Continuous casting inclusion analysis and control method, device and program product

Abstract

The invention provides a continuous casting inclusion analysis and control method and device and a program product, and relates to the technical field of ferrous metallurgy intelligent manufacturing. The analysis and control method comprises the steps of multi-source data preprocessing and feature screening, including data acquisition, data cleaning, abnormal data enhancement, feature screening and weight calculation; performing association rule offline analysis, including modeling a time sliding window, performing data time sequence coding by adopting a bidirectional long-short-term memory network, and performing strong association rule mining and dynamic optimization based on a support degree-confidence joint optimization method and a rule dynamic updating mechanism; the online control and real-time optimization of the continuous casting production line comprises data acquisition, process calculation and quality judgment of an embedded DeepSeek lightweight inference engine, and further comprises real-time optimization control of a dynamic PID-DeepSeek hybrid control method. According to the method, a DeepSeek feature optimization engine is introduced, adaptive optimization of a parameter sensitive interval is realized based on a DeepSeek reinforcement learning framework, and the operation speed is increased through online-offline collaborative reasoning.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing of iron and steel metallurgy, and in particular to a method, device and program product for analyzing and controlling continuous casting inclusions. Background Art

[0002] The steel industry is facing stringent requirements from customers on the surface quality and mechanical properties of steel, and inclusions (such as alumina, protective slag, MnS segregation bands, etc.) are the core quality issues in continuous casting production.

[0003] Chinese patent CN118703736A uses multi-stage deoxidation and desulfurization process and parameter optimization to dynamically adjust the amount of deoxidizer added and stirring intensity, reduce the continuous casting flocculent rate, control the size of weathering steel inclusions below 5μm, and improve production efficiency by 15%. Chinese patent CN115846608A uses a laser rangefinder to monitor the nozzle offset in real time, and adjusts the continuous casting process parameters based on the model simulation results, reducing the slag roll rate caused by nozzle offset by 40%, the inclusion defect rate of the ingot by 25%, and the risk of steel leakage by 50%.

[0004] None of the above quality management modes have been modeled in-depth into the mechanism model to achieve inclusion control, and there are the following bottlenecks: 1. Data siloing: The control plan (CP) is scattered in the L1-L3 level system, and the process parameters and quality data lack dynamic linkage. 2. Model limitations: Traditional statistical methods cannot handle multivariate nonlinear relationships, such as the interaction of parameters such as superheat, aluminum content, and mean drawing speed. 3. Response delay: Abnormal handling relies on manual intervention, and the response cycle exceeds 30 seconds, which cannot block the spread of defects. Based on this, a solution that integrates multi-source data and intelligent algorithms is urgently needed to break through the bottleneck of traditional technology. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a method, device and program product for analyzing and controlling continuous casting inclusions. The present invention proposes an inclusion control method system integrating DeepSeek large model technology to perform offline analysis and online control of continuous casting inclusions. The core modules of the present invention include:

[0006] Multimodal data fusion platform: Integrate process data (including high-frequency signals), customer complaint lists and extreme operating condition data of converter, continuous casting and hot rolling processes to build a cross-process quality traceability link.

[0007] Offline analysis model: Based on the DeepSeek feature optimization engine and improved association rule algorithm, the key parameters and threshold ranges of inclusion formation are mined.

[0008] Online control model: Embedded with L2 quality judgment system, the DeepSeek lightweight inference engine is used to achieve millisecond-level dynamic optimization of process parameters.

[0009] Closed-loop feedback mechanism: Combining incremental learning technology, the model parameters are updated weekly to improve prediction accuracy and generalization ability.

[0010] Term description: The casting control system is usually divided into three levels: L1, L2, and L3, corresponding to different levels of functions and data types, specifically:

[0011] L1: Basic automation level. The corresponding data mainly comes from sensors and actuators of on-site equipment, including real-time process data such as molten steel temperature, flow rate, mold level, and straightening machine speed. These data are used to achieve real-time control and monitoring of continuous casting production equipment, ensuring the continuity and stability of the production process.

[0012] L2: Process control level. The corresponding data includes model calculation results, parameter setting values, quality tracking information, etc. of the casting machine. These parameters can control the execution of the L1-level system and can also trace and analyze product quality.

[0013] L3: Production management level. The corresponding data involves production plans, product coordination, and product tracking information. These parameters can reflect the production progress and status of products, track and adjust the plan arrangements, and coordinate the production rhythm and logistics connection between continuous casting and other production processes (such as steelmaking, refining, hot rolling, etc.).

[0014] In the first aspect, the present invention provides a method for analyzing and controlling inclusions in continuous casting, including the following steps:

[0015] Multi-source data preprocessing and feature screening, including data acquisition, data cleaning, abnormal data enhancement, feature screening, and weight calculation;

[0016] Offline analysis of association rules, including modeling time sliding window, using bidirectional long short-term memory network for data time series encoding, and strong association rule mining and dynamic optimization method based on support-confidence joint optimization method and rule dynamic update mechanism;

[0017] Online control and real-time optimization of continuous casting production line, including data acquisition, process calculation, and quality determination by embedding the DeepSeek lightweight inference engine, and also including real-time optimization control using the dynamic PID-DeepSeek hybrid control method.

[0018] As a further improvement of the present invention, the multi-source data preprocessing and feature screening includes the following steps:

[0019] Multi-source data acquisition, including acquiring data at different levels in the casting control system, and also including acquiring data outside the casting control system;

[0020] Data cleaning, including imputing missing values, standardizing, and removing outliers from the acquired data;

[0021] Abnormal data enhancement, including generating samples under extreme working conditions;

[0022] Feature screening and weight calculation, including calculating feature weights through a multi-head attention layer and performing parameter screening based on the feature weights.

[0023] As a further improvement of the present invention, the data cleaning includes the following steps:

[0024] Missing value imputation: Using the moving window mean method for missing value imputation;

[0025] Standardization: Using the Z-score method to eliminate the dimensional difference;

[0026] Outlier removal: Removing data beyond the mean ± 3 times the standard deviation.

[0027] As a further improvement of the present invention, the abnormal data enhancement includes: Using a generative adversarial network to generate samples under extreme working conditions.

[0028] As a further improvement of the present invention, the modeling time sliding window includes the following steps:

[0029] By defining the time window length and the sliding step, dividing the continuous data stream into overlapping windows;

[0030] For the window edge data, using weighted smoothing to reduce the truncation effect;

[0031] Through interpolation processing, aligning heterogeneous data to a unified timestamp;

[0032] Setting the lag time between parameters according to process knowledge and adjusting the data alignment within the window.

[0033] As a further improvement of the present invention, the data time series encoding, its encoder architecture includes:

[0034] Input layer: The parameter matrix within the window, which includes the numerical values of any parameter at any moment within the window;

[0035] Bidirectional LSTM layer: That is, the bidirectional long short-term memory network layer, capturing the forward and reverse temporal dependencies in the sequence data;

[0036] Self-attention mechanism: Calculating the weights of the time steps within the window and adaptively capturing the dependencies between different time steps within the window.

[0037] As a further improvement of the present invention, the data time series encoding includes encoding output and rule mining, specifically including:

[0038] Feature vector generation: Concatenate the window encoding result with the original parameter statistics to obtain an enhanced feature vector; the parameter statistics include the mean, variance, extreme values, etc.

[0039] Association rule mapping: Use a fully connected layer to map the enhanced feature vector to an item set probability distribution.

[0040] As a further improvement of the present invention, the support-confidence joint optimization method includes:

[0041] Determine the window stability according to the fluctuation standard deviation of relevant parameters, and adjust the minimum support according to the window stability;

[0042] Based on the Granger causality test, correct the rule confidence:

[0043] ;

[0044] Among them,

[0045] Confidence corrected : The corrected confidence;

[0046] Confidence: The original confidence, ;

[0047] CausalStrength: The F statistic of the Granger causality test.

[0048] As a further improvement of the present invention, the rule dynamic update mechanism includes:

[0049] Incremental learning framework: For each newly added window, adopt a sliding window-based incremental update method to dynamically adjust the support;

[0050] Confidence decay model: Introduce a time decay factor and adopt a weighted moving average method to dynamically adjust the confidence of the rule;

[0051] Decision tree model training and optimization, including setting splitting criteria, setting pruning strategies, and performing dynamic identification of sensitive intervals.

[0052] As a further improvement of the present invention, the decision tree model training and optimization includes:

[0053] Based on the Gini coefficient and DeepSeek feature importance weighting, set the splitting criteria;

[0054] Adopt cost complexity pruning to remove branches with contribution less than the set standard;

[0055] Dynamic identification of sensitive intervals, based on the Q-Learning reinforcement learning model of DeepSeek, construct a reward function:

[0056] ;

[0057] Among them, R: reward function; s: state feature set; a: parameter adjustment action; λ: trade-off coefficient;

[0058] Through training, find the strategy that maximizes the reward function value.

[0059] As a further improvement of the present invention, the on-line control and real-time optimization of the continuous casting production line includes the following integrations:

[0060] Data pipeline docking: Obtain the L1 data stream in real time through the OPC UA protocol;

[0061] Model inference service: Deploy a decision tree model accelerated by TensorRT on the L2 server;

[0062] Issuance of control instructions: Link the L3 to execute parameter adjustment.

[0063] As a further improvement of the present invention, the DeepSeek lightweight inference engine includes:

[0064] Model quantization and compression: Convert the decision tree model from FP32 to FP16 precision, and remove the branches with a contribution degree <0.5% in the decision tree;

[0065] TensorRT accelerated deployment: Build a decision tree model inference engine based on the Python API of TensorRT, and optimize feature calculation and rule matching in combination with CUDA parallel computing.

[0066] As a further improvement of the present invention, the real-time optimization control includes:

[0067] Data acquisition: Subscribe to the L1 data stream through the OPC UA unified architecture, and cache the recent short-term data in a circular buffer;

[0068] Feature calculation: Parallelly calculate statistics such as mean, variance, and trend slope;

[0069] Model inference: Input the feature vector into the TensorRT engine, and output the risk level and optimization suggestions;

[0070] Instruction issuance: Trigger control instructions according to the risk level.

[0071] As a further improvement of the present invention, the dynamic PID-DeepSeek hybrid control method includes:

[0072] Use the superheat PID controller as the basic control;

[0073] Generate the superheat deviation compensation value through LSTM time series prediction;

[0074] Superimpose the superheat deviation compensation value on the basis of the basic control to obtain the dynamic PID-DeepSeek hybrid control output.

[0075] As a further improvement of the present invention, the triggering of the control instruction according to the risk level includes an abnormal blocking logic,

[0076] When the risk level reaches the warning level, trigger an alarm instruction;

[0077] When the risk level reaches the severe level, trigger the speed reduction and waste cutting instructions.

[0078] In a second aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of the method in the first aspect.

[0079] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the method in the first aspect.

[0080] In a fourth aspect, the present invention provides a computer program product, and when the computer program is executed by a processor, it implements the steps of the method in the first aspect.

[0081] Compared with the prior art, the present invention has the following obvious technical improvement effects:

[0082] (1) Introduce the DeepSeek feature optimization engine: Integrate the attention mechanism to dynamically extract high-order features in multi-source data, replacing traditional SHAP values and mutual information analysis.

[0083] (2) Multivariable dynamic sensitive interval identification: Based on the DeepSeek reinforcement learning (RL) framework, realize the adaptive optimization of the parameter sensitive interval. For example, the superheat dynamic interval is adjusted to 12-38°C (original 15-35°C).

[0084] (3) Online-offline collaborative reasoning: The offline model generates a rule base, and the online model realizes a response within 5 seconds through the DeepSeek TensorRT acceleration engine, with an 80% speed increase compared to traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a flowchart of a method for analyzing and controlling inclusions in continuous casting disclosed by the present invention.

[0086] Figure 2 It is an architecture diagram of a quality determination system disclosed by the present invention. Specific Embodiments

[0087] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the present invention will, in conjunction with the accompanying drawings, clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. Among them, steps S1, S2,... in the embodiments described in the present invention do not limit the only execution steps of the present invention; various models, simulation environments, and software described in the present invention are not the only limiting means of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0088] In the present invention, a computer device / equipment refers to relevant entities applied to a computer, such as hardware, a combination of hardware and software, software, or software in execution, etc. Specifically, for example, software includes, but is not limited to, processes running on a processor, processors, objects, executable software, execution threads, programs, and / or computers. Also, application programs or script programs running on a server, and the server can both be software. One or more software can be in the process and / or thread of execution, and the software can be localized on one computer and / or distributed between two or more computers, and can be run by various computer-readable media.

[0089] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0090] In a first aspect, the present invention provides an embodiment of a continuous casting inclusion analysis and control method, as Figure 1 shown, the specific process can be as follows:

[0091] S1: Multi-source data preprocessing and feature screening.

[0092] S11: Multi-source data acquisition. Exemplarily, acquiring different types of data includes:

[0093] L1-level high-frequency data: average casting speed (1Hz), standard deviation of liquid level fluctuation, mold water flow rate (N / S side).

[0094] L2 / L3-level data: superheat, pure degassing time of RH, total oxygen content of RHOB, aluminum content, coil thickness.

[0095] External data: customer complaint list (HD / HQ block record), extreme working conditions (mold water flow rate > 530 m³ / h).

[0096] S12: Data cleaning.

[0097] In an embodiment of the present invention, the data cleaning process includes:

[0098] (1) Missing value imputation: The moving window mean method (window = 5 minutes) is adopted, and the formula is:

[0099]

[0100] Where,

[0101] : The calculation result of imputing the missing value;

[0102] k: The number of data points within the moving window. In this embodiment, k = 5;

[0103] t: Represents the time point at which the missing value is to be imputed currently;

[0104] : The data value at the i-th time point before the current time point t.

[0105] (2) Standardization: The Z-score method is adopted to eliminate the dimensional difference, and the formula is:

[0106]

[0107] Where,

[0108] Z: The data value after Z-score standardization;

[0109] X: The data value that needs to be standardized;

[0110] μ: The average value of all data values in the data set;

[0111] σ: The standard deviation of the data set.

[0112] The data after standardization follows a standard normal distribution with a mean of 0 and a standard deviation of 1, preparing for the next data processing.

[0113] (3) Outlier removal: Based on the 3σ principle, the data exceeding the mean ± 3 times the standard deviation is removed.

[0114] S13: Abnormal data enhancement.

[0115] Based on the DeepSeek synthetic data technology, a generative adversarial network (GAN) is used to generate extreme working condition samples to solve the data imbalance problem. In an embodiment of the present invention, the following method is adopted:

[0116] (1) Generator loss function:

[0117]

[0118] Where,

[0119] L G : Generator loss value;

[0120] : Expectation of subsequent calculations on z sampled from the noise distribution p z in;

[0121] G(z): Data generated by the generator G with noise z as input;

[0122] D(G(z)): Probability that the discriminator judges the generated data as real data;

[0123] (2) Discriminator loss function:

[0124]

[0125] Among them,

[0126] L D : Discriminator loss value;

[0127] : Expectation of subsequent calculations on X sampled from the real data distribution;

[0128] D(X): Probability that the discriminator considers the input data X as real data;

[0129] The remaining symbols are the same as those of the generator loss function.

[0130] S14: Feature screening and weight calculation.

[0131] DeepSeek attention mechanism: Calculate feature weights through the multi-head attention layer, formula:

[0132]

[0133] Among them,

[0134] Q: Query matrix;

[0135] K: Key matrix;

[0136] V: Value matrix;

[0137] d k : Dimension of the key matrix K;

[0138] softmax: Activation function.

[0139] In an embodiment of the present invention, through weight calculation, the top 3 key parameters are: liquid level fluctuation (weight 4.2), superheat degree (weight 3.8), aluminum content (weight 2.1).

[0140] S2: Offline analysis of association rules.

[0141] For association rule mining, the present invention proposes an improved Apriori-Temporal algorithm, and the specific improvement is as follows: introducing a time sliding window (5 minutes) and DeepSeek time series encoding to mine strong association rules.

[0142] The improvement mechanism is described as follows:

[0143] The limitation of the traditional Apriori algorithm lies in that: this algorithm generates association rules by searching for frequent item sets layer by layer, but its core assumption is that the data is independent and identically distributed (i.i.d.), ignoring the temporal correlation of process parameters.

[0144] For example: there is a strong coupling between the average continuous casting speed and the liquid level fluctuation in time (such as the liquid level fluctuation reaches the peak 10 seconds after the casting speed mutation); the parameter distributions in different production stages (starting casting, steady state, end slab) are significantly different, and mixed analysis will introduce noise rules.

[0145] Based on the above limitations and combined with the application requirements of the existing industrial scenarios, the improved algorithm proposed by this method aims to solve the following problems:

[0146] (1) Rule redundancy: Among the rules generated by the traditional method, 30% have a falsely high confidence due to time mismatch;

[0147] (2) Dynamic adaptability: The rule threshold needs to be adjusted in real time for process parameter fluctuations (such as superheat ±5°C);

[0148] (3) Computational efficiency: When processing 100,000 high-frequency data per hour, the rule update needs to be completed within a short time (exemplarily, within 5 minutes).

[0149] The specific implementation method of this step includes the following steps:

[0150] S21: Mathematical modeling and implementation of the time sliding window.

[0151] S211: Window segmentation strategy, including:

[0152] (1) Window definition: Divide the continuous data stream into overlapping windows,

[0153]

[0154] Among them,

[0155] T w : Time window length, exemplarily, T w = 5 minutes;

[0156] S: Sliding step, exemplarily, S = 1 minute.

[0157] This design can capture the gradual change process of process parameters (such as the liquid level fluctuation delay caused by the slow change of drawing speed).

[0158] (2) Edge data processing: For window edge data (such as the first minute and the last minute), weighted smoothing (EWMA) is used to reduce the truncation effect:

[0159]

[0160] Among them,

[0161] X edge : The window edge data value after weighted smoothing processing;

[0162] α: Smoothing coefficient. Exemplarily, α = 0.8;

[0163] : The data value at the starting moment of the window;

[0164] : The data value at the ending moment of the window.

[0165] S212: Parameter alignment within the window, including:

[0166] (1) Multi-source data synchronization:

[0167] For heterogeneous data such as the average drawing speed (1Hz), liquid level fluctuation (1Hz), and coil thickness (0.2Hz), using the slab number as an index, interpolate and align them to a unified timestamp:

[0168]

[0169] Among them,

[0170] : The heterogeneous data value aligned after interpolation processing at timestamp t;

[0171] LinearInterpolation: Linear interpolation function;

[0172] x(t 1 ), x(t 2 ): The original heterogeneous data values at timestamps t 1 and t 2 .

[0173] (2) Time series lag compensation:

[0174] Set the lag time between parameters according to process knowledge and adjust the data alignment within the window:

[0175]

[0176] Among them,

[0177] : Parameter lag time.

[0178] Exemplarily, according to the change of the drawing speed lagging behind the liquid level fluctuation, set this parameter's = 10 seconds.

[0179] S22: Data time series encoding.

[0180] S221: Encoder architecture, including:

[0181] (1) Input layer: Parameter matrix within the window:

[0182]

[0183] Among them,

[0184] X: Parameter matrix within the window;

[0185] T w : Time window length, exemplarily, T w = 300;

[0186] d: Parameter dimension.

[0187] (2) Bidirectional LSTM layer, that is, using a bidirectional long short-term memory network, combining forward and reverse time series information processing to more comprehensively capture the time series dependencies in the sequence data:

[0188]

[0189] Among them,

[0190] : Hidden state of the forward LSTM layer at time t;

[0191] : Hidden state of the reverse LSTM layer at time t;

[0192] : Sequence data input to the LSTM layer at time t.

[0193] (3) Self-attention mechanism: Calculate the weights of time steps within the window, adaptively capture the dependencies between different time steps within the window, so that the model can more flexibly focus on important information when processing sequence data:

[0194]

[0195] Among them,

[0196] α t : Attention weight at time step t;

[0197] softmax: Activation function;

[0198] W q and W k : Learnable parameter matrix;

[0199] h t : Hidden state vector at time step t;

[0200] H = [h 1 , h 2 , …, h Tw ;

[0201] Z: Feature vector obtained after processing by the self-attention mechanism.

[0202] S222: Encoding output and rule mining.

[0203] (1) Feature vector generation: Concatenate the window encoding result with the original parameter statistics (mean, variance, extreme values) to obtain an enhanced feature vector:

[0204]

[0205] (2) Association rule mapping: Use a fully connected layer to map Z enhanced to an itemset probability distribution:

[0206]

[0207] where,

[0208] Wf: Learnable weight matrix, ;

[0209] I: All possible item sets;

[0210] sigmoid: Activation function;

[0211] bf: Learnable bias vector.

[0212] S23: Strong association rule mining and dynamic optimization.

[0213] S231: Support-confidence joint optimization.

[0214] In the present invention, support is used to measure the frequency of occurrence of an item set in a dataset.

[0215] (1) Set the dynamic support threshold: Adjust the minimum support according to the window stability (fluctuation standard deviation σ):

[0216]

[0217] Among them,

[0218] support min : Minimum support;

[0219] σ: Standard deviation of fluctuations of relevant parameters covered by the window.

[0220] (2) Causal confidence correction: Based on Granger causality test, correct the confidence of the rule:

[0221]

[0222] Among them,

[0223] Confidence corrected : Corrected confidence;

[0224] Confidence: Original confidence, ;

[0225] CausalStrength: F statistic of Granger causality test.

[0226] This step proposes a confidence correction method based on causal strength, combining the traditional association rule confidence Confidence with the results of Granger causality test. By using the exponential function processing: Reduce the importance of rules with high confidence but lacking causal support; Retain rules with both high confidence and strong causal support; Amplify the confidence of rules lacking causal support to make them easier to identify.

[0227] S232: Rule dynamic update mechanism.

[0228] S2321: Incremental learning framework: Every time a new window Window k+1 is added, update the rule base:

[0229]

[0230] Among them,

[0231] : Support of item set X updated after adding the new window Window k+1 ;

[0232] k: Number of windows existing before adding the new window Window k+1 ;

[0233] : Support calculated based on the previous k windows;

[0234] : Itemset X in a new window Window k+1 The support in .

[0235] S2322: Confidence decay model: Introduce a time decay factor to dynamically adjust the confidence of old rules so that the rules can better adapt to data changes.

[0236]

[0237] in,

[0238] : The confidence of the rule after attenuation adjustment;

[0239] : Rule confidence before attenuation adjustment;

[0240] : The confidence calculated based on the latest data at the current moment;

[0241] λ: time decay factor, illustratively, λ=0.9.

[0242] Rule example: Under the conditions of {average pulling speed>9.7, coil thickness<3.5}, the probability of slag inclusion increases by 82%; the support degree of this rule is 0.25 and the confidence degree is 0.85.

[0243] S2323: Decision tree model training and optimization.

[0244] (1) DeepSeek-CART algorithm:

[0245] Splitting criteria: based on the Gini coefficient and DeepSeek feature importance weighting,

[0246]

[0247] in,

[0248] W i : DeepSeek weight;

[0249] : Gini coefficient;

[0250] p ik : The proportion of samples of the kth category in the node.

[0251] (2) Pruning strategy: Simplify the model structure and improve the model’s generalization ability on new data.

[0252] Exemplarily, cost complexity pruning (CCP) with α=0.005 is used to remove branches with contribution <0.5%.

[0253] In one embodiment of the present invention, the following performance indicators are achieved through training and optimization of the decision tree model: when the decision tree depth is 8, the inclusion class F1-score reaches 0.62 (precision 0.96, recall rate 0.45).

[0254] (3) Dynamic identification of sensitive areas.

[0255] DeepSeek reinforcement learning framework: Construct a Q-Learning model to optimize parameter thresholds, and the reward function is designed as:

[0256]

[0257] in,

[0258] R: reward function;

[0259] s: state feature set; exemplary, including production process parameters, production history information, equipment status information, etc.;

[0260] a: parameter adjustment action; exemplary, including various parameter threshold adjustments caused by process adjustment;

[0261] λ: trade-off coefficient, the optimal strategy is obtained through training.

[0262] Through training, we find a strategy that maximizes the reward function value: a parameter threshold adjustment method that can effectively reduce the probability of inclusions and reasonably control the cost of process adjustment.

[0263] In one embodiment of the present invention, the dynamic range output obtained through the above training is: the superheat sensitive range is adaptively adjusted to 12-38°C (originally 15-35°C), and the slag inclusion probability is reduced by 18%.

[0264] S3: Online control model deployment and real-time optimization.

[0265] S31: L2 quality determination system architecture and integration solution.

[0266] like Figure 2 As shown in the figure, the L2 (Level 2) system is the core control system of the continuous casting production line, responsible for process parameter setting, quality judgment and equipment linkage. Its core modules include:

[0267] Data acquisition layer: receives high-frequency sensor data (1Hz~100Hz) from L1 (basic automation);

[0268] Process calculation layer: Generate set values ​​based on preset rule base (such as superheat PID control);

[0269] Quality judgment layer: detect defects in real time and trigger alarms / disposal instructions.

[0270] Embed the DeepSeek lightweight inference engine into the L2 system to achieve the following integration:

[0271] (1) Data pipeline docking: Obtain the L1 data stream (such as average drawing speed, liquid level fluctuation, aluminum content, etc.) in real time through the OPC UA protocol.

[0272] (2) Model inference service: Deploy a decision tree model accelerated by TensorRT on the L2 server. In an embodiment of the present invention, the response time ≤ 5 ms.

[0273] (3) Control instruction issuance: Link L3 (MES system) to execute parameter adjustment (such as drawing speed reduction, tundish heating power adjustment).

[0274] S32: Design of the DeepSeek lightweight inference engine.

[0275] S321: Model quantization and compression. The specific methods include:

[0276] (1) FP16 quantization: Convert the decision tree model from FP32 to FP16 precision, reducing memory occupancy by 50% and doubling the inference speed:

[0277] MemoryFP16 = 0.5 × MemoryFP32

[0278] LatencyFP16 = 0.5 × LatencyFP32

[0279] (2) Pruning optimization: Remove branches with a contribution degree < 0.5% in the decision tree (such as rare working conditions with superheat < 10 °C), reducing the model size by 30%.

[0280] S322: TensorRT accelerated deployment.

[0281] (1) Engine construction: Use the Python API of TensorRT to convert the decision tree model into a highly optimized inference engine;

[0282] (2) CUDA core optimization: Utilize parallel computing to accelerate feature calculation and rule matching; In an embodiment of the present invention, the single inference time consumption ≤ 3 ms.

[0283] The performance indicators achieved by the lightweight engine design in step S32 are as follows in the table:

[0284]

[0285] S33: Real-time optimization control strategy.

[0286] S331: Millisecond-level data processing pipeline.

[0287] Data acquisition: Subscribe to the L1 data stream through the OPC UA unified architecture and cache the data of the most recent 5 seconds (5000 sampling points) in a circular buffer.

[0288] Feature calculation: Calculate statistics (mean, variance, trend slope) in parallel, with a time consumption of ≤1 ms.

[0289] Model inference: Input the feature vector into the TensorRT engine, and output the risk level and optimization suggestions, with a time consumption of ≤3 ms.

[0290] Instruction issuance: Trigger control instructions (such as speed reduction, heating power adjustment) according to the risk level, with a time consumption of ≤1 ms.

[0291] S332: Dynamic PID-DeepSeek hybrid control.

[0292] (1) Use the superheat PID controller as the basic control, and the formula is:

[0293]

[0294] Among them,

[0295] : Proportional term, where is the error at time t, and K p is the proportional coefficient;

[0296] : Integral term, where is the integral of the error from the initial time to the current time t, and K i is the integral coefficient;

[0297] : Differential term, where is the rate of change of the error with time, and K d is the differential coefficient.

[0298] In an embodiment of the present invention, the proportional coefficient K p is set to 0.8, the integral coefficient K i is set to 0.2, and the differential coefficient K d is set to 0.05.

[0299] (2) DeepSeek compensation amount: The superheat deviation compensation value Δu predicted by the model is generated through LSTM time series prediction:

[0300]

[0301] Among them,

[0302] Δu(t): The superheat deviation compensation value at time t output by the long short-term memory network;

[0303] LSTM: Long Short-Term Memory Network, a neural network that is good at processing time series data;

[0304] : Input data series related to superheat.

[0305] Combining the above two items (1) and (2), we get the hybrid control output:

[0306]

[0307] in,

[0308] final(t): the final control output at time t;

[0309] u PID (t): output value of PID controller at time t;

[0310] : The superheat deviation compensation value output by LSTM at time t.

[0311] S333: Abnormal blocking logic.

[0312] Risk grading: The risk level is determined based on the probability of abnormalities in the production status obtained through analysis, as well as other specific indicators. For example, risk determination and related operations are shown in the following table:

[0313]

[0314] Instruction priority: Severe risk instructions are interrupted with low priority control to ensure response time ≤ 5ms.

[0315] The technical effects of the present invention include:

[0316] (1) Accurate prediction capability: The accuracy of inclusion detection is improved to 96% (F1-score 0.62), and the false alarm rate is reduced to 4%.

[0317] (2) Real-time control efficiency: Online model response time is ≤ 3 seconds, and process parameter adjustment accuracy is improved by 30%.

[0318] (3) Economic benefits: Annual quality losses reduced by more than RMB 6 million, and customer satisfaction increased by 40%.

[0319] The specific implementation data of the present invention are as follows:

[0320] (1) In a certain continuous casting production line embodiment of the present invention:

[0321] Input data: 100,000 high-frequency data (average pulling speed, liquid level fluctuation, steel coil thickness).

[0322] Rule generation: One of the rules obtained through the analysis of the input data is:

[0323] Under the condition of {average casting speed > 97 mm / min, coil thickness < 3.5 mm}, the slag inclusion probability increases by 82%; based on the input data, the support degree of this rule is 0.25, and the confidence degree is 0.85.

[0324] The running effects of the offline analysis model proposed by the present invention and the traditional model are compared as follows in the table:

[0325]

[0326] (2) Example of the treatment of extreme working conditions in a certain iron and steel base by the present invention:

[0327] Abnormal data enhancement: For the extreme working condition of mold water flow > 530 m³ / h, synthetic data is generated and the rule is updated: Under the condition of {water flow > 530 m³ / h}, the HD blocking probability increases by 65%; the support degree of this rule is 0.18, and the confidence degree is 0.82.

[0328] Blocking effect: The blocking rate of HD-type defects is increased from 80% to 95%, and the annual waste reduction is 500 tons.

[0329] In the second aspect, the present invention provides an example of a computer device, including a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of the method described in the first aspect.

[0330] In the third aspect, the present invention provides an example of a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0331] In the fourth aspect, the present invention provides an example of a computer program product, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

Claims

1. A method for analyzing and controlling continuous casting inclusions, characterized in that: The following steps are involved: Multi-source data preprocessing and feature screening, including data acquisition, data cleaning, abnormal data enhancement, feature screening and weight calculation; Offline analysis of association rules, including modeling time sliding windows, using bidirectional long short-term memory networks for data temporal encoding, and strong association rule mining and dynamic optimization methods based on support-confidence joint optimization methods and rule dynamic update mechanisms; Online control and real-time optimization of continuous casting production lines, including data collection, process calculation, and quality determination embedded with the DeepSeek lightweight inference engine, as well as real-time optimization control using a dynamic PID-DeepSeek hybrid control method.

2. The method according to claim 1, characterized in that The multi-source data preprocessing and feature screening includes the following steps: Multi-source data acquisition, including acquisition of data at different levels in the casting control system, and also acquisition of data outside the casting control system; Data cleaning, including missing value interpolation, standardization, and outlier removal of acquired data; Abnormal data enhancement, including generation of extreme condition samples; Feature screening and weight calculation, including calculating feature weights through multi-head attention layers and performing parameter screening based on feature weights.

3. The method according to claim 2, characterized in that The data cleaning, The following steps are involved: Missing value interpolation: Sliding window mean method is used to interpolate missing values; Standardization: Z-score method is used to eliminate dimensional differences; Outlier removal: data exceeding the mean ± 3 times the standard deviation are removed.

4. The method according to claim 2, characterized in that: The abnormal data enhancement includes: using a generative adversarial network to generate extreme working condition samples.

5. The method according to claim 1, characterized in that The modeling time sliding window includes the following steps: Split the continuous data stream into overlapping windows by defining the time window length and sliding step size; For window edge data, weighted smoothing is used to reduce truncation effects; Align heterogeneous data to a unified timestamp through interpolation processing; Set the lag time between parameters based on process knowledge and adjust the data alignment within the window.

6. The method according to claim 1, characterized in that The data timing coding, the encoder architecture includes: Input layer: the parameter matrix within the window, which includes the value of any parameter at any time within the window; Bidirectional LSTM layer: a bidirectional long short-term memory network layer that captures the forward and reverse temporal dependencies in the sequence data; Self-attention mechanism: Calculate the weights of the time steps within the window and adaptively capture the dependencies between different time steps within the window.

7. The method according to claim 1, characterized in that The data time series coding includes coding output and rule mining, specifically including: Feature vector generation: concatenate the window encoding result with the original parameter statistics to obtain an enhanced feature vector; the parameter statistics include mean, variance, and extreme value; Association rule mapping: Use a fully connected layer to map the enhanced feature vector to the item set probability distribution.

8. The method according to claim 1, characterized in that The support-confidence joint optimization method comprises: Determine the window stability based on the standard deviation of fluctuations of relevant parameters, and adjust the minimum support based on the window stability; Based on the Granger causality test, the rule confidence is modified: ； in, Confidence corrected : Corrected confidence; Confidence: original confidence, ; CausalStrength: F statistic for Granger causality test.

9. The method according to claim 1, characterized in that: The rule dynamic update mechanism includes: Incremental learning framework: Every time a window is added, a sliding window-style incremental update method is used to dynamically adjust the support; Confidence decay model: introduces the time decay factor and uses the weighted moving average method to dynamically adjust the confidence of the rule; Decision tree model training and optimization, including setting splitting criteria, setting pruning strategies, and dynamically identifying sensitive intervals.

10. The method according to claim 9, characterized in that The decision tree model training and optimization includes: Set the splitting criteria based on the Gini coefficient and DeepSeek feature importance weighting; Use cost complexity pruning to remove branches whose contribution is less than the set standard; Dynamic identification of sensitive areas, based on DeepSeek's Q-Learning reinforcement learning model, to build a reward function: ； Where R: reward function; s: state feature set; a: parameter adjustment action; λ: trade-off coefficient; Through training, a strategy is found that maximizes the value of the reward function.

11. The method according to claim 1, characterized in that: The online control and real-time optimization of the continuous casting production line includes the following integration: Data pipeline connection: obtain L1 data stream in real time through OPC UA protocol; Model inference service: deploy TensorRT-accelerated decision tree models on L2 servers; Control instructions are issued: Link L3 to perform parameter adjustment.

12. The method according to claim 1, characterized in that The DeepSeek lightweight inference engine includes: Model quantization and compression: Convert the decision tree model from FP32 to FP16 precision and remove branches with contribution < 0.5% in the decision tree; TensorRT accelerated deployment: Build a decision tree model inference engine based on TensorRT's Python API, and combine CUDA parallel computing methods to optimize feature calculation and rule matching.

13. The method according to claim 1, characterized in that The real-time optimization control includes: Data collection: Subscribe to L1 data streams through the OPC UA unified architecture and use a ring buffer to cache the most recent short-term data; Feature calculation: parallel calculation of statistics, including mean, variance, and trend slope; Model reasoning: Input feature vectors to the TensorRT engine and output risk levels and optimization suggestions; Instruction issuance: trigger control instructions based on risk level.

14. The method according to claim 1, characterized in that The dynamic PID-DeepSeek hybrid control method comprises: Use superheat PID controller as basic control; Generate superheat deviation compensation value through LSTM time series prediction; The superheat deviation compensation value is superimposed on the basic control to obtain the dynamic PID-DeepSeek hybrid control output.

15. The method according to claim 13, characterized in that The control instructions are triggered according to the risk level, including abnormal blocking logic, When the risk level reaches the warning level, the alarm command is triggered; When the risk level reaches a serious level, speed reduction and waste elimination instructions are triggered.

16. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1-15.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 15 are implemented.

18. A computer program product, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 15 are implemented.

Images