectors according to their spatial position and motion sequence.

[0077] In this embodiment, motion parameters of the mixing tank and auxiliary doors, including the rotational speed of the rollers and the opening degree of each auxiliary door, are acquired through sensors or existing control interfaces. The acquired data is then processed synchronously and time-stamped. Specifically, the roller speed sensor signal is sampled and filtered to ensure the real-time performance and accuracy of the data. Simultaneously, the potentiometer or encoder output of the auxiliary door opening is read and anomalies are eliminated. Subsequently, each motion parameter is mapped and encoded according to its spatial position and operational sequence within the tank, forming a sub-vector of the equipment's action state.

[0078] Step S104: Obtain the current machine speed of the sintering machine and the current material level in the trough, and align them with the previously collected material status at the feeding and mixing points to generate a sub-vector that associates the material level history with the instantaneous status.

[0079] In this embodiment, the current machine speed of the sintering machine and the real-time material level of the mixing trough are acquired through a data acquisition device, and the acquired data is time-marked according to the control cycle. Subsequently, the real-time data is time-aligned with the previously acquired feed status sub-vectors of each feeding source and the material status sub-vector of the mixing point to ensure that data from different sources have a consistent operating condition reference at the same point in time. The specific operation includes sampling, filtering, and anomaly removal of the signals from the sintering machine speed sensor and the material level detector, and using the control cycle index to associate them with the previously acquired feed and mixing point data. Through this processing, a material level history and instantaneous status correlation sub-vector reflecting the current instantaneous state of the trough and the historical material level change trend is generated.

[0080] Step S105: Combine and normalize the generated sub-vectors according to the working condition code and time series to construct the original state vector representing the current comprehensive operating state of the mixing tank.

[0081] In this embodiment, the feed state sub-vector, mixing point material state sub-vector, equipment action state sub-vector, and material level history and instantaneous state association sub-vector generated in steps S101 to S104 are combined according to the control cycle order and working condition code to form a dataset with time series logic and working condition differentiation features. The specific operation includes normalizing each sub-vector so that data of different dimensions can be represented at the same scale, and identifying and encoding the data according to the working condition category to retain the impact information of different operating states on material level dynamics. Subsequently, the normalized sub-vectors are concatenated in time series order to construct the original state vector representing the current comprehensive operating state of the mixing tank. This vector reflects the comprehensive characteristics of feeding, mixing, equipment action, and historical material level evolution.

[0082] Step S2: Input the original state vector into the pre-trained Koopman network to map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a material level prediction trajectory by combining the material level balance. The material level balance represents the coupled influence of the total amount of material fed, equipment parameters and auxiliary gate opening on the material level increase or decrease trend. The linear prediction model satisfies the one-step prediction error, multi-step rolling prediction error and linear evolution consistency constraints.

[0083] The specific steps of step S2 are as follows:

[0084] like Figure 2 As shown, step S201: The constructed original state vector is input into the pre-trained Koopman network to generate an upgraded state representation in the high-dimensional observable space, and the nonlinear features of different working conditions are distinguished and labeled.

[0085] In this embodiment, the pre-trained Koopman network adopts a symmetrical architecture of encoder-Koopman operator-decoder. The encoder consists of three fully connected layers with 512, 1024, and 2048 neurons respectively. Each layer is followed by a batch normalization layer and a ReLU activation function. A residual connection is added between the second and third layers to alleviate the gradient vanishing problem. The Koopman operator is a 2048×2048 learnable real matrix. The decoder is symmetrical to the encoder, with 1024 and 512 neurons respectively. The output layer dimension is exactly the same as the original state vector dimension. The activation function is Sigmoid to match the normalized value range [0,1] of the original state vector. This network is trained offline using historical operating data of the mixing tank over the past 12 months. The training dataset includes... The dataset contains approximately 120 million original state vector samples across 256 different operating conditions. The training objective is to minimize the weighted total loss, which consists of reconstruction loss, one-step prediction loss, and operating condition classification loss. The reconstruction loss is the mean squared error between the decoder output and the input original state vector, with a weighting coefficient of 0.7. The one-step prediction loss is the mean squared error between the next-time state reconstructed by the decoder after the Koopman operator is applied to the upgraded state and the actual next-time state, with a weighting coefficient of 0.2. The operating condition classification loss is the cross-entropy loss between the encoder output and the actual operating condition encoding, with a weighting coefficient of 0.1. The AdamW optimizer is used for training, with an initial learning rate of 0.2, a batch size of 256, and 100 training epochs. Training is terminated early when the validation set loss stops decreasing after 10 consecutive epochs.

[0086] When the original state vector constructed in step S105 is input into the Koopman network, the dimension of the input vector is first verified to ensure that it matches the dimension of the network input layer. Vectors with mismatched dimensions are discarded and marked as input anomalies. After successful verification, the vector is input into the encoder. Through a nonlinear transformation of three fully connected layers, the low-dimensional original state vector is mapped to a 2048-dimensional high-dimensional observable space, generating an upgraded state representation. This upgrade process decouples the nonlinear features coupled in the original state vector, such as feeding, mixing, equipment, and material level, into a high-dimensional representation using a nonlinear activation function. The system consists of mutually orthogonal linear features in a 2048-dimensional space. The first 512 dimensions correspond to the feeding state features, the middle 512 dimensions correspond to the material state features at the mixing point, the next 512 dimensions correspond to the equipment operation state features, and the last 512 dimensions correspond to the correlation features between material level history and instantaneous state. When distinguishing and labeling the nonlinear features of different operating conditions, the 2048-dimensional up-dimensional state representation output by the encoder is first extracted and input into a condition classification head trained synchronously with the encoder. This classification head is a single fully connected layer with an input dimension of 2048 and an output dimension of 256. The activation function used is S0. `oftmax` outputs the probability distribution vectors of 256 working conditions corresponding to the current upgraded state representation. The working condition code with the highest probability value is taken as the main working condition label. At the same time, the cosine similarity between the current upgraded state representation and the feature centers of the 256 pre-stored standard working conditions is calculated. The feature centers of the standard working conditions are calculated by the arithmetic mean of all upgraded state representations under each working condition in the training set. When the similarity is greater than 0.8, it is marked as a stable state of the working condition. When the similarity is between 0.5 and 0.8, it is marked as a transitional state of the working condition. When the similarity is less than 0... At time 5, the condition is marked as an unknown working condition and the upgraded state is stored in the abnormal feature library. The final upgraded state is represented as a 2048-dimensional real number vector, with each element taking values ​​between [-1, 1]. The corresponding working condition differentiation label includes an 8-bit binary working condition code, a 256-dimensional working condition probability distribution vector, a cosine similarity value, and a state stability identifier. The state stability identifier uses 0 to represent a stable state, 1 to represent a transitional state, and 2 to represent an unknown working condition. All the labeling information is appended to the header of the upgraded state representation to form a complete high-dimensional state feature vector.

[0087] exist Figure 2 In the process, the initial state vector of the mixing tank is first... As input, the vector includes low-dimensional key process parameters such as material level, feed rate, and sintering machine speed, arranged in a time series. Then, an encoder is used to perform a nonlinear up-dimensional mapping of the original low-dimensional state. Convert to high-dimensional observation vector This high-dimensional observation vector contains complex nonlinear dynamic characteristics and potential interaction information. Then, the up-dimensional state representation is linearly evolved in the high-dimensional space using the Koopman operator K to generate the future state. This maps the nonlinear dynamic process into a linearly tractable form. The decoder then uses dimensionality reduction to reconstruct the mapping. Future state Reconstructing back to the original low-dimensional space yields the predicted material level state. This is used as the input for the next control cycle to achieve continuous prediction and dynamic adjustment of the material level in the mixing tank, where R... n Let R represent an n-dimensional real vector space. m Let R represent an m-dimensional real vector space. m×m Let n represent the space of m rows and m columns of real numbers, where n and m represent the dimensions.

[0088] Step S202: Establish a linear prediction model in the upgraded state representation, generate the future material level evolution sequence by linear combination and time recursion of each upgraded state representation, and record the prediction confidence interval information for each time step.

[0089] In this embodiment, the linear prediction model is constructed based on the linear evolution characteristics of the Koopman operator. Its core is the 2048×2048-dimensional real-valued Koopman operator matrix K, which is converged through offline training in step S201. This matrix learns the linear evolution law of the mixing tank system in the high-dimensional observable space and satisfies the linear recursive relationship through 120 million historical samples. The time step is completely consistent with the sampling time step of the original state vector. When generating the future material level evolution sequence, the 2048-dimensional up-dimensional state representation x(t) of the current time t output in step S201 is first extracted as the initial state. It is then multiplied with the Koopman operator matrix K to obtain the up-dimensional state prediction value x(t+1) at time t+1. The above operation is repeated with x(t+1) as the new initial state, and t+2, t+3 and up to t+12 are generated in sequence. A complete up-dimensional state prediction sequence with a 0-second time limit is generated, with a prediction duration of 120 seconds matching the maximum material residence time in the mixing tank, ensuring coverage of the entire process from material entry into the tank to discharge. To suppress the accumulation of errors caused by long-term recursion, a rolling prediction mechanism is adopted. Every 10 seconds, the predicted state at the corresponding time moment is automatically replaced with the latest acquired measured up-dimensional state representation as a new starting point to restart the recursion. At the same time, when the prediction step length exceeds 60 seconds, a decay coefficient of 0.995 is applied to the Koopman operator to reduce the weight of long-term prediction. The up-dimensional state prediction values ​​of each time step are input into the decoder trained synchronously in step S201, and mapped back to the original state space through the inverse transformation of the symmetric fully connected layer to obtain the original state vector prediction value at the corresponding time moment. Then, the value of the material level parameter dimension is extracted from the original state vector to form the material level evolution sequence for the next 120 seconds.

[0090] When recording the prediction confidence interval information at each time step, a linear propagation method based on the training error covariance is adopted. First, the one-step prediction error covariance matrix of all samples in the training set is calculated offline. This matrix is ​​a 2048×2048 symmetric positive definite matrix, representing the inherent prediction uncertainty of the Koopman operator. During online prediction, the initial state uncertainty is obtained by linearly transforming the measurement error covariance matrix of the original state vector using the encoder. Subsequently, the prediction covariance matrix at the k-th time step is calculated using the covariance propagation formula, which is: , Let i represent the prediction covariance matrix at the k-th time step, and let i represent the index variable. This represents the transpose of the Koopman operator matrix, and adaptive adjustments are made based on the condition differentiation markers output in step S201: the predicted covariance matrix remains unchanged under stable conditions, multiplied by a magnification factor of 1.5 under transitional conditions, and multiplied by a magnification factor of 3.0 under unknown conditions. Finally, the upper and lower confidence limits of the predicted material level at each time step are calculated based on a 95% confidence level. The calculation formula is: Upper confidence limit = predicted material level + 1.96 × ( _k (material level dimension)) 1 / 2 Confidence lower limit = predicted material level - 1.96 × ( _k (material level dimension)) 1 / 2 , _k (material level dimension) represents the prediction covariance matrix of the material level dimension at the k-th time step; finally, a material level evolution sequence is generated, containing complete data for 120 time steps.

[0091] Step S203: Combine the feeding amount, roller speed, auxiliary door opening and sintering machine speed with the dimension-upgrading status representation, and use the material level balance to perform coupled analysis on the material level increase and decrease trend to generate material level balance adjustment items.

[0092] In this embodiment, when aligning the upgraded state representation with the currently collected data on the feeding quantity, roller speed, auxiliary door opening, and sintering machine speed for time and condition matching, the timestamp accuracy is unified to the microsecond level. All parameters are mapped to the same time axis with a fixed time step of 1 second. For parameters with different sampling frequencies, arithmetic mean downsampling or cubic spline interpolation is used to complete the data, ensuring that the time base of all data is aligned. Subsequently, the 8-bit binary condition code output in step S201 is extracted and its cosine similarity is calculated with 128 pre-stored typical material level balance condition templates. The condition templates are composed of stable material level fluctuations of less than ±0.1 m / min from historical operating data. The samples under fixed working conditions are clustered to generate the initial coupling coefficient of the corresponding template when the similarity is greater than 0.85. When the similarity is between 0.6 and 0.85, the coupling coefficient of the adjacent templates is fused by linear interpolation. When the similarity is less than 0.6, the general basic template is used. Finally, a combined state vector is generated, which is composed of a 2048-dimensional up-dimensional state representation, 17-dimensional normalized key operating parameters (including the instantaneous feeding amount of 6 feeding sources, the rotation speed of 2 rollers, the opening degree of 8 auxiliary doors, and the speed of 1 sintering machine), an 8-bit working condition code, and a 1-bit working condition matching degree identifier. The total dimension is 2074, and all numerical parameters have been mapped to the [0,1] interval.

[0093] When performing level balance analysis on a combined state vector, a linear baseline model is first established based on the fundamental equations of material dynamic balance: ,in, This represents the instantaneous rate of change of the material level in the mixing tank; a positive value indicates a rise in the material level, and a negative value indicates a fall in the material level. Let j be the instantaneous feeding amount of the j-th feeding source. The instantaneous discharge rate of the mixing tank is given by N, where N represents the total number of feed sources and A is the effective cross-sectional area of ​​the tank. Calculated using empirical formulas verified in industrial settings: =0.82×n×α×v×W×h des Where 0.82 is the discharge correction coefficient, n is the roller speed (r / min), α is the auxiliary door opening (%), v is the sintering machine speed (m / min), W is the effective width of the trolley (m), and h desTo design the material layer thickness (m), feature sub-dimensions corresponding to each influencing factor are extracted from the upgraded state representation: 32-dimensional sub-features corresponding to the 6 material sources are extracted from the first 512-dimensional feeding features; 64-dimensional sub-features corresponding to the 2 rollers are extracted from the middle 512-dimensional mixed features; next, 32-dimensional sub-features corresponding to the 8 auxiliary gates and 64-dimensional sub-features corresponding to the sintering machine speed are extracted from the 512-dimensional equipment features. The Pearson correlation coefficient between each sub-feature and the historical material level change sequence over the past 300 seconds is calculated. The correlation coefficient is multiplied by the coefficient in the linear baseline model to obtain the corrected coupling coefficient. A 17×120-dimensional material level coupling relationship matrix C is constructed, where the matrix element C(i1,k1) represents the i1th... The contribution coefficients of each influencing factor to the material level change at the k1th future time step are calculated. Simultaneously, a 17-dimensional weight distribution vector is generated, obtained by row-wise summation and normalization of the material level coupling matrix. Each element represents the contribution weight of the i1th influencing factor to the total material level change in the next 120 seconds, with weight values ​​ranging from 0 to 1, and the sum of the contribution weights of all influencing elements being 1. It should be noted that the coefficients in the linear baseline model refer to the initial sensitivity coefficients obtained after first-order linearization of the 17 key operating parameters based on the material dynamic balance equation. These include the forward feeding sensitivity coefficients corresponding to each feeding source, the discharge sensitivity coefficient corresponding to the roller speed, the discharge sensitivity coefficient corresponding to the auxiliary door opening, and the discharge sensitivity coefficient corresponding to the sintering machine speed. For the feeding source, its linear baseline coefficient is 1 / A; for the roller speed, auxiliary door opening, and sintering machine speed, their linear baseline coefficients are based on Q. out =0.82×n×α×v×W×h des The partial derivative with respect to the corresponding variable is then divided by the effective cross-sectional area A of the tank to obtain the result.

[0094] When generating the material level balance adjustment term based on the material level coupling relationship matrix, a weighted aggregation calculation method is used: ,in This represents the material level balance adjustment amount at the k-th future time step. Let p be the normalized value of the p-th influencing factor at the current time t. Let C(p,k) be the normalized value of the previous time step t-1, and let C(p,k) represent the contribution coefficient of the p-th influencing factor to the material level change at the k-th future time step. A 120-dimensional material level balance adjustment term vector is calculated, with each element corresponding to the material level balance adjustment amount at the k-th future time step. The material level balance adjustment term is weighted according to the prediction confidence output in step S202. When the confidence is lower than 0.6, the material level balance adjustment term is multiplied by a decay coefficient of 0.5 to avoid over-correction under low confidence.

[0095] Step S204: Integrate the future material level evolution sequence with the material level balance adjustment term to form a material level prediction trajectory, wherein the material level prediction value at each time step simultaneously includes nonlinear dynamic response and coupled balance constraint information.

[0096] In this embodiment, when performing time-correspondence matching between the future material level evolution sequence and the material level balance adjustment term, the future material level evolution sequence with 120 time steps generated in step S202 is first extracted. Each time step element includes a prediction timestamp, the original material level prediction value of the Koopman network, the upper and lower limits of the 95% confidence interval, and the prediction confidence level. At the same time, the 120-dimensional material level balance adjustment term vector generated in step S203 is extracted. Each element corresponds to the material level correction amount and correction confidence level of the same time step. By comparing the timestamp fields of the two, a one-to-one correspondence matching is completed, and finally a sequence consisting of 120 prediction combination units is formed. Each prediction combination unit includes the original Koopman material level prediction value, the material level balance adjustment term, the confidence level of both, and a unified timestamp for the corresponding time step.

[0097] When performing weighted fusion on each prediction combination unit, a dynamic weighted fusion algorithm based on dual confidence is adopted. First, the weight w_k of the upgraded state and the weight w_b of the material level balance constraint are calculated, satisfying w_k + w_b = 1, where the weight w_k of the upgraded state is α1 × c_k, and the weight w_b of the material level balance constraint is β × c_b. α1 and β are the adaptive coefficients of the working conditions, determined by the state stability indicator output in step S201: α1 = 0.6 and β = 0.4 under stable working conditions, α1 = 0.4 and β = 0.6 under transitional working conditions, and α1 = 0.2 and β = 0.8 under unknown working conditions. c_k is the prediction confidence of the Koopman network, ranging from 0 to 1, obtained by trace normalization of the prediction covariance matrix in step S202; c_b is the confidence of the material level balance adjustment term, ranging from 0 to 1, obtained by the process in step S203. The confidence interval is calculated using the weighted average of the matching degree and the data quality of each influencing factor. When the confidence level of any item is lower than 0.3, the corresponding weight is reset to 0 and the other weight is renormalized to ensure that the fused weights always satisfy the constraint that the sum is 1. The formula for calculating the fused single-step material level prediction value is: H_fusion(k)=w_k×H_koopman(k)+w_b×(H_koopman(k)+ΔH_balance(k)), where H_koopman(k) is the Koopman original material level prediction value at the k-th time step, and ΔH_balance(k) is the material level balance adjustment term at the corresponding time step. At the same time, the confidence interval after fusion is calculated using the variance weighted method, and the half-width of the confidence interval is σ_fusion(k)=(w_k²×σ_k²+w_b²×σ_b²). 1 / 2 , where σ_k is the half-width of the confidence interval for the original Koopman level prediction, and σ_b is the half-width of the confidence interval for the level balance adjustment term. Finally, a single-step prediction vector containing the fused level value, the upper and lower limits of the fused confidence interval, and the fused confidence level is generated.

[0098] The single-step predicted value vectors of each time step are sequentially connected in chronological order to form a continuous material level prediction trajectory with a time span of 120 seconds. During the rolling verification of the formed material level prediction trajectory, a complete verification process is executed every 10 seconds. First, an evolution consistency verification is performed: the material level change rate dH / dt of adjacent time steps is calculated. When the change rate exceeds the physically reasonable threshold of ±0.05 m / s, it is determined to be an evolution anomaly, and the abnormal value is replaced by the linear interpolation result of the two preceding and following normal time steps. Second, a coupling constraint compliance verification is performed: the single-step predicted value of each time step is substituted into the basic equation of material dynamic balance to calculate the predicted... The deviation between the material quantity and the actual output quantity is judged as a constraint violation when the deviation exceeds ±2%. The theoretical material level value and the fused prediction value are used for secondary weighted fusion, and the correction ratio is 50% of the deviation value. Finally, extreme value verification is performed: when the material level prediction value exceeds the upper limit of the tank design of 4.0 meters or is lower than the lower limit of 0.2 meters, it is forcibly corrected to the corresponding extreme value and marked as a warning state. After the above verification and adjustment, the final material level prediction trajectory is generated. This trajectory contains complete data of 120 time steps. Each data point is marked with a timestamp, material level prediction value, 95% confidence interval, fusion confidence level, data quality label and anomaly correction mark.

[0099] Step S205: Perform a rolling test on the predicted material level trajectory to ensure that it meets the one-step prediction error constraint, the multi-step rolling prediction error constraint, and the high-dimensional space linear evolution consistency constraint.

[0100] In this embodiment, the material level prediction trajectory generated in step S204 is subjected to rolling verification according to a time series to verify the accuracy and consistency of the prediction results within a continuous control cycle. Specifically, this involves comparing the predicted material level value at each time step in the material level prediction trajectory with the corresponding actual historical material level data, calculating the one-step prediction error, and accumulating and analyzing the prediction errors of multiple consecutive time steps to evaluate the multi-step rolling prediction error. Simultaneously, a consistency check is performed on the state evolution of the material level prediction trajectory in the high-dimensional observable space. By verifying the linear mapping relationship between the higher-dimensional state representations, it is ensured that the evolution of the material level prediction trajectory conforms to the linear prediction model. If the prediction error or linear evolution does not meet the preset threshold, the preset threshold refers to the error allowable range pre-set based on the offline training of the model and the on-site process constraints. It is used to determine whether the material level prediction trajectory meets the prediction accuracy requirements and the Koopman linear evolution requirements. The one-step prediction error threshold is set to 0.05m, the multi-step rolling prediction error threshold is set to 0.10m, and the high-dimensional space linear evolution consistency threshold is set to 0.05. Then, the material level prediction value of the corresponding time step is corrected or weighted and adjusted to generate a corrected material level prediction trajectory that meets the one-step prediction error, multi-step rolling prediction error and high-dimensional space linear evolution consistency constraints.

[0101] Step S3: Based on the predicted material level trajectory, the comprehensive feeding quantity adjustment is obtained by using a predictive composite PID strategy, and then sent to the feeding execution unit to maintain the material level within the target range. The predictive composite PID strategy introduces the predicted material level change trend as a feedforward correction into the PID control process.

[0102] like Figure 3 As shown, the specific steps of step S3 are as follows:

[0103] Step S301: Based on the predicted material level trajectory, extract the predicted material level value, the direction of material level change, and the trend information of the material level approaching the target range boundary in the order of prediction time, and generate the predicted trend input quantity.

[0104] In this embodiment, when extracting the predicted trend input based on the final material level prediction trajectory that satisfies the triple constraints generated in step S205, the preset material level control target range is first defined: the target material level under normal production conditions is set to 2.0 meters, the allowable fluctuation range is 1.8 meters to 2.2 meters, the first-level warning boundary is 1.7 meters and 2.3 meters, and the second-level alarm boundary is 1.5 meters and 2.5 meters. All boundary values ​​are based on the effective volume of the mixing tank of 40 cubic meters, the rated capacity of the sintering machine of 450 tons / hour, and the material residence time of 120 seconds. The process parameters for each second are calculated and determined. When extracting the predicted material level value, the predicted material level values ​​for the current time t and five key time nodes (10 seconds, 30 seconds, 60 seconds, and 120 seconds in the future) are extracted sequentially from the final predicted material level trajectory according to a fixed time step of 1 second. When calculating the direction of material level change, a sliding linear regression method with a window size of 5 seconds is used to fit the local slope of the predicted material level trajectory. The fitting formula is H(t) = a × t + b, where H(t) is the fitted material level value corresponding to time t within the sliding window, and the slope a is the instantaneous change in material level. The rate, in meters per second, is used. b is the fitted baseline material level value at the beginning of the sliding window (t=0). The regression coefficients are solved using the least squares method, and the regression determination coefficient R² is calculated as a confidence index for the direction of change. R² ranges from 0 to 1; the closer R² is to 1, the more significant the trend. Based on the slope value, the direction of material level change is divided into 5 levels and coded as integers: a slope greater than 0.02 meters per second is marked as a rapid rise (code 4), and a slope between 0.005 meters per second and 0.02 meters per second is marked as a slow rise. An upward trend (code 3, including 0.005 m / s and 0.02 m / s) is marked as stationary (code 2, including -0.005 m / s) with a slope between -0.005 m / s and 0.005 m / s, a slow downward trend (code 1, including -0.02 m / s) with a slope between -0.02 m / s and -0.005 m / s, and a rapid downward trend (code 0) with a slope less than -0.02 m / s. When the regression coefficient of determination R² is less than 0.5, the trend is marked as uncertain (code 5).

[0105] When extracting trend information of material level approaching the target range boundary, first calculate the absolute distance between the current predicted material level and the upper and lower control boundaries: d_upper = 2.2 - H_pred(t), d_lower = H_pred(t) - 1.8, where d_upper represents the absolute distance between the current predicted material level and the upper control boundary, d_lower represents the absolute distance between the current predicted material level and the lower control boundary, 2.2 is the upper control boundary value of the mixing tank material level, 1.8 is the lower control boundary value of the mixing tank material level, and H_pred(t) represents the predicted material level at the current time t. Then, based on the current material level... The rate of change is used to calculate the predicted time to reach the boundary (TTA), including the predicted time to reach the upper control boundary (TTA_upper = (2.2 - H_pred(t)) / a) and the predicted time to reach the lower control boundary (TTA_lower = (H_pred(t) - 1.8) / (-a)). When the rate of change a is 0 or the calculated TTA is negative, the arrival time of the corresponding boundary is marked as infinity. Simultaneously, the upper and lower limits of the boundary arrival time are calculated using a 95% confidence interval. The upper confidence interval is used to calculate the shortest time to reach the upper control boundary, and the lower confidence interval is used to calculate the shortest time to reach the lower control boundary. Based on the arrival time, the boundary is then... The urgency level near the boundary is divided into four levels: arrival time less than 10 seconds is marked as an emergency state (code 3), arrival time between 10 and 30 seconds is marked as a warning state (code 2, including 10 seconds), arrival time between 30 and 60 seconds is marked as a state of concern (code 1, including 30 and 60 seconds), and arrival time greater than 60 seconds is marked as a safe state (code 0). The final generated predictive trend input is a 15-dimensional one-dimensional vector, with the elements arranged in the following order: the current normalized predicted material level value (mapped to the [0,1] interval), the normalized predicted material level value for the next 10 seconds, and the normalized predicted material level value for the next 30 seconds. Level value, normalized predicted level value for the next 60 seconds, normalized predicted level value for the next 120 seconds, level change rate within a 5-second window (m / s, 4 decimal places), level change direction code, confidence level of change direction (R² value, 2 decimal places), absolute distance from the upper control boundary (m, 3 decimal places), absolute distance from the lower control boundary (m, 3 decimal places), predicted time to reach the upper control boundary (seconds, 1 decimal place), predicted time to reach the lower control boundary (seconds, 1 decimal place), urgency code of approaching the upper control boundary, urgency code of approaching the lower control boundary, and overall data quality identifier.

[0106] Step S302: Compare the predicted trend input with the preset target range to determine the material level deviation type within the current control cycle. The material level deviation type includes material level rise deviation, material level fall deviation, full material approach deviation, and material shortage approach deviation.

[0107] In this embodiment, when comparing the predicted trend input generated in step S301 with the preset target range, the preset material level control parameter system is first defined: the control cycle is set to 10 seconds, synchronized with the verification cycle in step S205; the target material level center value is 2.0 meters; the normal control range is 1.8 meters to 2.2 meters; the first-level warning boundary is 1.7 meters and 2.3 meters; the second-level alarm boundary is 1.5 meters and 2.5 meters; the confidence threshold for deviation judgment is set to 0.6; only when the overall data quality of the predicted trend input is identified as normal and the confidence level R² of the change direction is ≥0.6, the complete deviation type judgment process is executed; otherwise, the deviation type of the previous control cycle remains unchanged and is marked as a low confidence judgment. Extract the core parameters from the predicted trend input: the current normalized predicted level value, the 5-second window level change rate 'a', the level change direction code, the prediction time to reach the upper control boundary (TTA_upper), the prediction time to reach the lower control boundary (TTA_lower), the urgency code for approaching the upper control boundary, and the urgency code for approaching the lower control boundary. Inversely normalize the current normalized predicted level value to the actual predicted level value H_pred. Calculate the absolute deviation ΔH = H_pred - 2.0 between the actual predicted level value and the target center value. Then, determine the level deviation type in descending order of priority: full material approach deviation > short material approach deviation > level rise deviation > level. The descent deviation, of which the full material approach deviation is determined by the following conditions: (1) the current material level prediction value H_pred ≥ 2.1 meters and the material level change direction code is 3 (slow rise) or 4 (rapid rise), and the prediction time to reach the upper control boundary TTA_upper ≤ 60 seconds; (2) the urgency level of the upper control boundary is coded as 2 (early warning state) or 3 (emergency state); (3) the current material level prediction value H_pred ≥ 2.2 meters and the material level change direction code is not 0 (rapid descent) or 1 (slow descent); the material shortage approach deviation is determined by the following conditions: (1) the current material level prediction value H_pred ≤ 1.9 meters and the material level change direction code is not 0 (rapid descent) or 1 (slow descent). (1) The code is 0 (rapid descent) or 1 (slow descent), and the predicted time to reach the lower control boundary is TTA_lower≤60 seconds; (2) The urgency level of the lower control boundary is coded as 2 (early warning state) or 3 (emergency state); (3) The current material level prediction value H_pred≤1.8 meters and the material level change direction code is not 3 (slow rise) or 4 (rapid rise); When the judgment conditions for the deviation of the full material approach and the material shortage approach are not met, the material level rise deviation and the material level fall deviation are judged: The judgment conditions for the material level rise deviation are: the current material level prediction value H_pred is in the range of 1.8 meters to 2.2 meters, and the material level change direction code is 3 (slow rise) or 4 (rapid rise), and the absolute deviation ΔH>0.The material level descent deviation is determined as follows: the current predicted material level H_pred is within the range of 1.8 meters to 2.2 meters, the material level change direction is encoded as 0 (rapid descent) or 1 (slow descent), and the absolute deviation ΔH < -0.05 meters; when none of the deviation type determination conditions are met, the material level is determined to be in a stable state, and the deviation type code is set to 0; the final output material level deviation information includes: an 8-bit binary deviation type code (where the high 2 bits represent the deviation category, 00 represents a stable state, 01 represents an upward deviation, and 10 represents a downward deviation). The deviation is categorized as follows: 11 indicates a boundary approach deviation; the lower 6 bits indicate the specific deviation subtype: 000001 indicates a level rise deviation, 000010 indicates a level fall deviation, 000100 indicates a full-material approach deviation, and 001000 indicates a material shortage approach deviation. The data includes the current absolute deviation ΔH (in meters, rounded to 3 decimal places), the level change rate a (in meters per second, rounded to 4 decimal places), the estimated time to reach the corresponding boundary (in seconds, rounded to 1 decimal place), the deviation confidence level (value 0-1, rounded to 2 decimal places), and the deviation priority indicator.

[0108] Step S303: Generate a feedforward correction amount according to the material level deviation type, and introduce the feedforward correction amount into the PID control process, and fuse it with the feedback control amount formed based on the current actual material level deviation to obtain the predicted composite PID control amount.

[0109] In this embodiment, when factoring the predicted material level change trend based on the material level deviation type output in step S302, the 5-second window material level change rate a, the prediction time to reach the upper control boundary TTA_upper, the prediction time to reach the lower control boundary TTA_lower, and the confidence level R² of the material level change direction are first extracted from the predicted trend input generated in step S301. The overall trend is then decomposed into four orthogonal components using a linear decomposition method: the upward trend component F_up=max(a,0)×R², the downward trend component F_down=max(-a,0)×R², the full-material approach component F_full=min(1 / TTA_upper,0.1)×R² (F_full=0 when TTA_upper is infinite), and the short-material approach component F_empty=min(1 / TTA_lower,0.1)×R² (F_empty=0 when TTA_lower is infinite). The value range of each component is [0,0.1], and the material level deviation is determined according to the material level deviation. For the difference type, non-corresponding components are zeroed out: only F_up is retained when the material level rises, only F_down is retained when the material level falls, both F_up and F_full are retained when the material level approaches full, and both F_down and F_empty are retained when the material level approaches short. When generating feedforward correction values ​​for all trend components, the formula for calculating the feedforward correction value for the material feeding amount is ΔQ_ff=K_ff×(F_up×K_up-F_down×K_down-F_full×K_fu The formula is: 11 + F_empty × K_empty) × Q_rated, where ΔQ_ff represents the feedforward correction amount, K_ff is the global feedforward gain coefficient with a value of 1.2, K_up, K_down, K_full, and K_empty are the specific gain coefficients of each component with values ​​of 0.8, 0.8, 1.5, and 1.5 respectively, and Q_rated is the rated feedforward amount of the mixing tank with a value of 7.5 tons / minute. The calculated feedforward correction amount is in tons / minute.

[0110] When generating the comprehensive feedforward correction, the feedforward corrections corresponding to the upward trend component, downward trend component, full-material approach component, and short-material approach component are first mapped to the corresponding control cycles according to the prediction time axis, so that each feedforward correction is consistent with the time window in which it actually exerts its control effect. Then, based on the predicted amplitude, duration, proximity to the target material level boundary, and distance from the corresponding control cycle of each trend component, weights are assigned to each feedforward correction. The weight of the current or near-end control cycle is greater than that of the far-end control cycle, and the weights of the full-material approach component and the short-material approach component are higher than those of the general upward or downward trend component. Then, the feedforward corrections within the same control cycle are normalized and weighted according to the adjustment direction. For components with opposite adjustment directions, they are offset or the dominant direction is retained according to the material level deviation priority, finally obtaining the comprehensive feedforward correction for the current control cycle.

[0111] When matching the comprehensive feedforward correction with the control cycle, a time-weighted allocation method is used to evenly distribute the predicted trend influence within the next 10-second control cycle to the current control cycle. Simultaneously, based on the average lag time of 30 seconds from each material source to the mixing tank outlet, a three-control-cycle advance compensation is applied to the comprehensive feedforward correction to ensure that the feedforward action is synchronized with the actual material arrival time. When calculating the feedback adjustment based on the current actual material level deviation, an incremental PID control algorithm is used. The PID parameters are adaptively adjusted according to the operating conditions: under stable operating conditions, the proportional coefficient Kp = 2.0, integral coefficient Ki = 0.1, and derivative coefficient Kd = 0.5; under transient operating conditions, Kp = 3.0, Ki = 0.15, and Kd = 0.8; and under boundary warning operating conditions, Kp = 4.0, Ki = 0.2, and Kd = 1.0. The formula for calculating the feedback adjustment ΔQ_fb is as follows:

[0112] ΔQ_fb=Kp×(ΔH(k)-ΔH(k-1))+Ki×ΔH(k)+Kd×(ΔH(k)-2ΔH(k-1)+ΔH(k-2)), where ΔH(k) is the deviation between the actual predicted material level and the target material level at time step k, and ΔH(k-1) and ΔH(k-2) are the deviations of the previous time step and the two previous time steps, respectively.

[0113] When the comprehensive feedforward correction and feedback adjustment are weighted and fused, the fusion weight is dynamically determined based on the prediction confidence and deviation priority. The feedforward weight w_ff = α2 × β2, and the feedback weight w_fb = 1 - w_ff, where α2 is the prediction confidence coefficient, which is equal to the change direction confidence R² in step S301, and β2 is the deviation priority coefficient. Under steady state, β2 = 0.3, under rising / falling deviation, β2 = 0.5, and under full / short material approach deviation, β2 = 0.7. When w_ff exceeds 0.8, it is forcibly limited to 0.8. The final generated predicted composite PID adjustment ΔQ_total = w_ff × ΔQ_ff + w_fb × ΔQ_fb, in tons / minute, is used as the comprehensive feeding adjustment input for the next 10-second control cycle.

[0114] Step S304: Perform process boundary processing on the predicted composite PID adjustment amount to generate a comprehensive feed rate adjustment amount. The process boundary processing includes upper and lower limits of the comprehensive feed rate, feed rate variation range constraints, adjustment direction consistency verification, and adjustment cycle limit matching the sintering machine speed change state.

[0115] In this embodiment, when performing process boundary processing, the preset process boundary parameter system for the feeding amount is first defined: the lower limit of the total feeding amount is set to 3.0 tons / minute, corresponding to the material consumption when the sintering machine operates at its lowest speed of 0.5 m / min. If it is lower than this value, the material level will drop rapidly and cause a material shortage risk. The upper limit of the total feeding amount is set to 10.0 tons / minute, corresponding to the sum of the maximum feeding rate of the mixing tank and the rated feeding capacity of the 6 material sources. If it is exceeded, the tank will overflow. During the material change transition, the upper limit is temporarily lowered to 8.0 tons / minute to prevent uneven mixing of different batches of materials. The predicted composite PID adjustment amount generated in step S303 is compared bit by bit with the above boundary parameters. When the adjustment amount is lower than 3.0 tons / minute, it is forcibly constrained to 3.0 tons / minute. When the adjustment amount is higher than 10.0 tons / minute, it is forcibly constrained to 10.0 tons / minute. At the same time, it is marked as a boundary constraint state in the data quality identifier, forming a preliminary constraint adjustment value.

[0116] When verifying the change range of the initial constraint adjustment value, the actual execution feed amount Q_prev of the previous control cycle (10 seconds) is extracted, and the change range of the adjustment amount ΔQ=|Q_current-Q_prev| is calculated. Q_current represents the feed amount to be adjusted in the current week. The maximum allowable change range under normal working conditions is preset to 1.0 tons / minute / control cycle, which is relaxed to 1.5 tons / minute / control cycle under boundary warning conditions, and further relaxed to 2.0 tons / minute / control cycle under emergency conditions (the time to reach the boundary prediction is less than 10 seconds). When ΔQ exceeds the allowable range of the corresponding working condition, smoothing is performed. The compressed adjustment amount Q_compressed=Q_prev+sign(Q_current-Q_prev)×ΔQ_max, where ΔQ_max is the maximum allowable change range under the current working condition, sign(∙) is the sign function, and ΔQ_max is the maximum allowable change range under the current working condition. The compression process keeps the adjustment direction unchanged and only limits the adjustment range to avoid fluctuations in the mixture composition and equipment impact caused by sudden changes in the feed amount.

[0117] When performing a direction consistency check, a historical queue of adjustment directions for three consecutive control cycles is established. If the current adjustment direction is opposite to the adjustment direction of the previous two cycles and the change exceeds 0.5 tons / minute, it is determined to be a potential direction conflict. At this time, a direction switching dead zone of 0.3 tons / minute is introduced, and the adjustment amount is corrected to Q_prev+sign(Q_current-Q_prev)×0.3 tons / minute. The adjustment direction is then re-evaluated in the next control cycle to prevent system oscillations caused by frequent direction changes. When performing sintering machine speed cycle matching, the real-time machine speed v of the sintering machine collected in step S104 and the past 30... The machine speed change rate dv / dt is measured in seconds. When the absolute value of the machine speed change rate is less than 0.01 m / min², a standard adjustment cycle of 10 seconds is maintained. When the machine speed change rate is between 0.01 m / min² and 0.05 m / min², the adjustment cycle is shortened to 5 seconds, and the allowable change range is increased by 20%. When the machine speed change rate is greater than 0.05 m / min², a fast response mode is triggered, the adjustment cycle is shortened to 2 seconds, and the allowable change range is increased by 50%, ensuring that the change in feed rate and the change in sintering machine output rate remain synchronized. Finally, a comprehensive feed rate adjustment value is generated as the total feed rate command value, in tons per minute.

[0118] Step S305: Send the comprehensive feeding quantity adjustment amount to the feeding execution unit, and after sending it, record the predicted trend input amount, feedforward correction amount, feedback adjustment amount and material level feedback data after execution within the corresponding control cycle.

[0119] In this embodiment, the comprehensive feed rate adjustment calculated by the predictive composite PID strategy is sent to the feed execution unit through the control interface to achieve real-time adjustment of the mixing tank feed equipment. During the adjustment process, the predicted material level trajectory, feedforward correction, and feedback adjustment based on the current actual material level deviation are synchronously recorded according to the control cycle. At the same time, the actual material level data after the adjustment is executed is collected. Specifically, various parameter data are stored in the control recording unit in chronological order, and a control cycle index is established. By jointly recording the input, feedforward correction, and feedback adjustment, the effect of the comprehensive feed rate adjustment is accurately tracked, and a reliable historical data basis is provided for subsequent Koopman network and online iterative correction of material level balance, thereby ensuring the continuity and predictability of material level control under multiple operating conditions and multiple disturbances.

[0120] like Figure 3 As shown, in step 1, the actual material level, predicted material level trajectory, and target material level range of the current mixing tank are collected to form the basic input information for prediction and feedback. In step 2, the trend of the predicted trajectory is extracted, and the material level deviation type is identified. Then, in step 3, a feedforward correction amount is generated based on the material level deviation type, and in step 4, a feedback adjustment amount is formed by combining the real-time material level deviation. In step 5, the feedforward correction amount and the feedback adjustment amount are fused to obtain the predicted composite PID adjustment amount. Next, in step 6, the predicted composite PID adjustment amount is processed by the process boundary, including upper and lower limits of feeding amount, variation range limit, consistency verification of adjustment direction, and adjustment cycle limit, to ensure that the adjustment amount can be safely executed. Finally, in step 7, the comprehensive feeding amount adjustment amount is sent to the feeding execution unit, and the material level feedback data is recorded in real time to form a closed-loop feedback, providing correction input for the next control cycle, and realizing dynamic control and optimization adjustment of the material level.

[0121] Step S4: After adjusting the overall feeding amount, collect the actual material level and key process data in real time, calculate the prediction deviation, and perform online iterative correction of the Koopman network and material level balance based on the prediction deviation.

[0122] The specific steps of step S4 are as follows:

[0123] Step S401: After adjusting the overall feeding amount, collect the actual material level and feeding amount in the mixing tank, the rotation speed of the roller, the opening degree of the auxiliary door, and the speed of the sintering machine in real time, and perform time marking and working condition coding according to the control cycle sequence.

[0124] Step S402: Compare the real-time collected data with the material level prediction trajectory, calculate the prediction deviation at each time step, and generate a prediction deviation vector based on the deviation magnitude and trend.

[0125] Step S403: Based on the predicted deviation vector, perform online iterative correction on the up-dimensional state mapping parameters in the Koopman network and the coupling weights in the level balance.

[0126] Step S404: Use the corrected Koopman network as the prediction input for the next control cycle, and simultaneously record the deviation data, adjustment parameters and corresponding process data of the current control cycle for rolling optimization and adaptive reconfiguration of operating conditions.

[0127] In this embodiment, after the comprehensive feed rate adjustment is executed, the actual material level in the mixing tank and key process parameters, including feed rate, roller speed, auxiliary door opening, and sintering machine speed, are collected in real time by a high-speed acquisition unit. Simultaneously, the data is time-stamped according to the control cycle and encoded for different operating conditions, forming a complete time-series data stream. Subsequently, the collected data is compared step-by-step with the pre-generated material level prediction trajectory to calculate the prediction deviation. A prediction deviation vector is generated based on the deviation magnitude and trend to characterize the difference between the dynamic response and the prediction of the material level. Based on this prediction deviation vector, the upgraded state mapping parameters in the Koopman network and the coupling weights in the material level balance are iteratively updated online, enabling the network to capture changes in nonlinear dynamic characteristics and reflect the impact of operating condition disturbances, while correcting the coupling relationships of various control parameters in the material level balance. Finally, the corrected Koopman network is used as the prediction input for the next control cycle, and the deviation data, parameter adjustment status, and process data of the current control cycle are recorded. This supports rolling optimization, adaptive reconstructing of operating conditions, and dynamic improvement of prediction accuracy, thereby achieving adaptive closed-loop adjustment of material level control under continuous production conditions.

[0128] Example 2:

[0129] Please see Figure 4 Another embodiment of the present invention provides a mixing tank level control system based on Koopman network, comprising: a vector construction module, a level prediction module, a level control module, and an iterative correction module;

[0130] The vector construction module is used to collect key process data of the mixing tank and construct the original state vector of the material level control. The key process data includes the feeding amount, the amount of material at each mixing point, the rotation speed of the roller, the opening degree of the auxiliary door, the speed of the sintering machine, and the current material level of the tank.

[0131] The material level prediction module is used to input the original state vector into a pre-trained Koopman network, map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a material level prediction trajectory in combination with the material level balance. The material level balance represents the coupled influence of the total amount of material fed, equipment parameters and auxiliary gate opening on the material level increase or decrease trend. The linear prediction model satisfies the one-step prediction error, multi-step rolling prediction error and high-dimensional linear evolution consistency constraints.

[0132] The material level control module is used to solve the comprehensive feeding quantity adjustment based on the predicted material level trajectory using a predictive composite PID strategy, and then send it to the feeding execution unit to maintain the material level within the target range. The predictive composite PID strategy introduces the predicted material level change trend as a feedforward correction into the PID adjustment process.

[0133] The iterative correction module is used to collect actual material level and key process data in real time after the comprehensive feeding amount adjustment is executed, calculate the prediction deviation, and perform online iterative correction of the Koopman network and material level balance based on the prediction deviation.

[0134] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0135] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling the material level in a mixing tank based on a Koopman network, characterized in that, include: Collect key process data of the mixing tank and construct the original state vector for material level control. The key process data includes feeding amount, material amount at each mixing point, roller speed, auxiliary door opening, sintering machine speed and current material level in the tank. The original state vector is input into a pre-trained Koopman network to map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a material level prediction trajectory by combining the material level balance. The material level balance represents the coupled influence of the total amount of material fed, equipment parameters and auxiliary gate opening on the material level increase and decrease trend. The linear prediction model satisfies the one-step prediction error, multi-step rolling prediction error and linear evolution consistency constraints. Based on the predicted material level trajectory, a predictive composite PID strategy is used to solve for the comprehensive material feeding adjustment amount, and then the adjustment amount is sent to the feeding execution unit to maintain the material level within the target range. The predictive composite PID strategy introduces the predicted material level change trend as a feedforward correction into the PID control process. After adjusting the overall feed rate, the actual material level and key process data are collected in real time, the prediction deviation is calculated, and the Koopman network and material level balance are iteratively corrected online based on the prediction deviation.

2. The mixing tank level control method based on Koopman network as described in claim 1, characterized in that, The key process data of the mixing tank are collected, and the original state vector for material level control is constructed, including: The feeding amount of each feeding source in the mixing tank is collected in real time, and the data of each feeding source is synchronously arranged according to the timestamp to form a feeding status sub-vector; Collect the material volume of each mixing point in the mixing tank, and associate and mark it according to the material flow direction and mixing sequence to construct a sub-vector of material state at the mixing point; Obtain the motion parameters of the trough and auxiliary doors, including the rotational speed of the rollers and the opening degree of each auxiliary door, and map them into equipment action state sub-vectors according to their spatial position and motion sequence; Get the current machine speed of the sintering machine and the current material level in the trough, and align them with the collected material status at the feeding and mixing points in time to generate a sub-vector that associates the material level history with the instantaneous status. The generated sub-vectors are combined and normalized according to the working condition code and time series to construct the original state vector representing the current comprehensive operating state of the mixing tank.

3. The mixing tank level control method based on Koopman network as described in claim 1, characterized in that, The original state vector is input into a pre-trained Koopman network to map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establishing a linear prediction model, and generating a predicted material level trajectory by combining the material level balance, including: The constructed original state vector is input into a pre-trained Koopman network to generate an upgraded state representation in a high-dimensional observable space, and to distinguish and label the nonlinear features of different working conditions. A linear prediction model is established in the up-dimensional state representation. The future material level evolution sequence is generated by linear combination and time recursion of each up-dimensional state representation, and the prediction confidence interval information of each time step is recorded. By combining the feeding amount, roller speed, auxiliary door opening and sintering machine speed with the dimension-upgrading status, and using the material level balance to perform coupled analysis on the material level increase and decrease trend, a material level balance adjustment item is generated. The future material level evolution sequence is fused with the material level balance adjustment term to form a material level prediction trajectory, wherein the material level prediction value at each time step simultaneously includes nonlinear dynamic response and coupled balance constraint information. The predicted material level trajectory is subjected to rolling verification to ensure that it meets the one-step prediction error constraint, the multi-step rolling prediction error constraint, and the linear evolution consistency constraint.

4. The mixing tank level control method based on Koopman network as described in claim 3, characterized in that, The method combines the feeding amount, roller speed, auxiliary door opening, and sintering machine speed with the dimensional state representation, and uses material level balance to perform coupled analysis on the material level increase / decrease trend to generate material level balance adjustment items, including: The upgraded state representation is time-aligned and operating condition-matched with the currently collected data on the feeding amount, roller speed, auxiliary door opening, and sintering machine speed to generate a combined state vector; A material level balance analysis is performed on the combined state vector. By mapping the contribution of each feeding point, equipment operation and auxiliary door opening change to the material level increase / decrease trend space, a material level coupling relationship matrix is ​​formed, and the weight distribution of different influencing factors is marked. Based on the material level coupling relationship matrix, the contributions of each influencing factor are aggregated and calculated to generate material level balance adjustment terms, where each adjustment term corresponds to a prediction correction vector of the material level change trend in future time steps.

5. The mixing tank level control method based on Koopman network as described in claim 4, characterized in that, By integrating the future material level evolution sequence with the material level balance adjustment term, a material level prediction trajectory is formed, including: The future material level evolution sequence is unfolded according to time steps and matched with the material level balance adjustment item in time to form a time-aligned prediction combination unit; For each time step within the prediction combination unit, a weighted fusion is performed according to the coupling relationship between the dimensionality-upgraded state weights and the material level balance constraints to generate a fused single-step prediction value vector. The single-step prediction vectors fused from each time step are connected sequentially to form a continuous material level prediction trajectory. The material level prediction value at each time step simultaneously reflects the nonlinear dynamic evolution and the coupling balance constraint information of each control parameter on the material level. The generated material level prediction trajectory is rolled over and verified. By checking the evolution consistency and coupling constraint compliance between consecutive time steps, the predicted material level values ​​in the material level prediction trajectory are adjusted to generate the final material level prediction trajectory.

6. The mixing tank level control method based on Koopman network as described in claim 1, characterized in that, The process of using a predictive composite PID strategy based on the predicted material level trajectory to obtain the comprehensive feeding rate adjustment is then sent to the feeding execution unit to maintain the material level within the target range, including: Based on the predicted material level trajectory, the predicted material level value, the direction of material level change, and the trend information of the material level and the boundary of the target range are extracted in the order of prediction time to generate the predicted trend input. The predicted trend input is compared with the preset target range to determine the material level deviation type in the current control cycle. The material level deviation type includes material level rise deviation, material level fall deviation, full material approach deviation, and material shortage approach deviation. A feedforward correction amount is generated based on the material level deviation type, and the feedforward correction amount is introduced into the PID control process. It is then fused with the feedback control amount formed based on the current actual material level deviation to obtain the predicted composite PID control amount. The predicted composite PID adjustment is subjected to process boundary processing to generate a comprehensive feed rate adjustment. The process boundary processing includes upper and lower limits of comprehensive feed rate, feed rate change range constraint, adjustment direction consistency verification, and adjustment cycle limit matching the sintering machine speed change state. The comprehensive feeding quantity adjustment is sent to the feeding execution unit, and after the issuance, the predicted trend input, feedforward correction, feedback adjustment and material level feedback data within the corresponding control cycle are recorded.

7. The mixing tank level control method based on Koopman network as described in claim 6, characterized in that, A feedforward correction amount is generated based on the material level deviation type, and this feedforward correction amount is introduced into the PID control process. It is then fused with the feedback control amount formed based on the current actual material level deviation to obtain a predicted composite PID control amount, including: Based on the material level deviation type, the predicted material level change trend is decomposed by factor, and the trend components corresponding to different material level deviation types are labeled as upward trend component, downward trend component, full material approach component and short material approach component, respectively. For each trend component, a feedforward correction is generated. The predictive effect of each trend component is matched with the corresponding control period in time to form a comprehensive feedforward correction. The comprehensive feedforward correction amount and the feedback adjustment amount calculated based on the current actual material level deviation are linearly or weighted and fused to form a fused adjustment vector; The fusion adjustment vector is defined as the predicted composite PID adjustment amount.

8. The mixing tank level control method based on Koopman network as described in claim 7, characterized in that, The predicted composite PID control quantity is subjected to process boundary processing to generate a comprehensive feed rate control quantity, including: The predicted composite PID adjustment is compared with the preset upper and lower limits of the feed amount. Boundary constraints are applied to the adjustment amount that exceeds the upper and lower limits to form the initial constraint adjustment value. The change range of the initial constraint adjustment value is compared with that of the feeding amount in the previous control cycle, and the adjustment amount that exceeds the allowable change range is compressed. Check the consistency between the adjustment amount after amplitude compression and the feeding direction, correct the conflict of adjustment direction, and match the adjustment rhythm according to the change of sintering machine speed to generate an adjustment plan that conforms to the process rhythm. The adjustment plan is defined as the comprehensive feeding quantity adjustment amount and is issued to the feeding execution unit.

9. The mixing tank level control method based on Koopman network as described in claim 1, characterized in that, After adjusting the overall material feeding rate, the system collects actual material level and key process data in real time, calculates the prediction deviation, and performs online iterative correction of the Koopman network and material level balance based on the prediction deviation, including: After adjusting the overall feeding amount, the actual material level and feeding amount in the mixing tank, the rotation speed of the roller, the opening degree of the auxiliary door, and the speed of the sintering machine are collected in real time, and time stamps and working condition codes are performed according to the control cycle sequence. The real-time collected data is compared with the material level prediction trajectory to calculate the prediction deviation at each time step, and a prediction deviation vector is generated based on the deviation magnitude and trend. Based on the predicted deviation vector, the upgraded state mapping parameters in the Koopman network and the coupling weights in the level balance are iteratively corrected online. The modified Koopman network is used as the prediction input for the next control cycle, while the deviation data, adjustment parameters and corresponding process data of the current control cycle are recorded at the same time.

10. A mixing tank level control system based on a Koopman network, used to implement the mixing tank level control method based on a Koopman network as described in any one of claims 1-9, characterized in that, include: The module includes a vector construction module, a material level prediction module, a material level control module, and an iterative correction module. The vector construction module is used to collect key process data of the mixing tank and construct the original state vector of the material level control. The key process data includes the feeding amount, the amount of material at each mixing point, the rotation speed of the roller, the opening degree of the auxiliary door, the speed of the sintering machine, and the current material level of the tank. The material level prediction module is used to input the original state vector into a pre-trained Koopman network, map the nonlinear dynamic process of the material level in the mixing tank to a high-dimensional observable space, establish a linear prediction model, and generate a material level prediction trajectory in combination with the material level balance. The material level balance represents the coupled influence of the total amount of material fed, equipment parameters and auxiliary door opening on the material level increase or decrease trend. The linear prediction model satisfies one-step prediction error, multi-step rolling prediction error and linear evolution consistency constraints. The material level control module is used to solve the comprehensive feeding quantity adjustment based on the predicted material level trajectory using a predictive composite PID strategy, and then send it to the feeding execution unit to maintain the material level within the target range. The predictive composite PID strategy introduces the predicted material level change trend as a feedforward correction into the PID adjustment process. The iterative correction module is used to collect actual material level and key process data in real time after the comprehensive feeding amount adjustment is executed, calculate the prediction deviation, and perform online iterative correction of the Koopman network and material level balance based on the prediction deviation.

Images