A continuous casting inclusion analysis and control method, device and program product

Abstract

The present invention provides a method, device and program product for analyzing and controlling inclusions in continuous casting, which relates to the technical field of intelligent manufacturing of iron and steel metallurgy. The analysis and control method includes: 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 window, using bidirectional long short-term memory network for data time series encoding, strong association rule mining and dynamic optimization method based on support-confidence joint optimization method and rule dynamic update mechanism; online control and real-time optimization of continuous casting production line, including data collection, process calculation and quality judgment embedded in DeepSeek lightweight inference engine, and also including real-time optimization control using dynamic PID-DeepSeek hybrid control method. The present invention introduces DeepSeek feature optimization engine, based on DeepSeek reinforcement learning framework, to achieve adaptive optimization of parameter sensitive intervals, and improves running speed 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 customer requirements for steel surface quality and mechanical properties, and inclusions (such as alumina, protective slag, MnS segregation bands, etc.) are the core quality issues in continuous casting production.

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

[0004] None of the aforementioned quality management models delve into mechanistic modeling to achieve inclusion control, resulting in the following bottlenecks: 1. Data silos: Control plans (CPs) are dispersed across L1-L3 systems, lacking dynamic linkage between process parameters and quality data. 2. Model limitations: Traditional statistical methods cannot handle multivariate nonlinear relationships, such as the interaction of parameters like superheat, aluminum content, and mean casting speed. 3. Response delays: Abnormal handling relies on manual intervention, with response cycles exceeding 30 seconds, making it impossible to prevent defect propagation. Consequently, a solution that integrates multi-source data with intelligent algorithms is urgently needed to overcome the bottlenecks of traditional technologies. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention proposes a continuous casting inclusion analysis and control method, device and program product. The present invention proposes an inclusion control method system that integrates 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: Integrates process data (including high-frequency signals), customer complaint lists, and extreme operating condition data from converters, 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, it mines the key parameters and threshold ranges of inclusion formation.

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

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

[0010] Terminology: 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. Data primarily comes from sensors and actuators in on-site equipment, including real-time process data such as molten steel temperature, flow rate, mold level, and straightener speed. This data is used to achieve real-time control and monitoring of continuous casting equipment, ensuring the continuity and stability of the production process.

[0012] L2: Process control level. Corresponding data includes casting machine model calculation results, parameter settings, quality tracking information, etc. These parameters control the execution of the L1 system and enable traceability and analysis of product quality.

[0013] L3: Production management level. This level includes data related to production planning, product coordination, and product tracking. These parameters reflect the production progress and status of products, track and adjust plans, and coordinate the production rhythm and logistics connection between continuous casting and other production processes (such as steelmaking, refining, and hot rolling).

[0014] In a first aspect, the present invention provides a method for analyzing and controlling continuous casting inclusions, comprising 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 windows, using bidirectional long short-term memory networks for data temporal encoding, and strong association rule mining and dynamic optimization methods based on a support-confidence joint optimization method and a dynamic rule update mechanism;

[0017] Online control and real-time optimization of continuous casting production lines, including data acquisition, process calculation, and quality determination embedded in the DeepSeek lightweight inference engine, as well as real-time optimization control using a 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 acquisition of data at different levels within the casting control system, as well as acquisition of data outside the casting control system;

[0020] Data cleaning, including missing value interpolation, standardization, and outlier removal of acquired data;

[0021] Abnormal data enhancement, including generating extreme operating condition samples;

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

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

[0024] Missing value interpolation: Sliding window mean method is used to interpolate missing values;

[0025] Standardization: Z-score method was used to eliminate dimensional differences;

[0026] Outlier removal: data exceeding the mean ± 3 times the standard deviation are removed.

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

[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 sliding step size, the continuous data stream is divided into overlapping windows;

[0030] For window edge data, weighted smoothing is used to reduce truncation effects;

[0031] Align heterogeneous data to a unified timestamp through interpolation processing;

[0032] Set the lag time between parameters based on process knowledge and adjust the data alignment within the window.

[0033] As a further improvement of the present invention, the data timing coding, its encoder architecture includes:

[0034] Input layer: the parameter matrix within the window, which includes the value of any parameter at any time within the window;

[0035] Bidirectional LSTM layer: a bidirectional long short-term memory network layer that captures the forward and reverse temporal dependencies in sequence data;

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

[0037] As a further improvement of the present invention, the data time series coding includes coding 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 mean, variance, extreme value, etc.

[0039] Association rule mapping: Use a fully connected layer to map the enhanced feature vector into 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 based on the standard deviation of fluctuations of relevant parameters, and adjust the minimum support based on the window stability;

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

[0043] ;

[0044] in,

[0045] Confidence corrected : Corrected confidence;

[0046] Confidence: original confidence, ;

[0047] CausalStrength: F statistic for Granger causality test.

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

[0049] Incremental learning framework: Each time a window is added, a sliding window-style incremental update method is used to dynamically adjust the support;

[0050] Confidence decay model: introduces a time decay factor and uses 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 dynamic identification of sensitive intervals.

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

[0053] Set the splitting criterion based on the Gini coefficient and DeepSeek feature importance weighting;

[0054] Use cost complexity pruning to remove branches whose contribution is less than the set standard;

[0055] Dynamic identification of sensitive areas, based on DeepSeek's Q-Learning reinforcement learning model, to build a reward function:

[0056] ;

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

[0058] Through training, a strategy is found that maximizes the value of the reward function.

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

[0060] Data pipeline connection: Real-time acquisition of L1 data streams via the OPC UA protocol;

[0061] Model Inference Service: Deploy TensorRT-accelerated decision tree models on L2 servers;

[0062] Control instructions are issued: Link L3 to perform 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 branches with contribution less than 0.5% in the decision tree;

[0065] TensorRT accelerated deployment: Build a decision tree model inference engine based on TensorRT's Python API, and combine CUDA parallel computing to optimize feature calculation and rule matching.

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

[0067] Data collection: Through the OPC UA unified architecture, subscribe to the L1 data stream and cache the latest short-term data in a ring buffer;

[0068] Feature calculation: parallel calculation of statistics such as mean, variance, trend slope, etc.

[0069] Model inference: Input feature vectors to the TensorRT engine and output risk levels and optimization suggestions;

[0070] Instruction issuance: trigger control instructions based on risk level.

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

[0072] Use superheat PID controller as basic control;

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

[0074] The superheat deviation compensation value is superimposed on the basic control to obtain the dynamic PID-DeepSeek hybrid control output.

[0075] As a further improvement of the present invention, the control instruction is triggered according to the risk level, including abnormal blocking logic,

[0076] When the risk level reaches the warning level, the alarm instruction is triggered;

[0077] When the risk level reaches a serious level, speed reduction and waste elimination instructions are triggered.

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

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

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

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

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

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

[0084] (3) Online-offline collaborative reasoning: The offline model generates a rule base, and the online model responds within 5 seconds through the DeepSeek TensorRT acceleration engine, which is 80% faster than traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 The present invention discloses a flow chart of a method for analyzing and controlling continuous casting inclusions.

[0086] Figure 2 This is an architecture diagram of the quality determination system disclosed in the present invention. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the present invention will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, wherein steps S1, S2... in the embodiments described in the present invention do not limit the only execution steps of the present invention; the various models, simulation environments, and software described in the present invention are not the only way to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0088] In the present invention, a computer device / equipment refers to a related entity applied to a computer, such as hardware, a combination of hardware and software, software, or software in execution. Specifically, for example, software includes, but is not limited to, a process running on a processor, a processor, an object, executable software, an execution thread, a program, and / or a computer. Furthermore, an application or script running on a server, or a server, can also be software. One or more software programs can be in an execution process and / or thread, and software can be localized on a single computer and / or distributed between two or more computers, and can be executed from various computer-readable media.

[0089] It should be noted that, unless there is any conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0090] In a first aspect, the present invention provides an embodiment of a method for analyzing and controlling inclusions in continuous casting, such as Figure 1 As shown, the specific process can be as follows:

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

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

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

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

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

[0096] S12: Data cleaning.

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

[0098] (1) Missing value interpolation: Sliding window mean method (window = 5 minutes) is used, formula:

[0099]

[0100] in,

[0101] : calculation results of interpolation missing values;

[0102] k: the number of data points in the sliding window. In this embodiment, k=5;

[0103] t: represents the time point at which missing values are to be interpolated;

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

[0105] (2) Standardization: Use the Z-score method to eliminate dimensional differences. The formula is:

[0106]

[0107] in,

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

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

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

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

[0112] The standardized data obeys the standard normal distribution with a mean of 0 and a standard deviation of 1, preparing for the next step of data processing.

[0113] (3) Outlier elimination: Based on the 3σ principle, data exceeding ±3 times the standard deviation of the mean are eliminated.

[0114] S13: Abnormal data enhancement.

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

[0116] (1) Generator loss function:

[0117]

[0118] in,

[0119] L G : Generator loss value;

[0120] :Based on the noise distribution p z The expectation of subsequent calculation of z obtained by sampling;

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

[0122] D(G(z)): The probability that the discriminator determines that the generated data is real data;

[0123] (2) Discriminator loss function:

[0124]

[0125] in,

[0126] L D : Discriminator loss value;

[0127] : The expectation of subsequent calculations on X sampled from the true data distribution;

[0128] D(X): The probability that the discriminator believes that the input data X is real data;

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

[0130] S14: Feature screening and weight calculation.

[0131] DeepSeek attention mechanism: feature weights are calculated through a multi-head attention layer, formula:

[0132]

[0133] in,

[0134] Q: query matrix;

[0135] K: bond matrix;

[0136] V: value matrix;

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

[0138] softmax: activation function.

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

[0140] S2: Offline analysis of association rules.

[0141] For association rule mining, this paper proposes an improved Apriori-Temporal algorithm. The specific improvements are: introducing a time sliding window (5 minutes) and DeepSeek temporal coding to mine strong association rules.

[0142] The improvement mechanism is explained as follows:

[0143] The limitation of the traditional Apriori algorithm is that it generates association rules by searching frequent item sets layer by layer, but its core assumption is that the data are independent and identically distributed (iid), ignoring the temporal correlation of process parameters.

[0144] For example, the mean casting speed and the liquid level fluctuation are strongly coupled in time (for example, the liquid level fluctuation reaches its peak 10 seconds later after a sudden change in casting speed); the parameter distributions in different production stages (start of casting, steady state, and tail billet) vary significantly, and hybrid analysis will introduce noise rules.

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

[0146] (1) Rule redundancy: 30% of the rules generated by traditional methods have inflated confidence levels due to time mismatches;

[0147] (2) Dynamic adaptability: Fluctuations in process parameters (e.g., superheat ±5°C) require real-time adjustment of rule thresholds;

[0148] (3) Computational efficiency: 100,000 high-frequency data items must be processed per hour, and rule updates must be completed within a short period of time (for example, within 5 minutes).

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

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

[0151] S211: Window segmentation strategy, including:

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

[0153]

[0154] in,

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

[0156] S: sliding step size, illustratively, S=1 minute.

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

[0158] (2) Edge data processing: For window edge data (such as the first 1 minute and the last 1 minute), weighted smoothing (EWMA) is used to reduce the truncation effect:

[0159]

[0160] in,

[0161] X edge : Window edge data value after weighted smoothing;

[0162] α: smoothing coefficient, for example, α=0.8;

[0163] : The data value at the start time of the window;

[0164] : The data value at the end 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 pulling speed (1Hz), liquid level fluctuation (1Hz), and coil thickness (0.2Hz), the slab number is used as the index and aligned to a unified timestamp through interpolation:

[0168]

[0169] in,

[0170] : heterogeneous data values aligned after interpolation at timestamp t;

[0171] LinearInterpolation: linear interpolation function;

[0172] x(t1), x(t2): original heterogeneous data values at timestamps t1 and t2.

[0173] (2) Timing lag compensation:

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

[0175]

[0176] in,

[0177] : Parameter lag time.

[0178] For example, according to the change of liquid level fluctuation lag pulling speed, the parameter is set =10 seconds.

[0179] S22: Data timing encoding.

[0180] S221: Encoder architecture, including:

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

[0182]

[0183] in,

[0184] X: Window parameter matrix;

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

[0186] d: parameter dimension.

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

[0188]

[0189] in,

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

[0191] : The 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: Calculates the weights of time steps within a window and adaptively captures the dependencies between different time steps within the window, allowing the model to more flexibly focus on important information when processing sequence data:

[0194]

[0195] in,

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

[0197] softmax: activation function;

[0198] Wq and W k : learnable parameter matrix;

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

[0200] H=[h1, h2,…, h Tw ];

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

[0202] S222: Coding output and rule mining.

[0203] (1) Feature vector generation: Window encoding result Combined with the original parameter statistics (mean, variance, extreme value), we get the enhanced feature vector:

[0204]

[0205] (2) Association rule mapping: Use the fully connected layer to map Z enhanced Mapping to item set probability distribution:

[0206]

[0207] in,

[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 how frequently an item set appears in a data set.

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

[0216]

[0217] in,

[0218] support min : Minimum support;

[0219] σ: standard deviation of fluctuation of the relevant parameter covered by the window.

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

[0221]

[0222] in,

[0223] Confidence corrected : Corrected confidence;

[0224] Confidence: original confidence, ;

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

[0226] This step proposes a confidence correction method based on causal strength, which combines the traditional association rule confidence with the results of Granger causality test. The treatment is: reduce the importance of rules with high confidence but lack of 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: Dynamic rule update mechanism.

[0228] S2321: Incremental learning framework: Each new window k+1 , update the rule base:

[0229]

[0230] in,

[0231] :Item set X is added to the new window Window k+1 Support for later updates;

[0232] k: Add a new window k+1 The number of existing windows;

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

[0234] :Item set X in new window Window k+1 The support in .

[0235] S2322: Confidence decay model: Introduces 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 casting speed > 9.7, coil thickness < 3.5}, the probability of slag inclusion increases by 82%; the support level of this rule is 0.25, and the confidence level is 0.85.

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

[0244] (1) DeepSeek-CART algorithm:

[0245] Split criterion: based on 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: to simplify the model structure and improve the model's generalization ability on new data.

[0252] For example, 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 F1-score of the inclusion class reaches 0.62 (precision 0.96, recall 0.45).

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

[0255] DeepSeek reinforcement learning framework: Build 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 resulting from process adjustments;

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

[0262] Through training, a strategy to maximize the reward function value is found: a parameter threshold adjustment method that can effectively reduce the probability of inclusion 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: Generates set values based on a preset rule base (such as superheat PID control);

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

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

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

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

[0273] (3) Control instructions are issued: Link L3 (MES system) to perform parameter adjustments (such as speed reduction and tundish heating power adjustment).

[0274] S32: DeepSeek lightweight inference engine design.

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

[0276] (1) FP16 quantization: Converting the decision tree model from FP32 to FP16 precision reduces memory usage by 50% and increases inference speed by 2 times:

[0277] MemoryFP16=0.5×MemoryFP32

[0278] LatencyFP16 = 0.5 × LatencyFP32

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

[0280] S322: TensorRT accelerated deployment.

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

[0282] (2) CUDA core optimization: using parallel computing to accelerate feature calculation and rule matching; in one embodiment of the present invention, a single inference takes ≤3ms.

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

[0284]

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

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

[0287] Data collection: Through the OPC UA unified architecture, subscribe to the L1 data stream and cache the last 5 seconds of data (5000 sampling points) in a ring buffer;

[0288] Feature calculation: parallel calculation of statistics (mean, variance, trend slope), taking ≤ 1ms;

[0289] Model inference: Input feature vectors to the TensorRT engine, output risk levels and optimization suggestions, taking ≤3ms;

[0290] Instruction issuance: Control instructions (such as speed reduction and heating power adjustment) are triggered according to the risk level, taking ≤1ms.

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

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

[0293]

[0294] in,

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

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

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

[0298] In one embodiment of the present invention, the proportionality coefficient K is set p =0.8, integral coefficient K i =0.2, differential coefficient K d =0.05.

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

[0300]

[0301] in,

[0302] Δu(t): 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 occurring in the production status obtained through analysis, as well as other specific indicators. For example, risk assessment 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 are reduced by more than 6 million yuan, and customer satisfaction is increased by 40%.

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

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

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

[0322] Rule generation: Through analysis of input data, one of the rules obtained is:

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

[0324] The comparison of the operating effects of the offline analysis model proposed in this invention and the traditional model is shown in the following table:

[0325]

[0326] (2) Example of the present invention for treating extreme working conditions at a steel base:

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

[0328] Blocking effect: The blocking rate of HD defects increased from 80% to 95%, reducing waste by 500 tons annually.

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

[0330] In a third aspect, the present invention provides an embodiment of a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described in the first aspect when executed by a processor.

[0331] In a fourth aspect, the present invention provides an embodiment of a computer program product, which implements the steps of the method described in the first aspect when executed by a processor.

Claims

1. A method for analyzing and controlling continuous casting inclusions, characterized in that: The following steps are involved: Step 1: Data acquisition and preprocessing, including: Multi-source data acquisition, including acquisition of data at different levels within the casting control system, as well as acquisition of data outside the casting control system; Multi-source data preprocessing and feature screening, including data cleaning, abnormal data enhancement, feature screening and weight calculation; Step 2: Offline analysis of association rules, including: By setting the window segmentation strategy and aligning the parameters within the window to model the time sliding window, Data temporal encoding, whose encoder architecture includes input layer, bidirectional LSTM layer, and self-attention mechanism. Strong association rule mining and dynamic optimization, including support-confidence joint optimization method, rule dynamic update mechanism based on incremental learning framework and confidence decay model; Step 3: Online control and real-time optimization of the continuous casting production line, including: Embed the DeepSeek lightweight inference engine into the L2 system, integrating data pipeline docking, model inference services and control command issuance. Real-time optimization control strategy, including millisecond-level data processing pipeline, dynamic PID-DeepSeek hybrid control and abnormal blocking logic.

2. The method according to claim 1, characterized in that The multi-source data preprocessing and feature screening includes the following steps: Data cleaning, including missing value interpolation, standardization, and outlier removal of acquired data; Abnormal data enhancement, including generating extreme operating condition samples; Feature screening and weight calculation, including calculating feature weights through a multi-head attention layer 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 was 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, comprising modeling a time sliding window by setting a window segmentation strategy and aligning parameters within the window, characterized in that The setting window segmentation strategy includes: By defining the time window length and sliding step size, the continuous data stream is divided into overlapping windows; For window edge data, weighted smoothing is used to reduce truncation effects; The parameters within the alignment window include: 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, comprising data temporal encoding, wherein the encoder architecture comprises an input layer, a bidirectional LSTM layer, and a self-attention mechanism, characterized in that: The input layer includes: a parameter matrix within the window, which includes the value of any parameter at any time within the window; The bidirectional LSTM layer, i.e., the bidirectional long short-term memory network layer, captures the forward and reverse temporal dependencies in the sequence data; The self-attention mechanism includes: calculating the weights of time steps within the window and adaptively capturing 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 into an item set probability distribution.

8. The method according to claim 1, characterized in that The support-confidence joint optimization method includes: 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 level; Confidence: original confidence, ; CausalStrength: F statistic for Granger causality test.

9. The method according to claim 1, comprising a rule dynamic update mechanism based on an incremental learning framework and a confidence decay model, characterized in that: The incremental learning framework includes: each time a window is added, a sliding window incremental update method is used to dynamically adjust the support; The confidence decay model includes: introducing a time decay factor and dynamically adjusting the confidence of the rule using a weighted moving average method; The dynamic rule update mechanism also includes: decision tree model training and optimization, including setting splitting criteria, setting pruning strategies, and dynamic identification of sensitive intervals.

10. The method according to claim 9, characterized in that The decision tree model training and optimization includes: Set the splitting criterion 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, comprising embedding the DeepSeek lightweight inference engine into the L2 system, integrating data pipeline docking, model inference services and control instruction issuance, characterized in that: The data pipeline connection includes: obtaining L1 data stream in real time through the OPC UA protocol; The model inference service includes: deploying a TensorRT-accelerated decision tree model on an L2 server; The control instruction is issued including: linking 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 less than 0.5% in the decision tree; TensorRT accelerated deployment: Build a decision tree model inference engine based on TensorRT's Python API, and combine it with CUDA parallel computing methods to optimize feature calculation and rule matching.

13. The method according to claim 1, wherein The millisecond-level data processing pipeline includes: Data collection: Through the OPC UA unified architecture, subscribe to the L1 data stream and cache the latest short-term data in a ring buffer; Feature calculation: parallel calculation of statistics, including mean, variance, and trend slope; Model inference: 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, wherein The dynamic PID-DeepSeek hybrid control includes: 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 1, wherein The abnormal blocking logic includes: When the risk level reaches the warning level, the alarm instruction 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, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 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