Automatic analysis and disposal method for anomalies in titanium sponge production process based on MES (Manufacturing Execution System)

Abstract

The invention relates to the technical field of sponge titanium production processes, in particular to a sponge titanium production process anomaly automatic analysis and disposal method based on an MES (Manufacturing Execution System), which comprises the following steps of: firstly, integrating DCS (Distributed Control System), LIMS (Laser Information Management System) and quality management system data through the MES, and collecting key process parameters and quality data to form a complete data chain; establishing a dynamic early warning model, and combining SPC, machine learning and ARIMA models to realize abnormal real-time early warning and trend prediction; abnormal root causes are automatically analyzed by means of an AI model, a report is generated, and treatment measures are accurately pushed through a graph database knowledge base; and finally, monitoring-diagnosis-disposal-knowledge precipitation closed-loop management is formed. Advanced prevention and control of production abnormity are achieved, diagnosis accuracy and disposal efficiency are improved, repeated occurrence of quality defects and similar problems is reduced, and the intelligent level and operation efficiency of titanium sponge production are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of sponge titanium production technology, specifically to an automatic analysis and handling method for abnormalities in the sponge titanium production process based on a MES system. Background Technology

[0002] Titanium sponge, as a core raw material for titanium materials and titanium alloys, directly determines the performance of downstream high-end products through its purity and quality. Globally, over 90% of titanium sponge is produced using the mature Kroll process. This process involves several key steps. First, titanium-rich raw materials are mixed with petroleum coke and reacted with chlorine in a fluidized bed reactor to produce crude titanium tetrachloride. Then, the crude titanium tetrachloride is purified by distillation to remove impurities such as silicon, iron, and vanadium, yielding high-purity refined titanium tetrachloride. Finally, in a reducing distillation furnace filled with inert gas, liquid refined titanium tetrachloride is injected into molten magnesium metal, undergoing a magnesothermic reduction reaction to produce sponge-like metallic titanium and the byproduct magnesium chloride. After the reaction, the product undergoes vacuum distillation, gate cutting, crushing, and sieving to finally obtain granular titanium sponge.

[0003] Although the Kroll process is technologically mature, the production of sponge titanium is essentially a complex physicochemical reaction, characterized by a long process flow, high energy consumption, harsh operating conditions, and significant time lags. In actual production, multiple stages are highly susceptible to abnormalities due to improper control of process parameters or external interference. In the chlorination and refining processes, fluctuations in the temperature and pressure of the chlorination furnace can lead to incomplete reactions, furnace scaling, or equipment corrosion. Inaccurate control of the temperature and reflux ratio in the distillation column can cause excessive levels of key impurities in the product, affecting the final grade of sponge titanium. The reduction distillation process is a core step; excessively high reaction temperatures can lead to product sintering and hardening, resulting in reduced yield, while excessively low temperatures will slow down the reaction rate and reduce production efficiency. A mismatch between the titanium tetrachloride feeding rate and the reaction heat can cause violent splashing or stagnation of the reaction. Abnormal system vacuum or internal pressure can cause excessive chlorine content in the product, rendering it unusable. At the same time, some raw materials are highly corrosive, and there is a risk of leakage in reactors, pipelines, and valves. The entire system must be strictly isolated from air; otherwise, serious safety accidents such as product oxidation, nitriding, or even combustion and explosion may occur.

[0004] Currently, the monitoring and management of sponge titanium production processes largely rely on distributed control systems for basic data collection, combined with manual judgment and intervention based on operator experience. This approach has significant drawbacks. Massive amounts of process data cannot be effectively correlated and analyzed; early signs of anomalies are easily buried by normal data, making timely detection difficult; by the time anomalies are detected, they have often already caused substantial damage to product quality. Furthermore, due to process delays and lags, tracing the root cause of anomalies is extremely difficult, heavily reliant on the experience of technical personnel, and lacks systematic analytical methods. For anomalies that have already occurred, there is a lack of an integrated knowledge base to quickly match and recommend effective countermeasures, resulting in low processing efficiency, and similar problems may recur in different shifts or batches.

[0005] Manufacturing Execution Systems (MES), as information management systems connecting the enterprise planning and shop floor control layers, have been widely applied in many process industries for production order management, material tracking, and performance analysis. However, their functionality has not been fully explored in the specific field of titanium sponge production. Existing MES applications largely remain at the level of data display and report generation, lacking integrated capabilities for intelligent analysis of real-time data across the entire production process, anomaly warnings, automatic root cause diagnosis, and proactive dissemination of countermeasures. Therefore, there is an urgent need in this field for a solution that can deeply integrate MES with the titanium sponge production process, enabling a shift from passive response to proactive warning, and from experience-driven to data-driven approaches, thereby effectively improving product quality, production safety, and operational efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic analysis and handling method for anomalies in the production process of sponge titanium based on a MES system, so as to solve the problems of difficult early warning, difficult root cause tracing, and low handling efficiency in the existing sponge titanium production process mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An automated analysis and handling method for anomalies in the sponge titanium production process based on a MES system includes the following steps:

[0009] S1: Construct a production process data integration layer, which collects key process parameters from the distributed control system (DCS) in real time through the OPCUA protocol or API interface, and obtains product impurity content and quality grade data from the laboratory information management system (LIMS) through a database intermediate table or WebService interface; the key process parameters include at least the reaction temperature T of the reduction distillation furnace, the system pressure P, the vacuum degree V, and the titanium tetrachloride feeding rate F.

[0010] S2: Establish a dynamic process parameter early warning model. Based on historical production data and a process knowledge base, use the Statistical Process Control (SPC) method to calculate the process capability index Cpk of historical qualified batch data at each production stage. Combine the 3σ principle to determine the dynamic normal operating range and early warning threshold of key process parameters. The formula for calculating the process capability index Cpk is as follows:

[0011] ;

[0012] USL is the upper limit of parameter specifications, LSL is the lower limit of parameter specifications, μ is the parameter mean, and σ is the parameter standard deviation; or, through machine learning algorithms, cluster analysis is performed on time series data under normal operating conditions, and the cluster boundary is used as a dynamic threshold. The MES system monitors the incoming data in real time, and triggers a real-time alarm when the parameter continuously or instantaneously exceeds the corresponding warning threshold.

[0013] S3: Construct an anomaly-quality correlation analysis module. After the production batch is completed, the quality results of this batch of products are compared with the complete historical data of the production process recorded in S1. The anomaly pattern matching algorithm based on sliding window is used, or the mutual information I(X,Y) and Pearson correlation coefficient r(X,Y) of process parameters and quality indicators within the time period of the anomaly event are calculated to quantify the correlation strength between parameters and quality indicators and calculate the influence weight. The formula for calculating mutual information is:

[0014] ;

[0015] The formula for calculating the Pearson correlation coefficient is:

[0016] ;

[0017] Where Cov(X,Y) is the covariance of parameter X and quality index Y, σX is the standard deviation of parameter X, and σY is the standard deviation of quality index Y.

[0018] S4: Integrates an AI-driven root cause analysis and report generation module. It inputs the abnormal event sequence and corresponding working condition data identified by S3 into a pre-trained artificial intelligence analysis model. The model outputs the root cause probability distribution and automatically generates a structured natural language analysis report based on a preset template.

[0019] S5: Build a knowledge base and recommendation engine for disposal measures. Based on the root cause analysis results of S4, match disposal measures from the knowledge base through graph traversal query, and push them according to the historical effectiveness probability, forming a closed-loop management of monitoring-diagnosis-disposal-knowledge accumulation.

[0020] Preferably, the production stage described in S2 is automatically identified by the process status signal in the DCS, or divided by analyzing the trajectory characteristics of the key parameters, temperature change rate k and cumulative feed amount Q, wherein the formula for calculating the temperature change rate is:

[0021] ;

[0022] The formula for calculating the cumulative amount of material fed is:

[0023] ;

[0024] Where ΔT is the temperature change, Δt is the time interval, q(t) is the instantaneous feeding rate, and t is the time.

[0025] Preferably, the machine learning algorithm described in S2 employs the K-means clustering algorithm, which determines the cluster boundaries by calculating the Euclidean distance d between the sample data and the cluster centers. The formula for calculating the Euclidean distance is as follows:

[0026] ;

[0027] x i For sample data points, c j The parameter range corresponding to the cluster center and the cluster boundary is the dynamic early warning threshold.

[0028] Preferably, the abnormal pattern matching algorithm for the sliding window described in S3 uses a sliding window of size n with a window movement step of s. Abnormal event points are identified by calculating the similarity S between the data within the window and the normal pattern template. The similarity calculation formula is as follows:

[0029] ;

[0030] When the similarity S is lower than the preset threshold S0, it is identified as an abnormal event point. w For data within the window, x m For normal mode template data, max(x) m ) represents the maximum value of the normal mode template data.

[0031] Preferably, the pre-trained artificial intelligence analysis model described in S4 is a classification model based on a temporal convolutional network (TCN). The model training data includes historical batch production process time-series data, operation event records, equipment status data, final quality reports, and manual diagnostic conclusions. The model feature engineering extracts the abnormal event types, durations, amplitudes, and temporal relationship features between abnormal events.

[0032] Preferably, the root cause probability distribution output by the AI ​​model in S4 is calculated using the softmax function, and the calculation formula is as follows:

[0033] ;

[0034] Where z i y is the score of the i-th node in the model's output layer. i For the i-th root cause category, P(y) i ) represents the probability of the i-th root cause category.

[0035] Preferably, the expert knowledge base described in S5 is constructed using a graph database. The knowledge graph includes abnormal phenomena, root causes, remedial measures, and related equipment entities. Entities are connected by edges that lead to solutions, and the weights of these edges are set based on historical remedial effects.

[0036] .

[0037] As a preferred embodiment, the graph traversal query of the recommendation engine described in S5 adopts a depth-first search algorithm with a search depth of 3 layers. During the traversal, a comprehensive score for the processing measures is calculated using a formula, which is:

[0038] ;

[0039] W represents the weight of the relation edge, and P represents the probability of the corresponding root cause. The response measures are pushed out in descending order of the comprehensive score.

[0040] Preferably, the dynamic early warning model in S2 also includes a parameter trend prediction function, which uses an ARIMA model to make short-term predictions of key process parameters. The prediction formula is as follows:

[0041] ;

[0042] Where X t Let φ be the predicted value at time t, φ0 be a constant term, and φ i For autoregressive coefficients, θ j ε is the moving average coefficient. t This is a random error term that triggers an early warning when the predicted value exceeds the warning threshold.

[0043] Preferably, the method further includes an effect verification step. After implementing the measures pushed by S5, the effect of the measures is verified by calculating the quality improvement rate η. The formula for calculating the quality improvement rate is:

[0044] ;

[0045] When η ≥ preset threshold, the abnormal event, root cause analysis results, handling measures and effect data are stored in the knowledge base, and the relation edge weights and historical effectiveness probabilities of handling measures in the knowledge base are updated.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] The production process data integration layer constructed in this invention achieves seamless integration between the MES system and DCS, LIMS, and quality management systems through multiple protocols and interfaces. It collects key process parameters such as reaction temperature and vacuum level in real time, and synchronously correlates them with data such as product impurity content and quality grade, forming a complete data chain covering the entire production process. This integrated data integration model breaks down the barriers of traditional decentralized data management, avoids the problem of disconnect between process parameters and quality data, provides a comprehensive, coherent, and accurate data foundation for anomaly analysis, and completely solves the pain point of the inability to effectively correlate and utilize massive amounts of production data, allowing information from every production link to be fully mined and applied.

[0048] II. Dynamic intelligent early warning system for proactive prevention and control

[0049] Compared to traditional fixed-threshold early warning methods, the dynamic process parameter early warning model established in this invention has stronger adaptability and accuracy. By calculating the process capability index using the SPC method combined with the 3σ principle, or by determining the dynamic threshold using machine learning cluster analysis, the early warning range can be automatically adjusted according to different production stages such as the initial reduction stage and the peak distillation period, accurately capturing abnormal fluctuations in parameters. Simultaneously, the ARIMA model is introduced to achieve short-term trend prediction of key parameters, identifying potential abnormal risks in advance. This transforms traditional post-event remediation into pre-event early warning, allowing operators to intervene in time before parameters exceed safe ranges and cause irreversible damage to product quality, significantly reducing the incidence of quality defects.

[0050] III. AI-driven intelligent analysis improves diagnostic efficiency and accuracy.

[0051] This invention utilizes an AI model-based root cause analysis module, enabling a leap from experience-driven to data-driven anomaly diagnosis. By inputting anomaly event sequences and operational data, a pre-trained temporal convolutional network model automatically outputs root cause probability distributions. Combined with quantitative tools such as mutual information and Pearson correlation coefficients, it accurately pinpoints the core causes of quality problems, avoiding the drawbacks of traditional manual diagnosis that rely on personal experience, are inefficient, and produce inconsistent conclusions. Simultaneously, the system automatically generates structured natural language analysis reports, clearly presenting production overviews, anomalies, root cause conclusions, and quality impact assessments, significantly shortening problem investigation time and improving the objectivity and reliability of diagnostic results.

[0052] IV. Closed-loop knowledge management to continuously optimize production efficiency

[0053] This invention constructs a knowledge base and recommendation engine for handling measures, combining a graph database knowledge graph to achieve accurate matching and ranking of handling measures, ensuring that operators can quickly obtain targeted solutions. More importantly, the system calculates the quality improvement rate through an effectiveness verification step, automatically accumulating effective handling cases into the knowledge base, continuously updating the relational edge weights and the probability of measure effectiveness, forming a complete closed loop of monitoring-diagnosis-handling-knowledge accumulation. This cyclical optimization model not only avoids the recurrence of similar anomalies but also continuously improves the enterprise's process management level and intelligence. Simultaneously, through automated report generation and measure recommendation, it reduces unplanned shutdowns, ensures stable and efficient production operation, and significantly improves overall operational efficiency. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0055] Figure 1 This is a flowchart of the present invention;

[0056] Figure 2 This is a comparison chart of correlation indicators among various embodiments of the present invention;

[0057] Figure 3 This is a comparison chart of the scores of the treatment measures in the three embodiments of the present invention. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0059] Example 1: Analysis and Handling of Temperature Anomalies in the Reduction Distillation Stage

[0060] Data Integration: The MES system communicates with the DCS via the OPCUA protocol to collect real-time data on reaction temperature, distillation column pressure, vacuum degree, and titanium tetrachloride feeding rate during the reduction distillation stage. It also obtains data on the oxygen and nitrogen content of sponge titanium from the LIMS system and the product purity level data from the quality management system through the WebService interface. The sampling frequency is 1 time / second.

[0061] Dynamic early warning model construction: Based on historical data from 500 batches of qualified production, the process capability index Cpk for reaction temperature is calculated using the SPC method. The upper limit for reaction temperature is set at 950℃, the lower limit at 900℃, the mean at 925℃, and the standard deviation at 8℃. The formula is as follows:

[0062] Cpk=min[(950-925) / (3×8),(925-900) / (3×8)];

[0063] Cpk = 1.04 was calculated, and the warning threshold range was determined to be 899℃-951℃ based on the 3σ principle. Simultaneously, the normal operating condition time series data was clustered using the K-means clustering algorithm, resulting in 5 clusters, and the Euclidean distance formula was applied.

[0064] ;

[0065] Determine cluster boundaries and dynamically adjust thresholds. Trigger an alarm when the reaction temperature remains above 951°C for 3 consecutive seconds.

[0066] Anomaly-Quality Correlation Analysis: The oxygen content of this batch of products is 0.18% (exceeding the standard). The cross-information between data and oxygen content during the abnormal temperature period was calculated.

[0067] ;

[0068] The mutual information value is 0.72, and the Pearson correlation coefficient is calculated using the formula:

[0069] ;

[0070] The calculated value is 0.81, and the weight of the strong correlation between temperature anomalies and excessive oxygen content is determined to be 0.85.

[0071] Root cause analysis: Abnormal sequences and operating condition data are input into a pre-trained TCN model. The model contains 6 stacked blocks, each with 4 layers of convolutional kernels of size 6. size Set to 64, the output root cause probability distribution is: heating element power fluctuation 68%, temperature control algorithm parameter drift 22%, sensor failure 10%, and a structured analysis report is automatically generated.

[0072] Recommended remedial measures: Through graph database traversal and query, the matching remedial measure is to adjust the heating element power supply voltage to 380V±3V and optimize the PID temperature control parameters, according to the comprehensive score formula:

[0073] ;

[0074] The calculated scores were 89.67 and 2.3, respectively, and were pushed out in descending order. After execution, the oxygen content decreased to 0.12%, and the quality improvement rate was calculated using the formula:

[0075] η = (0.18% - 0.12%) / 0.18% × 100% = 33.3%, which meets the preset threshold of 15%. The data is then stored in the knowledge base to update the relation edge weight to 91.

[0076] Example 2: Coordinated handling of abnormal feed rate and pressure during chlorination stage

[0077] Data integration: The MES system interacts with the DCS in real time through the API interface to collect data on pressure, titanium tetrachloride feeding rate, chlorine flow rate, and reaction temperature in the chlorination furnace. It also links the data on titanium chloride purity, impurities, and iron content from the LIMS system through an intermediate table in the database. The sampling frequency is 2 times / second.

[0078] Dynamic early warning model construction: Production stages are divided by analyzing the temperature change rate and cumulative feed amount. The temperature change rate is calculated using the formula:

[0079] ;

[0080] Calculate the cumulative feeding amount using the formula

[0081] ;

[0082] Calculation; The feeding rate is predicted using the ARIMA model, and the prediction formula is as follows:

[0083] ;

[0084] With φ0=120, φ1=0.7, and θ1=0.3, the feeding rate for the next 10 minutes is predicted, and an early warning is issued when the predicted value is below 100 kg / h or above 150 kg / h. Simultaneously, a dynamic pressure threshold is determined using K-means clustering, with cluster centers c... j The values ​​are 0.08, 0.10, and 0.12 MPa, respectively. An Euclidean distance d less than 0.02 MPa is within the normal range.

[0085] Anomaly-quality association analysis: A sliding window algorithm with a window size of n=20 and a step size of s=5 was used, based on the similarity formula.

[0086] ;

[0087] The similarity between the data within the window and the normal template is calculated, and anomalies are identified when S is less than 0.75. The mutual information between the feeding rate and the purity of titanium chloride during the abnormal period is calculated to be 0.68, the Pearson correlation coefficient is 0.76, and the association weight is 0.78.

[0088] Root cause analysis: The TCN model takes abnormal data and equipment status records as input and outputs the following root cause probability distribution: unstable feed pump flow rate 72%, chlorine pressure fluctuation 20%, valve internal leakage 8%. The root cause probability is calculated using the softmax formula.

[0089] ;

[0090] The calculations show that z_i are 3.2, 1.8, and 0.9 respectively.

[0091] Recommended remedial measures: The graph database knowledge graph contains the abnormal phenomenon of fluctuating feeding rate as the root cause of unstable feeding pump flow. Remedial measures include cleaning the pump body filter and adjusting the inverter frequency. The relation edge weight W=88, and according to the comprehensive score formula: Score=88×0.72=63.36, it is prioritized for push. After execution, the feeding rate stabilized at 125±5 kg / h, the purity of titanium chloride increased from 98.2% to 99.5%, the quality improvement rate was 26.3%, and the knowledge base weight was updated to 93.

[0092] Example 3: Traceability and Closed-Loop Handling of Impurity Anomalies in Vacuum Refining Stage

[0093] Data integration: The MES system collects vacuum degree refining temperature and holding time data through the OPCUA protocol, and associates it with the total impurity content of sponge titanium, the impurity data of each element in the LIMS system, and the final quality level of the quality management system. The sampling frequency is 1 time / second.

[0094] Dynamic early warning model construction: Based on the SPC method, the vacuum process capability index is calculated, with the upper limit set at 0.001 MPa, the lower limit at 0.0005 MPa, the mean at 0.0007 MPa, and the standard deviation at 0.00008 MPa, according to the formula:

[0095] Cpk=min[(0.001-0.0007) / (3×0.00008),(0.0007-0.0005) / (3×0.00008)]

[0096] The result is Cpk=1.25, and the warning threshold is 0.00047MPa-0.00103MPa; the refining temperature trend is predicted by the ARIMA model, and the risk of temperature drift is warned 15 minutes in advance.

[0097] Anomaly-quality correlation analysis: A sliding window algorithm with window size n=15 and step size s=3 is used. Anomalies are identified when the similarity S is less than 0.8. The mutual information between vacuum degree anomalies and total impurity content is calculated to be 0.75, and the Pearson correlation coefficient is 0.83. The correlation weight is determined to be 0.82.

[0098] Root cause analysis: The TCN model extracts the time-series relationship features of the duration and amplitude of anomalies, and outputs the root cause probability as follows: vacuum pump seal aging 65%, vacuum system leakage 25%, temperature control deviation 10%. An analysis report with data trend charts is automatically generated.

[0099] Recommended remedial measures: The graph traversal query matching remedial measures are to replace the vacuum pump seals and inspect the vacuum pipeline. The edge weights are 90 and 85 respectively. According to the comprehensive score formula...

[0100] ;

[0101] The calculated score is 58.521.25, prioritizing the replacement of seals. After execution, the total impurity content decreased from 0.22% to 0.10%, and the quality improvement rate, calculated using the formula: η=(0.22%-0.10%) / 0.22%×100%=54.5%, meets the preset threshold of 12%. This data is then stored in the knowledge base, updating the edge weight of the seal aging-related relationship to 95, thus forming a closed-loop management system.

[0102] The table below compares the data from the three sets of examples:

[0103] Dimension Example 1 Example 2 Example 3 Process Stage Reductive distillation Chlorination Vacuum refining Core monitoring parameters Reaction temperature, distillation column pressure, vacuum degree Chlorination furnace pressure, feeding rate, chlorine flow rate Vacuum degree, refining temperature, holding time Main Algorithm Combinations SPC, K-means clustering, TCN model K-means clustering, ARIMA, sliding window SPC, ARIMA, TCN models Root cause analysis results Heating element power fluctuation (68%) The feed pump flow rate is unstable (72%). Vacuum pump seals are aging (65%). Disposal measures Adjust heating element voltage and optimize PID parameters Clean the pump body filter and adjust the inverter frequency. Replace vacuum pump seals Quality improvement rate 33.3% 26.3% 54.5%

[0104] This invention deeply integrates the MES system with the sponge titanium production process to construct a full-process intelligent analysis and handling system. Its core advantages are reflected in four dimensions: data integration, accurate early warning, intelligent analysis, and closed-loop management, as detailed below:

[0105] I. Comprehensive data integration, breaking down information silos

[0106] The production process data integration layer constructed in this invention achieves seamless integration between the MES system and DCS, LIMS, and quality management systems through multiple protocols and interfaces. It collects key process parameters such as reaction temperature and vacuum level in real time, and synchronously correlates them with data such as product impurity content and quality grade, forming a complete data chain covering the entire production process. This integrated data integration model breaks down the barriers of traditional decentralized data management, avoids the problem of disconnect between process parameters and quality data, provides a comprehensive, coherent, and accurate data foundation for anomaly analysis, and completely solves the pain point of the inability to effectively correlate and utilize massive amounts of production data, allowing information from every production link to be fully mined and applied.

[0107] II. Dynamic intelligent early warning system for proactive prevention and control

[0108] Compared to traditional fixed-threshold early warning methods, the dynamic process parameter early warning model established in this invention has stronger adaptability and accuracy. By calculating the process capability index using the SPC method combined with the 3σ principle, or by determining the dynamic threshold using machine learning cluster analysis, the early warning range can be automatically adjusted according to different production stages such as the initial reduction stage and the peak distillation period, accurately capturing abnormal fluctuations in parameters. Simultaneously, the ARIMA model is introduced to achieve short-term trend prediction of key parameters, identifying potential abnormal risks in advance. This transforms traditional post-event remediation into pre-event early warning, allowing operators to intervene in time before parameters exceed safe ranges and cause irreversible damage to product quality, significantly reducing the incidence of quality defects.

[0109] III. AI-driven intelligent analysis improves diagnostic efficiency and accuracy.

[0110] This invention utilizes an AI model-based root cause analysis module, enabling a leap from experience-driven to data-driven anomaly diagnosis. By inputting anomaly event sequences and operational data, a pre-trained temporal convolutional network model automatically outputs root cause probability distributions. Combined with quantitative tools such as mutual information and Pearson correlation coefficients, it accurately pinpoints the core causes of quality problems, avoiding the drawbacks of traditional manual diagnosis that rely on personal experience, are inefficient, and produce inconsistent conclusions. Simultaneously, the system automatically generates structured natural language analysis reports, clearly presenting production overviews, anomalies, root cause conclusions, and quality impact assessments, significantly shortening problem investigation time and improving the objectivity and reliability of diagnostic results.

[0111] IV. Closed-loop knowledge management to continuously optimize production efficiency

[0112] This invention constructs a knowledge base and recommendation engine for handling measures, combining a graph database knowledge graph to achieve accurate matching and ranking of handling measures, ensuring that operators can quickly obtain targeted solutions. More importantly, the system calculates the quality improvement rate through an effectiveness verification step, automatically accumulating effective handling cases into the knowledge base, continuously updating the relational edge weights and the probability of measure effectiveness, forming a complete closed loop of monitoring-diagnosis-handling-knowledge accumulation. This cyclical optimization model not only avoids the recurrence of similar anomalies but also continuously improves the enterprise's process management level and intelligence. Simultaneously, through automated report generation and measure recommendation, it reduces unplanned shutdowns, ensures stable and efficient production operation, and significantly improves overall operational efficiency.

[0113] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic analysis and disposal of abnormality in the production process of titanium sponge based on an MES system, characterized in that, The method comprises the following steps: S1: constructing a production process data integration layer, collecting key process parameters in a distributed control system (DCS) in real time through an OPCUA protocol or an API interface, and obtaining product impurity content and quality grade data provided by a laboratory information management system (LIMS) through a database intermediate table or a WebService interface; the key process parameters at least include a reaction temperature T of a reduction distillation furnace, a system pressure P, a vacuum degree V, and a titanium tetrachloride feeding rate F; S2: establishing a dynamic process parameter early warning model, calculating process capability indexes Cpk of historical qualified batch data at each production stage based on historical production data and a process knowledge base, and determining a dynamic normal operation range and a pre-warning threshold of the key process parameters in combination with a 3σ principle, wherein a process capability index Cpk calculation formula is: ； USL is an upper limit of a parameter specification, LSL is a lower limit of the parameter specification, μ is a parameter mean value, and σ is a parameter standard deviation; or performing clustering analysis on time series data under a normal working condition through a machine learning algorithm, taking a clustering boundary as a dynamic threshold, and monitoring flowing data in real time through an MES system, and triggering a real-time alarm when a parameter continuously or instantaneously exceeds a corresponding pre-warning threshold; S3: constructing an abnormality-quality correlation analysis module, after a production batch ends, performing reverse correlation comparison between a product quality result of the batch and complete production process historical data recorded in S1, and performing abnormal pattern matching through a sliding window-based abnormal pattern matching algorithm, or calculating mutual information I(X, Y) and a Pearson correlation coefficient r(X, Y) between process parameters and quality indexes in an abnormal event time period, quantifying correlation strength between the parameters and the quality indexes and calculating an influence weight, wherein a mutual information calculation formula is: ； a Pearson correlation coefficient calculation formula is: ； wherein Cov(X, Y) is a covariance of a parameter X and a quality index Y, σX is a standard deviation of the parameter X, and σY is a standard deviation of the quality index Y; S4: integrating an AI-driven root cause analysis and report generation module, inputting an abnormal event sequence and corresponding working condition data identified in S3 into a pre-trained artificial intelligence analysis model, outputting a root cause probability distribution from the model, and automatically generating a structured natural language analysis report based on a pre-set template; S5: constructing a disposal measure knowledge base and a recommendation engine, matching disposal measures from the knowledge base through graph traversal query based on a root cause analysis result of S4, sorting and pushing according to a historical effectiveness probability, and forming a closed-loop management of monitoring-diagnosis-disposal-knowledge sedimentation.

2. The method according to claim 1, wherein the method is characterized in that, The production stage in S2 is automatically identified through a process state signal in the DCS, or is divided through analysis of a temperature change rate k and a cumulative feeding amount Q trajectory feature, wherein a temperature change rate calculation formula is: ； a cumulative feeding amount calculation formula is: ； wherein ΔT is a temperature change amount, Δt is a time interval, q(t) is an instantaneous feeding rate, and t is time.

3. The method according to claim 1, characterized in that, The machine learning algorithm in S2 adopts a K-means clustering algorithm, and a clustering boundary is determined by calculating a Euclidean distance d between sample data and a clustering center, wherein a Euclidean distance calculation formula is ； x i For sample data points, c j For cluster centers, the parameter range corresponding to the cluster boundary is the dynamic early warning threshold.

4. The method according to claim 1, characterized in that, The abnormal pattern matching algorithm of the sliding window in S3 adopts a sliding window with a window size of n and a window moving step of s, identifies abnormal event points by calculating the similarity S between the data in the window and the normal pattern template, and the similarity calculation formula is: ； When the similarity S is lower than a preset threshold S0, it is identified as an abnormal event point, x w is the data within the window, x m is the normal mode template data, max(x m ) is the maximum value of the normal mode template data.

5. The method according to claim 1, wherein the method is characterized in that, The pre-trained artificial intelligence analysis model in S4 is a classification model based on a time series convolution network (TCN), and the model training data includes historical batch production process time series data, operation event records, device state data, final quality reports, and artificial diagnosis conclusions. The model feature engineering extracts abnormal event type, duration, amplitude, and time sequence relationship features between abnormal events.

6. The method according to claim 1, wherein the method is characterized in that, The root cause probability distribution output by the AI model in S4 is calculated by a softmax function, and the calculation formula is: ； where z i is the score of the i-th node of the model output layer, y i is the i-th root cause class, P(y i ) is the probability of the i-th root cause class.

7. The method according to claim 1, wherein the method is characterized by, The expert knowledge base in S5 is constructed using a graph database, and the knowledge graph includes abnormal phenomena, root causes, disposal measures, and related device entities. The entities are connected by cause-solution relationship edges, and the relationship edge weights are set based on historical disposal effects: 。 8. The method according to claim 1, wherein the method is characterized by, The graph traversal query of the recommendation engine in S5 uses a depth-first search algorithm, and the search depth is set to 3 layers. In the traversal process, the disposal measure comprehensive score is calculated by the formula, and the comprehensive score calculation formula is: ； W is the relationship edge weight, P is the probability of the corresponding root cause, and the disposal measures are pushed in descending order of comprehensive score.

9. The method according to claim 1, wherein the method is characterized in that, The dynamic early warning model in S2 also includes a parameter trend prediction function, which uses an ARIMA model to perform short-term prediction on key process parameters, and the prediction formula is: ； where X t is the prediction value at time t, φ0is the constant term, φ i is the autoregressive coefficient, θ j is the moving average coefficient, and ε t is the random error term.

10. The method according to claim 1, wherein the method is characterized by, The method also includes an effect verification step. After executing the disposal measures pushed in S5, the disposal effect is verified by calculating the quality improvement rate η, and the quality improvement rate calculation formula is: ； When η is greater than or equal to a preset threshold, the current abnormal event, root cause analysis result, disposal measure, and effect data are stored in the knowledge base, and the relationship edge weight and disposal measure historical effectiveness probability in the knowledge base are updated.

Images